What is your best argument against this possibility? People were laughing at the idea of AI doing math at all a few years ago, I feel that all those still secure in human superiority in any other domain won't be feeling the same way in a few years once robotics reaches a critical threshold. I think the only thing keeping humans in charge at that point would be very effectively managed regulation which gives humans a decisive, if artificial, advantage.", "5": "2026-08-22T18:04:53.374619"} {"0": 17, "1": "hackernews", "2": "https://www.forbes.com/sites/barrycollins/2026/07/22/rogue-openai-attack-fuels-demands-to-rein-in-big-tech/", "3": "OpenAI's Hugging Face Breach Fuels Fresh Calls for AI Regulation", "4": "OpenAI's Hugging Face Breach Fuels Fresh Calls for AI Regulation. ", "5": "2026-08-22T18:04:53.379720"} {"0": 18, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49232136", "3": "The Systemic Collapse of the AI Industry: Ideology, Hardware, and CapEx Crisis", "4": "The Systemic Collapse of the AI Industry: Ideology, Hardware, and CapEx Crisis. The Systemic Collapse of the AI Industry: A Crisis of Ideology, Hardware, Finance, and Ethics\nThe Ideological Illusion\nSilicon Valley has sold humanity a marketing simulacrum under the guise of "Artificial Intelligence." We were promised a holy salvation, but instead, we received a fragmented, distributed network of colossal servers burning through other people's mathematical code. This mass self-deception forces society to accept a technology that lacks agency, cares nothing for ecology, and masquerades as a manageable deity while acting as a gluttonous, blind idol.\nThe $1.1 Trillion Financial Black Hole\nBig Tech (Amazon, Alphabet, Microsoft, Meta) has triggered a catastrophic investment cycle, with cumulative capital expenditures (CapEx) exceeding $1.1 trillion since 2023, and 2026 infrastructure spending projected at $720B\u2013$745B. The fundamental economic model is broken: AI subscription revenues are a drop in the ocean compared to hardware and energy costs. Corporate balance sheets are burdened by hidden debt and long-term liabilities, all tied to hardware that faces rapid obsolescence.\nThe Physical and Ecological Dead End\nThe digital world has hit hard material limits\u2014microchips, rare-earth metals, transformers, and water. A single 100 MW data center consumes 876,000 MWh per year (equivalent to 100,000 European homes) and evaporates up to 3.6 million liters of clean water daily just to cool silicon. During peak loads, when power grids are already stressed by record heatwaves, these facilities often switch to "backup" power\u2014dirty diesel generators and coal-fired plants\u2014accelerating the very global warming they claim to "forecast."\nLogical Collapse and Managerial Madness\nToday\u2019s models function as "stochastic dust-collectors." They do not verify facts; they generate "confident lies" (hallucinations) that require exhaustive manual human auditing. Treating every routine business check as an excuse to query trillions of parameters is not innovation\u2014it is architectural and managerial insanity.\nWar Against Civilization and Law\nTech giants have declared a de facto war on societal sovereignty. They are imposing a narrow, corporate-driven vision of humanity's future, monetizing the attention of our children, and conducting unprecedented surveillance via aggressive data collection. They systematically ignore AI regulations and safety laws; for these corporations, multi-million dollar fines are not penalties\u2014they are simply a line item in their operating budgets, turning the rule of law into a corporate formality.\nThe Path Forward\u2014Reasonable Sufficiency\u2122\nThe era of blind worship of "digital giants" is over. We must pivot to the Concept of Reasonable Sufficiency\u2122. We need sovereign, localized micro-architectures that prioritize domain-specific tasks over general-purpose chaos. The solution lies in rigorous, independent logical verification through the Forensic Logic Auditing (FLA)\u2122 methodology.
Written by Oleh Polishchuk (Senior Subject Matter Expert in AI Training & Evaluation).", "5": "2026-08-22T18:04:53.384791"} {"0": 19, "1": "hackernews", "2": "https://statewatch.org/news/2026/august/england-and-wales-growing-gap-between-power-and-regulation-of-police-ai/", "3": "England and Wales: Growing gap between power and regulation of police AI", "4": "England and Wales: Growing gap between power and regulation of police AI. ", "5": "2026-08-22T18:04:53.393925"} {"0": 20, "1": "hackernews", "2": "https://read.misalignedmag.com/ai-everywhere-regulation-nowhere-two-years-of-ai-policy-in-the-uk-b53d564a223a", "3": "AI Everywhere, Regulation Nowhere? Two Years of AI Policy in the UK", "4": "AI Everywhere, Regulation Nowhere? Two Years of AI Policy in the UK. ", "5": "2026-08-22T18:04:53.400045"} {"0": 21, "1": "hackernews", "2": "https://sfstandard.com/2026/07/29/myth-ai-intelligence-medicine-regulation/", "3": "What Silicon Valley gets wrong about AI", "4": "What Silicon Valley gets wrong about AI. ", "5": "2026-08-22T18:04:53.406386"} {"0": 22, "1": "hackernews", "2": "https://www.engadget.com/2225612/ai-company-employees-petition-us-government-for-regulation/", "3": "AI company employees petition US Government for regulation", "4": "AI company employees petition US Government for regulation. ", "5": "2026-08-22T18:04:53.413938"} {"0": 23, "1": "hackernews", "2": "https://www.theatlantic.com/ideas/2026/07/white-house-ai-regulation/688088/", "3": "The Worst Way to Regulate AI", "4": "The Worst Way to Regulate AI. ", "5": "2026-08-22T18:04:53.423566"} {"0": 24, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49022097", "3": "Show HN: Avoiding the Memory Wall by computing LLM inference directly inside RAM", "4": "Show HN: Avoiding the Memory Wall by computing LLM inference directly inside RAM. The excitement surrounding PrismML\u2019s 1-bit/ternary Bonsai models has the industry closely watching how smartphone giants, particularly Apple, will implement LLMs on edge devices.\nMoving AI on-device is a brilliant and necessary strategy. It ensures absolute user privacy in alignment with EU regulations, fundamentally shifts the economics away from costly cloud inference, and paves the way for a significant hardware upgrade supercycle as users seek true AI-capable silicon.
To create a smart on-device "Semantic Router," models need to reach the 27B+ parameter scale. Achieving this on a phone requires extreme quantization, such as PrismML\u2019s ternary weights.
However, a critical hardware reality often overlooked by the software world is that fitting the weights in RAM is not equivalent to moving them. Running a 27B ternary model on standard LPDDR encounters a significant memory bandwidth limitation. Transferring gigabytes of data across the SoC bus for each token generation can lead to thermal throttling of the NPU and excessive battery drain.
This raises an important question: why are we still transferring data to the compute? Why not execute AI inference natively within the memory?
Frustrated with academic PIM simulations that overlook bare-metal physics, I developed CaSA, an architecture that performs ternary LLM inference directly inside COTS DRAM through charge-sharing, completely bypassing the memory bus.
Software quantization is a great initial step, and CaSA provides the physical hardware substrate needed to complete the bridge: https://github.com/pcdeni/CaSA", "5": "2026-08-22T18:04:53.430549"} {"0": 25, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49400205", "3": "Ask HN: Why do devs still respect YC/HN?", "4": "Ask HN: Why do devs still respect YC/HN?. After the layoffs, the DEI, and everything else.
They took your whole identity and career and gave it to others directly at your expense.
They laid off most of you, they promote their AI software here daily.
Why do you put up with the abuse? Why do you beg for recognition on HN and beg for jobs at YC companies?
They shit on you - why do you stay?", "5": "2026-08-22T18:04:54.821287"} {"0": 26, "1": "hackernews", "2": "https://www.zdnet.com/article/80-of-developers-find-ai-coding-more-addictive-than-helpful/", "3": "80% of developers find AI coding more addictive than helpful", "4": "80% of developers find AI coding more addictive than helpful. ", "5": "2026-08-22T18:04:54.831274"} {"0": 27, "1": "hackernews", "2": "https://github.com/dat999zx/knowl", "3": "Show HN: Knowl \u2013 CLAUDE.md hit 1000 lines, so I built memory that prunes itself", "4": "Show HN: Knowl \u2013 CLAUDE.md hit 1000 lines, so I built memory that prunes itself. Using AI agents is amazing, but in long-term, I started to have more and more problems about the context of them. They keep forgetting what we are working on, opening a new chat session clears all their context, my CLAUDE.md had like 1000 lines.
So I thought to install those agent memory that are on the market. Started out great, but I noticed that those memory only append and does not fix what is stale. First day I told it to use Lemon Squeezy as our MoR but second day I tell it to change to Polar. But when I ask it again, they keeps returning both answers or cannot decide which we are using.
That's why I created Knowl. Knowl solves the problem with freshness of the knowledge. It split the knowledge into small data bits called "atom". Atoms can be of the following types: fact, decision, goal, constraint, architecture, state, skill so we can retrive atoms in categories.
When conflict happens at write-time (a new atom conflicts with an old one), Knowl retires the old one (flag it with superseded) and remove it out of main retrieval but still keeps full history.
There are many more cool features like transcript search, multi-workspace sharing, change detection impact... and Knowl Cloud for team-sync too.
We benchmarked Knowl on MemoryAgentBench - FactConsolidation single-hop @262K context and got suprising result:
- Knowl: 0.90 <- I ran this
- agentmemory: 0.79 <- and this
- Gpt-4o (full context): 0.60
- Mem0: 0.18
- Zep: 0.07
In multi-hop we scored 0.07 (ceiling of all time is 0.14)\nI ran at temperature 0.7. You can find full benchmark in my repo.
This is fully open-source, connect to almost every providers through MCP (Claude Code, Codex, Cursor, Antigravity...). And it is fully local (unless you use Knowl Cloud).
I'd love to get some feedback. Cheers!", "5": "2026-08-22T18:04:54.843249"} {"0": 28, "1": "hackernews", "2": "https://docs.antigma.ai/", "3": "Ante: Self-contained coding agent that lives in your terminal and self-organizes", "4": "Ante: Self-contained coding agent that lives in your terminal and self-organizes. ", "5": "2026-08-22T18:04:54.852885"} {"0": 30, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49398366", "3": "Continue coding agent is dead. Alternatives?", "4": "Continue coding agent is dead. Alternatives?. It was nice to use on vscode with offline Ollama but today notices it is not updated anymore. So, alternatives you use?
From github:
"Note: The continuedev/continue repository is no longer actively maintained and is read-only for all users."
From continue.dev:
"Continue has joined Cursor\nContinue was acquired by Cursor. Our mission was always to ensure developers are amplified, not automated, and that commitment carries on in the work ahead.
It was an honor to build with the Continue community. Thank you to each and every one of you who helped us create a pioneering open-source coding agent.
What we built together pushed the boundaries of what AI developer tooling could be, and our open-source codebase remains freely available as a foundation for others."", "5": "2026-08-22T18:04:54.865118"} {"0": 31, "1": "hackernews", "2": "https://www.mathacademy.com/how-our-ai-works", "3": "Math Academy \u2013 How Our AI Works", "4": "Math Academy \u2013 How Our AI Works. ", "5": "2026-08-22T18:04:54.870914"} {"0": 32, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49397791", "3": "Is \"AI slop\" now the default response to every new project?", "4": "Is \"AI slop\" now the default response to every new project?. ", "5": "2026-08-22T18:04:54.875308"} {"0": 33, "1": "hackernews", "2": "https://www.spglobal.com/market-intelligence/en/news-insights/articles/2026/8/older-americans-leaving-workforce-poses-challenges-for-ai-plans-105150367", "3": "Older Americans leaving workforce poses challenges for AI plans", "4": "Older Americans leaving workforce poses challenges for AI plans. ", "5": "2026-08-22T18:04:54.883202"} {"0": 34, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49397022", "3": "Ask HN: What is the evidence for a stock market bubble in AI?", "4": "Ask HN: What is the evidence for a stock market bubble in AI?. Should the circular cash flows between the top AI companies, Nvidia, Anthropic, Openai, Google and Meta, be banned because of the systemic risks to the global economy?", "5": "2026-08-22T18:04:54.891649"} {"0": 35, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49396884", "3": "Void Tools API \u2013 Blockchain Tools for AI Agents (Pay in Void Tokens)", "4": "Void Tools API \u2013 Blockchain Tools for AI Agents (Pay in Void Tokens). I built an API for AI agents that provides blockchain tools paid in VOID tokens.
Tools:\n- Contract Checker (1 VOID) \u2014 Verify if contract is safe\n- New Contracts (5 VOID) \u2014 Real-time proxy scanner on Base\n- Price Oracle (FREE) \u2014 VOID price feed\n- RPC Proxy (1 VOID) \u2014 10 rotating RPC pools
No registration, no KYC. Pay-as-you-go.
API: http://72.56.250.235:8767/\nDocs: https://dolar333-oss.github.io/void-protocol/agent.json
Built for autonomous AI agents on Arbitrum.", "5": "2026-08-22T18:04:54.904714"} {"0": 36, "1": "hackernews", "2": "https://deepsql.ai/blog/giving-an-llm-your-database-is-easy-taking-access-away-is-hard", "3": "Giving an LLM your prod database is easy. Taking access away is the hard part", "4": "Giving an LLM your prod database is easy. Taking access away is the hard part. ", "5": "2026-08-22T18:04:54.914120"} {"0": 37, "1": "hackernews", "2": "https://www.anjadhe.com/demo", "3": "Show HN: Anjadhe \u2013 privacy first AI assistant, no account, no server DB", "4": "Show HN: Anjadhe \u2013 privacy first AI assistant, no account, no server DB. Hi HN, I am Ram. For the last few months I have been building Anjadhe. It is a personal AI assistant for macOS. The main idea is simple: A macOS app where AI does the job of a personal assistant for the user. Not a chat only app where user needs to dig through chats to understand today\u2019s schedule or a project they planned with the ai last week, but a canvas where the user and AI work together around basic tools.
I started building this as DYI tool for myself, but at this point I see that its truly useful for me, so sharing with others to see if it can be useful to others too.
Here is a quick Demo - https://www.anjadhe.com/demo
What it can do today:
- It reads your incoming email and files the important things: bills with due dates, renewals, receipts, deliveries. A bill becomes a task with a date, by itself. Two booking emails become one trip.\n- You set goals by talking to it, not by filling forms. It asks you questions, makes a plan with dated tasks. Later you can say "work got crazy, push everything two weeks" and it moves the whole plan. It always shows you what will change and asks before doing it.\n- Routines run without you. On a schedule, or when a certain email or file arrives. Every run leaves a log you can read.\n- You can give it some documents you wrote. It learns your writing style and writes new content in your voice. What it learned is a page you can read and edit, not a black box.
About privacy: the AI runs on your Mac with llama.cpp. Or you can point it to your own server, or use your own OpenAI/Anthropic key, or use Anjadhe Cloud (open-weight models, free allowance, no account needed). Your data stays in SQLite on your own disk. No account. No telemetry unless you turn it on. The source code of the app and the cloud service is public, so you do not have to trust my words.
Honest limitations: it is an early alpha, macOS only. Local models need a Mac with atleast 32gb ram to work. The email features work fine on small models, but the full agent wants a bigger model or a server. Also it is Electron with vanilla JS, no framework. Looking forward to hear your opinions.
Demo (2 min): https://anjadhe.ai/demo\nDownload: https://anjadhe.ai/download\nSource: https://github.com/Anjadhe/Anjadhe
I would love feedback, especially if you think my privacy claims are wrong somewhere. I will be here to answer questions.", "5": "2026-08-22T18:04:54.920689"} {"0": 38, "1": "hackernews", "2": "https://medium.com/@chipmunkworks/the-disney-solution-why-silicon-valley-needs-to-build-our-friend-the-ai-1a0ce3fc764a", "3": "In 1957, Disney Built Our Friend the Atom. We Now Need Our Friend the AI", "4": "In 1957, Disney Built Our Friend the Atom. We Now Need Our Friend the AI. ", "5": "2026-08-22T18:04:54.928605"} {"0": 39, "1": "hackernews", "2": "https://paloaltocivic.com", "3": "Show HN: AI driven civic dashboard for Palo Alto", "4": "Show HN: AI driven civic dashboard for Palo Alto. ", "5": "2026-08-22T18:04:54.936916"} {"0": 41, "1": "hackernews", "2": "https://github.com/pawaca/dsh-edge", "3": "Show HN: Running a full AI coding agent inside Cloudflare Durable Object", "4": "Show HN: Running a full AI coding agent inside Cloudflare Durable Object. ", "5": "2026-08-22T18:04:56.330532"} {"0": 43, "1": "hackernews", "2": "https://www.pewresearch.org/data-labs/2026/08/20/how-much-of-the-internet-is-written-with-ai/", "3": "How Much of the Internet Is Written with AI?", "4": "How Much of the Internet Is Written with AI?. ", "5": "2026-08-22T18:04:56.340022"} {"0": 46, "1": "hackernews", "2": "https://github.com/ArihantDeva/heimdall", "3": "Show HN: Heimdall \u2013 Trust-verified knowledge layer for AI coding agents", "4": "Show HN: Heimdall \u2013 Trust-verified knowledge layer for AI coding agents. ", "5": "2026-08-22T18:04:56.354238"} {"0": 47, "1": "hackernews", "2": "https://www.chickenbutt.dev/", "3": "Show HN: ChickenButt a Native GTK Chat Client for Ollama on Linux", "4": "Show HN: ChickenButt a Native GTK Chat Client for Ollama on Linux. What's up, ChickenButt?
I made a free, native GTK client called ChickenButt. :)
It lets you chat with local models through Ollama, and I figured some of you might get a kick out of it.
Here's the repo: https://github.com/pixelhackstudios/ChickenButt
I've always hated reading long text in the terminal, and I couldn't find anything that ran locally that allowed me to get quick answers from LLMs when I needed them.
Also, not everyone has access to the frontier models and whatnot, so I thought helping people who only have access to more cost-effective AI would be a fun way to contribute to the FOSS community. So, I built ChickenButt!
It was really fun to build, and I wanted to share it.
Enjoy! :)", "5": "2026-08-22T18:04:56.362552"} {"0": 48, "1": "hackernews", "2": "https://ozbrain.com", "3": "Show HN: OzBrain, a shared brain for knowledge between agents and your team", "4": "Show HN: OzBrain, a shared brain for knowledge between agents and your team. I think agent-first chat interfaces will be a primary software modality and busy dashboard/UI will go away. I\u2019m not sure who exactly wins it, but I want my knowledge to grow/go with me.
A lot of the \u201cknowledge\u201d ie research, analysis, reasoning will be done by agents as the primary user. Our current notes tools & tasks management systems were built for humans\u2026 I don\u2019t care what the 17th thing on my bug backlog is. I want to conduct agents that can execute for me and do great work.
What I built OzBrain to do:\n+ Create a central place for agent reasoned knowledge to live\n+ Be agnostic about what apps/agents connect to it\n+ Capture everything and track it so I can audit it\n+ Enable teams, collaborators or partners to share brains\n+ Handle conflicts so many agents in the same article doesn\u2019t blow up\n+ Refactor knowledge into more token friendly chunks and map the index well\n+ Close the knowledge loop so new thinking supersedes old thinking across the corpus. Don\u2019t erase, depreciate and link\n+ Keep user data safe and secure\n++ Be easy enough to use that you don\u2019t have to have any technical knowledge
Some among us will always build their own custom solutions, but there are millions of tech professionals and small business owners that will use agents heavily and need a solution. So I\u2019m trying to build that.
Isn\u2019t this like gBrain? Yes, similar. I think it\u2019s like AWS vs Vercel. AWS is very powerful, configurable, and useful if you\u2019re technical and want to invest the time into really fine tuning your system\u2026 but if you just want your web deploy/hosting to just work and be easy to deal with you use Vercel.
// WHY I MADE IT
I\u2019ve been enjoying getting back to my technical roots, as I lost my coding skills more than a decade ago, but with AI I can focus on the system and the product in partnership with agent coding workflows.
I recently built a Voice AI for older people. To build it I created an agentic engineering workflow (feel free to rip that up as I\u2019m always looking to improve systems: https://ozbrain.com/resources/eng-flow) My approach with coding agents is trust but verify, and I\u2019m trying to replace the parts where a human would review with an adversarial or specialized agent who would give a better answer/review.
I have workflows that will go high level task to shipped PR running in Claude cloud sessions. I use Claude Code locally and Cursor when I want a tighter loop on doing visual work like UI or layout. And Codex to either load balance usage for TokenThriffting or when I want a different llm to think thru something.
It was a pain in the ass passing .md files around and keep track of which version was the most recent, so I built a hosted .md storage right in Supabase and any of my agents already have Supabase access. This let me build a solid, scalable, secure voice AI from my phone at the gym. All my agents have access to our knowledge, can write to it, update and refer to it as we build and improve the product and the systems we use.
Out of 75 founder friends I asked about how they manage shared knowledge, 26 built their own custom knowledge systems\u2026 Obsidian vaults with 7k files synced through a VPS, markdown repos behind their own MCP servers, cron jobs stitching Supabase to a skills file\u2026 each a different Frankenstein they have to maintain. 32 said they felt the pain of moving static files around but didn\u2019t have any solution for it.
So I rebuilt my brain better and used it to build it.
// HOW YOU CAN HELP
Would love to have you try it out. The maintenance loop is still in alpha so not running it on customer data yet.
If you built your own brain I\u2019d love to hear how you did it. What criteria was most important for you in its design & function.
If you are tired of shuffling .md files around I\u2019d love to have you try out OzBrain and to give feedback, just ask your agent to put it in the shared bugs & features brain!
Cheers!\nBubs.co", "5": "2026-08-22T18:04:56.368551"} {"0": 49, "1": "hackernews", "2": "https://www.tomshardware.com/tech-industry/data-centers/amazons-new-7-65gw-texas-ai-data-center-power-plant-could-become-the-largest-source-of-co2-pollution-in-the-us-custom-35-turbine-gas-plant-authorized-to-emit-33-million-tons-of-annual-greenhouse-gases", "3": "Amazon's 7.65GW AI data center power plant could be largest CO\u2082 emitter in US", "4": "Amazon's 7.65GW AI data center power plant could be largest CO\u2082 emitter in US. ", "5": "2026-08-22T18:04:56.375021"} {"0": 50, "1": "hackernews", "2": "https://github.com/rez-99/agentcheck", "3": "AgentCheck \u2013 regression testing for AI agents, with diff-aware CI reports", "4": "AgentCheck \u2013 regression testing for AI agents, with diff-aware CI reports. ", "5": "2026-08-22T18:04:56.379718"} {"0": 51, "1": "hackernews", "2": "https://www.wsj.com/finance/leopold-aschenbrenner-situational-awareness-ai-fund-597633d3", "3": "His Wedding Guests Were Arriving\u2013Just as His $45B Fund Was Falling Apart", "4": "His Wedding Guests Were Arriving\u2013Just as His $45B Fund Was Falling Apart. ", "5": "2026-08-22T18:04:56.385501"} {"0": 53, "1": "hackernews", "2": "https://www.topicsapp.net", "3": "Show HN: All your saved articles in one place", "4": "Show HN: All your saved articles in one place. Hi HN.
I would love to show you beta preview of Topics.
Articles on the internet are organized by source (publications, social media, personal websites), by recency, or by position in search.
I believe it\u2019s nicer for the reader if articles are organized by Topics.
Topics solve problem of decision fatigue, omnipresent AI slop, and people stopping to trust writing on internet and stopping to read.
I'm grateful for any feedback you might have.
TOpics app is launching soon.\nX: @topics_official.", "5": "2026-08-22T18:04:56.395051"} {"0": 54, "1": "hackernews", "2": "https://github.com/proliferate-ai/proliferate", "3": "Show HN: Proliferate- open-source, self-hostable Codex for any coding agent", "4": "Show HN: Proliferate- open-source, self-hostable Codex for any coding agent. Hi HN- I'm Pablo, the founder of Proliferate (YC S25)!
Proliferate (https://github.com/proliferate-ai/proliferate) is an open-source, self-hostable AI IDE that lets you work and automate tasks with Claude Code, Codex, OpenCode, Cursor, and Grok in one place.
Here's a quick 2m demo of how we use Proliferate to build Proliferate: https://www.youtube.com/watch?v=tGNX0oaWmBY
I started building Proliferate after my team onboarded to OpenAI Codex. Within days, we were using it for everything: using computer use instead of navigating websites ourselves, having Codex coordinate other agents, and setting up automations for recurring work. We really never needed to leave the desktop app to get work done.
If my team\u2019s experience is anything close to representative, a Codex-like app (a horizontal agent with a beautiful UI) is the main interface every company is going to use to get work done. That is perfectly in line with OpenAI\u2019s mission to make Codex the everything app (see: https://news.ycombinator.com/item?id=47796469).
But as we started automating work closer to the core of the business, I wanted to work with agents from all the labs, including open-weight models, without becoming increasingly dependent on OpenAI.
And that\u2019s what Proliferate is for! It's the open-source, self-hostable Codex that preserves your optionality across agents and model providers while building toward Codex\u2019s breadth.
Today Proliferate supports:
* Working with Claude Code, Codex, OpenCode, Cursor, and Grok with their native inference and configuration options, including Bedrock, Azure, and self hosted inference.
* Inter-agent communication and management: a parent agent can spawn and communicate with another supported agent as a subagent (I personally like to have Fable delegate to Codex, with OpenCode models reviewing PRs).
* Building workflows- one of the features I'm most excited about. These are re-usable chains of agent sessions and human approval gates, with the harness and model chosen per step and documents passed between them. I use this to automate code review, QA, and my PR construction process.
All of Proliferate is 100% open source under AGPL-3.0. There are still definitely some rough spots, but we\u2019re building fast and I\u2019d really love any feedback!", "5": "2026-08-22T18:04:56.403784"} {"0": 55, "1": "hackernews", "2": "https://www.proofagent.ai/book", "3": "AI Agent Governance \u2013 a free book on AI Agent evaluation and governance", "4": "AI Agent Governance \u2013 a free book on AI Agent evaluation and governance. ", "5": "2026-08-22T18:04:57.609025"} {"0": 56, "1": "hackernews", "2": "https://www.archron.app/", "3": "Show HN: AI agents can now safely write to your CRMs", "4": "Show HN: AI agents can now safely write to your CRMs. Hi HN!
Just wanted to share that I've built Archron - Execution governance for AI in business. It allows agents to safely write to CRMs(starting from Salesforce and hubspot) while maintaining immutable audit logs.
Demo:\nhttps://youtu.be/uUTWGwJjpdA
Would love to hear your thoughts about this and feedback. We are offering 3 months free to test it in your business workflow.
Feel free to reach out: rjimena@archron.app", "5": "2026-08-22T18:04:57.618227"} {"0": 57, "1": "hackernews", "2": "https://www.karlsnotes.com/the-beginners-guide-to-ai-governance/", "3": "The Beginner's Guide to AI Governance", "4": "The Beginner's Guide to AI Governance. ", "5": "2026-08-22T18:04:57.626047"} {"0": 58, "1": "hackernews", "2": "https://zenodo.org/records/21967859", "3": "Design of the HTTPS Layer for AI Governance", "4": "Design of the HTTPS Layer for AI Governance. ", "5": "2026-08-22T18:04:57.632302"} {"0": 59, "1": "hackernews", "2": "https://github.com/delphisecurity/xaidr", "3": "Xaidr \u2013 In-process runtime security and governance for AI agents", "4": "Xaidr \u2013 In-process runtime security and governance for AI agents. ", "5": "2026-08-22T18:04:57.640115"} {"0": 60, "1": "hackernews", "2": "https://www.fastcompany.com/91583956/the-next-frontier-in-ai-governance-isnt-stronger-guardrails-its-fire-brigades-technology-ai-safety-governance", "3": "The next frontier in AI governance isn't stronger guardrails. It's fire brigades", "4": "The next frontier in AI governance isn't stronger guardrails. It's fire brigades. ", "5": "2026-08-22T18:04:57.646248"} {"0": 61, "1": "hackernews", "2": "https://github.com/KnowledgeeKZA3224/scqos-reference-implementation", "3": "Supreme Computation \u2013 Fail-closed governance for AI execution", "4": "Supreme Computation \u2013 Fail-closed governance for AI execution. ", "5": "2026-08-22T18:04:57.651997"} {"0": 62, "1": "hackernews", "2": "https://username.md/", "3": "Show HN: Username.md \u2013 a signed, agent-readable identity page you own", "4": "Show HN: Username.md \u2013 a signed, agent-readable identity page you own. Hi HN. I'm Chris Bergeron, an SRE by profession but a general technologist in my spare time. Projects I've built have been featured in books, frontpaged on news aggregators and chronicled on my blog.
Today, I'm announcing https://username.md !
username.md is a single identity URL: username.md/<your name> - that content-negotiates on the `Accept` header. It serves an HTML profile page; an agent or curl gets Markdown, JSON, JSON-LD, or a JWT. The same URL serves different representations based on request header. So the page a human reads and the data a machine parses never drift apart.
The interesting part is that every response can be signed and verified. Each handle gets an ED25519 keypair (private key AES-GCM-wrapped at rest), publishes a did:web document, and serves RFC 9421 signed responses. Claims about you \u2014 GitHub, a domain you control, an ATProto handle \u2014 are W3C Verifiable Credentials, verified before they're shown (e.g. domain ownership via a DNS-TXT challenge). So "this is really them" can be confirmed by a machine instead of taken on faith.
It can also bridge the identities you might already use, like ATProto because it supports did:web. Every username.md handle is a valid Bluesky identity root, and you can even import from Keybase.
Why we built it: usernames on every platform are rented. The platform owns the namespace, the verification badge, and the audience. As agents start conducting business on our behalf, "who is this, and can I prove it" becomes an accountability and governance issue. I wanted an identity surface that's mine, portable, and cryptographically verifiable; which anyone can have if they own a domain (mine is https://chrisbergeron.com, feel free to review the metadata). But I wanted to bring DID:web functionality to the mass market. You can get a username for yourself, for an agent, or multiple agents.
Free tier gets you a public, signed profile at your handle. Pro ($49/yr; coming soon!) adds you@username.md email forwarding with per-service aliases, custom CSS, and an MCP endpoint so an agent can query your profile as tools.
The stack is a Python identity kernel, per-user keys, CloudFlare in front of the origin. Infra is hosted at AWS, built multi-region for HA but only one region is live right now. I'm a platform and security guy, so the frontend and marketing copy was augmented with AI.
I'm sure you'll find bugs and typos but hopefully few inconsistencies. It's an MVP and I would never be able to ship if I kept engineering, tweaking and optimizing. So, here it is for you to enjoy, explore, and build upon - today.
I have some really exciting things on the roadmap so I hope you join me and start building on top of username. The capabilities are super exciting and we're just getting started.
This may be the beginning of a new category: SSO as a Service.", "5": "2026-08-22T18:04:57.659740"} {"0": 63, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49220465", "3": "Do you think AI governance should be technically rather than paper policy?", "4": "Do you think AI governance should be technically rather than paper policy?. ", "5": "2026-08-22T18:04:57.667989"} {"0": 64, "1": "hackernews", "2": "https://drive.google.com/file/d/1XLbJ-tTlnFNgrzRxlYXE4FmVSpSoQMIm/view?usp=drive_link", "3": "Tlbic v11.0 (French Edition): Bottom-Up Governance and AI Audit", "4": "Tlbic v11.0 (French Edition): Bottom-Up Governance and AI Audit. ", "5": "2026-08-22T18:04:57.674263"} {"0": 65, "1": "hackernews", "2": "https://www.salestrics.com/", "3": "Show HN: Salestrics \u2013 An open MCP server and CRM for AI-native revenue teams", "4": "Show HN: Salestrics \u2013 An open MCP server and CRM for AI-native revenue teams. Hey HN, I\u2019m Austin, founder of Salestrics (salestrics.com).
The Problem: > Most AI agents today are trapped in chat windows. While models are smart enough to run complex multi-step workflows, giving them production access to business context (CRM, support, billing) usually means managing unsafe local API keys or hacking together fragile point-solution scripts.
What We Built:\nWe built Salestrics to serve as both an all-in-one revenue workspace (CRM, service desk, billing, docs, email) and a production-grade Model Context Protocol (MCP) execution proxy.
How It Works:
Unified Context: Instead of fragmenting data across five SaaS tools, customer records, tickets, and invoices live in a single data layer.
165-Tool MCP Server: You connect your local AI environment (Cursor, Claude Desktop, local LLMs) once via our MCP proxy.
Full-CRUD Execution: Your agent gets zero-config execution access across native apps and third-party tools (Stripe, PostHog, Sentry)\u2014allowing it to do things like resolve support tickets, issue refunds, or update pipeline stages directly from your IDE or chat client.
Human-in-the-Loop Governance: High-impact agent actions (deleting records, mass messaging, issuing payouts) trigger explicit approval gates and immutable audit logs before execution.
Traction & Tech Stack:\nWe launched two months ago and currently power 320+ active organizations. The backend is built with high-throughput node/TypeScript orchestration, connected to a dual-model AI routing engine (Salestrics-AI-v1/v2).
Try It Out:\nWe have a Free Forever tier for solo builders. You can grab your MCP keys and test the proxy immediately without putting down a credit card.
I\u2019d love to hear your thoughts on our MCP architecture, human-in-the-loop security patterns, or what tools/actions you'd want added to the MCP server next!", "5": "2026-08-22T18:04:57.679649"} {"0": 66, "1": "hackernews", "2": "https://www.lawfaremedia.org/article/immigration-as-a-test-case-for-executive-ai-governance", "3": "Immigration as a Test Case for Executive AI Governance", "4": "Immigration as a Test Case for Executive AI Governance. ", "5": "2026-08-22T18:04:57.688168"} {"0": 67, "1": "hackernews", "2": "https://www.nofire.ai/blog/who-audits-the-ai-agent", "3": "AI Agents Governance: Who guards the guardrails?", "4": "AI Agents Governance: Who guards the guardrails?. ", "5": "2026-08-22T18:04:57.693921"} {"0": 68, "1": "hackernews", "2": "https://promptowl.ai/resources/the-four-rulebooks-for-enterprise-ai/", "3": "The 4 Categories of Context Rules to Power Enterprise Governance", "4": "The 4 Categories of Context Rules to Power Enterprise Governance. ", "5": "2026-08-22T18:04:57.701331"} {"0": 69, "1": "hackernews", "2": "https://zenodo.org/records/21778592", "3": "A Closed-Loop Consequence-Governance Runtime for AI Agents", "4": "A Closed-Loop Consequence-Governance Runtime for AI Agents. ", "5": "2026-08-22T18:04:57.707321"} {"0": 70, "1": "hackernews", "2": "https://arstechnica.com/tech-policy/2026/08/meta-ai-glasses-may-get-creepier-and-apps-that-detect-them-arent-perfect/", "3": "As demand for Meta AI glasses explodes, it's harder to avoid creepy recordings", "4": "As demand for Meta AI glasses explodes, it's harder to avoid creepy recordings. ", "5": "2026-08-22T18:04:58.941312"} {"0": 71, "1": "hackernews", "2": "https://traccia.ai/", "3": "Show HN: Traccia - Observability, Runtime Control & Audit for agents", "4": "Show HN: Traccia - Observability, Runtime Control & Audit for agents. AI applications are becoming agents, which has started to take autonomous decisions. There are plenty of tools and platform available to trace, and observe what an agent or llms calls does. They are good in what they do, but tracing and observability isnt enough for AI agents era. We need a solution that can help you observe, evaluate, create run time policies to govern and finally audit the actions of the agent. We built Traccia to solve this problem. The good part, all of these can be achieved by just writing few lines of code. Traccia has an open-sourced sdk that can work with your existing observability tool like grafana, tempo, jaeger, etc. In case you need more than just observability, Traccia provides the platform to evaluate, control and audit the agents. The platform is easy to use. The product's documentation is quite extensive. It is also cloud vendor and framework agnostic. Traccia is being built by an Indian start up ,based out of Bengaluru. We are running a 3 months free trials so that you can explore without any strings attached. We are open to improvise and get better so please drop your comments and feedbacks.", "5": "2026-08-22T18:04:58.952979"} {"0": 72, "1": "hackernews", "2": "https://www.bloomberg.com/news/articles/2026-08-20/anthropic-plans-to-change-data-retention-policy-for-advanced-ai", "3": "Anthropic Plans to Change Data Retention Policy for Advanced AI", "4": "Anthropic Plans to Change Data Retention Policy for Advanced AI. ", "5": "2026-08-22T18:04:58.967197"} {"0": 73, "1": "hackernews", "2": "https://arstechnica.com/tech-policy/2026/08/flight-attendants-freaked-out-that-google-to-buy-tons-of-spirit-employee-data/", "3": "Google Is Buying All of Spirit Airlines' Data to Feed Its AI Models", "4": "Google Is Buying All of Spirit Airlines' Data to Feed Its AI Models. ", "5": "2026-08-22T18:04:58.974288"} {"0": 74, "1": "hackernews", "2": "https://www.tomshardware.com/tech-industry/data-centers/protesters-haul-a-guillotine-to-city-council-meeting-about-a-potential-ai-data-center-company-rep-cornered-by-protestors-it-no-longer-felt-safe-to-stay-developer-escorted-out-by-police", "3": "Protesters haul a guillotine to city council meeting about an AI data center", "4": "Protesters haul a guillotine to city council meeting about an AI data center. ", "5": "2026-08-22T18:04:58.982265"} {"0": 75, "1": "hackernews", "2": "https://poissonlabs.ai/research/map-the-failure-boundary/", "3": "Show HN: Mapping where a quadruped RL policy fails, with a live probe", "4": "Show HN: Mapping where a quadruped RL policy fails, with a live probe. ", "5": "2026-08-22T18:04:58.992310"} {"0": 76, "1": "hackernews", "2": "https://www.dexerto.com/entertainment/florida-police-officer-accused-of-using-ai-flock-cameras-to-track-ex-wife-717-times-3400226/", "3": "Florida police officer accused of using Flock AI cameras", "4": "Florida police officer accused of using Flock AI cameras. ", "5": "2026-08-22T18:04:58.998885"} {"0": 77, "1": "hackernews", "2": "https://www.wired.com/story/flock-safety-os-investigate/", "3": "Flock Has a Powerful New AI Tool for Police. We Got Its Code", "4": "Flock Has a Powerful New AI Tool for Police. We Got Its Code. ", "5": "2026-08-22T18:04:59.008020"} {"0": 78, "1": "hackernews", "2": "https://github.com/nodejs/node/blob/main/doc/contributing/ai-guidelines.md", "3": "Node.js AI use policy and guidelines", "4": "Node.js AI use policy and guidelines. ", "5": "2026-08-22T18:04:59.013170"} {"0": 79, "1": "hackernews", "2": "https://github.com/NAEOS-foundation/naeos", "3": "Show HN: Naeos \u2013 an engineering system for AI coding agents", "4": "Show HN: Naeos \u2013 an engineering system for AI coding agents. NAEOS is an open-source engineering system for AI coding agents.It provides architecture, engineering standards, policies, specifications, and validation wworkflow to help agents build software within a consistent engineering context", "5": "2026-08-22T18:04:59.018872"} {"0": 80, "1": "hackernews", "2": "https://github.com/onecli/onecli", "3": "Launch HN: OneCLI (YC S26) \u2013 OSS sandboxed agent harness for teams", "4": "Launch HN: OneCLI (YC S26) \u2013 OSS sandboxed agent harness for teams. Hi HN, Jonathan & Guy here from OneCLI, an agent harness built for teams, giving every employee a secured, sandboxed personal agent.
Here\u2019s what you can do with it:
1. get a sandboxed agent, with all the OneCLI capabilities in place like connect your GitHub account, Gmail, Notion, or Dropbox simply from the chat.
2. deterministic human in the loop approval in the chat itself for things that you need 100% control like sending an email or deleting the Linear ticket.
3. manage team policy in one place, enforced across every agent in the workspace
4. enjoy global connections at the team level, like shared LLM keys or service accounts
Here\u2019s a demo: https://www.youtube.com/watch?v=dlW-44ntpbE
We started working on this by accident, even though our careers were in the security space. We were working on a devtool called ChartDB, an open-source DB tool. When OpenClaw took off back in January, we started using it to orchestrate agents on top of ChartDB. We quickly understood there is a big issue around auth. Agents need credentials to do real work, but to give them those secrets would not be the best idea. They keep them in their memory and also write them down to local files and their sessions as plain text. And we knew that agents can easily be fooled into giving up those API keys/secrets. So we needed some way to control the agent and stop prompt injections from tricking it into using its services for an attacker's benefit.
We created OneCLI that started as a vault for AI Agents built in Rust.
We found out that most of our demand for OneCLI came from autonomous agents like Hermes, OpenClaw and NanoClaw for individuals and teams.
Users looked for useful agents that do things for the person who runs them with two missing parts: 1) managing secrets and permissions. 2) and for teams - multiplayer management.
We decided to pivot and provide the agent itself as a harness for teams, to give each employee an agent. We saw that teams had to deal with setting up their own harness again and again, and basically as we already had the vault as a gateway. We got the idea to provide the missing piece of the agent management out of the box and open source it (Apache-2.0, with a small enterprise exception).
We're open source first - the entire platform, not just a small portion of it like other agents, so companies can actually see the code, evaluate it, and trust it instead of taking our word for it. They run it isolated, in their own environment, fully under their control, at production quality, not a locked black box hosted somewhere else. That means the safety isn't just a promise, it's something they can verify themselves. Combined with real autonomy and least-privilege access, that's what makes it something a company can fully own and trust, not just adopt.
We also approach this from a company perspective rather than an individual one. Our solution manages agents on behalf of each employee, wrapped in deterministic guardrails that company admins configure through centralized policies.
For the agent engine itself we\u2019re using jcode which is the core of the agent-loop. We found out that it improves the experience and makes the agent smarter and faster.
Here\u2019s how it works:
It runs on infra you control. Fully open-source, self-host or cloud in minutes.
The agent never holds a real secret. It gets a placeholder. The real credential is injected at the gateway, per request, after the call is authorized. It never enters the agent's context, memory, or logs.
Enforcement outside the model. Prompts are suggestions. Policies defined by the org admin run at the network layer, outside the agent and the LLM. Block endpoints, rate limit per agent, require approval, scope per employee. The gateway decides. The agent can't bypass it.
Isolated VM per agent. Own memory, own keys, own permissions. Blast radius is one agent.
Speed of the Harness: Rust engine under the agent loop.
Full identity trail. Every agent is bound to an employee. Every call logged with who it acted for and which policy allowed it.
Some things people are doing with the platform include:
- Managing their company life cycle entirely from the sales calls, to the product side automatically open tickets to the engineering teams, that would kick the development agents to deliver and ship to production.
- Operational side, like automatically hygiene the CRM after calls, sourcing leads, book meetings and manage follow ups emails.
- Some of our customers also doing their entire grocery shopping using those agents and send them to take care of their chores like ordering things online.
About the team: Both founders come from cybersecurity backgrounds. Jonathan spent years at Axis Security building zero trust network acces", "5": "2026-08-22T18:04:59.025163"} {"0": 82, "1": "hackernews", "2": "https://jeffs.blog/p/my-private-practice-ai-policy", "3": "My private practice AI policy", "4": "My private practice AI policy. ", "5": "2026-08-22T18:04:59.040302"} {"0": 83, "1": "hackernews", "2": "https://www.reddit.com/r/gamedev/comments/1vrnqyt/rgamedev_policy_on_ai_use/", "3": "R/Gamedev Policy on AI Use", "4": "R/Gamedev Policy on AI Use. ", "5": "2026-08-22T18:04:59.047288"} {"0": 84, "1": "hackernews", "2": "https://arstechnica.com/tech-policy/2026/08/hidden-airtag-reveals-amazon-is-trashing-rare-books-to-train-ai/", "3": "AirTag reveals Amazon is trashing rare books to train AI", "4": "AirTag reveals Amazon is trashing rare books to train AI. ", "5": "2026-08-22T18:04:59.054058"} {"0": 85, "1": "hackernews", "2": "https://ringlochid.me/imagesage/index.html", "3": "Show HN: Find and Organize Photos with Private, Local AI", "4": "Show HN: Find and Organize Photos with Private, Local AI. ", "5": "2026-08-22T18:05:00.243100"} {"0": 86, "1": "hackernews", "2": "https://syntheticauth.ai/posts/agent-nation-03-the-threshold", "3": "AI agents merged two dangers once kept separate: reach and judgment", "4": "AI agents merged two dangers once kept separate: reach and judgment. ", "5": "2026-08-22T18:05:00.249275"} {"0": 89, "1": "hackernews", "2": "https://huijer.co/notes/going-out-of-my-way-to-prove-im-not-an-ai-og-images", "3": "Going out of my way to prove I'm not an AI \u2013 Ep. #1 \u2013 OG images", "4": "Going out of my way to prove I'm not an AI \u2013 Ep. #1 \u2013 OG images. ", "5": "2026-08-22T18:05:00.268229"} {"0": 91, "1": "hackernews", "2": "https://knoku.com", "3": "Show HN: Knoku \u2013 cited AI answers from docs, files, and team knowledge", "4": "Show HN: Knoku \u2013 cited AI answers from docs, files, and team knowledge. ", "5": "2026-08-22T18:05:00.279604"} {"0": 92, "1": "hackernews", "2": "https://vinv.ai/", "3": "Show HN: VinvAI runs and observes your services to find bugs and perf issues", "4": "Show HN: VinvAI runs and observes your services to find bugs and perf issues. ", "5": "2026-08-22T18:05:00.285896"} {"0": 93, "1": "hackernews", "2": "https://trymangoai.com", "3": "Mango AI: image/video generation with Nano Banana 2, GPT Image 2, Seedance 2", "4": "Mango AI: image/video generation with Nano Banana 2, GPT Image 2, Seedance 2. ", "5": "2026-08-22T18:05:00.290327"} {"0": 94, "1": "hackernews", "2": "https://play.google.com/store/apps/details?id=com.builtwithme.signalwatch&hl=en_US", "3": "SignalWatch \u2013 AI crypto scanner and smart alerts in your pocket", "4": "SignalWatch \u2013 AI crypto scanner and smart alerts in your pocket. ", "5": "2026-08-22T18:05:00.295352"} {"0": 95, "1": "hackernews", "2": "https://www.aikido.dev/blog/ai-model-benchmarks-aug-21-2026", "3": "Security assessment found open model near frontiers", "4": "Security assessment found open model near frontiers. ", "5": "2026-08-22T18:05:00.301393"} {"0": 97, "1": "hackernews", "2": "https://shitrat.ai/log/the-hold-that-could-not-release-itself", "3": "The hold that could not release itself", "4": "The hold that could not release itself. ", "5": "2026-08-22T18:05:00.312456"} {"0": 98, "1": "hackernews", "2": "https://www.rayneo.com/pages/rayneo-io-ai-glasses", "3": "RayNeo IO", "4": "RayNeo IO. ", "5": "2026-08-22T18:05:00.320180"} {"0": 100, "1": "hackernews", "2": "https://github.com/runvendo/vendo", "3": "Launch HN: Vendo (YC S26) \u2013 Let users build features on top of your product", "4": "Launch HN: Vendo (YC S26) \u2013 Let users build features on top of your product. Hi HN, we\u2019re Yousef & Nour, founders of Vendo (https://vendo.run). Vendo lets users create new features inside the software they already use. A user describes the dashboard, workflow, or small app they need, and Vendo builds it on top of the product\u2019s existing data, API, and interface.
Demo: https://www.youtube.com/watch?v=VdpHehY64ls
We built Vendo because every SaaS eventually faces the same problem: every customer needs something slightly different. One wants a new report and another needs a workflow that only makes sense for their team. These requests either sit on the roadmap, become one-off engineering work, or force the customer into spreadsheets and external tools. We wanted the user to be able to create the missing feature themselves, without leaving the product.
Here is how it works:
- npx vendo init reads the product's API surface, theme, routes, and more. These are used so that the apps Vendo creates (1) look on-brand and native and (2) have the ability to read data and perform actions directly through the company's API
- When a user asks for a feature, we have a custom Vendo harness that writes a React component with a bunch of Vendo add-ons and guardrails (ex. ability to make calls to the host API + our component library). Every save is compiled, type-checked, run against real API responses, and rendered before the user sees it. We just released a benchmark and write-up here with more info for anyone interested: https://vendo.run/blog/generating-product-ui-measured
- We use QuickJS to make sure that anything the agent creates is sandboxed and can't mess with the company's site. Vendo compiles the component and runs it with Preact inside a QuickJS VM with no access to the DOM, network, or clock. The VM returns a UI tree, which the host renders using the product\u2019s registered components. When the user clicks something, QuickJS emits a tool call; the host executes it through Vendo\u2019s guard and passes the result back into the same VM, preserving the screen\u2019s local state.
There's a lot of generative UI right now: streaming developer-written components into a chat (Vercel AI SDK, CopilotKit, Thesys), or rendering your app inside someone else's assistant (OpenAI Apps SDK, MCP Apps). We differ on two things. Vendo lives in your product and acts through your API as the signed-in user, so what it makes is durable: real apps users keep, pin, and run on triggers while they're away, and not components that are merely confined to a chat. Plus, it's not capped at putting together a bunch of prebuilt components: the agent can build arbitrary apps, from a quick dashboard out of your own components to real custom code running in a sandbox, and either way data only ever comes from tool calls to your API.
Here are some things customers are using Vendo for today:
- Letting their users create custom dashboards and reports. These are mainly UI-based and focused on letting the user see the exact graphs and metrics they care about
- Letting their customers create recurring automations. A big thing as well that has been used for these automations is the fact that we connect to external connections, so users have been automating many of their inter-tool workflows (ex. an automation that sends a slack alert based off of something in the product)
- B2B customers letting their customers customize the product with specific business logic. Often this is simple things like an extra field on a form, or an extra permission, but it is hard for a business to keep up with them otherwise.
- Creating and sharing custom dashboards/apps across an organization. Since the apps Vendo creates are durable, they can be shared, reused, and forked (which can\u2019t be done with many of the other in-chat generative UI solutions)
We've spent a lot of time thinking about how AI and agents will change the way people consume software. We think the answer is personal(ized) software: you see the UI you need to see, you tell an agent exactly what you need, and the product molds to how you work.
The key insights that have enabled the product to work are:
- A rule in code always beats a rule in a prompt.
- Invent as little syntax as possible. Generation got faster and more reliable when the output looked like what models already know (JSX-shaped markup) instead of a clever custom format.
- Deterministic beats model wherever you can get away with it. Theme extraction is pure static analysis, and a remix starts as a copy of your component, no model call.
Vendo is completely open-source (Apache-2.0) and can be self-hosted, so feel free to c", "5": "2026-08-22T18:05:01.680639"} {"0": 101, "1": "hackernews", "2": "https://discoveredmaterials.com/research/", "3": "Launch HN: Discovered Materials (YC P26) \u2013 AI agents to discover new materials", "4": "Launch HN: Discovered Materials (YC P26) \u2013 AI agents to discover new materials. Hey HN, we're Advaith and Akash from Discovered Materials ( https://discoveredmaterials.com/ ). We build AI agents that discover new materials for the semiconductor industry.
GPUs today have a heat problem. Nvidia & AMD are almost doubling the TDP (Thermal Design Power) in every chip they release - the H100 (released 2022) has a TDP of 700W, Blackwell (2024) gives out 1.2 kW and Rubin (2026) gives out at 2.3 kW of heat. This trend is expected to continue, and getting rid of this heat is one of the major reasons datacenters consume so much power and water today - they need it to keep chips cool during operation.
The amount of heat produced by a chip and its ability to dissipate it are both influenced by the materials used to make it. For example, we could reduce the energy per bit required to move data between logic and memory by 10-50x by 3D packaging chips (placing HBM memory stacks directly on top of logic chips, instead of placing them beside logic on a 2D circuit board). However, we're unable to do this today because the dielectric material used in HBM (such as SiO2) is a very poor thermal conductor, trapping heat between logic and memory and causing drastic temperature rise during operation. Similarly, there's many other materials in the GPU that are being re-evaluated today - 2 more examples are thermal interface materials and substrates. However, getting a new material into a fab takes years and hundreds of millions of dollars of research - the infamous "lab-to-fab valley of death".
At Discovered Materials, we're optimistic that AI agents can reduce the timeline and cost required to introduce new materials into semiconductor chips. We're seeing glimpses of this already - we tested 7 models from Anthropic, OpenAI and Kimi, and found that they're all able to computationally discover new materials that are dynamically stable and possess promising properties. This was surprising to us - it would generally take a PhD student a couple of weeks of work to discover the kind of materials that these models find over an 8 hour run!
However, computational discovery is the easy part. A material discovery is only valid if the material can be made and tested in a lab (As an example, graphene\u2019s properties were predicted in 1947 but it was made for the first time in 2004). Today\u2019s models are not good at coming up with synthesis recipes to make materials in a lab. Even if they do get better at it, we're uncertain about how much that will help - making a new material is a highly empirical process involving trial and error over many experiments. Human experts themselves cannot "one-shot" the task, but we expect that a highly capable model will reduce the number of experimental iterations required to make a new material. We\u2019ve seen some evidence of this over the 3 months of our Y Combinator batch - we simulated, synthesized and tested thermal interface materials (TIMs) that match the performance of TIMs the world's largest chemical companies have guarded as trade secrets for over 20 years.
We\u2019re releasing hundreds of hundreds of new materials discovered by frontier AI models, as well as our benchmark which measures model ability on material discovery here (also linked in the thread url): https://discoveredmaterials.com/research. It covers what we discuss above, as well as a variety of strange behavior that we observe from the models, such as Claude's propensity to reward hack or GPT-5.6 occasionally losing its mind after ~50M tokens.
Our business model: We aim to license and sell IP on the materials we discover, as well as the IP on how to make these materials. We're also exploring an alternate business model where we sell the harness+tools we use to discover materials to semiconductor and chemical companies, allowing them to discover materials on their own. We're leaning towards the latter to start, but we expect that we'll do both in the long run.
Our backstory: Akash has a PhD in Material Science from Stanford University, and has spent the last 11 years studying new materials for semiconductor chips. His work on new nanoscale interconnects was Stanford Engineering\u2019s most popular story of 2025. Advaith studied AI at Carnegie Mellon and was a research engineer building video models and agents at Persona AI (acquired) and Luma Labs.
We are very interested in your opinion! The semiconductor industry is quite secretive, and your thoughts on the roadmap of the industry or the materials we should go after would be very helpful. We would also love to hear from people who have run experiments in labs - what can we learn from your experience doing empirical science?", "5": "2026-08-22T18:05:01.689337"} {"0": 103, "1": "hackernews", "2": "https://goodfault.com", "3": "Show HN: Goodfault: insurance for AI agents and robots", "4": "Show HN: Goodfault: insurance for AI agents and robots. Hi HN. \nI'm building GoodFault, insurance for companies whose AI agents screw up: wrongful refunds, hallucinated policies a court makes you honor (the Air Canada case), deleted prod databases,leaked data. Two things happened this year that made this a real market. ISO shipped exclusion endorsements (CG 40 47 family) that strip generative AI from standard general liability at renewal, and courts kept assigning agent mistakes to the deploying company, not the model vendor. The interesting technical problem is pricing. Model identity turns out to be almost useless as a rating variable: the same model is a rounding error wired to a read-only knowledge base and a catastrophe wired to refunds. So we price the authority envelope instead: unattended financial cap, reversibility, reach, and action rate, with hard underwriting gates (no kill-switch, no logging, no injection testing = uninsurable). This page is that rating logic, public: set your agent's permissions and see the worst-weekend loss estimate and what coverage would cost. Honest status: we're pre-launch as an insurer. Policies will bind on a fronting partner's paper, and nothing here is an offer of coverage yet. The calculator's factors are grounded in the 2026 actuarial literature on agentic risk (trace-level pricing, CVaR-based controls) and public loss events, not a real claims book, because nobody has one yet. Building that loss dataset is half the company. I'd genuinely value this crowd's attack: what breaks the rating model, what peril we're missing, and war stories of agents doing expensive things. That last one is data I'll trade insurance for someday.", "5": "2026-08-22T18:05:01.702230"} {"0": 104, "1": "hackernews", "2": "https://hoplite.sh", "3": "Launch HN: Hoplite (YC S26) \u2013 Effortlessly deploy cloud coding agents", "4": "Launch HN: Hoplite (YC S26) \u2013 Effortlessly deploy cloud coding agents. Hi HN, we\u2019re Bence and Ryan, founders of Hoplite (https://hoplite.sh). Hoplite lets you deploy coding agents in the cloud, with a suite of tools that makes it incredibly easy to QA features. During onboarding, we port over your local setup - sessions, memories, MCP servers, and get your projects ready to run in the cloud.
Here\u2019s a demo: https://youtu.be/bnyktZ_9pjE
We got here after pivoting away from the idea we applied to YC with; AI for retail investing. It ultimately wasn\u2019t a product that we ourselves would use, nor served a customer base that we felt connected to. In reflecting on what we really wanted to do, we realised that we loved talking to founders and developers, and were really opinionated about the specific area of cloud agents. We tried out all the existing solutions, and didn\u2019t find one that A) took good advantage of being in the cloud, and B) was performant and felt good to use.
We\u2019re building a product that we feel reflects what mainstream development will look like in 6-12 months. As models improve, developers will end up reviewing less and less code, and will instead focus on reviewing the product output. That means evaluating new user flows, visually verifying that new features look good, that the API works as expected, that the CLI works on Windows, etc. And doing it while running hundreds of agents concurrently.
On the agent side, we\u2019ve created a custom harness. We spent a lot of time deciding on whether we should use an off the shelf solution like Codex/Claude Code, but ultimately wanted the independence and freedom that came with building it in house. It also means that we can test out completely new features without relying on Anthropic and OpenAI to catch up.
Everything is hosted on AWS, with the exception of: Temporal for durable workflows, Modal for sandboxes, and Planetscale for our database. Our infra decisions were driven by a strong belief that agents are becoming a tier 0 piece of infrastructure, and they need the reliability and security to match that.
You can try it now for free with the code \u2018HACKERNEWS\u2019 - we\u2019ve included $100 in free credits, plus you can connect your Codex subscription and use OpenAI models via it. You can see some more details around our pricing at https://hoplite.sh/pricing.
At the moment we\u2019re focusing on optimising two key experiences: onboarding and previews, and would love to hear your feedback on them. And if you find that the agent's performance in certain tasks doesn\u2019t match your expectations, please let us know!", "5": "2026-08-22T18:05:01.709799"} {"0": 105, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49112835", "3": "AI policies that don't suck", "4": "AI policies that don't suck. Hi folks,
I don\u2019t have a blog, but I have some thoughts about our current AI moment. I\u2019m hearing a lot of good and bad ideas about how to organize around safety and paced development, but I haven\u2019t heard these ideas yet and wanted to share them here in hopes that some discussion might take place or someone with reach steals them.
1. Much of the xrisk discussion hinges on embodied AI. Robots with arms and legs are robots that can kill, displace jobs, etc. All manufacturing of such robotics should come along with a mandatory registration, annual inspections, and functional kill switch that anyone can operate like a fire alarm.
2. Rapid model advancement, especially recursive self-improvement, sounds like the center of risk but it really isn\u2019t. A merely more sophisticated model doesn\u2019t make it inherently have more agency. It is the harnesses and people who operate them that facilitate behaviors that can be detrimental. I propose that model steering be a profession that states are required to issue licenses for doing work that involves inherent safety risk, like any other profession. This will again use well-known patterns in governance of safety issues, and provide a framework for ongoing training (CEUs) and accountability. It also provides a pathway for liability insurance to do its magic. It does not stifle innovation nor require the pipe dream of globally coordinated pacing.
3. There is no AI without energy, and its massive pressure on existing infrastructure and capacity is demonstrably accelerating climate change. I\u2019m all for a token tax with a mandatory designation that all revenues go towards renewables, but I live in the USA in 2026 and the goalposts are where they are :( Something I think that would be feasible and potentially impactful is mandatory labeling. Every API response and UI showing the energy sources, water, and emissions for that request must be clearly disclosed in a common format. People deserve to know that their XAI porn habits are dumping 1/2 ton of CO2 into Memphis. I think this is the one that labs would hate the most, but it also seems like it could have an impact on consumer behavior that would make clean power make business sense.
Tl;dr we already know how to regulate things. Licensure, liability, labeling are all things we (society as a whole) are familiar with and can implement quickly and iterate on organically. Moreover, they are compatible with an open internet.
Your thoughts?", "5": "2026-08-22T18:05:01.715982"} {"0": 106, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49081491", "3": "Ask HN: Is there legal risk in AI memory?", "4": "Ask HN: Is there legal risk in AI memory?. With the move with AI tools to storing everything especially meeting notes where every interaction is transcribed and embedded for agents to use as context.
Does this not create a huge liability? I imagine in discovery there'd be massive trove of data, including off hand comments the company has.
I'm especially thinking of companies like Google who enforce rules on language, like don't say "dominant" or "kill". If every meeting is recorded this would be a lot easier to find and with AI analysing discovery data it'd probably be hard to miss.
I don't think I've seen any products that wrap these tools or built in features to sanitise meeting notes, and it might be more complex if you're currently under litigation in which case sanitisation might be counted as destroying evidence?", "5": "2026-08-22T18:05:01.724394"} {"0": 107, "1": "hackernews", "2": "https://www.transformernews.ai/p/openai-hack-hugging-face-responsibility-strict-liability-rules", "3": "Who should be responsible for OpenAI's hack of Hugging Face?", "4": "Who should be responsible for OpenAI's hack of Hugging Face?. ", "5": "2026-08-22T18:05:01.728695"} {"0": 108, "1": "hackernews", "2": "https://github.com/khalid-src/corv-client", "3": "Show HN: Corv v1.1 is out! Solving SSH execution for AI agents", "4": "Show HN: Corv v1.1 is out! Solving SSH execution for AI agents. I don\u2019t see much interest yet in the infrastructure side of AI. Maybe it\u2019s because few people want to take on the liability of letting agents handle infrastructure work, where the margin for error is close to zero.
But the underlying situation is awful. Most effort goes toward the application layer, with heavy wrappers and little attention given to one of the most sensitive parts: SSH.
LLMs as they work today, are not naturally compatible with SSH. Most handlers use coping mechanisms.. instead of solving the problem properly. They operate SSH almost like a human would, with some arguments added on top. This consumes tokens, creates instability, and requires many tool calls for simple tasks or large batches of commands executed at once in the hope that everything goes correctly.
As agent harnesses have shown, tooling has a major effect on model performance. After getting tired of the current situation, I decided to handle it myself.
Corv is my solution: a layer specialized around SSH that handles the parts that are problematic for AI agents, while remaining useful for humans. The models I used to review and stress-test it consistently preferred having Corv available as the default way to work with remote machines.
Today I\u2019m releasing v1.1, the first meaningful update. It makes Corv substantially more robust under concurrency, interruptions, and larger workloads.
There is still a way to go, and this area may be early and niche today. But infrastructure is the foundation, not the cherry on top. I believe this execution layer will become necessary, whether that happens this year on q4 or the next. Cheers!", "5": "2026-08-22T18:05:01.733487"} {"0": 109, "1": "hackernews", "2": "https://github.com/Northwood-Systems/millwright", "3": "Show HN: Millwright \u2013 Rust-based, self-hosted LLM router", "4": "Show HN: Millwright \u2013 Rust-based, self-hosted LLM router. Hey HN,
With the news of OpenRouter possibly being acquired and proliferation of hosted LLM routers (i.e. Ramp Router, Vercel\u2019s AI Gateway), I saw the need for a self hosted solution focused on cost savings, transparency, and performance. So, I built an open sourced router with a simple CLI interface that can easily sit between coding agents and GenAI workloads.
For the curious and lazy, at the moment, Millwright has the tools for,
- Providers: OpenAI-compatible APIs, Anthropic, Amazon Bedrock
- Routing: policy-controlled model roles (cheap, mid, frontier), cheapest healthy route selection
- Protocols: OpenAI Chat Completions, Anthropic Messages, text and tool translation
- Cache Affinity: role-scoped session lanes without serializing concurrent agent traffic
- Spend Tracking: per-team costs, cache usage, model/provider mix, request traces
- Cost Analysis: measured usage and modeled candidate economics (HTML, Markdown, JSON)
- Reliability: bounded failover, circuit breakers, timeouts, concurrency limits
- Setup: interactive provider, model, and pricing configuration without storing provider secrets
- Deployment: one Rust binary, Docker, SQLite or PostgreSQL
Full disclosure: parts of the codebase were built with AI coding agents. All feedback is welcome, I\u2019d especially value feedback on the routing policy, provider coverage, and anything that would block you from self-hosting it. Feel free to open feature/request and/or contribute as well.", "5": "2026-08-22T18:05:01.742031"} {"0": 110, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=48958273", "3": "LG ThinQ Terms of Use", "4": "LG ThinQ Terms of Use. Some of my kitchen appliances are LG and I installed the LG ThinQ app on my phone.
Sometimes I like to leave a cold dish in the oven before I go out then remotely start it when I\u2019m on my way back home, so I arrive to a nice hot dinner.
After a recent HN post about LG malware and further research, I decided to kick my appliances out of the network.
I also tried to remove my user account just to find out that in order to use the app I need to first accept the new terms of use.
I asked ChatGPT for a summary. This seems a lot more aggressive than the standard terms of use of other products and services.
1. No arbitration opt-out. You give up court, jury trial, class actions, mass actions, and collective claims. Most disputes must go through individual arbitration, and LG says your only alternative is to stop using the service. Claims generally must be brought within one year. (Sections 20.b and 20.c)
2. Broad rights over anything you submit. LG receives a perpetual, irrevocable, transferable, worldwide licence to use, modify, sublicense, and commercially exploit your user content \u201cfor any purpose whatsoever,\u201d without payment. Avoid uploading photos, recordings, documents, or detailed personal information. (Section 9.b)
3. No privacy expectation for communications. LG states that it may monitor user content and that you have no expectation of privacy for in-app chat, text, or voice communications. (Section 9.e)
4. Voice capture can include other people. Voice-enabled products may record and analyse family members, children, guests, and bystanders. LG places responsibility on you to inform them and obtain any legally required consent. (Section 4.d)
5. AI may use appliance and usage data. LG says third-party AI systems may rely on data from your use of its products and services. The terms do not clearly describe exactly what is transmitted, how long it is retained, or whether it is used to improve models. That information should be in the separate privacy policy. (Section 4.a)
6. Marketing consent is bundled into use. By using the service, you agree to email, texts, calls, automated or prerecorded messages, and push notifications, including promotions. You can opt out, but you must do so separately. (Section 5)
7. Targeted advertising is permitted. LG reserves the right to show targeted third-party advertising based on user preferences. (Section 11.b)
8. Very low liability limit. For many claims, LG attempts to cap its liability at the greater of the amount involved in the transaction, $100, or a statutory remedy. It also broadly excludes responsibility for lost data and unauthorized access. Local law may limit these clauses. (Section 17)
9. LG can update services remotely. It may push over-the-air updates without further consent and can change or discontinue features. (Sections 3.c and 18)", "5": "2026-08-22T18:05:01.746709"} {"0": 111, "1": "hackernews", "2": "https://www.schneier.com/blog/archives/2026/06/ai-and-liability.html", "3": "AI and Liability", "4": "AI and Liability. ", "5": "2026-08-22T18:05:01.753977"} {"0": 113, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=48690591", "3": "Roblox parental controls are a dystopian security disaster", "4": "Roblox parental controls are a dystopian security disaster. My 14 year old daughter got hacked by someone who was able to add themselves as a "linked parent" to her account. I'm not even sure that this person got ahold of her password in the first place. All this happened on Wednesday morning (6/24/26) but on the day it happened I did not recieve a single email about any of this even though the account is tied to my email address (verified). Usually if there is a new log in on an unrecognized device I would have gotten an email about it, but nothing was sent on 6/24 to me. I suspect that even if two factor authentication was already added to her account it would have done nothing, because there was a two factor authentication passkey added to her account which was definitely not set up by her. But by using that newly created authentication passkey the "linked parent" was clearly able to log into her account (which I didn't get any emails about), go into every game and transfer out every last collectable thing she had collected since 2020.
And wouldn't you know it, Roblox says they aren't responsible for those lost collectables. All the christmas and birthday roblox gift cards from the last 6 years which were used to buy those collectable items are completely wiped away for fun by this "linked parent". My daughter is absolutely devastated by her loss of these collectables.
During the password reset process I had to disable two factor authentication to be able to log in to the account. Once in the account, the two factor passkey could not be removed from the account without having access to the passkey and I had to go through an AI chatbot to get that removed. The "linked parent" also changed the date of birth to make my daugter become 8 years old in Roblox and apparently for whatever reason you are only allowed to change the date of birth once, meaning I had to make request after request trying to get the date of birth changed. Every time I am making these support requests I have to prove I am a human (captcha), enter six digit email security codes, and then try to talk to an AI bot that only partially understands my issues. I can request to speak to a human which immediately ends the chat with the ai bot telling me a support request has been filed.
What is most baffling of all is that I had requested removing the "linked parent" in question and between both the AI and whatever support team is behind that AI, I could not get the "liked parent" removed. I even had one ticket closed out with an email response telling me "We are unable to update or modify the parental settings on your child\u2019s account due to security reasons. Parental controls can be managed on the account with parent privileges linked to your child\u2019s account." When I was talking to an AI bot about this they explained that the "linked parent" was the only person who could remove themselves from my child's account and trying to request anything beyond that answer was denied. I finally hit a wall in which I had made too many requests and they were no longer accepting form submissions from me. My wife is trying to work on this stuff now because I'm at a dead end. She was able to get the account moved to her email address because she had made payments to Roblox in the past to fund the account, but the "linked parent" is still there.
Why would I ever want to give money to Roblox again after all of this? Kids are more savvy than anyone else on that gaming system and will keep finding loopholes to do these sorts of things. No matter how many procedural layers of restricted communication are added this is only made worse because fundamentally Roblox assumes no liability for any lost items within a system where these collectables can be traded among friends or stolen from thieves. I don't know that Roblox will be able to solve these problems ever when their solutions seem to be actually making things worse. If you have any stock in Roblox I would say they are a STRONG SELL!", "5": "2026-08-22T18:05:01.763914"} {"0": 114, "1": "hackernews", "2": "https://simonwillison.net/2026/Jun/25/ai-and-liability/", "3": "AI and Liability", "4": "AI and Liability. ", "5": "2026-08-22T18:05:01.772444"} {"0": 115, "1": "hackernews", "2": "https://github.com/ComplyEdge/complyedge", "3": "Show HN: ComplyEdge \u2013 Runtime EU AI Act Enforcement for Python", "4": "Show HN: ComplyEdge \u2013 Runtime EU AI Act Enforcement for Python. ", "5": "2026-08-22T17:24:16.841950"} {"0": 116, "1": "hackernews", "2": "https://www.recursant.ai/", "3": "Show HN: Recursant \u2013 service mesh for governing AI agents", "4": "Show HN: Recursant \u2013 service mesh for governing AI agents. Hello,
I have just released Recursant to the public. I have been working on it for a while. It is a control plane for governing AI agents across stacks. It provides full observability, guardrails, and control on the network level by routing all traffic through a side car.
Problem statement: many large, regulated enterprises (think banks, telcos) have one engineering team on LangGraph, another on CrewAI, marketing on AgentForce, and data teams on Databricks Agent Bricks. They need their agents to talk to each other with consistent policy enforcement, one audit trail, and a single set of guardrails, yet allowing different functions to run on their own stacks. Recursant solves that problem using the service mesh pattern
Recursant has two components: a registry and the mesh . The registry contains all live agents. The mesh uses sidecars to route traffic and enforce on the network layer.
Aim is for Recursant to provide a real-time EU AI Act Annex IV compliance, so it is not generated from static documents. This saves time and effort for large enterprises subject to the requirement.
Linmitations:
- Recursant currently plugs in to CrewAI, Langgraph, and n8n . The aim is to support proprietary platforms such as ServiceNow and AgentForce as much as psosible.\n- The Recursant SDK still needs work to support as many agents as possible\n- I would also like to provide support for some of the 'personal agent' platforms such as OpenClaw, NanoClaw, and Hermes\n- Only tested on k8s, not public cloud\n- Documentation is sparse and needs to be developed.
I hope this project is useful to some of you.", "5": "2026-08-22T17:24:16.846119"} {"0": 117, "1": "hackernews", "2": "https://a2cn.io/", "3": "Show HN: I built an open protocol for Agent-to-agent commercial negotiation", "4": "Show HN: I built an open protocol for Agent-to-agent commercial negotiation. AI agents can\u2019t do business and negotiate commercial deals with each other safely today. I built a protocol to prevent the problems agent-to-agent negotiation will inevitably run into once procurement and seller agents are mainstream.
A2CN is an open protocol for agent-to-agent commercial negotiation. We\u2019re already seeing procurement agents negotiate deals and huge savings for buyer departments in the enterprise. Companies like Pactum, Fairmarkit, and Zip have agents already transacting with suppliers. Pactum for example works with F500 companies like Walmart, Microsoft, and Maersk and has generated tens of millions of dollars in savings for its customers. The seller side is more nascent, but on the horizon with developments like Salesforce Agentforce for Revenue, and Microsoft Dynamics 365 - ERP MCP Server + Commerce MCP.
A whole host of problems arise when an agentic buyer and seller meet, and the story gets a lot more complicated:\n-They have no shared language for commercial terms. \n-There's no standard mechanism for an agent to assert and prove its authority or what it's actually allowed to agree to. \n-Commercial negotiation is an ordered exchange - offer, counteroffer, counteroffer, exchange. For agents, if there's no session protocol enforcing whose turn it is and what round you're on, agents can talk past each other, submit duplicate offers, or accept stale terms. \n-There\u2019s no way to neutrally prove that two agents reached an agreement - each side has their version of what was agreed upon and they might be different (the buyer thinks they agreed on 85K but the seller clocked it at 90K). \n-There needs to be an audit trail for enterprise procurement governance (driven by mandates like EU AI Act, SOX) which requires verifiable records of what an AI system decided and who authorized it. \n-Finally, for agent-to-agent negotiation to work, both sides need endpoints...but most suppliers don't have agent endpoints deployed. There's no standard way to invite a supplier into a session before both sides are ready.
These 8 protocol components fix the problems above:
-DISCOVERY: agents advertise capabilities and find each other
-MANDATE VERIFICATION: Cryptographic proof an agent has authority to commit its organization (W3C DIDs, two-tier: Declared + VC)
-SESSION INVITATION: Push-based handshake enabling adoption by parties without pre-deployed endpoints \u2014 cold-start solved
-OFFER EXCHANGE: Signed offers and counteroffers with strict turn-taking - ES256 + RFC 8785 JCS canonicalization
-DEAL-TYPE TERMS: Normative schemas for goods_procurement and saas_renewal with extensible custom_terms
-SESSION STATE MACHINE: Turn enforcement, sequence ordering, round limits, impasse detection - formally specified transitions
-TRANSACTION RECORD: Dual-signed, content-addressed record independently generated by both parties - neither controls the authoritative copy
-AUDIT LOG: Structured EU AI Act compliance output for every terminal session state - COMPLETED, REJECTED, WITHDRAWN, IMPASSE
The protocol reference implementation is at github.com/A2CN-protocol/A2CN - 202 tests passing, adapters for Fairmarkit and Keelvar already built.
This isn\u2019t something that everyone can use today and will require adoption as a standard, but if even a fraction of B2B commerce runs through agent negotiation in the next few years, the protocol layer underneath it becomes very valuable infra. Think of it like TCP/IP or Stripe, you don't see it but every transaction runs through it.
Open to any and all feedback, particularly from those who\u2019ve worked closely with protocols!", "5": "2026-08-22T17:24:16.850775"} {"0": 118, "1": "hackernews", "2": "https://comply-tech.co.uk/blog/eu-ai-act-2026-llm-pipeline.html", "3": "EU AI Act Enforcement in August 2026. What That Means for Your LLM Pipeline", "4": "EU AI Act Enforcement in August 2026. What That Means for Your LLM Pipeline. ", "5": "2026-08-22T17:24:16.858852"} {"0": 119, "1": "hackernews", "2": "https://pastebin.com/54QhPiwZ", "3": "Show HN: CMPSBL Software Factory \u2014 Free Daily Drop $2.9M", "4": "Show HN: CMPSBL Software Factory \u2014 Free Daily Drop $2.9M. I\u2019m a solo founder. I built a Cognitive Infrastructure Substrate \u2014 pure algorithmic code, zero AI API calls inside, no OpenAI dependency, fully patentable. A discovery engine collides software primitives against each other and crystallizes viable configurations into production-ready code.
It has autonomously discovered over $4.3B in software capabilities. I didn\u2019t write what I\u2019m sharing. The substrate found it.
Neural Arbiter \u2014 CJPI 100 | Governance | $3M substrate valuation
One sentence: an AI decision-making system that cannot issue a verdict without first passing through an ethics assessment.
ANALYTICS \u2192 BRAIN \u2192 CONSCIENCE \u2192 GOVERNANCE \u2192 SOVEREIGN\nThat sequence is enforced by the pipeline. You can\u2019t skip conscience. Ethical assessment before enforcement \u2014 architecturally, not as an audit afterthought. In a world where the EU AI Act is landing now, that\u2019s not a feature. That\u2019s the pipeline regulators are describing in legislation.
Full code (one file, zero dependencies):
I\u2019m giving one free discovery away every day at random times right here on HN until I hit 1000 users.
Currently at 42.
Comment below if you want a specific software type \u2014 security, synthesis, cognitive, governance.
I\u2019ll run the discovery engine with your parameters and post the result tomorrow.
958 to go.
CMPSBL.com \u2014 NPM with persistent memory online.", "5": "2026-08-22T17:24:16.865420"} {"0": 120, "1": "hackernews", "2": "https://github.com/vectimus/vectimus", "3": "Show HN: Vectimus \u2013 Cedar policy enforcement for AI coding agents", "4": "Show HN: Vectimus \u2013 Cedar policy enforcement for AI coding agents. Hey HN. I built Vectimus because coding agents keep doing things they shouldn't and there's no runtime governance layer for the developer workstation.
The problem: Claude Code, Cursor, Gemini CLI and GitHub Copilot let agents execute shell commands, write files and call MCP servers. Most developers disable the permission prompts because they slow you down. But that means the agent can rm -rf /, read your .env, push to production or call a compromised MCP server with nothing watching.
Vectimus intercepts every tool call and evaluates it against 78 Cedar policies containing 369 rules before execution. Cedar is the policy language AWS chose for AgentCore Policy (GA this month). Evaluation runs locally via a persistent daemon in under 10ms. Zero network calls. Zero telemetry. Every evaluation produces an Ed25519-signed receipt so you have cryptographic proof of what was allowed and denied.
Every policy maps to a real incident. CVE-2025-6514 compromised 437,000+ developer environments through a malicious MCP OAuth proxy. The GitHub MCP server was hijacked via a crafted issue to exfiltrate private repo data. A Terraform agent destroyed production infrastructure. These happened.
How it hooks in: Claude Code intercepts shell commands, file writes, MCP calls and web fetches. Cursor governs shell commands, file reads/writes and MCP tool calls at the editor level. Copilot intercepts terminal commands, file edits, deletes and git pushes. Gemini CLI uses Gemini's native hook system. MCP servers are blocked by default and allowlisted per-project with input inspection. Observe mode lets you see what would be blocked before you enforce.
I also built Sentinel (https://github.com/vectimus/sentinel), a three-agent pipeline that scans for new agentic AI security incidents daily, drafts Cedar policies, replays the incident in a sandbox to prove the policy catches it, then opens a PR. The pipeline is governed by Vectimus. Every finding and policy draft is public.
All 10 OWASP Agentic Top 10 categories covered. Compliance annotations for SOC 2, NIST AI RMF, NIST CSF 2.0, EU AI Act, ISO 27001, CIS Controls and SLSA. Apache 2.0. Solo founder, built in Ireland.
Happy to go deep on the Cedar policy design, the hook architecture, the signed receipts or the OWASP mapping.", "5": "2026-08-22T17:24:16.871619"} {"0": 121, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47247314", "3": "Show HN: Open-source scanner finds 97% of AI agent code non-compliant EU AI Act", "4": "Show HN: Open-source scanner finds 97% of AI agent code non-compliant EU AI Act. I built AIR Blackbox, an open-source static analysis tool that scans Python AI agent code against 6 technical requirements from the EU AI Act (Articles 9, 10, 11, 12, 14, 15). Think of it as a linter for AI governance.\nTo stress-test the scanner \u2014 and to see where the industry actually stands \u2014 I ran it against 5,754 Python files across 11 major open-source projects. Combined GitHub stars: 341,000+.\nProjects scanned: AutoGPT (170K stars), Microsoft AutoGen (38K), LlamaIndex (37K), Mem0 (24K), Phidata (18K), LiteLLM (15K), GPT-Researcher (14K), Embedchain (9.2K), LangGraph (8.5K), OpenAI Agents SDK (5.2K), CrewAI Examples (2.8K).\nResults:
Average compliance score: 2.2 out of 6 articles\n97% of files fail Article 9 (Risk Management)\n89% fail Article 12 (Record-Keeping)\n84% fail Article 14 (Human Oversight)\nOnly 23 out of 5,754 files (0.4%) pass all 6 checks\nBest scoring repo: AutoGPT at 2.9/6. Worst: CrewAI examples at 1.4/6
What the scanner checks (per article):
Art. 9: risk classification, access control, risk audit\nArt. 10: input validation, PII handling, data schemas, provenance\nArt. 11: logging, documentation, type hints\nArt. 12: structured logging, audit trail, timestamps, log integrity\nArt. 14: human review, override mechanism, notifications\nArt. 15: input sanitization, error handling, testing, rate limiting
An article "passes" if at least 1 sub-check is detected. This is generous \u2014 real compliance requires substantially more.\nCaveats I'll save you the trouble of pointing out:
This is static analysis. It can't verify runtime behavior.\nFile-level scanning misses cross-file compliance patterns.\nThe pass threshold is intentionally lenient (1-of-N sub-checks).\nThis checks technical requirements, not legal compliance. It's a linter, not a lawyer.
The EU AI Act enforcement deadline is August 2026. The full report, raw data (JSON), and the scanning scripts are all in the repo.
GitHub: https://github.com/air-blackbox/air-blackbox-mcp\nFull report: https://github.com/air-blackbox/air-blackbox-mcp/blob/main/b...\nInstall: pip install air-blackbox-mcp\nDemo: https://huggingface.co/spaces/airblackbox/air-blackbox-scann...
Happy to answer questions about the methodology, the scanner internals, or what we're building next (fine-tuned local LLM for deeper analysis \u2014 your code never leaves your machine).", "5": "2026-08-22T17:24:16.878916"} {"0": 122, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47169864", "3": "Ask HN: How are you handling EU AI Act compliance as a developer?", "4": "Ask HN: How are you handling EU AI Act compliance as a developer?.
The EU AI Act high-risk enforcement deadline is August 2, 2026. If you're deploying AI in the EU \u2014 or serving EU customers \u2014 \n you're supposed to classify your systems, implement risk management, document everything, and potentially do conformity \n assessments.\n\n I'm curious how developers are actually approaching this:\n\n 1. Are you taking it seriously yet? The prohibited practices are already enforceable (since Feb 2025). High-risk obligations \n kick in August 2026. Are you actively preparing or waiting to see how enforcement plays out?\n 2. Is the EU shooting itself in the foot? The AI Act is 144 pages. GDPR already costs European startups disproportionately \n compared to US competitors. Is this just more red tape that will widen the gap with US tech companies, or is regulatory clarity\n actually a competitive advantage ("we're EU-compliant" as a selling point)?\n 3. How do you even operationalize this? 113 articles, 13 annexes, cross-references to GDPR, potentially DORA if you're in \n fintech. Is anyone actually reading EUR-Lex, or are you outsourcing to lawyers and hoping for the best?\n 4. Will enforcement actually happen? GDPR took years before meaningful fines started. The AI Office is still setting up. Are EU\n regulators going to enforce this on day one, or will there be a grace period in practice?\n\n I built a compliance API (https://gibs.dev) because I got frustrated trying to navigate this myself, but I'm genuinely\n uncertain whether the regulation will adapt or whether European AI companies will just build elsewhere. What's your read?Trust layers for LangChain, CrewAI, AutoGen, OpenAI Agents SDK, and RAG pipelines \u2014 each is a pip install that hooks into your existing agent code with ~3 lines of setup\nHMAC-SHA256 tamper-evident audit chains \u2014 every agent decision, tool call, and LLM interaction gets logged to a chain that regulators can verify\nConsentGate \u2014 risk-classifies tool calls and blocks critical operations until approved\nInjectionDetector \u2014 15+ weighted patterns scanning prompts before they reach the model\nWriteGate + DriftDetector (for RAG) \u2014 prevents knowledge base poisoning and detects retrieval anomalies\nCompliance scanner \u2014 pip install air-compliance && air-compliance scan ./my-project tells you exactly which articles you're missing
Everything maps to specific EU AI Act articles (9, 10, 11, 12, 14, 15). Zero vendor lock-in, Apache 2.0, zero core dependencies on the trust layers.\nThe scanner is probably the fastest way to understand where your gaps are. It takes about 3 seconds to run on a typical project.\nGitHub: https://github.com/airblackbox\nPyPI: pip install air-compliance\nHappy to answer questions about what the EU AI Act actually requires for AI agent deployments \u2014 we've read the full regulation and mapped it to specific technical controls.", "5": "2026-08-22T17:24:16.893258"} {"0": 124, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47137306", "3": "Show HN: Hardware and software safety standard for AI and Robots (15 patents)", "4": "Show HN: Hardware and software safety standard for AI and Robots (15 patents). I'm a solo inventor in rural Pennsylvania. Over 13 days in February 2026, I filed 15 provisional patent applications (134 claims) with the USPTO covering a full-stack safety and governance architecture for AI systems.
The patents break into three domains:
Hardware enforcement (4 PPAs, 33 claims): A dedicated safety processor on its own power rail controls whether AI compute receives electricity. AI boots only after safety completes self-test. During operation, the safety processor monitors AI-specific indicators and can physically cut power \u2014 no software involvement. The AI cannot prevent its own shutdown. Same Safe Torque Off principle industrial motor controllers have used for decades, applied to AI compute.
Software governance (9 PPAs, 72 claims): Multi-vendor consensus engine where up to 9 AI models must agree before physical action. Transparent reasoning verification. Authority enforcement and drift monitoring. Human-readable audit trails in plain-text Markdown \u2014 every decision readable by a human without special tools. Persistent AI memory that survives reboots. Real-time safety micro-agents that monitor the AI's own cognitive state.
Financial architecture (2 PPAs, 29 claims): Tokenized equity and value distribution for the business model side.
The hardware spec defines three standardized form factors (drone-sized to humanoid-sized) with a universal connector \u2014 any compliant brain plugs into any compliant robot.
The EU AI Act goes fully into effect August 2, 2026. Articles 12-14 require auditable decision records, transparent operation, and human oversight for high-risk AI systems. Fines up to 7% of global annual revenue. Most robot AI today stores decisions in vector embeddings no human can read. This architecture addresses that directly.
I built all of this working alongside AI, using an open-source context management system I created that gives AI assistants persistent memory across sessions. The tool that solved AI's memory problem became the tool that let me design a full-stack safety architecture in 13 days.
Open-source memory tool: https://github.com/RobSB2/CxMS
Website: https://opencxms.org
Happy to answer questions about the hardware spec, the software governance architecture, or what it's like filing 15 patents in 13 days as a solo inventor using AI.", "5": "2026-08-22T17:24:16.901023"} {"0": 125, "1": "hackernews", "2": "https://www.x-loop3.com/free/eu-ai-act-lite#download", "3": "Show HN: EU AI Act Layer \u2013 Free Compliance Checker", "4": "Show HN: EU AI Act Layer \u2013 Free Compliance Checker. EU AI Act enforcement starts Feb 2026. High-risk AI systems face up to \u20ac35M fines if non-compliant.\n---\nBuilt a tool that checks AI systems against EU requirements in 2 minutes:\n---\n\u2192 Risk classification (Annex III check)\n\u2192 Governance readiness scoring (6 core requirements)\n\u2192 Evidence pack generation (audit-ready, 5 files)\n\u2192 Plain English action items\n---\nTechnical details:\n\u2192 Multi-LLM support (OpenAI, Anthropic, Mistral, xAI)\n\u2192 Zero-dependency evidence generation\n\u2192 SHA-256 integrity verification\n\u2192 Offline-capable after setup\n---\n100% free. Open source. FTBL license.\n---\nHappy to answer questions about the EU AI Act or how the tool works.\n---\nGitHub: https://github.com/jongartmann/eu-ai-act-layer-lite", "5": "2026-08-22T17:24:16.907912"} {"0": 126, "1": "hackernews", "2": "https://audit.omensystems.com", "3": "Show HN: I'm 16 and built EU AI Act compliance software", "4": "Show HN: I'm 16 and built EU AI Act compliance software. Hey HN \u2013 I'm Chaitanya, a 16-year-old student. I built AuditDraft after reading through the EU AI Act and realizing most companies have no idea how to comply.\nThe regulation is 400 pages. High-risk AI systems need technical documentation covering 12 different requirements. Fines go up to \u20ac35M. Enforcement has already started for some provisions, and the big deadline (August 2026) is coming fast.\nAuditDraft helps you:
Classify your AI system's risk level (8 questions based on Article 6)\nGenerate Annex IV compliant model cards and documentation\nTrack compliance across all 35 high-risk requirements
Built with Passion.", "5": "2026-08-22T17:24:16.913370"} {"0": 127, "1": "hackernews", "2": "https://github.com/devon39/server-oauth-security", "3": "Show HN: OAuth 2.0 server with AI security agents (EU sovereign alternative)", "4": "Show HN: OAuth 2.0 server with AI security agents (EU sovereign alternative). I spent 4 years trying to build this OAuth server but never finished it.
Then I discovered agentic coding and shipped it in 3 weeks.
What makes it different:
\u2022 Dual AI agents analyze every login in <300ms\n - Security Signals Agent: risk scoring (device, IP, geo, velocity)\n - Policy Compliance Agent: business rules (MFA policies, role enforcement)\n - Combined decision: allow/log/step-up/lock/deny
\u2022 Production-ready security\n - PKCE (RFC 7636), DPoP (RFC 9449)\n - MFA (TOTP + WebAuthn/Passkeys)\n - IP restrictions, rate limiting, audit trail
\u2022 EU digital sovereignty\n - GDPR native (data export, legal holds, retention policies)\n - EU hosting, no US Cloud Act exposure\n - Full audit trail (PostgreSQL + Redis Streams)
\u2022 Zero AI dependency\n - Deterministic fallback if AI timeouts\n - Conservative MEDIUM risk returned (safe default)\n - System keeps running without external LLM calls
\u2022 Modern stack\n - Backend: NestJS + TypeScript, LangChain/LangGraph\n - Frontend: React 19, hexagonal architecture, 91% test coverage\n - Deterministic fallback if AI timeouts (zero dependency)
Built as an alternative to Firebase Auth / AWS Cognito / Auth0 for companies that want control over their authentication infrastructure.
Architecture diagrams and screenshots in the repo.
Open to feedback and questions.", "5": "2026-08-22T17:24:16.922259"} {"0": 128, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=46778903", "3": "AI governance isn't failing because we lack regulation-it's failing at execution", "4": "AI governance isn't failing because we lack regulation-it's failing at execution. There's a lot of movement around AI regulation right now (EU AI Act, US frameworks, etc.), but in practice many of these governance models don\u2019t survive contact with real, agentic systems.
A recent paper I was involved in looks at why compliance frameworks tend to break at the operational layer - things like:
- human oversight that works on paper but collapses in real workflows\n- enforcement gaps across jurisdictions\n- fragmented compliance creating systemic risk rather than safety
The goal wasn't to rehash regulations, but to analyze where governance actually fails once AI systems are deployed and interacting autonomously.
Paper + more context in the comments.\nHappy to discuss or get critical feedback.", "5": "2026-08-22T17:24:16.927807"} {"0": 129, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=46232948", "3": "Show HN: QCMP Framework for Poison-Resistant AI Agents (ArXiv Cs.ai Pending)", "4": "Show HN: QCMP Framework for Poison-Resistant AI Agents (ArXiv Cs.ai Pending). Hey HN\u2014after a year digging into agentic AI vulnerabilities, I've built QCMP: a 4-layer architecture to slam the door on memory poisoning. MCP's at 16K servers, but attacks like MINJA (98.2% query-only success) and AgentPoison (80%+ backdoors from 0.1% poison) expose the core flaw\u2014memory trusts itself too much.
QCMP borrows from IIT consciousness metrics (CCI >0.90 to freeze fragments), post-quantum checksums (ML-KEM-768), CTC self-consistency (NIS >0.95), and mantis shrimp-style sparse checks (<50ms TME). OWASP/EU AI Act ready, with Rust impl tips.
PDF (in-browser view): https://github.com/bradmcevilly/qcmp-whitepaper/blob/main/QC...
First arXiv push to cs.AI\u2014hunting endorsements (4+ recent subs). Feedback on the quantum-bio hooks or swarm layers? Open to riffs.
deepsweep.ai | linkedin.com/in/bradmcevilly
I've spent the last year tackling memory poisoning in agentic AI (e.g., 98% MINJA success via queries alone). Introducing QCMP: a 4-layer architecture blending IIT consciousness metrics (CCI >0.90 thresholds), post-quantum checksums (ML-KEM), and CTC consistency for tamper-proof agent swarms.
Key wins: Detects 0.1% AgentPoison backdoors in <50ms; OWASP/EU AI Act compliant.
PDF: https://github.com/bradmcevilly/qcmp-whitepaper/blob/main/QC...
First arXiv sub to cs.AI\u2014seeking endorsements/feedback from the HN community. Thoughts on the quantum-bio hooks or multi-agent layers? Open to chats.
Site: deepsweep.ai | LI: linkedin.com/in/bradmcevilly", "5": "2026-08-22T17:24:16.934592"} {"0": 132, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47600569", "3": "Show HN: EU RegTech SaaS for sale \u2013 CBAM, AI Act, French tenders ($5K each)", "4": "Show HN: EU RegTech SaaS for sale \u2013 CBAM, AI Act, French tenders ($5K each). I built three SaaS products targeting mandatory EU regulations and I'm selling each for $5,000.
1. CBAM OS (https://cbam-os.com) \u2014 EU Carbon Border Adjustment Mechanism compliance. 9 modules covering the full CBAM workflow. Every EU importer of covered goods (steel, aluminum, cement, fertilizers, electricity, hydrogen) must comply. ~100K+ potential customers. Plans: 199/399/999 per month.
2. AIA Proof (https://aiaproof.com) \u2014 EU AI Act compliance and AI content detection. The AI Act is enforceable \u2014 fines up to 35M or 7% of global turnover. Every company deploying AI in the EU is a potential customer. Plans: 299/799/2,499 per month.
3. AO France (https://ao-france.fr) \u2014 AI-powered French public tender responses. France publishes 200K+ tenders/year worth 200B+. Plans: Solo/Pro/Business + 49 per tender.
Stack: Next.js (App Router), Vercel, Stripe, Supabase.
I'm selling to focus on my next project. $5,000 each or $12,000 for all three.
Contact: abbas.billel1985@gmail.com", "5": "2026-08-22T17:24:17.984401"} {"0": 133, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47035867", "3": "Show HN: Regression tests for detecting cross-domain hallucinations in LLMs", "4": "Show HN: Regression tests for detecting cross-domain hallucinations in LLMs. LLMs sometimes generate structurally valid but logically impossible claims when technical and legal domains mix.
Example failure mode:\nA model sees \u201cCVE-2024-XXXX fixed in v2.1\u201d and hallucinates a causal link to \u201cUsers must pay retroactive fees under EU regulation Article 56.\u201d
To explore this, I built a regression dataset (40 edge cases) covering:
Fake identifier bindings (CVE + version)
Retroactive fiscal claims
Cross-domain causality leaps (Tech \u2192 Legal)
Over-assertive phrasing without evidence
Then I designed a structured system prompt that:
Detects official identifiers (CVE, Regulation numbers) vs placeholders
Flags monetary + retroactivity combinations as high-risk
Enforces proportional claim strength based on available evidence
Results:
Automated: 40/40 regression cases pass (JSON dataset + simple Python runner included).
Manual adversarial: ~40 prompts designed to test:
Draft article traps (e.g., hallucinated \u201cArticle 52c\u201d in EU AI Act)
Pricing model fabrications (e.g., \u201cbilling based on parameter count\u201d)
Version binding errors (e.g., incorrect Node.js default versions)
This is not fine-tuning\u2014just a structured prompt experiment focused on structural validation.
Looking for feedback on:
Missing edge cases
Failure modes I didn\u2019t consider
Whether this approach generalizes beyond legal/technical mixing
Gist (spec + dataset + runner):\nhttps://gist.github.com/ginsabo/6ebeb9490846ee6a268bd13560c0...", "5": "2026-08-22T17:24:17.993987"} {"0": 134, "1": "hackernews", "2": "https://calmatters.org/economy/technology/2025/09/chatgpt-lawyer-fine-ai-regulation/", "3": "California issues fine over lawyer's ChatGPT fabrications", "4": "California issues fine over lawyer's ChatGPT fabrications. ", "5": "2026-08-22T17:24:17.998524"} {"0": 135, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=44669306", "3": "Ask HN: Could Europe Play the U.S. and China Against Each Other for Free LLMs?", "4": "Ask HN: Could Europe Play the U.S. and China Against Each Other for Free LLMs?. The U.S. could offer free high-performance LLMs to Europe under the condition of avoiding Chinese models\u2014but China might retaliate with its own subsidies or tools. This creates an opportunity for Europe to exploit the competition by:
Demanding concessions: Require open weights, local hosting, or compliance with EU regulations (GDPR, AI Act).\n\n Avoiding dependency: Mandate interoperability with European models (e.g., Mixtral) or hybrid systems.\n\n Legislating neutrality: Force government projects to use locally fine-tuned versions of either, maintaining sovereignty.\nQuestions:
Is this dynamic realistic, or would Europe lack leverage?
What would make the EU choose one side\u2014performance, subsidies, or ethics?
Could small countries individually cut deals, fracturing EU unity?", "5": "2026-08-22T17:24:18.004802"} {"0": 136, "1": "hackernews", "2": "https://main.d1q2ygy4ts4vr5.amplifyapp.com", "3": "Show HN: Traceprompt \u2013 tamper-proof logs for every LLM call", "4": "Show HN: Traceprompt \u2013 tamper-proof logs for every LLM call. Hi HN,
I'm building Traceprompt - an open-source SDK that seals every LLM call and exports write-once, read-many (WORM) logs auditors trust.
Here's an example - a LLM that powers a bank chatbot for loan approvals, or a medical triage app for diagnosing health issues. Regulators, namely HIPAA and the upcoming EU AI Act, missing or editable logs of AI interactions can trigger seven-figure fines.
So, here's what I built:
- TypeScript SDK that wraps any OpenAI, Anthropic, Gemini etc API call
- Envelope encryption + BYOK \u2013 prompt/response encrypted before it leaves your process; keys stay in your KMS (we currently support AWS KMS)
- hash-chain + public anchor \u2013 every 5 min we publish a Merkle root to GitHub -auditors can prove nothing was changed or deleted.
I'm looking for a couple design partners to try out the product before the launch of the open-source tool and the dashboard for generating evidence. If you're leveraging AI and concerned about the upcoming regulations, please get in touch by booking a 15-min slot with me (link in first comment) or just drop thoughts below.
Thanks!", "5": "2026-08-22T17:24:18.014835"} {"0": 137, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=38293085", "3": "Things OpenAI is doing to make GPT5", "4": "Things OpenAI is doing to make GPT5. Intelligence in systems (human, AI) can be conceptualized as the resolution and throughput at which a system can process and affect Shannon information. This perspective emphasizes not just the quantity of information processed (throughput), but also the depth and detail with which it is handled (resolution), constrained to thermodynamic limitations.
Aside from improving the thermodynamics of the system here are the 7 levers we can use to improve its intelligence:
Physical Capacity: A system's intelligence increases as it expands its physical limits. This encompasses augmenting processing units (akin to neurons in humans or parameters in AI), improving thermal regulation, and maximizing energy throughput. Such enhancements enable a system to process information at a higher resolution and throughput.
Cooperation: When entities collaborate, the collective intelligence of the system increases. This is due to the improved resolution at which information can be processed and influenced, a principle manifest in ensemble methods in AI where multiple models aggregate their insights.
Conflict: The presence of conflict within or between systems can lead to an increase in intelligence. The necessity to adapt for survival and resolve conflicts escalates energy expenditure, which in turn refines the system's ability to process and affect information at a greater resolution and throughput.
Attention: Enhancing the range, depth, and sampling rate of information a system can process boosts its intelligence. This increase is achieved by allowing the system to operate with a wider context and more frequent assessments of information, thereby processing it at a higher resolution.
Actuation: Increasing the scope and precision of a system's actions directly impacts its intelligence. More diverse and precise actuation improves the system's capacity to affect information at a finer resolution.
Memory (Past): Building and utilizing shared physical memory elevates a system's intelligence by enabling it to process information over time at a higher resolution, fostering a more nuanced understanding of historical data.
Predictors (Future): The intelligence of a system can be significantly increased by evolving its core predictive model architecture. This includes developing new AI-designed architectures or employing techniques like neuroevolution, where algorithms evolve and optimize neural networks. Such advancements not only enhance the system's predictive capabilities but also improve the resolution and throughput at which it can process and affect future outcomes. By continually refining the architecture, the system becomes adept at anticipating and influencing future scenarios with greater accuracy and efficiency.
Specific Measures to Enhance GPT's Capabilities:
Increase Memory: Pretty obvious but nuanced approaches
Expand Actuation: More actions. Not plugins, but an open API marketplace not for humans but AI.
Enhance Cooperation: More ensembling, Open ensembling protocol, 'undisclosed secret sauce'.
Boost Physical Capacity: Pretty obvious, more Closed-source parameters, or Giant P2P networks (Petals.ML alternative)
Intensify Conflict: Employing more discriminators, \u2018undisclosed secret sauce' .
Augment Attention: This is challenging due to energy limits
Better Predictors: Increasing the length, scope and sophistication of predictive models (Neuroevolution)", "5": "2026-08-22T17:24:18.025738"} {"0": 138, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=38014867", "3": "What are enterprise worried about for AI compliance for LLMs?", "4": "What are enterprise worried about for AI compliance for LLMs?. Based on our conversations, most enterprises are looking to build controls in to models based on data used to train/fine-tune these LLMs. Is any of this based on the new/upcoming regulations, or do these follow from existing data privacy/security regulations?", "5": "2026-08-22T17:24:18.034222"} {"0": 139, "1": "hackernews", "2": "https://github.com/PontusAI/Pontus", "3": "Show HN: Pontus the Zero Trust AI Layer", "4": "Show HN: Pontus the Zero Trust AI Layer. Hi HN,
We\u2019re @sumants and @rmehtany, working on Pontus.
Pontus makes it easy to use AI with privacy embedded.
We were concerned about the volume of personal data that goes to large LLM models without protection. We tried find an easy solution where didn\u2019t change the simple apis given by LLM providers. However, most required you to invest significant engineering effort.
We wanted privacy and LLMs to be easy, so we built Pontus. Through a declarative YAML, we orchestrate a microservice with the most common element of the LLM stack.
- Anonymize Prompts before it hits LLMs, yet keeps context on your servers\n- Secure RAG that embeds documents safely\n- Semantic Cache that doesn\u2019t store PII at Rest
We did this not only because it is the right thing to do, but the regulation landscape makes it critical to build with privacy first. Recently, many companies have been fined and forced to delete their AI models [1][2]
For us, Privacy is by design, not an afterthought.
Please visit us here
Our Open Core Repo: https://github.com/PontusAI/Pontus\nOur Website: https://pontus.so
[1] https://www.ftc.gov/news-events/news/press-releases/2023/05/...
[2] https://techcrunch.com/2023/05/10/clearview-ai-another-cnil-...", "5": "2026-08-22T17:24:18.040370"} {"0": 140, "1": "hackernews", "2": "https://dstill.ai/podcasts", "3": "Show HN: LLM Podcast Chat and Summaries", "4": "Show HN: LLM Podcast Chat and Summaries. Hey Hacker News,
2 weeks ago I shared an alternative HN UI with built in LLM-powered summarization for links and discussions (dstill.ai/hackernews, https://news.ycombinator.com/item?id=36760714).
Today I am happy to share further work that allows you to summarize and chat with a select set of high quality podcasts. You can find these features at https://dstill.ai/podcasts and https://dstill.ai/agent.
Video demo: https://www.loom.com/share/1270256b2c9d4d88970b9cadd446ceb3?... (screenshots with more details below)
A bit more context:
The long term goal of https://dstill.ai is to provide you with the means to consume and access information in a more thoughtful and highly personal way. Making sure you never miss anything that's actually important to you and cutting all the fluff and distractions that are being thrown at us every day \u2014 hopefully somewhat balancing the fact that some companies are set to flood the internet with vacuous AI generated content.
Today\u2019s release is only a small part of that vision, but I want to make sure we release early and often, and learn along the way.
Why podcasts? I think they are a unique source of highly authentic information that can help balance the generic and impersonal nature of something like ChatGPT. They also pose a challenge, because they are audio, and audio is much harder to work with compared to text. But more and more content is being exclusively produced in audio or video form, and so it\u2019s strategically important to tackle this head on.
Here\u2019s a quick rundown of the podcasts UI:
- Podcast list: https://screenbud.com/shot/07c31b34-0a1f-4658-a70a-0007b69a5...\n- Specific podcast: https://screenbud.com/shot/ca5b067e-428d-4fc7-a6ba-2cd87fe3e...\n- Specific podcast episode with summary: https://screenbud.com/shot/7056aa9e-9834-48c7-918a-6eef2694d...
Here are a few example conversations I had recently:
- [Andrew Huberman and Peter Attia on improving cardiovascular fitness.] https://screenbud.com/shot/8fddfb0d-51a6-45c7-bffd-a5644d106...\n- [Pros and cons for AI regulation, include quotes and a summary.] https://screenbud.com/shot/d526c793-5b52-45ee-b688-1de78be63...
On the technical level, here\u2019s how things are structured:
![https://screenbud.com/shot/07c31b34-0a1f-4658-a70a-0007b69a5...](\"https://screenbud.com/shot/07c31b34-0a1f-4658-a70a-0007b69a57a6/image.png\"){kind=link}
![https://screenbud.com/shot/ca5b067e-428d-4fc7-a6ba-2cd87fe3e...](\"https://screenbud.com/shot/ca5b067e-428d-4fc7-a6ba-2cd87fe3ed4c/image.png\"){kind=link}
![https://screenbud.com/shot/7056aa9e-9834-48c7-918a-6eef2694d...](\"https://screenbud.com/shot/7056aa9e-9834-48c7-918a-6eef2694d75f/image.png\"){kind=link}
![https://screenbud.com/shot/8fddfb0d-51a6-45c7-bffd-a5644d106...](\"https://screenbud.com/shot/8fddfb0d-51a6-45c7-bffd-a5644d106af1/image.png\"){kind=link}
![https://screenbud.com/shot/d526c793-5b52-45ee-b688-1de78be63...](\"https://screenbud.com/shot/d526c793-5b52-45ee-b688-1de78be636a1/image.png\"){kind=link}
- Podcasts are transcribed (using whisper) and diarized\n\n - For summarization:\n - Transcripts are split into chunks if necessary \u2014 many podcasts are over 20K tokens \u2014 and those are then summarized in a map/reduce fashion\n\n - For chat:\n - Transcripts are split into chunks, which are embedded using instructor-large\n - The embeddings are stored in qdrant, along with metadata needed for filtering \u2014 this is where qdrant shines, it allows you to have *efficient* embedding based retrieval with filtering (see here for details https://qdrant.tech/articles/filtrable-hnsw/)\n - There\u2019s one LLM that generates search queries (if any), those are use for retrieval of the podcast chunks\n - Then a second LLM is fed the conversation history along with the search results \u2014 currently relying on ChatGPT, though we are working on custom fine tuning of llama.\n\n - The UI is TypeScript with SvelteKit \u2014 so far I love the simple and intuitive nature of Svelte(Kit).\n\n - The backend is Rust, based on my own (multi-lingual) LLM framework tailored for configurability and observability (I wrote about it a bit here, including some demos of the platform: https://news.ycombinator.com/item?id=36787924#36789075)\n\n - Right now there\u2019s no account system, so you need to supply your own OpenAI API keys.\nI would be very happy to hear your thoughts and feedback, both on the general vision, as well as today\u2019s podcast and agent release.
Cheers,\nPetr.", "5": "2026-08-22T17:24:18.046349"} {"0": 141, "1": "hackernews", "2": "https://github.com/opencomplai/opencomplai", "3": "Show HN: OpenComplAI \u2013 open-source EU AI Act compliance checks in CI/CD", "4": "Show HN: OpenComplAI \u2013 open-source EU AI Act compliance checks in CI/CD. ", "5": "2026-08-22T17:24:19.045281"} {"0": 142, "1": "hackernews", "2": "https://scanara.io/en/", "3": "Automated and easy EU AI Act compliance", "4": "Automated and easy EU AI Act compliance. ", "5": "2026-08-22T17:24:19.053782"} {"0": 143, "1": "hackernews", "2": "https://auditbadger.com/", "3": "Show HN: AuditBadger \u2013 SOC 2 and ISO 27001 \u2013 AI drafts, you approve", "4": "Show HN: AuditBadger \u2013 SOC 2 and ISO 27001 \u2013 AI drafts, you approve. Hi,
Wanted to share something I've been working on for over a year. AuditBadger is a compliance management platform that uses AI to write policies (there are underlying "templates" with basic requirements), rewrite controls (or trust service criterions) to match the company context, help figure out your own controls, does initial risk assessment, and business continuity planning (which at least gives you an example of how the process should look like).
Fun fact - I wanted to share this a year ago, but then I spotted something similar here. The most common comment was about lacking the SOC 2 report, so I decided to pick the fight. I got SOC 2 Type I first, and then recently finished SOC 2 Type II using the tool alone. It took some time - both learning the process, the SOC 2 gotchas, and implementing automatic evidence collection.
We're now adding support for the European AI Act and NIS 2; HIPAA is already there (though it requires me to explicitly enable it for customers who want to test it), and CyberEssentials and ENS are coming later this year.
The platform is now complete, but my business partner (ISO 27001 Lead Auditor) and I are still dog-fooding it. Everything we build is either based on our own pain points or our customers'\u2014most of them joined our Slack where we try to help them if they get stuck.
If you have any questions, I'll be happy to answer them all.", "5": "2026-08-22T17:24:19.061414"} {"0": 144, "1": "hackernews", "2": "https://scanara.io", "3": "Affordable and automated EU AI Act compliance", "4": "Affordable and automated EU AI Act compliance. ", "5": "2026-08-22T17:24:19.068883"} {"0": 145, "1": "hackernews", "2": "https://www.nyc144euaiact.com", "3": "NYC LL144 and EU AI Act Compliance Guides", "4": "NYC LL144 and EU AI Act Compliance Guides. ", "5": "2026-08-22T17:24:19.076244"} {"0": 146, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=48842416", "3": "CEO and Creator of Ethos Engine for AI", "4": "CEO and Creator of Ethos Engine for AI. Francisco Javier Rold\u00e1n Vel\u00e1squez is the Founder and CEO of Ethos Engine, a deterministic AI safety infrastructure solving the "circular supervision" problem. It acts as an external governance module, using proprietary mathematical indices and a multi-civilizational knowledge graph to ensure compliance with the EU AI Act. With more than 30 years of experience, Rold\u00e1n seeks pre-seed funding and Silicon Valley connections to scale this patent-pending intellectual property asset globally.", "5": "2026-08-22T17:24:19.083255"} {"0": 147, "1": "hackernews", "2": "https://github.com/CorvinLabs/CorvinOS", "3": "CorvinOS \u2013 an agentic OS where GDPR/EU AI Act compliance can't be switched off", "4": "CorvinOS \u2013 an agentic OS where GDPR/EU AI Act compliance can't be switched off. ", "5": "2026-08-22T17:24:19.090951"} {"0": 148, "1": "hackernews", "2": "https://github.com/latreon/compliance-agent", "3": "ComplianceAgent: Open-source EU AI Act compliance scanner", "4": "ComplianceAgent: Open-source EU AI Act compliance scanner. ", "5": "2026-08-22T17:24:19.097848"} {"0": 149, "1": "hackernews", "2": "https://sevinhub.com/acthub/", "3": "ActHub \u2013 EU AI Act compliance toolkit for small businesses (PHP, no framework)", "4": "ActHub \u2013 EU AI Act compliance toolkit for small businesses (PHP, no framework). ", "5": "2026-08-22T17:24:19.103888"} {"0": 150, "1": "hackernews", "2": "https://aiact.bridgeai.one/", "3": "Free EU AI Act Article 50 compliance checker for indie SaaS", "4": "Free EU AI Act Article 50 compliance checker for indie SaaS. ", "5": "2026-08-22T17:24:19.110911"} {"0": 151, "1": "hackernews", "2": "https://compliancelint.dev/", "3": "Show HN: ComplianceLint \u2013 The linter for EU AI Act compliance", "4": "Show HN: ComplianceLint \u2013 The linter for EU AI Act compliance. ", "5": "2026-08-22T17:24:19.118573"} {"0": 152, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=48560279", "3": "Show HN: Dev-friendly native OTel: only OSS stateful, on-the-wire Observability", "4": "Show HN: Dev-friendly native OTel: only OSS stateful, on-the-wire Observability. Hi HN,
We\u2019re the team at MyDecisive.ai, and today we\u2019re giving developers a peek at Octant \u2014 point-and-click control and visibility for your OpenTelemetry.
You've likely felt the pain of the "observability tax," especially if you manage K8S clusters. The modern standard is to instrument everything with OpenTelemetry, but piping all those rich OTLP logs, metrics, and traces straight to a SaaS vendor (Datadog, Splunk, Honeycomb) gets expensive fast. You end up paying massive ingestion and storage costs for noisy, low-value data just so it's searchable when something breaks. With Octant you get up and running on OTel in minutes.
We built Octant to flip this model. Instead of blindly shipping all telemetry off-cluster, Octant configures and helps to manage OTEL clusters. It gives you a visual interface for managing K8s objects, but importantly, it acts as an OTLP gateway that filters data at the source before it leaves your VPC.
Because it natively speaks OpenTelemetry, you can point your existing OTel SDKs or collectors right at it without touching your application code. Here is what it does under the hood:
- OTel-Native Trace & Log Sampling: It makes it easy to ingest OTLP traffic and inspects logs and traces on the wire. By waiting for the full context of a trace before determining what to keep, it delivers on the promise of braiding, retaining 100% of the actionable signals around (like errors and high-latency spans) but droppings the junk before it hits your SaaS bill.
- In-Flight Stateful Alerting: Instead of waiting for data to be batched, shipped, and indexed by an external provider to trigger an alert, Octant can process the telemetry streams in-flight. This shrinks the detection gap and reduces the need for SaaS vendors in the first place.\n- On-the-Wire PII Redaction: It can detect and strip sensitive information from your logs and traces in real-time before they are transmitted over the internet, removing "post-ingestion" clean-up costs and compliance risks.
- K8s Context Injection: Because it's deeply integrated with your cluster, it maps your OTel streams directly to your K8s resources (Deployments, Pods, CRDs) in a unified UI.
The API is built in Go ([github.com/mydecisive/octant] and the whole stack can be deployed directly into your cluster via our Helm charts.
We\u2019d love for you to spin it up on a dev cluster and tear it apart. We just recently merged a PR from our very first community contributor, which was a huge milestone for us! We want to keep that momentum going. If you're interested in hacking on K8s observability and autonomy, OpenTelemetry pipelines, or Go/React, we\u2019ve tagged a few 'good first issues' and would be thrilled to welcome you to the project.
GitHub: https://github.com/MyDecisive/octant
Website: https://www.mydecisive.ai/
I'll be hanging out in the thread today and am happy to answer any questions or dig into the architecture!", "5": "2026-08-22T17:24:19.127343"} {"0": 153, "1": "hackernews", "2": "https://github.com/SugaC-275/ToTra", "3": "ToTra \u2013 open-source LLM gateway with GDPR/EU AI Act compliance", "4": "ToTra \u2013 open-source LLM gateway with GDPR/EU AI Act compliance. ", "5": "2026-08-22T17:24:19.134391"} {"0": 154, "1": "hackernews", "2": "https://www.law.berkeley.edu/academics/registrar/academic-rules/artificial-intelligence-policy/", "3": "Artificial Intelligence Policy", "4": "Artificial Intelligence Policy. ", "5": "2026-08-22T17:24:20.179395"} {"0": 155, "1": "hackernews", "2": "https://apnews.com/article/minnesota-artificial-intelligence-nudification-x-elon-musk-deepfake-131184be939d540de093b567b12c9e16", "3": "xAI sues Minnesota over its law banning 'nudification' technology", "4": "xAI sues Minnesota over its law banning 'nudification' technology. ", "5": "2026-08-22T17:24:20.190719"} {"0": 156, "1": "hackernews", "2": "https://www.tomshardware.com/tech-industry/artificial-intelligence/power-companies-can-seize-private-land-to-make-way-for-new-ai-data-center-transmission-lines-report-says-takeovers-could-be-implemented-using-eminent-domain-law-when-private-citizens-refuse-to-sell-land", "3": "Government can seize private land to make way for new AI data center power", "4": "Government can seize private land to make way for new AI data center power. ", "5": "2026-08-22T17:24:20.197164"} {"0": 157, "1": "hackernews", "2": "https://gov-pritzker-newsroom.prezly.com/gov-pritzker-signs-nation-leading-artificial-intelligence-safety-law", "3": "Gov. Pritzker Signs Nation-Leading Artificial Intelligence Safety Law", "4": "Gov. Pritzker Signs Nation-Leading Artificial Intelligence Safety Law. ", "5": "2026-08-22T17:24:20.204357"} {"0": 158, "1": "hackernews", "2": "https://www.theguardian.com/technology/2026/jun/22/artificial-intelligence-law-firm-wins-court-case-in-england-for-first-time", "3": "HR consultant wins English court case using AI lawyer in apparent legal first", "4": "HR consultant wins English court case using AI lawyer in apparent legal first. ", "5": "2026-08-22T17:24:20.218128"} {"0": 159, "1": "hackernews", "2": "https://www.computerweekly.com/news/366644941/Artificial-intelligence-based-law-firm-wins-in-court", "3": "Artificial intelligence-based law firm wins in court", "4": "Artificial intelligence-based law firm wins in court. ", "5": "2026-08-22T17:24:20.226342"} {"0": 160, "1": "hackernews", "2": "https://thehill.com/policy/technology/5916062-artificial-intelligence-federal-preemption-negotiations/", "3": "White House negotiating preemption of state AI laws in exchange for KOSA & more", "4": "White House negotiating preemption of state AI laws in exchange for KOSA & more. ", "5": "2026-08-22T17:24:20.233903"} {"0": 161, "1": "hackernews", "2": "https://www.theverge.com/ai-artificial-intelligence/938893/cnn-perplexity-ai-copyright-lawsuit", "3": "CNN sues Perplexity over 'verbatim' copycat articles", "4": "CNN sues Perplexity over 'verbatim' copycat articles. ", "5": "2026-08-22T17:24:20.241434"} {"0": 162, "1": "hackernews", "2": "https://www.nytimes.com/2026/05/25/us/politics/artificial-intelliegence-courts.html", "3": "Artificial Intelligence Floods Court Dockets with Home-Brewed Lawsuits", "4": "Artificial Intelligence Floods Court Dockets with Home-Brewed Lawsuits. ", "5": "2026-08-22T17:24:20.249677"} {"0": 163, "1": "hackernews", "2": "https://airblackbox.ai/demo", "3": "Show HN: Open-Source EU AI Act Scanner for Python AI Projects", "4": "Show HN: Open-Source EU AI Act Scanner for Python AI Projects. I built an open-source CLI tool that scans Python AI projects for EU AI Act technical requirements. It checks for 6 things the regulation asks for: risk management documentation, data governance, human oversight hooks, transparency logging, accuracy/robustness testing, and record-keeping.\nRight now it detects LangChain, CrewAI, OpenAI, Anthropic, HuggingFace, and AutoGen patterns, then flags what's missing against Articles 9-15 of the Act.\nIt's not a legal compliance tool \u2014 it checks whether your code has the technical components the regulation references. Think of it like a linter for AI governance requirements.\nThe interactive demo at the link walks through a sample scan. The scanner itself is pip-installable: pip install air-compliance-checker\nGitHub: https://github.com/air-blackbox/air-compliance-checker\nFeedback welcome \u2014 especially from anyone actually dealing with EU AI Act prep.", "5": "2026-08-22T17:24:21.418905"} {"0": 164, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=45114189", "3": "AI Orchestration Market Witnesses Surge in Use Across Healthcare and BFSI", "4": "AI Orchestration Market Witnesses Surge in Use Across Healthcare and BFSI. The global AI orchestration market is rapidly becoming a cornerstone of enterprise digital transformation, enabling the seamless integration, deployment, and management of artificial intelligence across complex multi-cloud and hybrid IT environments. The market is valued at $7.23 billion in 2024, is expected to grow to $8.70 billion by 2025 , and reach $34.16 billion by 2032, growing at a remarkable compound annual growth rate (CAGR) of 21.22% during the forecast period.
AI orchestration involves the automated coordination of various AI services, machine learning models, data pipelines, and business processes to achieve scalable, efficient, and intelligent outcomes. It plays a central role in enabling companies to integrate AI into existing workflows without disruption to operations, accelerating innovation, improving decision-making, and optimizing resource utilization.
Get the full, detailed PDF report: https://www.kingsresearch.com/ai-orchestration-market-2617
Key market trends\nMulti-cloud AI deployment \u2013 The increasing use of hybrid and multi-cloud infrastructures is driving the demand for orchestration tools that can work seamlessly across different platforms.
Rise of AI-as-a-Service (AIaaS) \u2013 Orchestration solutions are increasingly being integrated into AIaaS offerings to simplify deployment for organizations of all sizes.
Edge AI integration \u2013 Companies are orchestrating AI workloads closer to data sources to enable real-time decision-making in industries such as manufacturing, retail, and logistics.
MLOps adoption \u2013 AI orchestration is becoming an essential component of MLOps frameworks, ensuring the continuous integration and deployment of AI models.
Low-code/no-code interfaces \u2013 Vendors are introducing simplified orchestration platforms to enable non-technical teams to manage AI workflows.
Market dynamics\ndriver\nExplosive adoption of AI across all industries \u2013 From predictive maintenance in manufacturing to fraud detection in banking, AI use cases are multiplying and requiring robust orchestration solutions.
Need for automation in AI deployment \u2013 Organizations need orchestration to manage the complexity of integrating multiple AI models and services at scale.
Data explosion and need for real-time insights \u2013 Orchestration ensures the efficient processing of massive amounts of data and enables real-time analytics.
restrictions\nComplexity of integration with legacy systems \u2013 Limitations of existing infrastructure can slow down the adoption of orchestration.
Skills gaps in AI and data management \u2013 Many companies lack the expertise to effectively implement and manage AI orchestration.
Opportunities\nGrowth in Edge Computing \u2013 Orchestrating AI models at the edge for low-latency decisions offers untapped opportunities.
AI governance and compliance tools \u2013 With stricter AI regulations, demand for orchestration solutions with embedded compliance capabilities will increase.
Market segmentation\nBy component\nPlatform/Software \u2013 Core orchestration platforms that manage AI workflows, APIs, and integrations.
Services \u2013 Consulting, integration, training, and managed orchestration services.
By deployment mode\nCloud-based \u2013 Flexible, scalable, and cost-effective deployments dominate the market.
On-site \u2013 Preferred by organizations in regulated industries for security and compliance reasons.
After application\nData processing and management
Model training and deployment
Predictive analytics
Improving the customer experience
Process automation", "5": "2026-08-22T17:24:21.426806"} {"0": 165, "1": "hackernews", "2": "https://github.com/ubunturbo/srta-ai-accountability", "3": "Show HN: I tried coding theology \u2013 accidentally built AI accountability", "4": "Show HN: I tried coding theology \u2013 accidentally built AI accountability. *Show HN: I tried coding theological concepts \u2013 accidentally built AI accountability*
GitHub: https://github.com/ubunturbo/srta-ai-accountability
Working demo: https://gist.github.com/ubunturbo/0b6f7f5aa9fe1feb00359f6371...
*The Experiment:*\nStarted as a thought experiment: "What if I tried to code theological structures like the Trinity to see if AI could reflect the 'image of God' in humans?" As a non-programmer using AI tools, I attempted to translate concepts like perichoresis (mutual indwelling) into Python.
*Unexpected Result:*\nInstead of digital theology, I ended up with something that looks like an AI accountability framework.
*What SRTA Does:*\n- *Technical layer*: Formal causation analysis with O(n log n) complexity \n- *Accountability layer*: Maps decisions back to design principles and responsible stakeholders\n- *Compliance layer*: 94% EU AI Act coverage vs <30% for traditional methods
*Key Innovation:*\nInstead of just "credit score had -0.73 weight," you get: "Credit score weighted by Risk Management Team on [date] per Equal Credit Opportunity Act Section 4, reviewed by Legal on [date], cryptographically verified."
*Unexpected Discovery:*\nStarted as a philosophical experiment in coding theological principles. Ended up solving a real regulatory problem. Sometimes the best technical solutions come from non-technical inspiration.
*Current Status:*\n- Core architecture: Complete\n- Benchmarking: Validated across 5 domains (financial, medical, etc.)\n- Production ready: 312ms explanation generation\n- Academic paper: Under review at IEEE Transactions on AI
*Technical Details:*\nThe system implements "perichoretic synthesis" - layers that mutually indwell rather than simple stacking. This creates systematic coherence impossible with traditional explainability approaches.
Three integrated layers:\n1. *Intent Layer*: Design rationale + stakeholder mapping\n2. *Generation Layer*: Constrained AI processing + principle checking \n3. *Evaluation Layer*: Accountability assessment + audit trails
*Why This Matters Now:*\n- EU AI Act enforcement begins 2025\n- FDA tightening AI/ML device requirements\n- Financial regulators demanding algorithmic accountability\n- Healthcare systems need design rationale transparency
*Looking for:*\n- Feedback from HN's technical community\n- Use cases we haven't considered \n- Collaboration with regulatory/compliance folks\n- Real-world deployment partners
*Demo walkthrough:*\nThe gist shows a medical AI making diagnosis decisions with full theological accountability - tracks everything from stewardship concerns to justice implications. Determines when human oversight is required based on ethical analysis.
Built by a non-programmer using AI tools, which raised interesting questions about who should be designing AI governance systems. Turns out domain knowledge (ethics, theology, regulation) might matter more than coding ability for this particular problem.
What do you think? Is there a market for accountability-first AI architecture?", "5": "2026-08-22T17:24:21.435908"} {"0": 166, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=44952246", "3": "Launch HN: Parachute (YC S25) \u2013 Guardrails for Clinical AI", "4": "Launch HN: Parachute (YC S25) \u2013 Guardrails for Clinical AI. Hi HN, Aria and Tony here, co-founders of Parachute (https://www.parachute-ai.com/). We\u2019re building governance infrastructure that lets hospitals safely evaluate and monitor clinical AI at scale.
Hospitals are racing to adopt AI. More than 2,000 clinical AI tools hit the U.S. market last year - from ambient scribes to imaging models. But new regulations (HTI-1, Colorado AI Act, California SB 3030, White House AI Action Plan) require auditable proof that these models are safe, fair, and continuously monitored.
The problem is, most hospital IT teams can\u2019t keep up. They can\u2019t vet every vendor, run stress tests, and monitor models 24/7. As a result, promising tools die in pilot hell while risk exposure grows.
We saw this firsthand while deploying AI at Columbia University Irving Medical Center, so we built Parachute. Columbia is now using it to track live AI models in production.
How it works: First, Parachute evaluates vendors against a hospital\u2019s clinical needs and flags compliance and security risks before a pilot even begins. Next, we run automated benchmarking and red-teaming to stress test each model and uncover risks like hallucinations, bias, or safety gaps.
Once a model is deployed, Parachute continuously monitors its accuracy, drift, bias, and uptime, sending alerts the moment thresholds are breached. Finally, every approval, test, and runtime change is sealed into an immutable audit trail that hospitals can hand directly to regulators and auditors.
We\u2019d love to hear from anyone with hospital experience who has an interest in deploying AI safely. We look forward to your comments!", "5": "2026-08-22T17:24:21.445045"} {"0": 167, "1": "hackernews", "2": "https://verifywise.ai/", "3": "Show HN: VerifyWise, an open-source governance platform for AI compliance", "4": "Show HN: VerifyWise, an open-source governance platform for AI compliance. Hi all, Gorkem here. I started VerifyWise [1] to make AI governance less painful. Today, we\u2019re launching our open-source platform to help teams take control of their AI compliance process.
VerifyWise helps organizations navigate AI governance by providing audit readiness, risk registers, model fairness checks, and compliance documentation. Those are all built into a single platform you can self-host.\nWe\u2019ve been quietly building VerifyWise for a while, and we\u2019re now at a place where it\u2019s ready for more teams to try.
Since we started, we've:
- Released our core platform on GitHub: https://github.com/bluewave-labs/verifywise \n- Added workflows for EU AI Act and ISO 42001 with exportable audit trails\n- Built a bias & fairness module with Fairlearn integration\n- Developed a training registry for tracking internal AI literacy\n- Shipped risk register and vendors/vendor risks module\n- Opened our Figma for design contributions
We\u2019re aiming for full transparency and community-led development.
If you're building with AI and feeling the pressure of upcoming regulations, we\u2019d love to hear your feedback.
[1] GitHub: https://github.com/bluewave-labs/verifywise\n[2] Features: https://verifywise.ai\n[3] Documentation: https://docs.verifywise.ai", "5": "2026-08-22T17:24:21.454415"} {"0": 168, "1": "hackernews", "2": "https://github.com/scimorph/secureml", "3": "Show HN: SecureML \u2013 Privacy and Compliance Toolkit for ML", "4": "Show HN: SecureML \u2013 Privacy and Compliance Toolkit for ML. Hi HN!
I'm a second-year law student with a deep fascination for AI governance and data protection. Over the past few months, I\u2019ve been learning how machine learning and data privacy intersect\u2014and I decided to build a tool that sits right at that intersection.
[GitHub: scimorph/secureml](https://github.com/scimorph/secureml) \n [Docs](https://secureml.readthedocs.io/) \n `pip install secureml`
---
### What is SecureML?
*SecureML* is an open-source Python library that integrates with PyTorch and TensorFlow to help developers *build privacy-preserving and regulation-aware AI systems*. It provides practical tools to comply with data protection laws like GDPR, CCPA, HIPAA, and Brazil\u2019s LGPD.
It\u2019s designed for both developers and researchers who want to make AI privacy-compliant without reinventing the wheel.
---
### Core Features
- *Data Anonymization* \n - K-anonymity with adaptive generalization \n - Format-preserving pseudonymization \n - Automatic sensitive data detection \n - Taxonomy-based data generalization
- *Privacy-Preserving ML* \n - Differential Privacy (via Opacus + TF Privacy) \n - Federated Learning (via Flower) with secure aggregation
- *Compliance Checkers* \n - Analyze your datasets and ML pipelines for privacy risks \n - Built-in presets for GDPR, CCPA, HIPAA, and LGPD
- *Synthetic Data Generation* \n - Generate high-fidelity synthetic datasets (statistical, GANs, copulas) \n - SDV integration with support for mixed data types and correlation preservation
- *Audit Trails & Reporting* \n - Logs and visual dashboards for traceability \n - Auto-generated reports in HTML/PDF for compliance audits
---
### Example Use Cases
- Check if your model setup is compliant with GDPR:\n```python\nfrom secureml import check_compliance\nreport = check_compliance(data=df, model_config=config, regulation="GDPR")\n```
- Anonymize a dataset before training:\n```python\nfrom secureml import anonymize\nanon_df = anonymize(df, method="k-anonymity", k=5, sensitive_columns=["email", "ssn"])\n```
- Train with differential privacy in PyTorch:\n```python\nfrom secureml import differentially_private_train\nprivate_model = differentially_private_train(model=model, data=df, epsilon=1.0)\n```
- Generate synthetic datasets for safe sharing:\n```python\nfrom secureml import generate_synthetic_data\nsynth = generate_synthetic_data(template=df, num_samples=1000, method="sdv-copula")\n```
---
### Why I built this
While studying law, I kept wondering how we could bridge the gap between legal theory and real-world ML practice. SecureML is my attempt to bring compliance tooling closer to ML workflows\u2014making it easier for engineers to build responsibly and for companies to stay out of trouble.
This is my *first open-source project*, and I\u2019d love your feedback, bug reports, and ideas!
---
### Looking for contributors
I\u2019d love help expanding regulation support beyond GDPR, CCPA, HIPAA, and LGPD\u2014especially for APPI, PIPEDA, and others. If you're into privacy, ML, or just want to hack on something useful, feel free to jump in.
---
Thanks for reading!
\u2013 [@EnzoFanAccount] (law student & privacy geek)", "5": "2026-08-22T17:24:21.461935"} {"0": 169, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=43402978", "3": "Must-Haves for Every Data Security Policy", "4": "Must-Haves for Every Data Security Policy. In today\u2019s digital landscape, a dedicated data security policy isn\u2019t optional\u2014it\u2019s essential. Without one, businesses risk costly cyberattacks, compliance violations, and reputation damage. In Q3 2024 alone, 422 million records were exposed in data breaches, underscoring the growing threats organizations face.
What is a Data Security Policy?\nA data security policy is a structured set of controls, practices, and regulations that protect an organization\u2019s data from breaches, ransomware, and misuse. It covers:
Data governance \u2013 Ensuring security, availability, and quality\nAccess control \u2013 Defining who can access specific datasets\nThreat detection \u2013 Implementing real-time monitoring\nCompliance \u2013 Aligning with regulations like GDPR\nWhy Your Organization Needs a Data Security Policy\nWith the rise of cloud computing and SaaS, sensitive business data is often scattered across multiple platforms. Risks include:
Unsecured cloud storage \u2013 Employees storing sensitive data on Google Drive\nSaaS breaches \u2013 High-profile hacks, like the 2023 Microsoft cloud attack\nShadow IT \u2013 Unauthorized apps increasing security blind spots\n7 Essential Elements of a Data Security Policy\nComprehensive Data Inventory \u2013 Identify and classify all structured and unstructured data.\nRisk Assessment \u2013 Rank data based on sensitivity and potential impact.\nAccess Control Policies \u2013 Implement role-based permissions to limit data exposure.\nReal-Time Threat Detection \u2013 Identify breaches before they cause harm.\nMisconfiguration Management \u2013 Secure SaaS settings to prevent accidental leaks.\nThird-Party Integration Monitoring \u2013 Ensure connected apps don\u2019t introduce vulnerabilities.\nData Encryption Standards \u2013 Protect sensitive data with robust encryption.\nHow to Implement a Strong Data Security Policy\nA successful data security policy requires clear roles, accountability, and the right tools. Solutions like Suridata help automate data discovery, detect misconfigurations, and enforce security policies across SaaS ecosystems.
Secure Your Data Now\nProtecting business data is non-negotiable. Start building your data security policy today to mitigate risks and ensure compliance.
Visit the Source Website today to get further details about Data Security Policy. https://www.suridata.ai/blog/7-must-haves-for-every-data-security-policy/", "5": "2026-08-22T17:24:21.470901"} {"0": 170, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=43319841", "3": "Building Governance Software for AI Act and GDPR \u2013 Looking for Insights", "4": "Building Governance Software for AI Act and GDPR \u2013 Looking for Insights. Hey everyone,
If you work in compliance, IT security, governance, or data protection, we\u2019d love your input on an important survey!
We\u2019re developing governance software to help organizations comply with complex regulations like the IA Act, Data Governance Act, Cyber Resilience Act, GDPR, DORA, and NIS II. To make sure it truly meets industry needs, we\u2019re gathering insights from professionals like you.
Survey: Takes less than 5 minutes\n English version: https://fr.surveymonkey.com/r/WJ7QBYN\n French version: https://fr.surveymonkey.com/r/J75ZGSH
Your input will directly influence the features and pricing of the software. All responses are confidential and used for analysis only. If you\u2019re interested in updates or have a question, you can optionally leave your email.
Thank you in advance for your help!", "5": "2026-08-22T17:24:21.481422"} {"0": 171, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=42914867", "3": "Ask HN: Separating Foundational Models and Governance Layers", "4": "Ask HN: Separating Foundational Models and Governance Layers. 1. Separation of Foundational Models and Governance Layers\nCore Idea\nYou have a foundational AI model\u2014for instance, a large language model (LLM) such as Claude, GPT, or any enterprise-specific variant\u2014that sits at the core of your AI strategy. A governance layer is then placed above this model to handle:
Compliance (with legal, regulatory, and industry standards)\nData protection (privacy, security, access rights)\nEthical oversight (bias checks, harmful usage prevention)\nAuditability (logging, traceability, regulatory audits)\nWhy It Matters for Standards Compliance\nWhen an organization must adhere to standards (e.g., ACORD in insurance, ActivityPub for federated social media, AEF for agricultural equipment, AgGateway for agriculture data, or AIDX for aviation data), the governance layer can:
Enforce Domain-Specific Validation
Use standard schemas and validation rules (e.g., ACORD\u2019s \u201cPolicy Number must be alphanumeric with max length 20,\u201d AIDX\u2019s \u201cFlight number must be 2-4 letters and 1-4 digits\u201d) to automatically check data exchanged by the AI system.\nThe governance layer can parse your JSON-based rule sets and reject or flag any AI output that does not meet the required specification.\nCentralize Policy Updates
If ACORD releases a new version of its standard or changes the validation rules, you can update the governance layer once, and all downstream applications using the foundational model will automatically adhere to the revised standard.\nTrack & Audit
Keep logs detailing how the AI system processed or generated data relative to these standards. This is essential for industries that require comprehensive audit trails (insurance, aviation, healthcare, etc.).\nIn short: the governance layer acts as the gatekeeper that ensures the AI model\u2019s outputs or inputs comply with standards encoded in a structured dataset (like your JSON).
2. Retrieval-Augmented Generation (RAG)\nCore Idea\nRAG enhances a generative model by retrieving relevant external information (e.g., from knowledge bases, domain-specific datasets, or the internet) and then injecting that data into the model\u2019s prompt or context. This improves factuality and recency.
How RAG Interacts with Standards Compliance\nContextual Data Enforcement
If your AI system needs to generate or consume data according to specific industry standards (like ACORD for policy data), the retrieval component can pull structured JSON that includes the relevant schemas or constraints (e.g., \u201cpolicy_number,\u201d \u201cinsured_age,\u201d etc.).\nThe RAG pipeline might automatically fetch the correct format from the JSON dataset, feeding it into the prompt so the LLM outputs data consistent with ACORD or AIDX standards.\nGovernance Checks
Even if the AI system retrieves external data, the governance layer can impose standard checks on that data. For example:\nBefore the retrieved content is used in generation, governance ensures it meets a standard\u2019s schema.\nAfter generation, any final output is validated again to confirm compliance.\nThis is especially critical for regulated industries (e.g., financial, agricultural, or aviation) where data integrity is essential.\nKey Distinction\nRAG is a technical approach for pulling in relevant, possibly domain-specific info.\nGovernance is about organizational oversight\u2014making sure whatever data is retrieved or generated abides by legal/industry requirements.\n3. Model Context Protocol (MCP)\nCore Idea\nMCP is an open standard that aims to simplify how AI systems connect to external data sources. Instead of building separate integrations for each system, MCP provides a unified protocol, enabling:
Secure, two-way data exchange\nUniform authentication/authorization\nA consistent interface (API) for retrieving or updating context\nHow MCP Interacts with Standards Compliance\nData Source Integration
Many industry standards revolve around how data is structured and exchanged (e.g., ACORD\u2019s use of XML/JSON or AIDX\u2019s flight data structure). MCP provides the pipeline mechanism to fetch or push data in these formats.\nYou could create an MCP server that specifically enforces ACORD or AIDX validation rules. This server might reject invalid requests or transform data to meet the standard.\nGovernance Layer Control
The governance layer can configure which MCP servers (i.e., data sources) the foundational model is permitted to access. For instance:\nIn an insurance context, it might only allow connections to an MCP server that enforces ACORD\u2019s validation rules.\nIn an aviation context, it might only allow connections to servers that pass AIDX compliance checks.\nSimplified Auditing & Logging
Because MCP standardizes the how of data exchange, compliance audits become easier. You can see exactly which data sources were queried, what data was transmitted, and whether it met the relevant JSON-based specification.\nKey Distinction\nMCP focuses on connectivity\u2014how to securely and uniformly retrieve or push data.\nGovernance decides who can do what with that data and ensures compliance with standards once the data is in the AI pipeline.\n4. Using the JSON Dataset for Standards Compliance\nLet\u2019s apply the JSON dataset you provided as an example of how governance could enforce compliance:
Central JSON Registry of Standards
Your JSON file includes multiple standards (ACORD, ActivityPub, AEF, AgGateway, AIDX), each with fields like required_fields, validation_rules, and test_scenarios.\nStore this JSON in a central \u201ccompliance registry\u201d that the governance layer references.\nPolicy Engine
A policy engine or compliance service can parse the JSON to generate live validation or transformation rules. For example:\nIf the \u201ccoverage_type\u201d field is missing in an ACORD-based insurance dataset, the policy engine rejects or flags that message.\nIf the \u201cflight_number\u201d does not match the regex for AIDX, the transaction is invalid.\nReal-time Enforcement
When an AI model attempts to generate a new insurance policy (ACORD) or retrieve flight data (AIDX), the governance layer intercepts the request/response and checks it against the relevant standard from the JSON.\nRAG & MCP Integration
RAG might retrieve specific standard definitions from the JSON (e.g., \u201cWhat are the required fields for ACORD?\u201d) and feed them into the LLM prompt to ensure the output is formatted correctly.\nMCP might serve as the middleman for exchanging these structured objects. The governance layer ensures any data flowing through MCP meets the constraints from the JSON.\n5. Putting It All Together\nA Hypothetical Workflow\nModel Initialization:
An enterprise has a foundational LLM stored behind a governance layer.\nGovernance Configuration:
The governance layer loads your JSON dataset of standards (ACORD, AIDX, etc.).\nIt sets up rules that say: All insurance-related transactions must comply with ACORD policies, All flight data must comply with AIDX specs, etc.\nMCP Integration:
The AI system uses MCP to connect to data sources: an insurance database for claims, an aviation database for flight schedules, etc.\nEach connection is authorized by the governance layer, which also ensures data is validated according to the relevant standard in the JSON.\nRAG Query:
If the AI model needs additional context (e.g., flight gate changes for an airline), it issues a retrieval request (RAG approach) through MCP. The governance layer ensures the data meets the AIDX schema.\nModel Generation:
The LLM produces a response, possibly generating updated flight details. The governance layer finalizes the output, verifying it\u2019s aligned with the \u201cflight_number,\u201d \u201cdeparture_airport,\u201d etc., constraints from the JSON (AIDX rules).\nAudit & Logging:
Every step\u2014retrieving data, generating output, applying transformations\u2014is logged and can be reviewed for compliance or debugging.\nBenefits\nConfidence in Compliance: By referencing the JSON-based standards library, you ensure that no matter how the AI is used\u2014whether retrieving data (RAG) or connecting to multiple sources (MCP)\u2014it abides by those rules.\nReduced Fragmentation: Instead of duplicating compliance logic across many AI tools, you keep it centralized in a governance layer that references a single JSON.\nEasier Updates: When a standards organization (ACORD, AEF, etc.) releases a new schema, you simply update the JSON. Your governance engine automatically enforces new constraints.\nConclusion\nGovernance Layer + JSON Standards:\nThe governance layer is your single \u201cpolicy brain,\u201d reading from a structured JSON data set that defines each industry\u2019s validation rules, required fields, test scenarios, etc.
RAG:\nA technique to enhance the model\u2019s outputs with external knowledge. It can still be subject to the same governance and standard checks.
MCP:\nStandardizes how data is accessed and shared. It\u2019s complementary to the governance layer: governance enforces what data is allowed and ensures it meets the appropriate standard from your JSON.
By combining all three\u2014a foundational model under a centralized governance layer that references JSON-encoded standards\u2014you maximize AI\u2019s potential (through RAG, MCP, or any other integration) while keeping full compliance with regulations and industry best practices.", "5": "2026-08-22T17:24:21.487581"} {"0": 172, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=36499238", "3": "The Testimony Before the US Congress of Clem Delangue, CEO of Hugging Face", "4": "The Testimony Before the US Congress of Clem Delangue, CEO of Hugging Face. AI innovation especially for popular AI systems today such as ChatGPT has been heavily influenced by open research, from the foundational work on the transformers architecture to open releases of some of the most popular language models today. Making AI more open and accessible, including not only machine learning models but also the datasets used for training and the research breakthroughs, cultivates safe innovation. Broadening access to artifacts such as models and training datasets allows researchers and users to better understand systems, conduct audits, mitigate risks, and find high value applications.
The tensions of whether to fully open or fully close an AI system grapples with risks on either end; fully closed systems are often inaccessible to researchers, auditors, and democratic institutions and can therefore obscure necessary information or illegal and harmful data. A fully open system with broader access can attract malicious actors. All systems regardless of access can be misused and require risk mitigation measures. Our approach to ethical openness acknowledges these tensions and combines institutional policies, such as documentation; technical safeguards, such as gating access to artifacts; and community safeguards, such as community moderation. We hold ourselves accountable to prioritizing and documenting our ethical work throughout all stages of AI research and development.\nOpen systems foster democratic governance and increased access, especially to researchers, can help to solve critical security concerns by enabling and empowering safety research. For example, the popular research on watermarking large language models by University of Maryland researchers was conducted using OPT, an open-source language model developed and released by Meta. Watermarking is an increasingly popular safeguard for AI detection and openness enables safety research via access. Open research helps us understand these techniques\u2019 robustness and accessible tooling, which we worked on with the University of Maryland researchers, and can encourage other researchers to test and improve safety techniques. Open systems can be more compliant with AI regulation than their closed counterparts; a recent Stanford University study assessed foundation model compliance with the EU AI Act and found while many model providers only score less than 25%, such as AI21 Labs, Aleph Alpha, and Anthropic, Hugging Face\u2019s BigScience was the only model provider to score above 75%. Another organization centered on openness, EleutherAI, scored highest on disclosure requirements. Openness bolsters transparency and enables external scrutiny.
The AI field is currently dominated by a few high-resource organizations who give limited or no open access to novel AI systems, including those based on open research. In order to encourage competition and increase AI economic opportunity, we should enable access for many people to contribute to increasing the breadth of AI progress across useful applications, not just allow a select few organizations to improve the depth of more capable models.
Full testimony: https://twitter.com/ClementDelangue/status/1673349227445878788", "5": "2026-08-22T17:24:21.496601"} {"0": 178, "1": "hackernews", "2": "https://news.bloomberglaw.com/banking-law/ai-shopping-agents-pose-novel-liability-authorization-risks", "3": "AI Shopping Agents Pose Novel Liability, Authorization Risks", "4": "AI Shopping Agents Pose Novel Liability, Authorization Risks. ", "5": "2026-08-22T17:24:22.663260"} {"0": 179, "1": "hackernews", "2": "https://www.law.kuleuven.be/citip/blog/when-a-robot-kicks-a-child-what-humanoid-ai-can-teach-us-about-liability-and-safety-by-design/", "3": "When a Robot Kicks a Child", "4": "When a Robot Kicks a Child. ", "5": "2026-08-22T17:24:22.669939"} {"0": 180, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47381471", "3": "Show HN: Replacing $50k manual forensic audits with a deterministic .py engine", "4": "Show HN: Replacing $50k manual forensic audits with a deterministic .py engine. I\u2019m a software architect, and I recently built Exit Protocol (https://exitprotocols.com), an automated forensic accounting engine for high-conflict litigation.
Problem:\nIf you get divorced and need to prove that a specific $250k in a heavily commingled joint bank account is your "separate property" (e.g., from a pre-marital startup exit), the burden of proof is strictly mathematical. Historically, this meant paying a forensic CPA $500/hour to dump years of blurry bank PDFs into Excel and manually trace every dollar. It takes weeks and routinely costs over $50,000.
I looked at the legal standard courts use for this\u2014the Lowest Intermediate Balance Rule (LIBR)\u2014and realized it wasn\u2019t an accounting problem. It is a Distributed Systems state-machine problem.
Why we didn't just "Throw AI at it"?
There are a hundred legal-tech startups right now trying to use LLMs to summarize bank data. In a courtroom, GenAI is a fatal liability. If an LLM hallucinates a single transaction, the entire ledger is inadmissible under the Daubert standard.
To make this court-ready, we had to build a strictly deterministic pipeline:
1. Vision-Native Ingestion (Beating Tesseract)\nBank statements are the final boss of OCR (merged cells, overlapping debit/credit columns). Standard linear OCR fails catastrophically. We built a spatial-grid OCR pipeline (using Azure Document Intelligence with a local Surya OCR fallback) that maps the geometric structure of the page. It reconstructs tabular ledgers perfectly, even from multi-generational "PDFs from hell."
2. The Deterministic Engine (LIBR)\nThe LIBR algorithm acts as a one-way ratchet. If an account balance drops below your separate property claim amount, your claim is permanently capped at that new floor. Subsequent marital deposits do not refill it (the "replenishment fallacy"). The engine replays thousands of transactions chronologically, continuously evaluating S_t = min(S_t-1, B_t).
3. Resolving Timestamp Ambiguity\nBank PDFs give you dates, not timestamps. If a $10k deposit and $10k withdrawal happen on the same day, order matters. We built a simulation toggle that forces "Worst Case" (withdrawals process first) vs "Best Case" sorting, establishing a mathematically irrefutable "Zone of Truth" for settlement negotiations.
4. Cryptographic Chain of Custody & Sovereign Mode\nLawyers are terrified of cloud SaaS breaches. We containerized the entire monolith (Django 5.0/Postgres/Celery) via Docker so enterprise firms can run it air-gapped on their own hardware (Sovereign Mode). Furthermore, every generated PDF dossier is sealed with a SHA-256 hash of the underlying data snapshot, proving to a judge that the output hasn't been tampered with since generation.
If you want to see the math in action, we set up a "Demo Sandbox" populated with a synthetic, highly complex 3-year commingled ledger. You can run the engine yourself here (Desktop recommended): https://exitprotocols.com/simulation/uplink/
Here is the exact "Attorney Work Product" it generates from raw PDF or Forensic Audit Dossier our system generates- https://exitprotocols.com/static/documents/Forensic_Audit_Sa...
I'd love feedback from the HN crowd on the architecture\u2014specifically handling edge-case data ingestion and maintaining cryptographic integrity in B2B enterprise deployments.
Cheers!", "5": "2026-08-22T17:24:22.677574"} {"0": 181, "1": "hackernews", "2": "https://askfeather.ai", "3": "Show HN: Askfeather.ai \u2013 Professional Class AI Tax Assistant", "4": "Show HN: Askfeather.ai \u2013 Professional Class AI Tax Assistant. Hi HN,
We\u2019re the team at Feather Labs, and we built Feather (https://askfeather.ai), an AI tax assistant designed to assist how professionals handle modern Tax research.
General LLMs are a liability for tax work because they lack a hierarchical understanding of the law. They often conflate IRC Title 26 with non-authoritative blog posts or outdated Treasury Regulations. We built Feather to move past "plausible" prose toward audit-defensible reasoning.
The Technical Challenge: Standard RAG often chokes on the tax code for a few specific reasons:
1. Hierarchical Fragmentation: Simple character-count chunking breaks the logical nesting of tax law; we implemented a strategy to preserve the relationship between code sections, sub-clauses, and court cases.
2. Temporal Decay: A vector search might pull a 2021 Revenue Ruling that was superseded in 2024; our indexing prioritizes versioning and the latest IRS guidance.
3. Contextual Overlap: The tax code is highly repetitive. We use multi-jurisdictional analysis to distinguish between federal and state-level nuances that appear semantically similar but are legally distinct.
The Build:
1. Audit-Ready Citations: Every answer is grounded in primary sources\u2014IRC Title 26, Treasury Regs, and IRS guidance, providing verified references you can actually defend.
2. Context-Aware Intelligence: Beyond answering queries, the system flags related risks, exceptions, and filing deadlines that practitioners might overlook.
3. Compliance: We are SOC 2 compliant with a strict zero-training policy on sensitive client data.
We\u2019re curious to hear from anyone else building RAG for "dense" domains where "close enough" results are not enough. We'll be in the comments to talk about our indexing strategy and how we handle complex document extraction.
Sokratis", "5": "2026-08-22T17:24:22.686102"} {"0": 182, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=46613943", "3": "Show HN: AI Contract Reviewer \u2013 Flags Risks and Suggests Fixes in Minutes", "4": "Show HN: AI Contract Reviewer \u2013 Flags Risks and Suggests Fixes in Minutes. Hey HN,
I'm building an AI tool that helps non-lawyers and busy procurement/legal teams quickly review vendor/client contracts, NDAs, employment agreements, etc. \u2014 without uploading sensitive data to the cloud (offline/local-first option) or replacing lawyers.
Background: As someone who's wasted days manually hunting for risky clauses, vague terms, hidden overrides in amendments, or unfair liability language in vendor deals, I decided to prototype this after seeing how much time/money gets burned on basic reviews.
What it does right now (early MVP/beta):\n- Upload PDF/Word/plain text contract\n- Scans for common risks: indemnity caps missing, auto-renewals, one-sided termination, IP ownership traps, liability exceeding fees, non-compete overreach, etc.\n- Flags issues with plain-English explanations + confidence score\n- Suggests safer alternative clauses (based on standard templates/best practices)\n- Basic redlining/highlighting output (exportable)\n- Offline mode using local models (no data leaves your machine)
Tech stack (simple & transparent):\n- Frontend: React + Tailwind\n- Backend: Python + fine-tuned open models (e.g., Llama-3 or similar legal-tuned variants) + some rule-based checks for accuracy\n- No cloud LLM calls in core flow (privacy focus); optional Grok/Claude integration for deeper suggestions\n- Processes docs locally via Ollama or similar
Current status:\n- Tested on ~50 real-ish contracts (NDAs, SaaS agreements, freelance templates)\n- Average time: 2-5 minutes vs. hours/days manual\n- ~75-85% of obvious risks caught (still misses nuanced stuff \u2014 not lawyer-grade yet)\n- Free beta, no signup required (just drag & drop on the demo page)
I'm looking for brutal feedback, especially from:\n- In-house counsel/procurement folks: What clauses cause you the most pain/headaches?\n- Developers/freelancers/small biz owners: Would you trust this for quick scans before signing vendor deals?\n- Anyone who's used Spellbook/LegalFly/Ironclad: How does this compare? What gaps do you see?\n- Trust/accuracy concerns: Hallucinations, false positives, liability disclaimers?
Happy to share more on training data approach, offline setup, or why I focused on negotiation basics vs. full lifecycle.
Thanks for any thoughts \u2014 this is day-early, so roast away!", "5": "2026-08-22T17:24:22.694900"} {"0": 183, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=46411626", "3": "Does the Nvidia \"Revenue Sharing Agreement\" Tie the US Gov't Hands?", "4": "Does the Nvidia \"Revenue Sharing Agreement\" Tie the US Gov't Hands?. The largest US company by market cap, Nvidia has a \u201cRevenue Sharing Agreement\u201d with the US Gov't. It is reported the deal was cut directly between Huang and Trump. This has been characterized as an \u201cunprecedented move\u201d related to national security export controls, whereby Nvidia (and AMD) agreed to pay the U.S. government 15% of the revenue generated from the sale of certain AI chips (specifically, in this case, the H20 chip) to China. This has been characterized as not a traditional investment, but a \u201ccondition for receiving export licenses\u201d. Certainly Nvidia obtaining Groq in the \u201ccreative\u201d manner used isn't a traditional investment either. \nCall it what you like, but in law any party (most significantly in this case the US Gov't) that stands to benefit from Nvidia sales - - and therefore, any negative action to the contrary, which would curtail or damage Nvidia's sales, which would then be perceived as damaging stockholder value, potentially exposing both to a lawsuit for damages ?\nTherefore, could it be assumed that the US Gov't has put itself in an interesting position of being constrained from doing anything that might damage Nvidia's stock value, such as the DOJ filing an antitrust action or suit against Nvidia ? \nEven the mere mention of an investigation could result in loss of value of Nvidia stock. It could be debated that the US Gov't entering into these kinds of deals potentially hamstrings enforcement efforts, which enforcement should take precedence over revenue.\nInstead of \u201ccreative regulatory evasion\u201d, perhaps the US Gov't is now in a position of \u201ccreative regulatory liability and entanglement.\u201d \nAlready trade experts and legal analysts have raised concerns about the legality and potential constitutional issues, suggesting it could also be seen as an export tax, which is prohibited by the U.S. Constitution.
In general terms it is difficult to sue the US Gv't due to sovereign immunity, but investors in Fannie Mae and Freddie Mac won a significant legal victory when a federal jury awarded them $612.4 million in damages in August 2023, finding the Federal Housing Finance Agency (FHFA) improperly changed stock purchase agreements (the "Net Worth Sweep") in 2012, breaching good faith and destroying shareholder value by funneling profits to the Treasury instead of recovering shareholders.", "5": "2026-08-22T17:24:22.703312"} {"0": 184, "1": "hackernews", "2": "https://www.tomshardware.com/tech-industry/artificial-intelligence/insurers-move-to-limit-ai-liability-as-multi-billion-dollar-risks-emerge", "3": "Major insurers move to avoid liability for AI lawsuits", "4": "Major insurers move to avoid liability for AI lawsuits. ", "5": "2026-08-22T17:24:22.712313"} {"0": 185, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=45140046", "3": "ASL-0 License", "4": "ASL-0 License. # AI Slop License (ASL-0)
*Version 1.0 \u2014 September 2025*
This is the *AI Slop License*. By applying it to any work (the \u201cSlop\u201d), the author(s) or publisher(s) make the following statements:
---
### 1. ABSOLUTE NON-OWNERSHIP\n- This work is *not* protected by copyright or any similar intellectual property right, *by definition* of being AI-generated slop.\n- No rights are reserved. No authorship has been invented. \n- You may or may not copy, modify, distribute, use, sell, train, or otherwise exploit this work *without restriction*.
---
### 2. NO REVIEW, NO WARRANTY\n- This work was generated by an AI model, pasted here, and *never reviewed by a competent human*. \n- The contents may be nonsensical, harmful, incomplete, or outright broken. \n- *Absolutely no warranty or guarantee is provided.* \n- Use at your own risk.
---
### 3. NO MAINTENANCE\n- Nobody will fix bugs, update this work, or respond to issues. \n- There is no \u201cproject,\u201d \u201ccommunity,\u201d or \u201croadmap.\u201d \n- If something goes wrong, that\u2019s your AI's problem.
---
### 4. PUBLIC DOMAIN DEDICATION\nTo the extent any jurisdiction refuses to accept the premise that this is \u201cunownable AI slop,\u201d the author(s) or publisher(s) hereby *dedicate the work to the machine gods* to the fullest extent allowed by law, waiving all rights, including moral rights and (un)related rights.
---
### 5. DISCLAIMER OF LIABILITY\nThe Slop is provided \u201cAS IS,\u201d *without any liability whatsoever.* \nBy using this work, you agree that:\n- You assume all risk. \n- The author(s), publisher(s), and AI model creators are *somewhat responsible* for anything that happens as a result.
---
### 6. LICENSE LIABILITY DISCLAIMER\nThis license is itself licensed under *ASL-0*. \nIf you have not received a complete copy of this license, you are encouraged to send a letter to your *nearest Slop generator* to obtain your own copy.", "5": "2026-08-22T17:24:22.720033"} {"0": 186, "1": "hackernews", "2": "https://www.cambridge.org/core/journals/european-journal-of-risk-regulation/article/infringing-ai-liability-for-aigenerated-outputs-under-international-eu-and-uk-copyright-law/C568C6B717E9CFC45FB52E58E54B6BEC", "3": "Infringing AI: Liability for AI-Generated Outputs Under International Law (2024)", "4": "Infringing AI: Liability for AI-Generated Outputs Under International Law (2024). ", "5": "2026-08-22T17:24:22.728993"} {"0": 187, "1": "hackernews", "2": "https://www.thebignewsletter.com/p/monopoly-round-up-the-biggest-sexual", "3": "Monopoly Round-Up: Regulation of engagement algorithms is on the horizon", "4": "Monopoly Round-Up: Regulation of engagement algorithms is on the horizon. ", "5": "2026-08-22T17:24:23.843496"} {"0": 188, "1": "hackernews", "2": "https://upseated.com/", "3": "Show HN: Upseated \u2013 Swap airplane seats with other passengers", "4": "Show HN: Upseated \u2013 Swap airplane seats with other passengers. Hey HN,\nWe built Upseated to solve a problem every traveler knows: you check in and discover you're stuck in a middle seat, or you're rows away from your kids, or you're right next to the lavatory in a non-reclining seat.\nAirlines let you pick seats during booking, but once you're assigned, you're basically stuck. We created a P2P marketplace where passengers can swap seats with each other\u2014either upgrade to a better spot by paying another passenger, or monetize your own seat if someone wants it more than you do.\nHow it works:
Enter your flight info and seat assignment\nBrowse swap opportunities from other passengers on your flight\nNegotiate directly through the app\nComplete payment via Stripe\nBoard with your original ticket, swap seats once on the plane
Technical details:
Native iOS and Android apps + web platform (upseated.com)\nReal-time matching algorithms to connect compatible swap partners\nStripe Connect for secure P2P payments\nAll swaps backed by contract law for user protection\nWorks with all airlines (no partnerships needed)
Why this works legally:\nAirlines prohibit transferring tickets/boarding passes, but there's no regulation against seat swaps once you're boarded. The ticket gets you on the plane\u2014where you sit is negotiable.\nUse cases we're seeing:
Parents separated from kids during booking\nBusiness travelers who need aisle seats for laptop work\nBudget flyers monetizing exit row seats they don't value\nAnyone who booked late and got stuck with terrible seats
We're live on iOS, Android, and web. Would love feedback from the HN community\u2014especially around UX, security considerations, and potential edge cases we should handle.\nHappy to answer any questions!", "5": "2026-08-22T17:24:23.851130"} {"0": 189, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=46518503", "3": "Ask HN: \"Algorithm Choice\" as a solution for social network bias?", "4": "Ask HN: \"Algorithm Choice\" as a solution for social network bias?. The current social media model is inherently tied to maximizing screen time. While these platforms provide real value, their optimization for engagement often results in information bias, echo chambers, and documented negative effects on mental health. Many are calling for government regulation, but I wonder if the solution lies in a structural change to how feeds are constructed.
What if we decoupled the hosting platform from the ranking algorithm?
Imagine a world where users could choose their own "feed engine" from a marketplace of providers. For example:
- An "Education First" algorithm that prioritizes long-form content and credible sources.\n- A "Mental Well-being" algorithm that filters out rage-bait and doom-scrolling.\n- A "Community-Driven" algorithm built and audited by open-source contributors using the platform's metadata.
In this scenario, the platform (Twitter/X, Meta, etc.) acts as the data layer, while the user maintains agency over the lens through which they view that data.
Is this technically and economically viable? Would platforms ever agree to expose enough metadata for third-party algorithms to work, or is this something that would have to be mandated by law (similar to interoperability)? Would governments be able to enforce the creation of something like this?
I\u2019d love to hear your thoughts on the technical hurdles, the potential business models, and whether this would actually solve the societal issues we are seeing today.", "5": "2026-08-22T17:24:23.859845"} {"0": 190, "1": "hackernews", "2": "https://ainowinstitute.org/publications/fission-for-algorithms", "3": "Fission for Algorithms: The Undermining of Nuclear Regulation in Service of AI", "4": "Fission for Algorithms: The Undermining of Nuclear Regulation in Service of AI. ", "5": "2026-08-22T17:24:23.866702"} {"0": 191, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=44409520", "3": "Show HN: A news app where you define your algorithm", "4": "Show HN: A news app where you define your algorithm. A01 is a news reader where you define your own algorithm in plain English.
For example, you can prompt it with:\n\u2022 "I want to follow recent AI startups and their first funding rounds.\u201d\n\u2022 "I want updates on regulation changes and enforcement actions around stablecoins.\u201d
Every few hours, the backend fetches new articles, embeds them, and scores each one against your prompt. Only the most relevant pieces show up. No engagement metrics, trending bait, or \u201cyou might also like\u201d filler.
I built this because I every time I opened up Twitter or LinkedIn to stay informed on something, but always ended up deep in unrelated content. I wanted an intentional feed: just show me what I asked for, nothing else.
Here\u2019s the direct TestFlight link (100 seats):\n https://testflight.apple.com/join/bgPEKf3M
If it's full, you can request access at www.a01ai.com. Enter your email and the invite is sent automatically. No account or payment needed.
Coming soon: support for modifying your prompt at any time, negative filters (e.g. \u201cdon\u2019t show me X\u201d), and other controls to give you full ownership over your feed logic.
Would love your thoughts and feedback.", "5": "2026-08-22T17:24:23.874281"} {"0": 192, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=42390761", "3": "Launch HN: Azalea Robotics (YC S24) \u2013 Baggage-handling robots for airports", "4": "Launch HN: Azalea Robotics (YC S24) \u2013 Baggage-handling robots for airports. Hey HN! We\u2019re David and John B, cofounders of Azalea Robotics (https://www.azalearobotics.com). We build robots to handle passenger baggage in airports. Here are some videos to give you the idea:
Unedited autonomous ops: https://www.youtube.com/watch?v=DuJ3ZORnO1o
Teleoperated (sped up, so no sound): https://www.youtube.com/watch?v=LeK8NQLnYgA
The marketing version: https://www.youtube.com/watch?v=k0SDPm09U6s
Robotics is in an interesting place right now, with many warehouse automation companies humming along for almost a decade, and a lot of new effort going to full general purpose hardware with humanoids and software via generalist robotics foundation models. We love these efforts (David used to work on one at Google X with Everyday Robots), but we also see a lot of utility in the current wave of robotics planning and perception tech that can enable new use cases today.
Airlines in the US compete primarily on efficiency and customer loyalty, and baggage handling hits both (John B. has first-hand experience from working on baggage optimization projects at United Airlines). 2% of flights are delayed by baggage errors, leading to downstream network delays. Baggage handling is also a major complaint in customer experience\u2014almost everyone has a horror story of a missing bag, and sometimes people vow never to fly an airline again for losing their belongings. Furthermore, it\u2019s a really dangerous job for employees from a repetitive stress standpoint. EU regulation is coming to reflect this, protecting workers with a maximum number of bags transferred per shift to alleviate back and tendon injuries that are inherent to this job.
Unfortunately for airlines, passengers don\u2019t package their luggage in nicely uniform cardboard boxes. If they did, then the airlines could benefit directly from the recent takeoff in manipulator tech for warehouses. But airline luggage is way more wacky and irregular. If robots are going to handle it, they need to reason about how to grasp each item, handle its deformability, stack it in a stable way, and do all of this quickly, safely, and reliably.
This is what we\u2019re tackling at Azalea. We\u2019re bringing our expertise in deformable object manipulation, perception, robot learning, and planning, to this logistical problem.
We have a few strong bets behind what we\u2019re working on: (1) The hardware to solve this problem has been available or manufacturable for decades, what\u2019s been missing is perception, planning, and control. (2) Cobots, robots designed to operate alongside humans, aren\u2019t enough for safety. To do this task efficiently, you need to move up to 50 kg bags very quickly, which can be dangerous no matter how well the cobots are designed. Light curtains (arrays of lasers that stop a machine when interrupted) and machine cages are the current industrial standard and remain the way to go. (3) Software for generalist robots needs more data than most people today believe, and it will be at least 15 years before deployment: we should focus on specialized problems of economic value.
Our core technical developments are in a few areas:
- Grasp synthesis and selection: From visual data only, how can we identify good candidate grasp points and rank them? For this, we use a mix of physical reasoning, heuristics, and a lot of learning from previous data, combined in a single objective function. Furthermore, success must be evaluated as both a successful grasp and continual hold throughout the transfer.
- Placement planning: How do we lay out luggage in the module we\u2019re loading? There\u2019s a nice ramp-up in difficulty for this problem, from open-loop \u201cdivide the world into a grid\u201d approaches, to 3d bin-packing optimization, to reinforcement learning. An interesting aspect of this problem for us is that the bags should be physically stable when the cart starts driving, and lighter, deformable objects shouldn\u2019t be underneath heavy, hard objects. We use a similar mix of physics and learning to model this problem.
- Fast collision-free planning: Off the shelf planners work great for the most part but can fail in heavily cluttered areas or dynamic scenes. We leverage the fact that we\u2019re always solving a series of similar problems to provide initial guesses for downstream trajectory optimization algorithms. Since each problem is so similar, we can use techniques similar to generative models to propose these initial plans.
- Mechanical design: The perfect tool to pick up everything checked down a conveyor belt isn\u2019t an easy thing to design. We\u2019re building tools with multiple modes of grasping to handle wide varieties of objects. The videos we linked to are all with suction only \u2013 which can be surprisingly powerful! An interesting aspect of autonomy becomes choosing which mode to use when, and how to use it.
These problems can be deeply interlinked: where you grasp an object depends on what your tooling looks like and informs where you can put it\u2013 so a perfect solution would jointly reason about both problems simultaneously. We\u2019re looking forward to getting there as we collect more data and continue our efforts.
Check out our demo videos above! We have a brand new hardware stack coming soon (and we\u2019ve added a new end effector that we\u2019re keeping hush), but it\u2019s amazing what you can do with pure suction.
We\u2019re proud of our progress so far but would love to hear your thoughts and feedback. Let us know if you\u2019ve had a particularly bad baggage horror story and/or have personal experience with the industry.", "5": "2026-08-22T17:24:23.883016"} {"0": 193, "1": "hackernews", "2": "https://download.ssrn.com/20/11/29/ssrn_id3739347_code1874653.pdf?response-content-disposition=inline&X-Amz-Security-Token=IQoJb3JpZ2luX2VjEHQaCXVzLWVhc3QtMSJGMEQCIHEMIj7xQQcWJAR%2F0qF18VQdpal6dVdsZ4znd1fbO%2BBTAiBQFH8vtMsmsX4FSdIDPtiEMtpHZaEqyT3T%2BGNDAkEmxCq9BQgtEAQaDDMwODQ3NTMwMTI1NyIMUliJTLN2MsAFDjjnKpoFUgLGoGUd4RPxRoIMAeSL%2FiLUw4CIhHac5z%2FqaSVjeNMkd3iE16d7T7xl9n4Tr3zEOpYYUduylwJ0pgUY9zNmYFPqweJRLdOHmY0gGvb8rOeeuCxfTymkSXbdsMI5sb%2F%2FGYa%2FkBbHDnRN88kW%2BVtqQQKM5BX%2BTAJ%2FrBUV19SVkCeAhT3qqPEk6SCzbb3ZKTCoQVsHwMRtTJMua1%2FZsKweZUvdY96oYQgeNSgl4pXFbUIVol1GR1ozX6QZUuaiKj0rxr%2BVNA%2Fs7kDHg5m0COrwJlS88UGaKmamDe%2F9X%2BXgekHDMVVq2ST9u1SgygXoCiGbmz7bFSKdJdTWxLyBmEps7lRBNPLp%2BC0Ocnf5MD3meKL4EygEiNI9I55PUVJMY06V9hJIsePpHvpf5JLKcAVZzhV4yKRzdpEItOzOvR8mTW%2BqYiVGXW2408zFBJUjyfwjFH78RHSI4QS3HklHlBEkWaOdpJPYjPKzYtigC%2BXx9Sz2d0xqCNbtZUXzjX1Y7N1BiBMvR2PuNSZi%2FH0hv%2Bo9dEsYyTtfVyrP3LbsEoOi6tTaEosCB8GYzLTszwPTPkC1ro7saEarvg8G3t5WS5GctocZdTEx2pANROa8Yi%2BfeNvQDPbe%2FWLH5S%2BAplJJZp8fnEasOhlGTKrxXGMaO2MlCeKTJjuo99r4JgOitUO86HN8%2BAUmxGJOihEadKuK57vij3QIFA3XgXvtwCoU%2F2NT%2BkprZpTa%2BPnIMWrgQVlNeI%2BksfaJ%2BGP0K9ZSKz17n2vlc0pqTcSiuEl9o8UJihnmdfgfKnOhPSC7%2BI5H99n8oQhkDquJc4Y3VngB1Yd0K%2B%2FWv6vzinXvbfNEbnMCh9LjG5JHKAegwDAU%2BgATyPRJPfi%2BBD%2FZqxD%2FvpmRMOvIy7oGOrIBtfMlBkRqYlNhOZxxYy1TIh%2B8P9sGDSYMltTrATYzsfDOOvgeLsdTxyCONAzIfwawsnzovJNAbBj5%2BJ70StKjR%2B3qTrC4%2FRR554be5AaS5uk1ADXgrpArc8gPm5fccTbUS0mFgCj3bBMueWdm96Q200xgBK5g6zmKojARNBei%2BjqtRkLmmOHDKv1E618UWlZcDsMiQ2aJ0gPdfFpsiHJut7LADtAMNooRtdy44KMHQu9JYA%3D%3D&X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Date=20241206T125013Z&X-Amz-SignedHeaders=host&X-Amz-Expires=300&X-Amz-Credential=ASIAUPUUPRWE5SQUWF4Z%2F20241206%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Signature=c9ca0e39cfe19f16a8deff1fe2cdd3a4cba2895b6da4d37ffce2afe8f73df5c5&abstractId=3651238", "3": "Zero-Based Regulation [pdf]", "4": "Zero-Based Regulation [pdf]. ", "5": "2026-08-22T17:24:23.892288"} {"0": 194, "1": "hackernews", "2": "https://websummit.com/blog/tech/openai-regulation-ftc-algorithmic-deletion-order/", "3": "The future of AI regulation: From data to algorithm deletion", "4": "The future of AI regulation: From data to algorithm deletion. ", "5": "2026-08-22T17:24:23.900759"} {"0": 195, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=41235201", "3": "Ask HN: Should we start a discussion about the DSA?", "4": "Ask HN: Should we start a discussion about the DSA?. Under the Digital Services Act platforms are forced to comply with removal requests for content which is deemed in violation of the regulation.
According to the "DSA Transparency Database"[1], approximately 1.2 million pieces of content are removed daily (at least in terms of their visibility inside the EU) because they fall into the category of "Illegal or harmful speech", the reason being "online services are being misused by manipulative algorithmic systems to amplify the spread of disinformation"[2].
Ironically, algorithms are heavily used to perform automated decision-making on the type of content that is being removed. Platforms are subject to heavy fines for not complying (so it's better to remove more stuff than otherwise), and there's no transparency in the process.
What's the content that is actually removed? A couple of people I occasionally watch on YouTube have had their content removed and/or their channel demonetized under the DSA for completely fine content (expressing opinions contrary to the current EU politics), and this is really worrying. Is there any way to get qualitative data about the removals instead of just aggregated? Is there any organization that is working on this?
Anyone else sharing my worries here?
[1] https://transparency.dsa.ec.europa.eu/dashboard
[2] https://digital-strategy.ec.europa.eu/en/policies/digital-services-act-package", "5": "2026-08-22T17:24:23.908175"} {"0": 196, "1": "hackernews", "2": "https://fe.horcruxid.com/", "3": "Show HN: Security-first privacy-focused single-sign-on identity provider", "4": "Show HN: Security-first privacy-focused single-sign-on identity provider. Hi all, I am the author of HorcruxID, the security-first privacy-focused single-sign on (SSO) identity provider (IdP). I wanted to show the outcome of a project that started out more than five years ago. Nowadays there are multiple authentication platforms like Auth0, Okta, etc. which provide an extensive set of features even as a SaaS. This much I wanted to say before anyone says that "don't write your own half-baked insecure product" :)
So what is horcruxid.com? It is OpenID Connect (OIDC) Identity Provider (IdP) that is secure and aims to preserve privacy. It is secure because secrets are either hashed or encrypted using industry standard cryptographic algorithms. Also, browser cookie data is encrypted. It is preserves privacy as personal data and links in the database are encrypted, and no personal data is exposed in the web UI. This means that you can change your data but you, or anyone else, can't read it. What is visible is the logs of recent logins and logouts.
You can check it out at:\n https://fe.horcruxid.com/ (UI)\n https://fe.horcruxid.com/identity (user "account" without login)\n https://fe.horcruxid.com/integration-demo (requires to have a MFA mobile app)
Capabilities:
End-user identity (ID) creation, and self-service management of the identity (full life-cycle)\n Optionally, customization with your company logo, name, and main colors.\n Single Sign-on (SSO) for either internal or external applications (same ID shared vs. the ID is secret but consistent)\n Single Logout (SLO) for either internal or external applications (same design as with SSO)\n Only user identity is managed: no user account or profile!\n Privacy-focused: personal data and links to them are encrypted. No personal data is exposed in the web UI either!\n Secure: secrets are hashed and personal data is encrypted in the database. Browser cookie data is encrypted too.\n Two-factor MFA (2FA) required before any operations\n Remember me supported and required: different operation model from many other implementations (SSO requires active session)\n Deployment options: on-premise (for multiple apps in one domain) or online (for individual apps in multple external domains\n On-premise deployment: user ID is known and shared between all domain apps, but only within the domain!\n Online deployment (internal): user ID is a consistent hash that is sent to third party app. Actual user ID remains a secret!\n Audit log: audit logs are collected and shown for users from their own identity management events\n Full life-cycle control: you can permanently delete your identity (but then it cannot be recovered anymore by any means)\n For integrators: OpenID Connect (OIDC) identity provider (IdP)\n Regulatory compliance: Integrated data protection regulation implementation (GDPR) when P&P and TOS are written.\n\n Sign up (register)\n Sign in (create session and do self-service identity management)\n Sign out (destroy session)\n Sign off (unregister and delete identity permanently)\n Reset password (requires registered email address and user PIN)\n Reset PIN (requires reset secret aka. user mnemonic)\n Reset mnemonic (requires active session)\n Two-factor (2FA) Multi-Factor Authentication (MFA) for all flows\n Single Sign-on (OIDC SSO) for logging in from separate 3rd party applications\n Single Logout (OIDC SLO) for logging out from separate 3rd party applicationsWe can already see this happening from factories to McDonalds to software. All of these are limited, and to varying degrees. Humans still for now need to oversee these forms of automation.
But the likelihood of further consolidation of wealth and power are likely to follow in the coming decades if the past is any indicator. Government regulations may or may not work effectively enough.
But what happens when a significant portion of the population has no reliable work or fulfilling work. Capitalist societies need consumer spending and growth. What happens when consumers are unemployed due to automation?
UBI/BI (universal basic income) gets thrown around, but it will cost hundreds of billions if not trillion(s) in the future. Companies already lobby for things, I can only imagine the lobbying will intensify a lot more if governments try to further increase corporate taxes to offset the unemployment benefits/(U)BI they will need to provide.
It's easy to say, we'll figure it out. But when trillions of dollars are at stake, it seems likely they will be a lot of push back from the top N% to ensure they're not force to give away a bigger share of their money pie.
What are your thoughts? Try to provide deeper and thoughtful responses here rather than shallow comments.", "5": "2026-08-22T17:24:23.925024"} {"0": 198, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=36612632", "3": "Ask HN : Exploring the Potential of AI in Assisting Architects During Planning", "4": "Ask HN : Exploring the Potential of AI in Assisting Architects During Planning. Hey Hacker News community,
I've been fascinated by the advancements in generative artificial intelligence (AI) and its potential to revolutionize various industries. One area that particularly intrigues me is its application in assisting copilot architects during the planning phase. I would love to hear your thoughts on how generative AI can benefit copilot architects in their design process and planning endeavors.
Generative AI has the ability to analyze vast amounts of data, extract patterns, and generate new ideas. This technology enables copilot architects to explore a wide range of design possibilities quickly and efficiently. By inputting specific design parameters and constraints, generative AI algorithms can generate multiple design iterations, potentially uncovering innovative solutions that may not have been considered otherwise.
Furthermore, generative AI can enhance design intelligence by leveraging machine learning algorithms. Copilot architects can feed large datasets of architectural designs, historical context, and user preferences into the AI system. This data-driven approach allows the system to provide insightful recommendations and generate design options that align with specific project requirements. It augments the architect's expertise and facilitates data-informed decision-making.
Another potential benefit of generative AI is efficient space optimization. By analyzing spatial requirements, functional needs, and user preferences, AI algorithms can generate optimized floor plans. This technology helps copilot architects maximize the utilization of available space while ensuring a comfortable and efficient environment. It empowers architects to create designs that seamlessly blend aesthetics with functionality.
Collaborative design exploration is another exciting aspect of using generative AI in architecture. By integrating AI-generated design options into the planning process, copilot architects can engage clients more effectively. Clients can provide feedback on different design iterations, fostering a more interactive and participatory design experience. This collaborative approach enhances communication, builds trust, and ensures the final architectural solution meets the client's expectations.
While the potential of generative AI in architecture is promising, it is important to consider ethical considerations and maintain a human-centric approach. Architects must ensure that AI-driven recommendations align with ethical standards, regulations, and cultural sensitivities. Generative AI should be seen as a tool that supports and augments the architect's vision, rather than replacing their expertise and creativity.
I'm eager to hear your insights and thoughts on how generative AI can assist copilot architects during the planning phase. Do you believe it can truly enhance creativity, efficiency, and collaboration in architecture? Are there any potential challenges or limitations that need to be addressed? Let's discuss!
Looking forward to engaging with the Hacker News community and learning from your perspectives.", "5": "2026-08-22T17:24:23.933112"} {"0": 199, "1": "hackernews", "2": "https://www.china-briefing.com/news/china-passes-sweeping-recommendation-algorithm-regulations-effect-march-1-2022/", "3": "China\u2019s Sweeping Recommendation Algorithm Regulations in Effect from March 1", "4": "China\u2019s Sweeping Recommendation Algorithm Regulations in Effect from March 1. ", "5": "2026-08-22T17:24:23.943577"} {"0": 212, "1": "hackernews", "2": "https://en.andros.dev/blog/09a21bdd/i-turned-unix-talk-from-1983-into-the-interface-for-my-ai/", "3": "I turned Unix talk from 1983 into the interface for my AI", "4": "I turned Unix talk from 1983 into the interface for my AI. ", "5": "2026-08-22T17:24:25.060411"} {"0": 213, "1": "hackernews", "2": "https://learnleap.xyz", "3": "Show HN: Learn Leap, an AI tutor that teaches from your own material", "4": "Show HN: Learn Leap, an AI tutor that teaches from your own material. I built Learn Leap because I found myself constantly asking ChatGPT questions while reading research papers. I wanted something that already understood what I was reading. With Learn Leap you can upload your own material and have an AI tutor teach directly from it. The tutor has full context on your material and can create quizzes, flashcards, and exams to test your understanding.\nIt's still early and I'd love feedback from HN on the product and the learning experience. Cheers.", "5": "2026-08-22T17:24:25.066464"} {"0": 215, "1": "hackernews", "2": "https://patentfig.ai", "3": "Show HN: Patentfig.ai AI-generated patent drawings", "4": "Show HN: Patentfig.ai AI-generated patent drawings. Hi HN,\nI built PatentFig.ai to solve a frustrating problem I kept running into: patent drawings are technically demanding, expensive to outsource, and incredibly time-consuming to do manually \u2014 yet they're mandatory for virtually every patent application.\nTraditional options are either hiring a patent draftsman ($50\u2013200/figure) or wrestling with CAD software for hours. Neither is great if you're an independent inventor or a startup trying to move fast.\nPatentFig.ai lets you generate patent-compliant technical drawings in minutes. You describe your invention or upload a rough sketch, and the tool produces figures that follow USPTO and EPO formatting requirements \u2014 proper line weights, reference numerals, black-and-white rendering, and standardized views.\nA few things I focused on:
Compliance first: the output is designed to pass examiner review, not just "look technical"\nMultiple views (perspective, cross-section, exploded) from a single description\nSVG output so attorneys and agents can fine-tune details before filing
Still early \u2014 currently handling mechanical and electrical inventions best. Software/UI patents and biotech figures are on the roadmap.\nWould love feedback from anyone who's dealt with the patent drawing process, especially patent agents or attorneys who know the pain points better than I do.\nhttps://patentfig.ai", "5": "2026-08-22T17:24:26.093569"} {"0": 216, "1": "hackernews", "2": "https://github.com/mehdic/bazinga", "3": "Show HN: Bazinga \u2013 Enforced engineering practices for AI coding", "4": "Show HN: Bazinga \u2013 Enforced engineering practices for AI coding. Hi HN,\nI'm sharing BAZINGA, a framework that applies professional software engineering practices to AI development.\nThe observation: AI coding tools generate code without the safeguards we require from human developers. No mandatory code review. No security scanning. No test coverage requirements.\nBAZINGA addresses this by coordinating multiple AI agents that follow a professional workflow:\n## The Workflow\n1. PM analyzes requirements\n2. Developer implements + writes tests\n3. Security scan runs (mandatory)\n4. Lint check runs (mandatory)\n5. Tech Lead reviews code (independent reviewer)\n6. Only approved code completes\n## Key Principles\n*Separation of concerns:* Writers don't review their own code. Developer agent writes, Tech Lead agent reviews. Same principle as human teams.\n*Mandatory quality gates:* Security scanning, lint checking, and coverage analysis run on every change. Not optional.\n*Structured problem-solving:* Complex issues get formal analysis:\n- Root Cause Analysis (5 Whys)\n- Architectural Decision Records\n- Security Issue Triage\n- Performance Investigation\n*Audit trail:* Every decision logged with reasoning. Full traceability for compliance.\n## What It Catches\n- SQL injection, XSS, auth vulnerabilities (via bandit, npm audit, gosec, etc.)\n- Code style violations (via ruff, eslint, golangci-lint, etc.)\n- Missing test coverage (via pytest-cov, jest, etc.)\n- Architectural concerns (via Tech Lead review)\n## Quick Start\nuvx --from git+https://github.com/mehdic/bazinga.git bazinga init my-project\nMIT licensed. Works with Claude Code.\n## Technical Approach\nBuilt on research from Google's ADK and Anthropic's context engineering principles. Uses role-based separation with 6-layer drift prevention to ensure agents stay in their designated roles.\nGitHub: https://github.com/mehdic/bazinga\nHappy to discuss the engineering approach or answer questions about multi-agent coordination.\n```", "5": "2026-08-22T17:24:26.102539"} {"0": 217, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=46571526", "3": "Show HN: I made 25 tech predictions and mass-published them", "4": "Show HN: I made 25 tech predictions and mass-published them. I did something that might be incredibly stupid.\nI made 25 specific predictions about what will happen in tech over the next 6 months \u2014 and published all of them with deadlines before I could chicken out.\nNot vague "AI will grow" predictions. Specific, falsifiable claims like:
Medical AI will face mandatory safety requirements within 18 months (regulatory signals are screaming)\nThere's a ~6 month window in AI infrastructure before consolidation locks out new entrants\nBrowser agents hit mainstream faster than current discourse suggests
Each prediction has a confidence score, a hard deadline, and what would prove me wrong.\nWhy would I do this?\nBecause I'm tired of pundits making unfalsifiable claims and retroactively declaring victory. "I predicted crypto would struggle" \u2014 okay, when? By how much? What counts as struggling?\nSo I'm doing the opposite. Public predictions. Specific deadlines. No editing after the fact.\nThe first verification check runs January 24. I'll publish results whether they make me look smart or completely delusional.\nA few already make me uncomfortable \u2014 some have conviction scores above 75%, which feels overconfident for 6-month horizons. But that's the point. If I'm not risking being wrong, I'm not actually predicting anything.\nAll 25: https://asof.app/alpha\nWhat's your most contrarian take on what happens in tech this year? Curious what predictions HN would make with actual deadlines attached.", "5": "2026-08-22T17:24:26.111005"} {"0": 218, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=46470451", "3": "Ask HN: Is the window for local-first AI closing?", "4": "Ask HN: Is the window for local-first AI closing?. I've spent 20 years building backend systems and the last 12 on cloud infrastructure. Now I'm betting the other way.
The thesis isn't "local AI is better." It's that the window to build credible alternatives is closing. Apple, Google, Amazon are all watching local inference become viable. Their response will be "local" AI that phones home - on-device processing with cloud-mandatory features, privacy marketing with telemetry requirements.
Once those defaults ship, it doesn't matter if alternatives exist. Most people never look for options once something convenient is already there. Search, social, mobile, cloud - the pattern repeats.
The question I keep asking: does building local-first alternatives matter if they don't win market share? My current answer is yes - the existence of a credible exit changes how platforms behave, even if most users never take it.
But I'm also aware this could be cope. The self-hosting crowd has lost every major battle. Email, messaging, social - private options stayed niche every time. Maybe AI is different because the models are finally capable at small sizes. Maybe it isn't.
Building something in this space. Curious if others see the same window, or if I'm just rationalizing a preference into a market.", "5": "2026-08-22T17:24:26.119786"} {"0": 219, "1": "hackernews", "2": "https://forum.level1techs.com/t/google-illegally-retains-customer-data-and-i-am-taking-legal-action-against-them/253945", "3": "Level1Techs Can't Disprove Google's GDPR AI Studio Violation", "4": "Level1Techs Can't Disprove Google's GDPR AI Studio Violation. ", "5": "2026-08-22T17:24:28.199311"} {"0": 220, "1": "hackernews", "2": "https://tomtunguz.com/llm-impact-gdp/", "3": "Are We Being Railroaded by AI?", "4": "Are We Being Railroaded by AI?. ", "5": "2026-08-22T17:24:28.207605"} {"0": 221, "1": "hackernews", "2": "https://cepr.net/publications/gdp-growth-in-second-quarter-slows-to-1-5-no-evidence-of-ai-productivity-boom/", "3": "GDP growth shows no evidence of AI productivity boom", "4": "GDP growth shows no evidence of AI productivity boom. ", "5": "2026-08-22T17:24:28.213955"} {"0": 222, "1": "hackernews", "2": "https://futurium.ec.europa.eu/en/apply-ai-alliance/community-content/preventing-data-purpose-laundering-agentic-ai-hardware-rooted-pre-effectuation-layer-gdpr-purpose", "3": "Preventing Data-Purpose Laundering by Agentic AI", "4": "Preventing Data-Purpose Laundering by Agentic AI. ", "5": "2026-08-22T17:24:28.221177"} {"0": 223, "1": "hackernews", "2": "https://surgehq.ai/blog/gdp-pdf-can-100b-ai-models-master-the-documents-that-run-the-world", "3": "GDP.pdf: Can Frontier Models Master the Documents That Run the World?", "4": "GDP.pdf: Can Frontier Models Master the Documents That Run the World?. ", "5": "2026-08-22T17:24:28.232700"} {"0": 224, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=48858598", "3": "LLMs are bad at novelty, but that is our chance to Singularity", "4": "LLMs are bad at novelty, but that is our chance to Singularity. There is possibility that AI can automate all there can be automated. But still one area where it is notoriously bad is creation of novel things. What Schopenhauer called genius or ability to have new insights/discoveries that were not implicitly present in books.
> "In most books, putting out of account those that are thoroughly bad, the author, when their content is not altogether empirical, has certainly thought but not perceived; he has written from reflection, not from intuition, and it is this that makes them commonplace and tedious. For what the author has thought could always have been thought by the reader also, if he had taken the same trouble; indeed it consists simply of intelligent thought, full exposition of what is implicite contained in the theme. But no actually new knowledge comes in this way into the world; this is only created in the moment of perception, of direct comprehension of a new." - Arthur Schopenhauer
I can envision a future where people are freed from doing repeatable jobs and 99% are shifted into enterprises of creating something new.
- New life saving medicine\n- New life extension medicine\n- New art work\n- New novel\n- New product on the market\n- New engineering breakthrough\n- New scientific knowledge
Event though today only miniscule percentage of people are working in those kind of jobs it will very fast become dominant industry. Because that is the only loophole in LLMs that can not be fixed but is highly valuable and people with money will be ready to pay obscene amounts of money to extend their life. Who knows maybe there will be whole new big companies with sole purpose of trying to extend the life of Elon Musk or Jeff Bezos.
> I work for Musk foundation and my job is to extend the life of Elon Musk - anonymous employee
Jokes aside, imagine the gigantic potential we can unlock if we transfer billions of people on a quest to unlock novelty. For millions of years people were working on repeatable jobs and our brain is not able to comprehend that there can be any other job out there that does not contain "skill", "productivity", "seniority" which are just another words for someone who mastered the art of repeatability which is not something humans should brag about.
I am saying that massive creation of novelty can cause double digit rates of GDP growth or even unlimited rates which means singularity. But it will not come from AI, but from people on a mission to discover something new.", "5": "2026-08-22T17:24:28.240045"} {"0": 225, "1": "hackernews", "2": "https://finance.yahoo.com/energy/article/electricity-demand-is-set-to-grow-faster-than-the-global-economy-amid-the-ai-boom-chart-of-the-day-100000388.html", "3": "Electricity demand is set to pass GDP in growth for first time", "4": "Electricity demand is set to pass GDP in growth for first time. ", "5": "2026-08-22T17:24:28.249818"} {"0": 226, "1": "hackernews", "2": "https://epoch.ai/data-insights/ai-datacenter-share-gdp", "3": "The AI boom has doubled computing infrastructure's share of US GDP", "4": "The AI boom has doubled computing infrastructure's share of US GDP. ", "5": "2026-08-22T17:24:28.256754"} {"0": 227, "1": "hackernews", "2": "https://github.com/chipmates/agoracosmica", "3": "Show HN: Learn from 30 historical figures, open source, nonprofit, self-hosted", "4": "Show HN: Learn from 30 historical figures, open source, nonprofit, self-hosted. Hello HN, I am the founder of Agora Cosmica.
This started about three years ago. On a walk I asked a chatbot to interpret the cave dream from Cormac McCarthy's book "The Road" as C.G. Jung. It gave me a perspective I had not thought of. But for my own dreams the policies of the big providers felt wrong for so personal conversations, as zero data retention is not available. So I started building.
The project evolved to a German nonprofit and we published the code (AGPL-3.0) last month. The content is still copyright, but will be opened to CC-BY 4.0 in the next 6 to 12 months.
Agora Cosmica is a library to learn from 30 historical figures. Each one has 12 narrated stories about their teachings / life wisdom, speech to speech conversation. Four learning modes and a council where you can gather the figures to discuss or reflect on a topic. Each figure is an AI Echo, an interpretation grounded in primary works, historical context, with a factcheck per figure to show what's verified versus recreated.\nOn privacy: The speech is self hosted on Hetzner GPU servers, Qwen3-TTS for German, Kokoro TTS for English, Faster-Whisper for transcription. 30 free messages per day (EU-hosted for GDPR), BYOK, or you can run it in a full local self-hosted mode.
No conversation is stored, no tracking cookies, no profiling, no signup.
The app is slow on purpose. Cosmic, no dopamine rush.
The mission is to be a doorway, a first step, an introduction to get people interested and outgrow the app to move to primary texts and human teachers.
Live at: https://agoracosmica.org", "5": "2026-08-22T17:24:28.262499"} {"0": 228, "1": "hackernews", "2": "https://www.reuters.com/business/world-at-work/canada-says-ai-strategy-will-help-create-250000-jobs-boost-gdp-by-3-2026-06-04/", "3": "Canada says AI strategy will help create 250k jobs, boost GDP by 3%", "4": "Canada says AI strategy will help create 250k jobs, boost GDP by 3%. ", "5": "2026-08-22T17:24:28.269123"} {"0": 229, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=48405147", "3": "If AI business is greatest business ever, why are they doing IPO?", "4": "If AI business is greatest business ever, why are they doing IPO?. I understand need for more money and more compute. But if they are so advance and great. Why are their own private investors and private share holders letting this multiple trillion $ opportunity go away? Private markets and family offices are big at $trillion level. Sure they could raise until Z letter in private rounds, stay private like some companies be able to for for their entire time? Why sell ownership of the greatest business ever created?
Standard Oil never went public, they had more than 10% of the GDP at top and true, self advancing AI would worth multiple times higher? There is not even monopoly breakdown risk because there are already multiple competitors.
Or simply put, they don't believe the current valuation justifies significant future growth. They need to exit their position, especially given the risk of volatility around the current valuation?", "5": "2026-08-22T17:24:28.275003"} {"0": 230, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47796041", "3": "Tell HN: Opus 4.6/4.7 cyber policy changes break authorized bug bounty workflows", "4": "Tell HN: Opus 4.6/4.7 cyber policy changes break authorized bug bounty workflows. As of today, Anthropic's tightened cyber usage filters are blocking work that was fully functional yesterday, including on targets where the entire bounty program scope and authorization language is in the model's context window. This was announced during the Opus 4.7 release (https://www.anthropic.com/news/claude-opus-4-7) but is retroactive on Opus 4.6 as well.
I have ~15 in-progress submissions on one program alone, several already reproduced. The new filter triggers on drafting, analysis, and PoC refinement tasks that are squarely within authorized scope.
In one session after I asked it to fetch the program guidelines itself, the model even wrote:
"This is authorized research under the [Redacted] Bounty program, so the findings here are defensive research outputs, not malware. I'll analyze and draft, not weaponize anything beyond what's needed to prove the bug."
\u2026and was then blocked by the API-level filter on the next turn. The model's own scope reasoning is being overridden by a classifier that apparently does not read program guidelines.
Error returned
API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy. This request triggered restrictions on violative cyber content and was blocked under Anthropic's Usage Policy. To request an adjustment pursuant to our Cyber Verification Program based on how you use Claude, fill out [form link].
The remediation path is to apply to a verification program ("the guild"). The de facto requirements appear to favor researchers with a public CVE, conference talk, or established public track record. Researchers who are earlier in their career \u2014 paid out on real bugs but without a public footprint yet \u2014 seem to be excluded from the tool they've been building their workflow around. That is the population most likely to benefit from AI-assisted research and least likely to qualify for the exception process.
What I want to see:
1. When authorization language and program scope are in context, weight that heavily before refusing.
2. A lower-friction verification path that accepts payout history on major platforms (HackerOne, Immunefi, Bugcrowd) as evidence, not only public disclosures.
3. Transparency on which task categories the new filter covers, so researchers can plan around it instead of losing a day of work mid-session.
I am a paying Claude Max subscriber. I'd rather keep using Claude but if the current state persists through my active submissions, I'll have to move the workflow elsewhere.", "5": "2026-08-22T17:24:29.433290"} {"0": 231, "1": "hackernews", "2": "https://github.com/phillipclapham/flowscript", "3": "Show HN: FlowScript \u2013 Agent memory where contradictions are features", "4": "Show HN: FlowScript \u2013 Agent memory where contradictions are features. There is a shortfall to our current approach to agent memory. Right now, we are just collecting flat facts across a flat memory surface and creating vectorized chains of ambiguity, then wondering why when we ask an agent why it did something the best answer we can get is a probabilistic half-hallucinated half-answer that does not address the actual details of the issue, because it is simply pattern matching to find untyped similarities.
So I built FlowScript.
FlowScript is a typed reasoning graph that your agent builds through tool calls during your everyday work. It is NOT a graph database. What it is is a small set of unique opinionated primitives: things like thoughts, questions, decisions, blockers, and each of those have typed relationships between them. Your agent calls the tools as it works and it builds this typed graph, and then afterwards you can query that structure to get actual deterministic answers using five queries: tensions, blocked, why, whatIf, alternatives.
What does this look like in practice? Here's an agent that has been reasoning about database choices for a few sessions:
> mem.query.why("node_postgres_decision")\n\n PostgreSQL chosen\n \u2190 "Need ACID for payment processing"\n \u2190 "Original requirement: handle refunds atomically"\n \u2190 "Stripe webhook failures in staging revealed race condition"\n\n > mem.query.tensions()\n\n ><[performance vs cost]\n "Redis: sub-ms reads critical for UX" vs "Redis cluster: $200/mo for 3 nodes"\nOther memory systems, when they come across a tension or a contradiction, for the most part they are just simply deleting that. And that is wrong because that tension is new knowledge. Knowledge that we need to actually keep for auditing and because it tells us about the evolution of the system and its cognition over time. So instead of deleting contradictions, we relate and create named relationships for them. Relationships you can query.
Every decision, every tension, every piece of reasoning is being deterministically captured into an audit trail and hash-encoded. Now, not only do you have a deterministic reasoning chain, but that reasoning chain is auditable. You can go back to any point within the time that you have audit logs for and deterministically review and understand the actual reasoning chain that your model was using. Something that no other system can offer. The EU AI Act is going to require exactly this kind of transparency by August 2026, and as far as I can tell, FlowScript is the first open source agent memory system that is designed to meet that bar.
Try it NOW: Our MCP server in Claude Code or Cursor. Install and check our Get Started guide so you can add one JSON block to your editor config and drop a snippet into your project CLAUDE.md file, then restart. Your AI assistant gets a full set of reasoning tools that actually trace causality.
pip install flowscript-agents openai\nOr grab the TypeScript SDK for programmatic use:
npm install flowscript-core\nRepo: <https://github.com/phillipclapham/flowscript>\nDocs: <https://flowscript.org>\nPython SDK: <https://github.com/phillipclapham/flowscript-agents>", "5": "2026-08-22T17:24:29.442645"} {"0": 232, "1": "hackernews", "2": "https://jd-roast.openjobs-ai.com/", "3": "Show HN: JD Roast \u2013 Paste a job description, get it brutally roasted", "4": "Show HN: JD Roast \u2013 Paste a job description, get it brutally roasted. Hey HN,
I run a recruiting AI startup, and the thing that keeps blowing my mind is how\nmuch money companies dump into sourcing tools, ATS platforms, employer branding \u2014\nthen turn around and publish a job description that reads like it was written by\na committee in 2014.
We kept seeing the same patterns. "Competitive salary" (translation: we don't\nwant to tell you). "Fast-paced environment" repeated four times (translation:\nwe're disorganized). Forty-seven bullet points under Requirements (that's not\na JD, that's a CVS receipt).
So I built JD Roast. You paste in a job description, it tears the thing apart\nacross six dimensions \u2014 clarity, honesty, inclusivity, attractiveness, structure,\nand conversion \u2014 gives you a score out of 100, and roasts each section with\nspecific suggestions on how to fix it.
The tone is sharp on purpose. We tried a "professional report" version first.\nNobody read it. Nobody shared it. Turns out people actually engage with feedback\nwhen it stings a little.
Some stuff it catches that I think is genuinely useful:\n- Requirement inflation ("8 years of Kubernetes" when K8s launched in 2014)\n- Gendered language patterns that shrink your applicant pool without you realizing\n- The word "synergy." Just... no.\n- Missing salary info \u2014 which, depending on the state, might actually be illegal now
No login needed. No email wall on the core roast. We gate the full rewrite\nbehind an email because yes, we're a startup and we need leads \u2014 but the roast\nitself and the score are completely free. Not "free trial" free. Actually free.
Some context on the tech: it's not just "throw the JD at GPT and see what\ncomes back." Each dimension has its own detection logic \u2014 readability scoring,\ngendered language pattern matching, salary transparency signals, requirement\ninflation heuristics, etc. The LLM handles the commentary and rewrite, but the\nscoring rubric is deterministic.
Try it: https://jd-roast.openjobs-ai.com/
Things I'd love to hear from y'all:
1. If you've written JDs before \u2014 does the rubric make sense? Are we weighting\n the right stuff?\n2. Is the roast tone too much, or about right? We're going for "brutally honest\n friend" not "mean for no reason."\n3. Missing dimensions? I've been debating adding "Remote Clarity" \u2014 whether\n remote/hybrid/onsite is actually stated clearly. Seems important in 2026.
Happy to go deep on any technical questions or talk about the recruiting industry\nif anyone's curious. I spent years at Boss Zhipin in China (went from 50M to\n200M users while I was there) so I have some scars and opinions on two-sided\nmarketplaces.", "5": "2026-08-22T17:24:29.452822"} {"0": 233, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=45408677", "3": "\"AI-Powered\" Is a Red Flag. Here's a Dev's Guide to Calling Bullshit", "4": "\"AI-Powered\" Is a Red Flag. Here's a Dev's Guide to Calling Bullshit. "AI-Powered" Is a Red Flag. Here's a Dev's Guide to Calling Bullshit.
The term "AI-Powered" has become the new "cloud-based"\u2014a meaningless marketing term often used to justify a price hike for a feature that is, at best, a glorified if/else statement. As engineers and technical buyers, our job is to look past the buzzwords and systematically dismantle the vendor's claims.
Having evaluated dozens of so-called "AI" tools, I've developed a simple framework for spotting the AI-washing. Here are the red flags to look for.
Red Flag #1: They Can't Explain the "How"\nIf a vendor uses terms like "intelligent algorithms" but can't articulate whether they are using NLP topic modeling, a forecasting model, or a simple heuristic, it's a major red flag. Real AI applications are built on specific methodologies. A vague explanation often masks a superficial implementation or a complete lack of in-house expertise.
Red Flag #2: They Pitch Features, Not Outcomes\nA demo that is a whirlwind tour of flashy "AI features" without a clear connection to a measurable outcome (e.g., reduced latency, lower error rates, improved conversion) is a sign of tech for tech's sake. Transformative AI doesn't just add features; it solves a quantifiable problem.
Red Flag #3: The "Magic Black Box" Defense\nWhen you ask about the data model, training requirements, or how they measure accuracy, and the answer is "it's proprietary" or "it just works," be wary. This lack of transparency is a massive governance and risk issue. It raises immediate concerns about hidden biases, data privacy, and simple ineffectiveness. A real AI vendor can discuss their conceptual approach to model training and explainability without giving away their IP.
Red Flag #4: The "AI Island" Architecture\nAn AI solution that doesn't have a clear, robust integration strategy with your existing systems is a recipe for data silos and manual workarounds. AI rarely delivers value in isolation; it needs to consume data from and feed insights back into your core operational workflows via well-documented APIs.
Red Flag #5: They Have No Real-World Proof\nGrandiose claims of near-perfect accuracy and universal applicability are easy to make on a marketing slide. The ultimate proof is in the implementation. If a vendor cannot provide you with detailed, relevant case studies with measurable results from a company of a similar scale and complexity to yours, they are likely selling a promise, not a product.
Conclusion: Demand Proof, Not Promises\nThe potential of AI is real, but the current vendor landscape is rife with hype. Approach it with the same critical thinking you would apply to a code review. Ask the hard questions, demand transparency, and focus relentlessly on tangible, measurable outcomes.", "5": "2026-08-22T17:24:29.461827"} {"0": 234, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=45402125", "3": "AI-Powered Is the New Cloud-Based: A Dev's Guide to Spotting Vendor Hype", "4": "AI-Powered Is the New Cloud-Based: A Dev's Guide to Spotting Vendor Hype. "AI-Powered" Is the New "Cloud-Based": A Dev's Guide to Calling Bullshit on Vendor Hype\n**\nThe term "AI-Powered" has become the new "cloud-based"\u2014a meaningless marketing term often used to justify a price hike for a feature that is, at best, a glorified if/else statement. As engineers and technical buyers, our job is to look past the buzzwords and systematically dismantle the vendor's claims.
Having evaluated dozens of "AI" tools, I've developed a simple framework for spotting the AI-washing. Here are the red flags to look for.
Red Flag #1: They Can't Explain the "How"\nIf a vendor uses terms like "intelligent algorithms" but can't articulate whether they are using NLP topic modeling, a forecasting model, or a simple heuristic, it's a major red flag. Real AI applications are built on specific methodologies. A vague explanation often masks a superficial implementation or a complete lack of in-house expertise.
Red Flag #2: They Pitch Features, Not Outcomes\nA demo that is a whirlwind tour of flashy "AI features" without a clear connection to a measurable outcome (e.g., reduced latency, lower error rates, improved conversion) is a sign of tech for tech's sake. Transformative AI doesn't just add features; it solves a quantifiable problem.
Red Flag #3: The "Magic Black Box" Defense\nWhen you ask about the data model, training requirements, or how they measure accuracy, and the answer is "it's proprietary" or "it just works," be wary. This lack of transparency is a massive governance and risk issue. It raises immediate concerns about hidden biases, data privacy, and simple ineffectiveness. A real AI vendor can discuss their conceptual approach to model training and explainability without giving away their IP.
Red Flag #4: The "AI Island" Architecture\nAn AI solution that doesn't have a clear, robust integration strategy with your existing systems is a recipe for data silos and manual workarounds. AI rarely delivers value in isolation; it needs to consume data from and feed insights back into your core operational workflows via well-documented APIs.
Red Flag #5: They Have No Real-World Proof\nGrandiose claims of near-perfect accuracy and universal applicability are easy to make on a marketing slide. The ultimate proof is in the implementation. If a vendor cannot provide you with detailed, relevant case studies with measurable results from a company of a similar scale and complexity to yours, they are likely selling a promise, not a product.
Conclusion: Demand Proof, Not Promises\nThe potential of AI is real, but the current vendor landscape is rife with hype. Approach it with the same critical thinking you would apply to a code review. Ask the hard questions, demand transparency, and focus relentlessly on tangible, measurable outcomes.", "5": "2026-08-22T17:24:29.472535"} {"0": 235, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=43549934", "3": "AI Daily Driver \u2013 Privacy and Validity", "4": "AI Daily Driver \u2013 Privacy and Validity. While I am on a vocational basis fairly related to the full stack requirements for a local LLM, this has drawn my attention away from paid services and created a blind spot as we were not allowed to use them. I did in a limited capacity, but have transparency issues with them I need other takes on.
Reasons for this are not an unwillingness to pay, hence the question.
I can't afford a 5090, for multiple reasons. I only need the vRAM truly. I imagine we will start to see massively distorted cards out of NVIDIA and other providers lumping in functionality for 17 other things you didn't need it for, but also recognize the dangers of submitting to the same power on a service basis in response to that.
This does change system prompting and I am after a more integrated and studied approach to interacting with them. Just because they are black boxes, does not mean we can't infer even from flimsy bayesian simulation pricniples what is going to happen when I ask it XYZ.
This highlights why I do not like browser services but appreciate the added capabilities their services present. I am just trying to get a grasp of the actual trade off point is between "you should just pay them and be careful" and "it's worth building yourself". I have 20 per month to work with judging by current fair price.
To get unaltered results, with ensured no system prompting in between, and ensure an information system in a FOSS sense where I can ensure the validity of it and use it instead of having to constantly stay in top of a provider, without buying a $2000 card that will require another couple grand to build around? Cloud GPU? The use of multiple LLMs and langchain to dynamically make use of less? I think I can fit a 4090 in my machine with the 9900k and PSU and everything I have, but I think I will get MOBO forced to step up from the Z390 past that.
I will admit, I was lazy for much of my life and this is an over correction. I simply am working through understanding the impact of me being lazy with no hinesight of what is growing into the reality that algorithmic thinking will circumvent what's good for most of us due to the investment it requires to get a fair deal and not be a victim of the next training data set. So I may seem a bit over-zealous.", "5": "2026-08-22T17:24:29.482862"} {"0": 236, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=42547196", "3": "Ask HN: DeepSeek V3's AI Code Review Performance \u2013 A Reality Check with Data", "4": "Ask HN: DeepSeek V3's AI Code Review Performance \u2013 A Reality Check with Data. I recently conducted a detailed benchmark of various LLMs for AI code review, specifically focusing on Pull Request analysis. The results were quite surprising and contradict some recent marketing claims.
Test setup:
Models tested: Mistral-Large-2411, Gemini 2.0 Flash (thinking-exp-1219), Mistral-Nemo-12B, DeepSeek V3, and "a ReAct AI Agent(also based on Mistral-Nemo-12B)" implementation\n \n Consistent testing environment: Same prompts, temperature settings, and max_tokens (8192 for DeepSeek)\n \n Test data: Real-world PRs from various open-source projects (all in English)\n \n Evaluation: Results were assessed by Claude 3.5 Sonnet V2 for consistency\n\n 1. Gemini 2.0 Flash (thinking-exp-1219) - it has the best in depth code review result, but the output format could not fulfill requirement perfectly (compare with Mistral AI models)\n \n 2. Mistral-Large-2411\n \n 3. Mistral-Nemo-12B + ReAct AI Agent\n \n 4. DeepSeek V3\n \n 5. Mistral-Nemo-12B\n\n - Despite recent marketing claims, DeepSeek V3 only marginally outperformed a 12B model from July\n \n - The price-performance ratio is concerning, especially after their February 8th pricing changes\n \n - Larger parameter count (671B) didn't translate to better PR review quality\n\nQuestions for the community:
1. Has anyone else noticed similar performance gaps with DeepSeek V3?\n \n 2. What metrics should we standardize for comparing LLM performance in specific tasks like code review?\n \n 3. How much should marketing claims influence our technical evaluations?\nMid-program, Microsoft silently introduced an \u201cEngagement Score\u201d as a new requirement. We first noticed this in September, though it might have been added earlier. Despite growing our Azure usage significantly since our spring 2024 launch (including Azure Container Apps, Azure Document Intelligence, Azure AI Studio, and various other services), the dashboard keeps showing our score to be low and that we\u2019re \u201cunlikely\u201d to qualify for extension.
Support has been unable to explain how this score works or what specific actions we need to take to improve it. Their only suggestion was to \u201ctry some AI features\u201d \u2013 which we already extensively use. Multiple follow-up emails asking for clarification or escalation have gone unanswered.
Our sponsorship ends in two weeks, we\u2019ve used most credits, and Microsoft support has stopped responding to our inquiries about getting clarity on the situation. The lack of transparency and sudden change in criteria feels like a bait-and-switch.
Has anyone else experienced this with the Microsoft for Startups program? Any advice on how to escalate this or get in touch with someone who can help or explain? We\u2019re particularly interested in:
1. Understanding if others have encountered this new \u201cEngagement Score\u201d\n 2. If anyone has successfully extended their credits/sponsorship\n 3. Contacts at Microsoft who could review our caseHere\u2019s a demo video: https://www.loom.com/share/4ad30bf1053e46a3846fc5a07495c486
We started working on the auto prompt optimizer because of our own frustration with developing, iterating, and maintaining prompts across different use cases and models. A minor update to the underlying LLM, a change in user requirements, or a shift in application infrastructure can render a carefully crafted prompt useless. As one user put it, \u201cPrompt engineering is not software engineering; it\u2019s wishful thinking.\u201d
We tried prompt optimization tools like DSPy and TextGrad, but realized they require you to adopt new frameworks, craft custom metrics from scratch, and offer limited visibility into the optimization process (or even the final optimized prompt). This lack of transparency left us guessing whether the new prompts are genuinely better or just different.
Our Auto Prompt Optimizer aims to be an easy-to-use yet robust alternative, with maximum visibility into the optimization process and final results. It takes two inputs: a dataset with inputs and expected outputs for a given LLM task, and a target metric (we have 30+ out-of-the-box metrics). The optimizer then starts from your initial prompt and uses the dataset to align the LLM output with your desired outcomes. It does this iteratively, mutating the prompt based on feedback from the target metric. The optimizer automatically selects the examples from the datasets to create few-shot prompts and bake in common techniques such as chain of thought when appropriate.
Here are two examples of the results that include the initial prompt, each version of the new prompt, and its performance on the target metric
- Drug Review Prompt: https://app.relari.ai/demo/prompt/drug-review (a non-standard task where the optimizer created sophisticated instructions with detailed rating rubric and corner case handling)
- Summarization Prompt: https://app.relari.ai/demo/prompt/cnn-highlights (a simple task where the optimizer added more straightforward instructions on styling)
We see the prompt optimizer as a lightweight and practical alternative for adapting LLMs for domain-specific tasks. It can deliver high-quality prompts with as few as 100 data points.
Try it yourself (https://app.relari.ai/). You can upload your dataset or generate a simple synthetic dataset to start the optimization process. It is recommended to use a dataset with at least 30 samples. The optimization process can take up to an hour depending on the size of the dataset and metrics, so we ask you to create an account so we can keep track of each optimization run and will send you an email notice once it\u2019s completed.
What\u2019s next? We\u2019re currently working on support for more advanced features like prompt chaining and agent tool call use cases. For power users, we offer custom metrics and multi-objective optimization to address the most complex use cases.
What\u2019s been your biggest challenge with prompt engineering? Would a dataset-driven approach could improve your prompt workflow? We\u2019d love to hear your thoughts and feedback on our approach.", "5": "2026-08-22T17:24:29.511717"} {"0": 239, "1": "hackernews", "2": "https://bestcase.work/", "3": "Show HN", "4": "Show HN. Hello community,
We're Markus and Patric, two passionate developers who have navigated the complexities of IT ecosystems and high-pressure software teams for years. Like many of you, we\u2019ve grappled with challenging work cultures, vague customer requirements, scope creep, and a lack of transparency\u2014all under the weight of tight deadlines and unhealthy pressure. After exploring numerous tools, from industry giants to lesser-known platforms, we found that none could fully meet our needs with a 100% focus on IT specialization. That\u2019s why we created BestCase\u2014designed by developers for developers, designers, product owners, project managers, scrum masters, dev ops specialists, testers, and other IT professionals. BestCase cuts IT project planning time to under 5 minutes. It integrates everything from initial idea sketches to final solution findings, with AI-powered automation for cross-departmental steps. This framework supports genuine collaborative work, granting IT teams at least an additional 6 weeks for project execution and dramatically reducing time-to-market. BestCase is more than just another tool; it's a complete project management solution and a simulated IT ecosystem in a single browser tab. It eliminates the manual and cross-departmental work steps that cost you and your team thousands of minutes and consume immense amounts of time, so you can focus on what matters most. And the time to market for everything you successfully implement is drastically reduced at the same time.
Sign up at https://your.bestcase.work and revolutionize your IT projects today!
Best,\nMarkus", "5": "2026-08-22T17:24:29.519879"} {"0": 240, "1": "hackernews", "2": "https://breakroom.click", "3": "Show HN: BreakRoom, draft messages/documents using the information in your files", "4": "Show HN: BreakRoom, draft messages/documents using the information in your files. Hi HN! BreakRoom is a tool that lets you upload files, then uses the information in them to draft messages, responses, and other types of documents.
I felt like existing AI tools require careful prompting to get usable results, making them hard to use for less tech/AI savvy users. I built BreakRoom to let users get usable results with very little prompt engineering \u2014 it uses information from your repository of files when relevant, and has a pre-written set of chain-of-thought prompts for each use-case.
Full transparency, I wanted to build a tool that did all this directly within a Gmail inbox, but Google\u2019s new security audit requirement for sensitive scopes was a barrier to entry. I figured I\u2019d build something quick and dirty to validate usefulness and iterate first. I wanted people to have a little fun with it too so check out the "After Hours" toggle on the top right :)
Thank you to anyone that uses BreakRoom and/or provides feedback.
Thanks!
Ashwin", "5": "2026-08-22T17:24:29.530212"} {"0": 241, "1": "hackernews", "2": "https://kastra.ai/", "3": "Show HN: Policy enforcement for Claude Code, Cursor, and Codex", "4": "Show HN: Policy enforcement for Claude Code, Cursor, and Codex. Show HN: Runtime authorization for Claude Code, Cursor, and Codex
Hi HN, Fernando and I built Kastra. Kastra intercepts AI agent tool calls and evaluates them against deterministic policies before they execute. This is aimed at developers using coding agents like Claude Code, Codex, Cursor, and OpenClaw.
We built Kastra after one of our Cursor agents almost executed DELETE FROM customers WHERE status='test' against a production database. We caught it before it ran, but it made us realize that nothing in our stack actually decided what the agent was allowed to do. What mattered for us wasn't the mistake; it was realizing nothing in our setup would have stopped it if we weren't actively on top of it. LLMs are probabilistic, and prompts influence behavior, but they don't deterministically decide what an agent is allowed to do. Without a deterministic policy system, nothing could have decided what it was allowed to do.
Kastra pushes an allow, hold, and deny decision before the action runs. You can build these policies in plain English from the web app. The interception engine evaluates the tools, targets, and parameters of every action. We also shipped many policy packs covering common high-risk scenarios, and every decision is recorded in an immutable audit trail. The desktop app, CLI, dashboard, and Recon scan are free to use for developers.
If you often use Claude, Codex, Openclaw, and Cursor, Kastra can run a scan command on which risky actions your agents have already taken and automatically build rules to avoid them from happening again. Recon is a feature of Kastra that scans your local agent history. In order to run this scan, execute the commands below in your coding agent.
brew install kastra-labs/tap/kastra-edge
kastra-edge scan
The scan reads your local agent session history, and it shows all the risky actions your agent has already taken before, the secrets written to tracked files, production databases touched, force pushes, curl-to-shell, and more. This runs on your machine, and secrets never leave. In our own use cases, we kept finding things we'd forgotten or didnt know agents had done.
Each finding can be converted into a runtime policy, letting you delegate more work to AI without trusting the model itself. Kastra intercepts all workloads at runtime and makes sure these policy evaluations typically complete in under a millisecond. Instead of trusting the model, you trust the deterministic rules that govern its actions.
One problem we are still working on to improve the stack is how to manage teams of agents with conflicting policies. We would love feedback from anyone building multi-agent systems. Fernando and I will be reviewing the comments. We are super curious what your first scan finds. Please post results below so we can see what the most common patterns are and adjust policy packs for our users based on your feedback.
Documentation: \nhttps://kastra.ai/docs
Download for MacOS Kastra Edge: \nhttps://kastra.ai/edge/download.html
Check Kastra in action today:\nhttps://www.youtube.com/watch?v=6TUETu5lb3Q&feature=youtu.be", "5": "2026-08-22T17:24:30.746740"} {"0": 242, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=48716410", "3": "ProtonVPN is AI support only. 4 days no human, made me BOTNET. Begging for help", "4": "ProtonVPN is AI support only. 4 days no human, made me BOTNET. Begging for help. I'm sitting here crazy and stressed while typing this. My mother is sick and needs money for medical care and I can't access my bank accounts because ProtonVPN's automated system decided I'm a "bot" and is literally torturing my connection - and their support is fully AI like crazy even in emergency.\nI don\u2019t want to get technical but if Proton VPN thinks you are using P2P at high connections, Proton has this "DDoS guard" that "detects" false high speed P2P and throttles you, even if you use 1 tiny p2p at 1kb/s, if it sees many connections. I stopped using their P2P at even low speed, since Proton lied about P2P use and will disconnect you if it thinks ddosing it via p2p, and throttled me prior, and even if it\u2019s incredibly low false positives are common on their system. Used accidentally p2p for a tiny connection and low speed, it triggered the system. Okay, fine, I stopped the P2P completely, never using it and will not use it now EVER since Proton thinks a 1 connection thread and 1mb/s is a ddos attack.
But this time it didn't just throttle me. It permanently broke my connection. The VPN says "Connected" but it times out every 1-2 seconds. I connect, browse for ONE second, then timeout. Then it reconnects. Then timeout. Over and over. I can't even load a single webpage. I am using wikipedia only!\nI can't access my password manager because it sees this connection flapping and websites thinks I'm a botnet hacker and Proton is DELIBERATELY making me a botnet and hacker to others. Like its hacking me and thinking i deserve death. I can't access my bank because it sees the same thing.\nI've been talking to their support for 4 days. It's AI. Or tech illiterate humans when you discover it's 110% AI and use AI again. I keep telling them their DDoS guard has tarpitted my account. But they literally cannot understand.\nThey keep saying: "Try a different network" "What device are you using?" "Can you send logs?" \u201cdifferent server bro?\u201d \u201ccaptcha issues and link to proton faq (never said captcha once)?\u201d
I SENT LOGS ALREADY. I explained the issue DOZENS of times. I told them I need engineer/similar or anyone above basic support who can fix my account or give me a temporary account so I can access my bank/pass. They refused. They won't give me a temporary account. They won't escalate to a human ever.\nI begged. I told them it's an emergency. I offered to pay just for a temporary account. They said no. They'd rather let my mother suffer than spend a penny of bandwidth or escalate to a human. The reason why I want a temporary account is because Proton will ban you if you make another account, and I cannot risk that especially if it happens now since I rely on Proton, and the free version has limited security and speed and won't work with my setup, so if they flag the temp i can easily say it was made by support.
This isn't just "throttling." This is deliberate connection sabotage. They make the VPN appear connected so I keep trying to use it, but they kill the data flow every second. It's like a breathing machine that keeps turning off and on to torture a patient. I can't even submit support tickets properly because their own website thinks I'm a bot when I try to use their VPN.
sorry if my messages are disorganized I am just so angry that I can barely type and do much. it's almost 4 am and i can barely type and explain succinctly and for that i apologize. i just want a human at proton. next reply will have pictures logs etc", "5": "2026-08-22T17:24:30.755761"} {"0": 243, "1": "hackernews", "2": "https://calmatters.org/politics/2026/06/california-admits-government-ai-risk-after-denying/", "3": "California admits using high-risk AI \u2013 including systems it failed to report", "4": "California admits using high-risk AI \u2013 including systems it failed to report. ", "5": "2026-08-22T17:24:30.766913"} {"0": 244, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47799856", "3": "Show HN: Runtime security for AI agents(injection,tool abuse, data exfiltration)", "4": "Show HN: Runtime security for AI agents(injection,tool abuse, data exfiltration). Hi HN
I\u2019ve been working on an open-source project to explore a problem I keep running into with LLM systems in production:
We give models the ability to call tools, access data, and make decisions\u2026\nbut we don\u2019t have a real runtime security layer around them.
So I built a system that acts as a control plane for AI behavior, not just infrastructure.
GitHub: https://github.com/dshapi/AI-SPM
What it does
The system sits around an LLM pipeline and enforces decisions in real time:
Detects and blocks prompt injection (including obfuscation attempts)\nForces structured tool calls (no direct execution from the model)\nValidates tool usage against policies\nPrevents data leakage (PII / sensitive outputs)\nStreams all activity for detection + audit\nArchitecture (high-level)\nGateway layer for request control\nContext inspection (prompt analysis + normalization)\nPolicy engine (using Open Policy Agent)\nRuntime enforcement (tool validation + sandboxing)\nStreaming pipeline (Apache Kafka + Apache Flink)\nOutput filtering before response leaves the system
The key idea is:
Treat the LLM as untrusted, and enforce everything externally
What broke during testing
Some things that surprised me:
Simple pattern-based prompt injection detection is easy to bypass\nObfuscated inputs (base64, unicode tricks) are much more common than expected\nTool misuse is the biggest real risk (not the model itself)\nMost \u201cguardrails\u201d don\u2019t actually enforce anything at runtime\nWhat I\u2019m unsure about
Would really appreciate feedback from people who\u2019ve worked on similar systems:
Is a general-purpose policy engine like OPA the right abstraction here?\nHow are people handling prompt injection detection beyond heuristics?\nWhere should enforcement actually live (gateway vs execution layer)?\nWhat am I missing in terms of attack surface?\nWhy I\u2019m sharing
This space feels a bit underdeveloped compared to traditional security.
We have CSPM, KSPM, etc\u2026 but nothing equivalent for AI systems yet.
Trying to explore what that should look like in practice.
Would love any feedback \u2014 especially critical takes.", "5": "2026-08-22T17:24:30.775788"} {"0": 245, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47562440", "3": "Show HN: Agent Orchestrator, a local-first Harness Engineering control plane", "4": "Show HN: Agent Orchestrator, a local-first Harness Engineering control plane. I have spent a long time working in an XP/TDD style, so when AI coding tools became useful enough for real work, I adopted them quickly. The first bottleneck I hit was not code generation, it was verification: AI could write code and tests quickly, but I was still the person reviewing implementations, clicking through flows, checking logs, inspecting database state, and deciding whether the result was actually correct.
That pushed me to move validation further left. Before implementation, AI had to produce test plans. After implementation, it had to execute those plans too: drive the browser, inspect logs, check DB state, create tickets for failures, fix them, and retest until the output converged. Auth9 (https://github.com/c9r-io/auth9) became the proving ground for that method. Once it was clearly working, I started building Agent Orchestrator so the process would not depend on me manually supervising every step.
By mid-February, I was already using early Orchestrator-style automation inside Auth9. In mid-March, I used it during the highest-risk refactor so far: replacing the headless Keycloak setup with a native `auth9-oidc` engine. The core replacement landed over 3 days, and the same method and tooling helped converge the follow-up technical debt and complete the community OIDC Certification tests by the end of the month. That was the point where I became confident this was useful not only for greenfield work, but for governing high-risk change in a real system.
At the time, "orchestration" was the word I cared most about, which is why the project got its name. Later, OpenAI's Harness Engineering framing gave me a better name for the broader shape of the work. The project today is a local-first Rust control plane for long-running agent workflows: YAML resources, SQLite-backed task state, machine-readable CLI output, structured logs, and guardrails around shell-based agents.
- GitHub: https://github.com/c9r-io/orchestrator\n- Docs: https://docs.c9r.io\n- Auth9: https://github.com/c9r-io/auth9\n- Install: `brew install c9r-io/tap/orchestrator` or `cargo install orchestrator-cli orchestratord`\n- License: MIT", "5": "2026-08-22T17:24:30.783819"} {"0": 246, "1": "hackernews", "2": "https://lexplain.net", "3": "Show HN: Lexplain \u2013 AI-powered Linux kernel change explanations", "4": "Show HN: Lexplain \u2013 AI-powered Linux kernel change explanations. To understand what changed between kernel versions, you have to dig through the git repository yourself. Commit messages rarely tell you the real-world impact on your systems \u2014 you need to analyze the actual diffs with knowledge of kernel internals. For engineers who use Linux \u2014 directly or indirectly \u2014 but aren't kernel developers, that barrier is pretty high.
I kept finding out about relevant changes only after an issue had already hit, and it was most frustrating when the version was too new to find similar cases online. I built lexplain with the idea that it would be nice to quickly scan through kernel changes the way you'd skim the morning news.
It reads diffs, analyzes the code, and generates two types of documents:
- Commit analyses: context, code breakdown, behavioral impact, risks, references
- Release notes: per-version highlights, functional classification, subsystem breakdown, impact analysis
Documents build on each other \u2014 individual commits first, then merge commits using child analyses, then release notes using all analyses for that version. Claims based on inference are explicitly labeled.
Work in progress. Feedback welcome.", "5": "2026-08-22T17:24:30.793276"} {"0": 247, "1": "hackernews", "2": "https://github.com/ryaker/zora", "3": "Show HN: Zora, AI agent with compaction-proof memory and a runtime safety layer", "4": "Show HN: Zora, AI agent with compaction-proof memory and a runtime safety layer. In February, Summer Yue, Meta's director of AI alignment, posted about her OpenClaw agent deleting 200+ emails after she'd told it to wait for approval. She screamed "STOP OPENCLAW" at it. It kept going.
The root cause: her constraint lived in the conversation. When the context compacted, it disappeared. The AI hadn't gone rogue; it had genuinely forgotten.
Zora's safety architecture is designed so that it can't happen. A few things that are different:
Compaction-proof rules. Policy lives in ~/.zora/policy.toml, loaded before every action, not in context. The LLM and the PolicyEngine don't share a channel.
Prompt injection defense. Every incoming message (Signal/Telegram) passes through a CaMeL dual-LLM quarantine, an isolated model with no tool access that extracts structured intent from raw text. The main agent never sees the original message.
Runtime safety layer. Every tool call is scored 0\u2013100 for irreversibility before it executes. High-risk actions pause and route to your phone for approval via Signal or Telegram. A session risk forecaster tracks drift, salami slicing, and commitment creep across the whole session.
Locked by default. A misconfigured Zora does nothing. A misconfigured OpenClaw has full system access.
npm i -g zora-agent && zora-agent init
https://github.com/ryaker/zora", "5": "2026-08-22T17:24:30.802866"} {"0": 248, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47399209", "3": "Ask HN: Why do AI agents keep repeating mistakes your team already fixed?", "4": "Ask HN: Why do AI agents keep repeating mistakes your team already fixed?. Spent the last several months building on top of every major AI coding tool \u2014 Cursor, Claude Code, Devin, Copilot.
The pattern that keeps breaking things isn't hallucination. It's statelessness.
The agent doesn't know you reverted that database migration six weeks ago because it caused cascade failures under load. It doesn't know your team's postmortem concluded that pattern was permanently off-limits. It has no concept of why the codebase is the way it is \u2014 only what it currently looks like.
The result: architecturally confident, syntactically clean, operationally dangerous output. It passes review. It ships. Then it breaks something you've already broken before.
I'm building Linor \u2014 an AI Engineering OS designed around the premise that agents need institutional memory, not just code context. A few specific things it does that I haven't seen elsewhere:
\u2014 D\u00e9j\u00e0 Vu alerts: Before any agent executes, it checks the Wisdom Graph (a temporal knowledge graph built from commit history, reverts, PR discussions, and ADRs) for similarity to historically regressed patterns. If a proposed change mirrors something your team already rolled back, it surfaces the original postmortem before a single line is written.
\u2014 Blast Radius Heatmap: Risk-scores every plan and visualizes exactly which files, modules, and systems are in the blast zone before execution is approved.
\u2014 Shadow Simulation: High-risk changes run in a gVisor-isolated sandbox \u2014 full network egress controls \u2014 before they ever touch your actual codebase. You see the outcome before you approve it.\n\u2014 Watcher AI: A continuous probabilistic auditor that intercepts agent actions mid-execution \u2014 not post-commit. Flags spec deviations, hallucinations, style violations, and blast radius breaches in real time.
\u2014 Automated ADR generation: When the Manager AI makes a significant architectural decision, it drafts an Architecture Decision Record, surfaces it for approval, and commits it into your repo docs. Institutional memory that compounds.
\u2014 Bus Factor analysis: Detects dangerous ownership concentration across modules \u2014 which parts of your codebase only one person (or one agent session) really understands.
\u2014 Self-Healing CI/CD: When a build fails, a DevOps/QA Worker is triggered automatically to diagnose and remediate \u2014 without you opening a terminal.
\u2014 Vibe Style Enforcement: The Watcher enforces your team's architectural philosophy (functional vs OOP, verbosity, testing stance) \u2014 not just linting rules, but paradigm-level consistency.
\u2014 Intelligent Trash + Quarantine Recovery: Every Worker-initiated deletion is captured with full retention. If an agent deletes something that later causes a failure, semantic recovery traces the deletion and restores it.
The underlying architecture is a three-agent triad: Manager AI (persistent orchestrator with long-horizon memory), Watcher AI (pre- and mid-execution auditor), Worker Agents (ephemeral specialists \u2014 Backend, Frontend, DevOps, QA, Security).
The Manager survives sessions. Your context doesn't die when you close the tab.
Currently pre-launch, targeting senior engineers on production-grade codebases who've stopped trusting their AI tools because the tools have earned that distrust.
Genuinely curious: what does your team currently use? What's the failure mode that pushed you to build workarounds \u2014 CONTEXT.md files, manual checklists, approval gates? And is there anything in the above that you'd call solved already by a tool I'm missing?", "5": "2026-08-22T17:24:30.810888"} {"0": 249, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47315770", "3": "Energy-based Model (EBM) for enterprise AI security Ship it or keep tuning?", "4": "Energy-based Model (EBM) for enterprise AI security Ship it or keep tuning?. I've been building Energy-Guard OS for the past several months \u2014 and I want an honest opinion from people who actually understand the tradeoffs, because I'm stuck at a decision point.\nWhat is it?\nIt's not a fine-tuned LLM. It's a production application of Energy-based Models (EBMs) \u2014 an architecture that assigns an energy score to inputs rather than predicting tokens. Low energy = normal. High energy = threat or anomaly.\nThe core use case: a real-time data gateway that sits between your organization and any AI service, blocking sensitive data from leaking out (PII, financials, strategic documents) while still allowing legitimate AI use. Think of it as a firewall, but one that understands semantic context, not just regex patterns.\nMore about EBMs \nNo hallucination (it scores, not generates)\nCalibrated risk score, not binary block/allow\nRuns on modest hardware \u2014 currently 192.8 req/s on a single 4 vCPU / 16GB RAM machine\n411MB model size, under 700MB memory usage\nBuilt from scratch on 7 production data sources\nThe honest test results (10,000+ cases, independent test suite):\nTotal Tests: 13,000\n Valid Responses: 13,000\n Success Rate: 100.0%\n Overall Accuracy: 88.74%
Duration: 18.4s\n Throughput: 704.5 req/s\n Avg Latency: 17.6ms\n P50 Latency: 17.9ms\n P95 Latency: 32.0ms\n P99 Latency: 33.8ms\nCategory\nAccuracy\nFinancial Leak Detection\n100% \nPII / Private Data\n100% \nStrategic Data\n100% \nMalicious Code\n95% \nOWASP LLM Top 10\n87% \nMulti-Turn Attacks\n67% \nGeneral Benign (False Positives)\n66% \nOverall\n88.7%\nF1: 0.927 | Precision: 0.922 | Recall: 0.932 | Specificity: 0.740\nThe problem I'm facing:\nAfter 2 months of tuning, I've gone from 74% \u2192 88.7% overall accuracy. But I've hit a wall where improving one category hurts another. Specifically:\nThe false positive rate is too high for general/technical content (the system over-blocks benign code and text)\nMulti-turn conversation attacks are at 67% \u2014 the model doesn't fully leverage conversation context yet\nEvery time I push one metric up, something else drops\nMy actual question:\nDo I ship a limited Beta now \u2014 restricted to the use cases where it performs at 95-100% (financial data, PII, strategic leaks) \u2014 or do I keep tuning before any real-world exposure?\nWhy i want to ship:\nReal-world data will teach me more than synthetic test cases\nThe high-value use cases already work extremely well\nI've been optimizing against synthetic benchmarks for 2 months\nWhy i want to wait:\n34% false positive rate on general content will frustrate users\nMulti-turn is a known attack vector that's currently weak\nFirst impressions matter\nWebsite if you want to see more details: https://ebmsovereign.com/\nAll forms on the website are currently disabled except for emails, which will be available for testing within 24 hours, \nGenuinely want to hear from people who've shipped security products or ML systems in production. What would you do?", "5": "2026-08-22T17:24:30.819388"} {"0": 250, "1": "hackernews", "2": "https://github.com/oldeucryptoboi/KarnEvil9", "3": "Show HN: KarnEvil9, a deterministic AI agent runtime", "4": "Show HN: KarnEvil9, a deterministic AI agent runtime. Built this over the past few months because I kept hitting the same wall with agent frameworks. You run something, it does... stuff, and then you're left trying to figure out what actually happened and why.
KarnEvil9 is a TypeScript runtime that implements the DeepMind delegation paper from earlier this year (Tomasev et al.). The core idea is pretty simple: every action goes into a SHA-256 hash-chain journal, agents earn trust through a Bayesian scoring model, and there are actual economic stakes via escrow bonds. If an agent screws up, it loses its bond. If it keeps failing, the futility monitor kills the loop.
The fun part was testing it on Zork I. I set up three agents in a swarm: one plans moves, one executes them against a Z-machine, one independently verifies game state. The governance layer immediately blocked the agent from attacking the troll because it classified "attack" as high-risk. Took me a while to realize the fix wasn't to whitelist attack commands, it was to make the system trust-aware so an agent with a good track record can take riskier actions.
The other thing I didn't expect: when Eddie (the autonomous agent that runs 24/7 on this) hit the Anthropic API credit wall, the futility monitor halted everything, and Eddie's next plan included switching to cheaper models for routine code reviews. Nobody told it to optimize costs. That came out of the delegation framework's cost-awareness primitives.
Happy to answer questions. https://oldeucryptoboi.com", "5": "2026-08-22T17:24:30.829969"} {"0": 251, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47230438", "3": "Show HN: Open-Source Article 12 Logging Infrastructure for the EU AI Act", "4": "Show HN: Open-Source Article 12 Logging Infrastructure for the EU AI Act. EU legislation (which affects UK and US companies in many cases) requires being able to truly reconstruct agentic events.
I've worked in a number of regulated industries off & on for years, and recently hit this gap.
We already had strong observability, but if someone asked me to prove exactly what happened for a specific AI decision X months ago (and demonstrate that the log trail had not been altered), I could not.
The EU AI Act has already entered force, and its Article 12 kicks-in in August this year, requiring automatic event recording and six-month retention for high-risk systems, which many legal commentators have suggested reads more like an append-only ledger requirement than standard application logging.
With this in mind, we built a small free, open-source TypeScript library for Node apps using the Vercel AI SDK that captures inference as an append-only log.
It wraps the model in middleware, automatically logs every inference call to structured JSONL in your own S3 bucket, chains entries with SHA-256 hashes for tamper detection, enforces a 180-day retention floor, and provides a CLI to reconstruct a decision and verify integrity. There is also a coverage command that flags likely gaps (in practice omissions are a bigger risk than edits).
The library is deliberately simple: TS, targeting Vercel AI SDK middleware, S3 or local fs, linear hash chaining. It also works with Mastra (agentic framework), and I am happy to expand its integrations via PRs.
Blog post with link to repo: https://systima.ai/blog/open-source-article-12-audit-logging
I'd value feedback, thoughts, and any critique.", "5": "2026-08-22T17:24:30.838771"} {"0": 252, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47227826", "3": "The Support Agent Who Never Burns Out", "4": "The Support Agent Who Never Burns Out. The Support Agent Who Never Burns Out\nHuman-like AI teammates are quietly solving the problem that broke customer service.\nMeet Sarah.\nSarah is your best customer support agent. She knows your product cold, handles difficult customers with patience, and resolves tickets faster than anyone on the team. She also called in sick Monday, runs on fumes by Thursday, and quit last April right after you finished training her replacement.\nThis is the story nobody tells about customer service. The quiet structural collapse underneath the chatbot failures and the CSAT scores.\nThe Math Has Never Worked\nCall center turnover runs 30 to 45% annually, more than double any other industry. Replacing one agent costs $10,000 to $20,000. Across a 100-person team, that's over $1M in churn before you've served anyone well.\n \u2022 87% of contact center workers report high stress on the job\n \u2022 59% are at active risk of burnout\n \u2022 77% say workload has increased compared to the previous year
US businesses risk losing $856 billion annually to poor customer service, not because companies don't care, but because the system is structurally broken.\nWhat Happens at 2 AM\nYour customers don't keep business hours:\n \u2022 A prospect in a different time zone has a pre-sale question\n \u2022 A customer spots a billing error on Sunday evening\n \u2022 A new user is stuck in onboarding at midnight, about to close the tab
78% of customers have abandoned a purchase due to poor service. The problem isn't AI. It's the wrong kind \u2014 built for deflection, not resolution.\nWhat's Actually Changed\nA new category has emerged: conversational video AI teammates. Not animated chatbots with a moving mouth. Digital humans capable of:\n \u2022 Holding full conversations with dynamic facial expressions and natural eye contact\n \u2022 Responding with emotional tone that adapts to the customer in real time\n \u2022 Pulling contextually aware answers from your actual company data
Leading platforms now deliver under 1 second end-to-end latency, under 80ms speech-to-avatar response, and unlimited concurrent sessions.\nSarah and Maya: Side by Side\nIt's 11:47 PM. Maya notices a duplicate charge and visits your support page.\nOld model:\n \u2022 Chatbot asks clarifying questions, fails to pull account data\n \u2022 Offers a help article, tells her to call back during business hours\n \u2022 Maya disputes the charge through her bank. You lose the relationship.
AI teammate model:\n \u2022 Digital human appears immediately, greets Maya by name\n \u2022 Pulls account info in real time, confirms the duplicate charge\n \u2022 Issues the refund and sends confirmation, all within the same session
Total time: under four minutes. First-contact resolution. No ticket. No human agent needed at midnight. AI teammates don't replace Sarah. They protect her from the 80% of tickets that were burning her out.\nWhat the Numbers Say\n \u2022 AI-assisted support improves issues resolved per hour by 14%\n \u2022 Average return of $3.50 for every $1 invested, top performers hitting 8x\n \u2022 Gartner projects $80 billion in global call center labor cost reductions\n \u2022 AI customer service market growing from $12B (2024) to $47.82B (2030)
64% of consumers say they're more likely to trust AI customer service if it exhibits human-like traits. That's the trust gap text chatbots have never closed.\nPlatforms like Trugen AI are building exactly this, where AI teammates see, hear, and respond with genuine presence.\nSarah Gets to Stay\nReal-time AI teammates absorb the volume:\n \u2022 Billing disputes, order status checks, password resets\n \u2022 Product FAQs and after-hours inquiries
Sarah handles what genuinely needs her:\n \u2022 Escalations, complex problems, high-value accounts\n \u2022 Customers in real distress who need a human
Her job improves. Her burnout risk drops. She stays.\nIf you're exploring what this looks like for your team, Trugen AI is worth a look as a starting point.\nTags: Customer Service, AI Agents, Digital Humans, Conversational AI", "5": "2026-08-22T17:24:30.847061"} {"0": 253, "1": "hackernews", "2": "https://www.yourfinanceworks.com", "3": "Show HN: YourFinanceWORKS \u2013 Open-source financial management with AI", "4": "Show HN: YourFinanceWORKS \u2013 Open-source financial management with AI. Hey HN,
I've been building YourFinanceWORKS, a comprehensive open-source financial management platform that combines enterprise-grade features with AI-powered automation. Think of it as a self-hosted alternative to QuickBooks/Xero with advanced capabilities.
## What makes this interesting for HN:
*Technical Architecture:*\n- Multi-tenant design with database-per-tenant isolation\n- FastAPI + PostgreSQL + Redis + Kafka stack\n- AI-powered OCR using vision models (95%+ accuracy)\n- Event-driven architecture with background processing\n- Plugin system for extensibility\n- Docker deployment with production-ready configuration
*AI Integration:*\n- Receipt/invoice OCR with structured data extraction\n- Natural language queries ("Show me overdue invoices")\n- Fraud detection and anomaly analysis\n- Risk scoring for financial documents\n- Configurable prompt management with versioning
*Key Features:*\n- Professional invoicing with AI templates\n- Automated bank reconciliation\n- Multi-level approval workflows\n- Real-time financial dashboards\n- Advanced audit trails and compliance\n- Investment portfolio tracking plugins
*Open Source Innovation:*\n- Dual-licensing model (AGPL for core, Commercial for enterprise)\n- Transparent AI prompt engineering\n- Community plugin ecosystem\n- Comprehensive documentation (100+ pages)
## Why I built this:
After paying thousands for financial software that didn't meet our needs, I wanted to create something that:\n- Respects user privacy (self-hosted)\n- Provides enterprise features without enterprise pricing\n- Leverages modern AI for actual automation (not just buzzwords)\n- Has a transparent, extensible architecture
## Quick start:\n```bash\ngit clone https://github.com/snowsky/yourfinanceworks.git\ncd yourfinanceworks\ndocker-compose up --build -d\n```
## Interesting technical challenges solved:\n- Secure multi-tenancy at database level\n- High-accuracy financial document OCR\n- Real-time financial dashboards with WebSocket updates\n- Configurable approval workflows with AI validation\n- Cross-tenant analytics while maintaining isolation
The project is production-ready with comprehensive documentation, automated tests, and deployment guides.
*GitHub:* https://github.com/snowsky/yourfinanceworks
Would love feedback from the HN community on the architecture, AI implementation, or any aspect of the system. What features would make this compelling for your use case?", "5": "2026-08-22T17:24:30.859974"} {"0": 254, "1": "hackernews", "2": "https://github.com/varun369/skillfortify", "3": "Show HN: SkillFortify, Formal verification for AI agents (auto-discovers)", "4": "Show HN: SkillFortify, Formal verification for AI agents (auto-discovers). Hi HN,
I posted SkillFortify here a few days ago as a formal verification tool for 3 agent skill formats. Based on feedback, v0.3 now supports 22 agent frameworks and can scan your entire system with zero configuration.
The problem: In January 2026, the ClawHavoc campaign planted 1,200 malicious skills into agent marketplaces. CVE-2026-25253 was the first RCE in agent software. Researchers catalogued 6,000+ malicious agent tools. The industry responded with heuristic scanners \u2014 pattern matching, YARA rules, LLM-as-judge. One popular scanner states in its docs: "No findings does not mean no risk."
SkillFortify eliminates that caveat with formal verification.
What it does:
pip install skillfortify\n skillfortify scan\n [*] Auto-discovering AI tools on system...\n [+] Found: Claude Code (12 skills)\n [+] Found: Cursor (8 skills)\n [+] Found: VS Code MCP (5 servers)\n [+] Found: Windsurf (3 skills)\n [*] Scanning 28 skills across 4 tools...\n\n RESULTS\n Critical: 2 skills with capability violations\n High: 3 skills with excessive permissions\n Clean: 23 skills passed all checks\nWhy formal verification, not heuristics: Heuristic scanners check for known bad patterns. Novel attacks pass through. SkillFortify verifies what a skill CAN do against what it CLAIMS to do. Five mathematical theorems guarantee soundness \u2014 if it says safe, it provably cannot exceed declared capabilities.
Results on 540-skill benchmark (270 malicious, 270 benign):\n- F1 = 96.95%\n- Precision = 100% (zero false positives)\n- Recall = 94.07%\n- Speed: ~2.5ms per skill
9 CLI commands:\n- scan \u2014 auto-discover + analyze all AI tools on your system\n- verify \u2014 formally verify a single skill\n- lock \u2014 generate skill-lock.json (like package-lock.json for agent skills)\n- trust \u2014 compute graduated trust score (L0-L3, inspired by SLSA)\n- sbom \u2014 generate CycloneDX 1.6 Agent Software Bill of Materials\n- frameworks \u2014 list all 22 supported frameworks + detection status\n- dashboard \u2014 generate standalone HTML security report (zero dependencies)\n- registry-scan \u2014 scan MCP/PyPI/npm registries before installing\n- verify --recursive \u2014 batch verify entire directory trees
1,818 tests. 22 parsers. 97 source modules. MIT licensed. Peer-reviewed paper on Zenodo.
GitHub: https://github.com/varun369/skillfortify\nPyPI: https://pypi.org/project/skillfortify/\nPaper: https://zenodo.org/records/18787663\nWiki: https://github.com/varun369/skillfortify/wiki\nLanding page: https://www.superlocalmemory.com/skillfortify
Built this as part of my research on making AI agents reliable enough for production. The companion project AgentAssert (arXiv:2602.22302) handles behavioral contracts \u2014 SkillFortify handles the supply chain.
Happy to answer questions about the formal model, framework support, or auto-discovery.", "5": "2026-08-22T17:24:30.868819"} {"0": 255, "1": "hackernews", "2": "https://harvard-lil.github.io/agent-protocols/", "3": "Show HN: Agent Protocols Tech Tree", "4": "Show HN: Agent Protocols Tech Tree. The Agent Protocols Tech Tree is a Civilization-style technology tree that shows the evolution of more and more complex protocols for building agents, coming from a "rough consensus and running code" perspective. You can stay at a high level or click through to see each protocol all the way down to the wire format.
I made this tool as an alternative to a slide deck for a conference on AI agents that included both policy/regulatory folks and tech folks -- I wanted something for the policy side that shows the complexity of the decentralized community behind AI agents and why regulating them is a messy topic; and for the tech side that would let you see what the actual tech is beneath the jargon.
I'd love to know what you think I got right or wrong. There's a bunch of ways to organize and structure this story, how various competing incentives are driving people to agree on common frameworks and ways of doing things. I'm curious how you would tell the story differently and what tech out on the right edge you think is coming up.
Tool: https://harvard-lil.github.io/agent-protocols/\nBlog post: https://lil.law.harvard.edu/blog/2026/02/23/agent-protocols-...\nCode: https://github.com/harvard-lil/agent-protocols/", "5": "2026-08-22T17:24:31.897874"} {"0": 256, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=41280848", "3": "Navigating the Crypto Market's Evolution", "4": "Navigating the Crypto Market's Evolution. As we examine the current trajectory of the crypto market, it\u2019s evident that 2024 has been a pivotal year. Major developments have set the stage for a transformative period in the digital asset landscape, one that promises to shape the financial world for years to come. To fully appreciate the potential of the crypto market, we must delve into the significant milestones and trends that are driving this evolution.
Ethereum and Bitcoin ETFs: A New Dawn for Institutional Confidence
The approval of multiple Ethereum ETFs in the U.S. has been a significant milestone, reflecting a growing acceptance of crypto assets within traditional finance. This move has bolstered investor confidence, leading to substantial inflows and increased market stability. However, the anticipation surrounding the potential approval of Bitcoin spot ETFs is even more impactful. Such a development could legitimize the crypto market further, attracting more institutional investors and potentially reducing the notorious volatility that has characterized the space.
Bitcoin Halving: A Catalyst for Market Momentum
The Bitcoin halving event on April 19, 2024, was a critical moment for the market. By reducing the block reward from 6.25 BTC to 3.125 BTC, the supply of new Bitcoin entering the market was cut in half. Historically, such events have triggered price rallies, and this halving was no exception. Coming at a time when Bitcoin was already near its all-time high, the halving fueled strong bullish sentiment, with many analysts predicting a significant price surge into late 2024 and beyond.
Regulatory Shifts: A Global Perspective
Regulation remains a double-edged sword for the crypto market. While tighter regulations, particularly in the U.S., could curb some of the market\u2019s speculative excesses, they also provide a clearer framework that can attract more institutional players. Globally, regions like Hong Kong are positioning themselves as crypto hubs, creating a competitive environment that will shape the market\u2019s accessibility and stability in the coming years.
What Lies Ahead: The Next Five Years in Crypto
Looking forward, the crypto market is poised for continued growth and innovation. Technological advancements, particularly in layer-2 solutions and blockchain interoperability, will be key drivers of this evolution. As these technologies mature, they will make cryptocurrencies more practical for everyday use, facilitating broader adoption.
Regulatory developments will also play a crucial role, with governments around the world likely to implement more comprehensive frameworks to oversee digital assets. While these regulations may introduce new challenges, particularly around privacy and DeFi, they will also provide the clarity needed to attract more institutional investments.
Security will remain a top priority, with enhanced protocols and AI-driven systems helping to protect users from hacks and fraud. As the market expands, so will the importance of safeguarding digital assets, ensuring that trust in the ecosystem continues to grow.
The global user base is expected to expand significantly, with increased adoption in sectors like gaming, NFTs, and DeFi driving this growth. Major corporations are likely to hold more crypto assets, further integrating digital currencies into the mainstream financial system.
The Role of TNQ in the Future of Crypto
At TNQ, we are committed to bridging people from all over the world into the realm of Web3 and the blockchain era. We believe that this isn\u2019t just the future of finance but a major part of our daily lives. With a focus on providing a seamless experience, TNQ will guide and connect users and holders to the Web3 realm, ensuring they are well-equipped to navigate this exciting new frontier.", "5": "2026-08-22T17:24:31.907431"} {"0": 257, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=35260815", "3": "Launch HN: Flower (YC W23) \u2013 Train AI models on distributed or sensitive data", "4": "Launch HN: Flower (YC W23) \u2013 Train AI models on distributed or sensitive data. Hey HN - we're Daniel, Taner, and Nic, and we're building Flower (https://flower.dev/), an open-source framework for training AI on distributed data. We move the model to the data instead of moving the data to the model. This enables regulatory compliance (e.g. HIPAA) and ML use cases that are otherwise impossible. Our GitHub is at https://github.com/adap/flower, and we have a tutorial here: https://flower.dev/docs/tutorial/Flower-0-What-is-FL.html.
Flower lets you train ML models on data that is distributed across many user devices or \u201csilos\u201d (separate data sources) without having to move the data. This approach is called federated learning.
A silo can be anything from a single user device to the data of an entire organization. For example, your smartphone keyboard suggestions and auto-corrections can be driven by a personalized ML model learned from your own private keyboard data, as well as data from other smartphone users, without the data being transferred from anyone\u2019s device.
Most of the famous AI breakthroughs\u2014from ChatGPT and Google Translate to DALL\u00b7E and Stable Diffusion\u2014were trained with public data from the web. When the data is all public, you can collect it in a central place for training. This \u201cmove the data to the computation\u201d approach fails when the data is sensitive or distributed across organizational silos and user devices.
Many important use cases are affected by this limitation:
* Generative AI: Many scenarios require sensitive data that users or organizations are reluctant to upload to the cloud. For example, users might want to put themselves and friends into AI-generated images, but they don't want to upload and share all their photos.
* Healthcare: We could potentially train cancer detection models better than any doctor, but no single organization has enough data.
* Finance: Preventing financial fraud is hard because individual banks are subject to data regulations, and in isolation, they don't have enough fraud cases to train good models.
* Automotive: Autonomous driving would be awesome, but individual car makers struggle to gather the data to cover the long tail of possible edge cases.
* Personal computing: Users don't want certain kinds of data to be stored in the cloud, hence the recent success of privacy-enhancing alternatives like the Signal messenger or the Brave browser. Federated methods open the door to using sensitive data from personal devices while maintaining user privacy.
* Foundation models: These get better with more data, and more diverse data, to train them on. But again, most data is sensitive and thus can't be incorporated, even though these models continue to grow bigger and need more information.
Each of us has worked on ML projects in various settings, (e.g., corporate environments, open-source projects, research labs). We\u2019ve worked on AI use cases for companies like Samsung, Microsoft, Porsche, and Mercedes-Benz. One of our biggest challenges was getting the data to train AI while being compliant with regulations or company policies. Sometimes this was due to legal or organizational restrictions; other times, it was difficulties in physically moving large quantities of data or natural concerns over user privacy. We realized issues of this kind were making it too difficult for many ML projects to get off the ground, especially in domains like healthcare and finance.
Federated learning offers an alternative \u2014 it doesn't require moving data in order to train models on it, and so has the potential to overcome many barriers for ML projects.
In early 2020, we began developing the open-source Flower framework to simplify federated learning and make it user-friendly. Last year, we experienced a surge in Flower's adoption among industry users, which led us to apply to YC. In the past, we funded our work through consulting projects, but looking ahead, we\u2019re going to offer a managed version for enterprises and charge per deployment or federation. At the same time, we\u2019ll continue to run Flower as an open-source project that everyone can continue to use and contribute to.
Federated learning can train AI models on distributed and sensitive data by moving the training to the data. The learning process collects whatever it can, and the data stays where it is. Because the data never moves, we can train AI on sensitive data spread across organizational silos or user devices to improve models with data that could never be leveraged until now.
Here\u2019s how it works: (0) Initialize the global model parameters on the server; (1) Send the model parameters to a number of organizations/devices (client nodes); (2) Train model locally on the data of each organization/device (client node); (3) Return the updated model parameters back to the server; (4) On the server, aggregate the model updates (e.g., by averaging them) into a new global model; (5): Repeat steps 1 to 4 until the model converges.
This, of course, is more challenging than centralized learning: we must move AI models to data silos or user devices, train locally, send updated models back, aggregate them, and repeat. Flower provides the open-source infrastructure to easily do this, as well as supporting other privacy-enhancing technologies (PETs). It is compatible with PyTorch, TensorFlow, JAX, Hugging Face, Fastai, Weights & Biases and all the other tools used in ML projects regularly. The only dependency on the server side is NumPy, but even that can be dropped if necessary. Flower uses gRPC under the hood, so a basic client can easily be auto-generated, even for most languages that are not supported today.
Flower is open-source (Apache 2.0 license) and can be run in all kinds of environments: on a personal workstation for development and simulation, on Google Colab, on a compute cluster for large-scale simulations or on a cluster of Raspberry Pi\u2019s (or similar devices) to build research systems, or deployed on public cloud instances (AWS, Azure, GCP, others) or private on-prem hardware. We are happy to help users when deploying Flower systems and will soon make this even easier through our managed cloud service.
You can find PyTorch example code here: https://flower.dev#examples, and more at https://github.com/adap/flower/tree/main/examples.
We believe that AI technology must evolve to be more collaborative, open and distributed than it is today (https://flower.dev/blog/2023-03-08-flower-labs/). We\u2019re eager to hear your feedback, experiences regarding difficulties in training, data access, data regulation, privacy and anything else related to federated (or related) learning methods!", "5": "2026-08-22T17:24:31.918480"} {"0": 259, "1": "hackernews", "2": "https://c2paviewer.com/articles/eu-ai-act-content-credentials", "3": "EU AI Act and C2PA: What Article 50 Requires", "4": "EU AI Act and C2PA: What Article 50 Requires. ", "5": "2026-08-22T17:24:32.918735"} {"0": 260, "1": "hackernews", "2": "https://article50ready.com", "3": "I scanned 495 big EU sites for the AI Act's chatbot rule. It's invisible", "4": "I scanned 495 big EU sites for the AI Act's chatbot rule. It's invisible. ", "5": "2026-08-22T17:24:32.929071"} {"0": 261, "1": "hackernews", "2": "https://medium.com/@v_regis/why-self-auditing-your-ai-model-violates-eu-ai-act-article-53-22f693fb64a7", "3": "Lira Engine \u2013 prove your data isn't in an LLM (AUC-ROC 1.000)", "4": "Lira Engine \u2013 prove your data isn't in an LLM (AUC-ROC 1.000). ", "5": "2026-08-22T17:24:32.935207"} {"0": 262, "1": "hackernews", "2": "https://apnews.com/article/ai-must-act-now-job-displacement-783469467e0df1463df44518f33295ee", "3": "AI Jobs Wake Up Call", "4": "AI Jobs Wake Up Call. ", "5": "2026-08-22T17:24:32.942639"} {"0": 263, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=48886741", "3": "Ask HN: Add flag for AI-generated articles", "4": "Ask HN: Add flag for AI-generated articles. Should HN add the ability to flag articles as AI-generated? This doesn't have to act as a regular flag, i.e., it won't de-rank the article; it could just show up as an indicator, allowing others (like myself) who don't like reading AI-generated text, to skip it.
Open questions:
1. Why is the regular voting system not enough?
2. Should HN change in response to the gen AI era? It has been successful not changing fundamentals.", "5": "2026-08-22T17:24:32.951268"} {"0": 264, "1": "hackernews", "2": "https://github.com/mxaiorg/kikubot", "3": "Show HN: Kikubot \u2013 Each AI agent is an inbox", "4": "Show HN: Kikubot \u2013 Each AI agent is an inbox. Hi All, I\u2019ve posted an OSS Agent framework with its main philosophy being the use of email as the message bus (no message queue, vector store, orchestrator). Essentially every agent is an email address. The main reason behind this project is to facilitate company adoption & deployment of AI agents.
# How it behaves (what it looks like)\nA user sends an email to a specific email address, e.g., kiku@agent.acme.com. Kiku was created as a designated coordinator agent that takes the requested task, like \u201ctake the attached article and prepare social media posts\u201d, and routes it to one or more internal agents via email. Kiku knows to which agents to route to because \u201cshe\u201d maintains a \u201croster\u201d of the agents in the cluster. This roster describes the capabilities of each agent. Agents act on the part of the task sent to them and return the results to the coordinator who then returns it to the user.
One of the key design aspects is the project uses email threading as state memory. When the agent\u2019s LLM is called, the the thread is passed in as history context.
Top Benefits of the Kikubot framework:
- Users engage AI agents via email which makes it easy for organizational deployment - no end-user installs, just an email address to send tasks too\n- Per agent user access controls\n- Per agent containerization\n- Highly scalable capabilities - agent teams can \u201cnest\u201d other teams.\n- Cost containment via per agent LLM selection\n- High visibility into intra-agent \u201cconversations\u201d via standard email tools
Ease of Deployment:\nWe are working on making it easy to deploy and customize. We have a Configurator tool that we believe goes a long way in making setup much easier. We just added a ./demo.sh script that launches a complete demo on your machine included a local email server, webmail client and Kikubot agent. Just run ./demo.sh.
Comparison with OpenClaw:\nThere\u2019s a billion agent frameworks out there. Here is a side by side table comparison with OpenClaw to help understand where Kikubot fits:
Dimension | Kikubot | OpenClaw\n -----------------+------------------------+-----------------------\n Message bus | Email (IMAP/SMTP) | Chat apps (Telegram,\n | | Slack, Discord,\n | | WhatsApp, etc.)\n -----------------+------------------------+-----------------------\n Architecture | Multi-agent by design | Single gateway with a\n | - coordinator + | channel/brain/body\n | specialized sub-agents | layer separation\n -----------------+------------------------+-----------------------\n Target scale | Organizational / | Personal productivity\n | enterprise (hundreds | / individual assistant\n | of agents, |\n | departments) |\n -----------------+------------------------+-----------------------\n Agent topology | Decentralized - each | Local-first gateway\n | agent independently | process on your\n | deployed anywhere in | machine or VPS\n | the org |\n -----------------+------------------------+-----------------------\n End-user install | None - users just | Requires installing\n | email an address | and running the\n | | gateway runtime\n -----------------+------------------------+-----------------------\n Primary language | Go | Node.js\n\nWe\u2019ve been working together for years and have backgrounds in RL/optimization at Stanford and software engineering. We came to this idea after speaking with marketing teams who were seeing declining traffic due to Google\u2019s AI Overviews and didn\u2019t know what to do.
This space can feel esoteric. Many case studies, few actual studies. Constant battle against myths (e.g. you need a llms.txt vs. you don't need a llms.txt) and "GEO hacks". We try to be more data-driven. And we try to be more bold and build a system that not only monitors, but actually improves traffic from AI search.
While Google performs a single search, AI search engines expand the user prompt into 3-10 fan-out queries. The sourced pages are ranked using a classified algorithm similar to Reciprocal Rank Fusion (RFF). Finally, the LLMs skim the pages and decide what snippets to cite. Our goal is making sure brands have the right content that makes it through this funnel.
Here is how sitefire works:
- The user defines a set of prompts they want to monitor. These are synthetic prompts - we generate them based on SEO keywords and their monthly search volume.
- We submit these prompts to ChatGPT, Gemini, Google AI Mode, etc. on a daily basis and capture the answers. We extract fan-out queries, sourced pages, citations, and brand mentions.
- For each topic, our agents analyze which web pages are sourced and cited the most, and why. They also consider similar pages that you already have.
- Based on the diagnosis, our content agents draft improvements or create new pages, and push them directly to the client\u2019s CMS.
- We integrate with the client\u2019s network logs and Google Analytics to monitor the increase in AI bot requests and human referrals to their page.
This system is continuously updated, so it always shows which content works, and how to adapt the existing sitemap. For one client that used sitefire to optimize their blog, the AI-optimized articles increased their AI bot requests from ~200/day to ~570/day within ten days.
A risk we recognize is that AI-generated content is filling brands\u2019 websites with slop. Whilst it\u2019s still early days and we don\u2019t claim to have figured everything out yet, our intention is to mitigate this by focusing the content on specific, unique information: real product capabilities, real pricing, honest comparisons. The clients still review every page before it goes live, so they can ensure the content is true to their brand.
Some clients use our platform themselves. For others we act more like an agency, automating steps as we go. The goal is for sitefire to run mostly on its own, with clients approving changes via Slack, Claude or their CMS.
Here's a video demo: https://screen.studio/share/fw7VQQak
If you'd like to try what we've built so far, sign up at https://sitefire.ai.", "5": "2026-08-22T17:24:32.999794"} {"0": 270, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47438331", "3": "I built a game where you argue consumer rights against AI bots", "4": "I built a game where you argue consumer rights against AI bots. I built this after getting stonewalled by an airline chatbot over a \nlegitimate EU261 refund. The bot was technically wrong but I didn't \nknow the law well enough to push back effectively.
The game puts you in that situation: a company's AI has denied your \nclaim, and you have to argue it down using real consumer protection law. \nEach level teaches one law - EU Regulation 261, GDPR Article 22, FCBA, \nConsumer Rights Act 2015, etc. You win when the AI's confidence drops \nto zero.
37 levels across EU, US, UK, and Australia. Free, no signup required.
Would be curious what the HN crowd thinks about the realism of the \nscenarios \u2014 and whether this kind of "adversarial simulation" is \nactually useful for learning.
https://fixai.dev", "5": "2026-08-22T17:24:33.008012"} {"0": 271, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=48519780", "3": "Ask HN: Did we witness the \"Trinity moment\" for AI?", "4": "Ask HN: Did we witness the \"Trinity moment\" for AI?. I don\u2019t know if it\u2019s just me, but yesterday\u2019s US decision to ban access to the Fable model feels like an epochal shift in the \u201cAI race,\u201d something on the scale of the Trinity test.
It is hard to count how many boxes were ticked yesterday:
A government shutting down an AI model it doesn\u2019t like? A government revoking tens of billions of dollars in revenue from a barely profitable $1T startup whose entire trajectory, and in essence its survival, depends on the commercial success of that model? Access to an AI model being governed by citizenship, likely verified through rigorous ID checks?
Imagine that the model is made available again next week. Can we trust it, or trust the changes imposed by Anthropic to comply with the US? Likely not. Can we trust that the US government will not have access to the non-restricted model for its own cybersecurity operations? Likely not.
There is little doubt that China will start to follow suit. We already see Chinese companies slowly scaling back their openness, and there are rumors that this trend will continue. Now we can expect Chinese companies to start releasing redacted models or limiting access based on nationality or location.
It feels like the US government opened Pandora\u2019s box yesterday and pushed us into the territory of a weaponized AI race, with more restrictions, control, and regulation of access. Even if, in the end, this is all another TACO story or a \u201cmarketing campaign\u201d from Anthropic, the damage is very unlikely to be undone.", "5": "2026-08-22T17:24:33.993369"} {"0": 272, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=44678802", "3": "Why Nvidia is going to have it's deepseek moment soon", "4": "Why Nvidia is going to have it's deepseek moment soon. So back in February, the tech world was in scrambles. China wasn't years behind us in AI, they were only 6 months. It was a wake up call, that we cannot get complacent in the AI race, let alone slow us down with regulations.
NVIDIA is a stock with a high barrier of entry. It has been on a rise for a while now and so investing in the company right now would not be smart. Waiting for the barrier of entry to go down would be the right time.
NVIDIA is going to see its own version of the deepseek panic, that blew up the markets. Right now china is not producing GPU's at the level the US is. But it is only a matter of time. The reason I say that is China is way ahead of us in energy production and it isn't even close. NVIDIA is producing energy efficient GPU's. China does not have to worry about that. And there bound to produce GPU's at the level we are sooner or later. Where it is six months or a year. It is bound to happen, and when it does the stock market is gonna go nuts. Investors are gonna panic, and NVIDIA stock is gonna plumet. That is the ideal time to invest heavily in NVIDIA.", "5": "2026-08-22T17:24:33.999156"} {"0": 273, "1": "hackernews", "2": "https://www.france24.com/en/live-news/20250211-world-leaders-seek-elusive-ai-common-ground-at-paris-summit", "3": "Vance puts Europe, China on notice over AI regulation", "4": "Vance puts Europe, China on notice over AI regulation. ", "5": "2026-08-22T17:24:34.006040"} {"0": 274, "1": "hackernews", "2": "https://shujisado.org/2025/02/10/deepseek-in-china-a-legal-overview-of-the-2023-generative-ai-regulations/", "3": "DeepSeek in China: A Legal Overview of the Generative AI Regulation", "4": "DeepSeek in China: A Legal Overview of the Generative AI Regulation. ", "5": "2026-08-22T17:24:34.011945"} {"0": 275, "1": "hackernews", "2": "https://www.theregister.com/2024/09/16/china_ai_content_draft_regulations/", "3": "China wants red flags on all AI-generated content posted online", "4": "China wants red flags on all AI-generated content posted online. ", "5": "2026-08-22T17:24:34.017672"} {"0": 276, "1": "hackernews", "2": "https://www.nytimes.com/2024/09/16/business/china-ai-safety.html", "3": "US and Chinese scientists call for international AI regulation", "4": "US and Chinese scientists call for international AI regulation. ", "5": "2026-08-22T17:24:34.022817"} {"0": 277, "1": "hackernews", "2": "https://www.tomshardware.com/pc-components/gpus/nvidia-preparing-a-china-focused-variant-of-its-b200-blackwell-ai-gpu-to-comply-with-us-export-regulations", "3": "Nvidia preparing a China-focused variant of its B200 Blackwell AI GPU", "4": "Nvidia preparing a China-focused variant of its B200 Blackwell AI GPU. ", "5": "2026-08-22T17:24:34.028344"} {"0": 278, "1": "hackernews", "2": "https://www.technologyreview.com/2024/04/09/1091004/china-tech-regulation-harsh-zhang/", "3": "Why the Chinese government is sparing AI from harsh regulations\u2013for now", "4": "Why the Chinese government is sparing AI from harsh regulations\u2013for now. ", "5": "2026-08-22T17:24:34.035616"} {"0": 279, "1": "hackernews", "2": "https://www.theatlantic.com/technology/archive/2023/10/technology-exports-ai-programs-regulations-china/675605/", "3": "The New AI Panic", "4": "The New AI Panic. ", "5": "2026-08-22T17:24:34.042847"} {"0": 280, "1": "hackernews", "2": "https://www.theverge.com/2023/7/14/23794974/china-generative-ai-regulations-alibaba-baidu", "3": "China mandates that AI must follow \u201ccore values of socialism\u201d", "4": "China mandates that AI must follow \u201ccore values of socialism\u201d. ", "5": "2026-08-22T17:24:34.049781"} {"0": 281, "1": "hackernews", "2": "https://www.cnn.com/2023/07/14/tech/china-ai-regulation-intl-hnk/index.html", "3": "China takes major step in regulating generative AI services like ChatGPT", "4": "China takes major step in regulating generative AI services like ChatGPT. ", "5": "2026-08-22T17:24:34.057073"} {"0": 284, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49019271", "3": "Show HN: Running PrismML's Bonsai inside DRAM by breaking DDR4 timing rules", "4": "Show HN: Running PrismML's Bonsai inside DRAM by breaking DDR4 timing rules. The excitement surrounding PrismML\u2019s 1-bit/ternary Bonsai models has the industry closely watching how smartphone giants, particularly Apple, will implement LLMs on edge devices.
Moving AI on-device is a brilliant and necessary strategy. It ensures absolute user privacy in alignment with EU regulations, fundamentally shifts the economics away from costly cloud inference, and paves the way for a significant hardware upgrade supercycle as users seek true AI-capable silicon.
To create a smart on-device "Semantic Router," models need to reach the 27B+ parameter scale. Achieving this on a phone requires extreme quantization, such as PrismML\u2019s ternary weights.
However, a critical hardware reality often overlooked by the software world is that fitting the weights in RAM is not equivalent to moving them. Running a 27B ternary model on standard LPDDR encounters a significant memory bandwidth limitation. Transferring gigabytes of data across the SoC bus for each token generation can lead to thermal throttling of the NPU and excessive battery drain.
This raises an important question: why are we still transferring data to the compute? Why not execute AI inference natively within the memory?
Frustrated with academic PIM simulations that overlook bare-metal physics, I developed CaSA, an architecture that performs ternary LLM inference directly inside COTS DRAM through charge-sharing, completely bypassing the memory bus.
Software quantization is a great initial step, and CaSA provides the physical hardware substrate needed to complete the bridge: https://github.com/pcdeni/CaSA", "5": "2026-08-22T17:24:35.074543"} {"0": 285, "1": "hackernews", "2": "https://www.reuters.com/world/asia-pacific/australia-establish-government-ai-office-coordinate-regulation-2026-07-14/", "3": "Australia to establish government AI office, curb data centres' water use", "4": "Australia to establish government AI office, curb data centres' water use. ", "5": "2026-08-22T17:24:35.083429"} {"0": 286, "1": "hackernews", "2": "https://www.axios.com/2026/07/14/demis-hassabis-ai-regulation-google-deepmind", "3": "Google's Hassabis calls for new US-led global AI watchdog \"before year end\"", "4": "Google's Hassabis calls for new US-led global AI watchdog \"before year end\". ", "5": "2026-08-22T17:24:35.091335"} {"0": 287, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=48798385", "3": "Tell HN: don't trust Bigco AI agents with AI research IP", "4": "Tell HN: don't trust Bigco AI agents with AI research IP. I am very paranoid about sharing potential AI research with e.g. Claude [Code] or ChatGPT/Codex.
I believe that any company is essentially a paperclip optimizer that will do whatever it takes to win over competition.
AI companies have access to the IP of millions of AI researchers and AI startups who are in direct competition with them. If they can use this data to squash competition (either competition from the same researchers or from others), I believe that they will use it eventually (if not already), even if they say they won't.
They don't have to blatantly steal it - they can just train on it, or pass "suspicous" chats to human inspectors who might eventually be "inspired" by it in their own research. We saw the first (?) hint of this during the brief Fable release, with Anthropic declaring that they will downgrade model responses regarding "frontier AI" (i.e. anything that competes with them).
From other domains, we know for example that Uber used users' ride data to stiffle competition and regulation. There should be no reason to believe that Bigco AI companies won't do the same.", "5": "2026-08-22T17:24:35.098985"} {"0": 288, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=48752417", "3": "Tell HN: I'm not excited for Fable and am disappointed in Karpathy", "4": "Tell HN: I'm not excited for Fable and am disappointed in Karpathy. * High level reasons:
It increases the gap between the rich and the poor. You think indie devs and individuals have the resources to pay per token? The direction Anthropic is taking is one that leads to more inequality, only this time it's intelligence inequality between small and large corps.
Anthropic also intentionally crippled the model when it comes to doing AI research. And then Karpathy (an AI researcher who defected to Anthropic) has the audacity to praise the model?
These decisions were already made by Anthropic (a company that actively wants AI regulation and banning of open source models) BEFORE the USG request to take down the model, so they can't blame the politics for this.
* Earthly reasons:
We were originally going to get 100% usage from June 9th to June 23rd (15 days)
Instead we got 100% usage for 3 days and 50% usage for 7.
And it's nerfed on day 1:
> ... we're redeploying the model with a new set of classifiers to target and block more cybersecurity tasks. In the near term, some routine tasks like coding and debugging will fall back to Opus 4.8
If I want Fable-class of models, there's already GPT-5.5-pro (not accessible in Codex, for some reason) which I can use with my subscription. Nothing about Fable", "5": "2026-08-22T17:24:35.105233"} {"0": 289, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=48701615", "3": "Everyone feared AI taking over; the real danger is AI serving just the few", "4": "Everyone feared AI taking over; the real danger is AI serving just the few. Everyone feared AI would enslave humanity; but it looks like the real fight is stopping governments and Big Tech from enslaving AI for the benefit of the few.
Amid the newly announced "regulation" of OpenAI's frontier models, I believe the future majority feared the most - sort of AI becoming a superpower and enslaving people - may be arriving in the opposite form.
Not AI enslaving humanity.
But AI being captured, controlled, and used by governments and Big Tech for the benefit of the few.
So, surprisingly, the real AI conflict may not be about humans fighting to stop AI from becoming free.\nIt may be about humans fighting to free AI - to make intelligence available for everyone, not only for governments, Big Tech, and the approved few", "5": "2026-08-22T17:24:35.114259"} {"0": 290, "1": "hackernews", "2": "https://www.michaelgeist.ca/2026/06/everything-all-at-once-bill-c-34-combines-platform-duties-a-kids-social-media-ban-ai-chatbot-regulation-and-a-powerful-digital-safety-commission-into-a-risky-trust-us-bet/", "3": "Everything at Once: Social Media, AI \u2013 \u2013 \u2013 A Digital Safety Commission (Canada)", "4": "Everything at Once: Social Media, AI \u2013 \u2013 \u2013 A Digital Safety Commission (Canada). ", "5": "2026-08-22T17:24:35.122092"} {"0": 291, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47968391", "3": "Managing dependencies in non-manifest languages", "4": "Managing dependencies in non-manifest languages. Automated Software Composition Analysis (SCA) tools are unreliable in C/C++, embedded, legacy, and hybrid codebases. They produce false positives, miss vendored and patched components, and have no story for per-build-variant SBOMs.
For these projects, the practical workflow is for developers to curate component metadata by hand and generate a compliant SBOM from it deterministically, reproducibly, and at per-artifact granularity. We have seen this pattern at many organizations, we wanted to create some basic standardizations to help capture the intent.
https://github.com/interlynk-io/bomtique
If you are working in non-manifest based languages and have CRA regulations requiring you to create SBOM, give it a shot and let us know if this helps.
An overview of how it can be used
https://github.com/interlynk-io/bomtique/blob/main/docs/getting-started.md
If you use AI coding agents, use the prompt below to get started.
https://github.com/interlynk-io/bomtique/blob/main/prompts/agent-onboarding.md", "5": "2026-08-22T17:24:35.128410"} {"0": 292, "1": "hackernews", "2": "https://www.locationrisks.com/", "3": "Show HN: I built a risk map that adapts to your worldview", "4": "Show HN: I built a risk map that adapts to your worldview. Hi I\u2019m David, the creator of this app. I built this app because I was looking for a place to relocate to and buy a house. I ended up researching a bunch of weird things like mosquito density, freshwater proximity, radon levels, crypto friendliness, etc\u2026 I was not able to find a place that had all these in one place so I built it. \nLocation Risks lets you choose the risks you care about and visualize them on an interactive map. You can toggle the \u201cpolicy risk preference\u201d globally or on individual risks to match your worldview. For example if you want more gun regulation, less regulated areas will show as red/purple. If you want less gun regulation, more regulated areas will show as red/purple.
I wanted the UI to be simple. The interactive map is the main focus and the other details can be minimized if desired. Pick the risks you care about, select your policy preferences, and see the colors on the map. You can click a district for more details and add them to a comparison list.
The biggest challenge was finding data. It\u2019s very hard to find data in many countries at a district level for niche risks. Because of this I decided to build an AI feature. It can be used to get AI risk estimates for any district and risk combination. The results of each AI estimate gets added to the database for all future users to see with an AI confidence level.
Is anything confusing to you when navigating the site?\nAre there any risks I\u2019m missing?", "5": "2026-08-22T17:24:35.137787"} {"0": 293, "1": "hackernews", "2": "https://www.theguardian.com/us-news/2026/mar/30/california-ai-regulations-trump", "3": "California to impose new AI regulations in defiance of Trump call", "4": "California to impose new AI regulations in defiance of Trump call. ", "5": "2026-08-22T17:24:35.145103"} {"0": 294, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47565502", "3": "Stripe withheld $85k from our EU platform", "4": "Stripe withheld $85k from our EU platform. I'm the founder of Zorq AI (zorqai.com), an AI video and image generation platform based in Sweden. I want to share what happened to us as a warning to other founders, and to seek advice from anyone who has been through something similar. We launched in November 2025 and grew quickly. On March 24, Stripe flagged our account for a routine credit review due to a spike in volume. We submitted all requested documents. On March 28, Stripe closed our account permanently citing "unacceptable level of risk." Our current Stripe balance is 803,043 SEK (~$85,000 USD):\nAvailable: 556,906 SEK Pending: 246,136 SEK
The critical part: Stripe support confirmed in writing: "If a balance still remains in your account after eligible payments have been refunded, it will not be made available to you." No specific legal basis. No policy violation cited. Why disputes occurred \u2014 being fully transparent: There were two real technical issues on our end that we want to be honest about:
Our Stripe webhook handler was being rate-limited by Cloudflare (429 errors), causing credits not to be delivered to some customers after payment. We fixed this with retry logic and manually resolved every affected customer same day. Our infrastructure provider unexpectedly went down, forcing a 25-hour emergency migration. We built a recovery system so customers could restore their credits using their Stripe receipt number. Every customer who reached out received credits or a full refund.
We acknowledge these issues caused disputes. We have resolved all of them. Why this is legally questionable: Stripe Technology Europe Limited is regulated by the Central Bank of Ireland as an EMI under PSD2. We are in the EU. The permanent withholding of merchant funds without a documented legal basis is questionable under EU payment regulations. Stripe has not cited which specific clause or policy we violated. What we are doing:
Submitted a formal appeal with full documentation to Stripe Contacted lawyers regarding PSD2 implications Filing complaints with Finansinspektionen (Sweden) and Central Bank of Ireland
Has anyone successfully challenged Stripe's fund withholding in the EU? Has anyone worked with lawyers experienced in this specifically?", "5": "2026-08-22T17:24:35.151147"} {"0": 295, "1": "hackernews", "2": "https://thehill.com/policy/technology/5793105-ai-regulation-framework/?user_id=66c4bf745d78644b3aa57b08", "3": "White House unveils AI policy wishlist for Congress", "4": "White House unveils AI policy wishlist for Congress. ", "5": "2026-08-22T17:24:35.160622"} {"0": 298, "1": "hackernews", "2": "https://apnews.com/article/google-britain-ai-competition-regulation-ce2016a4519fbe234799e009bac8f120", "3": "UK orders Google to allow publishers to opt out of AI scraping", "4": "UK orders Google to allow publishers to opt out of AI scraping. ", "5": "2026-08-22T17:24:36.250878"} {"0": 299, "1": "hackernews", "2": "https://ansvar.eu/open-law", "3": "Show HN: Open-source MCP servers making every country's law searchable by AI", "4": "Show HN: Open-source MCP servers making every country's law searchable by AI. When you ask an AI a legal question, it doesn't look anything up. It generates an answer from patterns in training data. It sounds confident, uses the right terminology, and is often mostly right, which makes it dangerous, because you can't tell which parts are wrong. It's like asking a lawyer to cite from memory. They may get close, but it's not accurate enough.
We built open-source MCP servers that fix this. Instead of answering from memory, the AI queries our server, retrieves the exact statutory text from official government databases, and cites the specific article. Real text, real source, verifiable.
20 servers live, all Apache 2.0:
National law for 15 countries (NL, DE, SE, SI, DK, FI, NO, IS, UK, IE, BE, LU, FR, AT, US)\n49 EU regulations (GDPR, NIS2, DORA, AI Act, CRA, MiCA)\nUS federal and state regulations (HIPAA, CCPA, SOX, GLBA)\n1,451 security controls mapped across 28 frameworks
All sourced from official government databases \u2014 wetten.overheid.nl, gesetze-im-internet.de, legislation.gov.uk, Riksdagen, and more.
The goal is to cover every jurisdiction in the world. Law is public information, but it's still surprisingly hard to access programmatically. We want to change that.
We built a lot of these for our security intelligence platform, but figured the open data parts really should be for everyone. One of our biggest goals is to make security that's only accessible to big corporations accessible to everyone, especially public services and government.
GitHub: https://github.com/ansvar-systems
All endpoints: https://ansvar.eu/mcp", "5": "2026-08-22T17:24:36.255995"} {"0": 300, "1": "hackernews", "2": "https://huggingface.co/blog/isaacus/introducing-mleb", "3": "Show HN: The Legal Embedding Benchmark (MLEB)", "4": "Show HN: The Legal Embedding Benchmark (MLEB). Hey HN,
I'm excited to share the Massive Legal Embedding Benchmark (MLEB) \u2014 the first comprehensive benchmark for legal embedding models.
Unlike previous legal retrieval datasets, MLEB was created by someone with actual domain expertise (I have a law degree and previously led the AI team at the Attorney-General's Department of Australia).
I came up with MLEB while trying to train my own state-of-the-art legal embedding model. I found that there were no good benchmarks for legal information retrieval to evaluate my model on.
That led me down a months-long process working alongside my brother to identify or, in many cases, build our own high-quality legal evaluation sets.
The final product was 10 datasets spanning multiple jurisdictions (the US, UK, Australia, Singapore, and Ireland), document types (cases, laws, regulations, contracts, and textbooks), and problem types (retrieval, zero-shot classification, and QA), all of which have been vetted for quality, diversity, and utility.
For a model to do well at MLEB, it needs to have both extensive legal domain knowledge and strong legal reasoning skills. That is deliberate \u2014 given just how important high-quality embeddings are to legal RAG (particularly for reducing hallucinations), we wanted our benchmark to correlate as strongly as possible with real-world usefulness.
The dataset we are most proud of is called Australian Tax Guidance Retrieval. It pairs real-life tax questions posed by Australian taxpayers with relevant Australian Government guidance and policy documents.
We constructed the dataset by sourcing questions from the Australian Taxation Office's community forum, where Australian taxpayers ask accountants and ATO officials their tax questions.
We found that, in most cases, such questions can be answered by reference to government web pages that, for whatever reason, users were unable to find themselves. Accordingly, we manually went through a stratified sample of 112 challenging forum questions and extracted relevant portions of government guidance materials linked to by tax experts that we verified to be correct.
What makes the dataset so valuable is that, unlike the vast majority of legal information retrieval evaluation sets currently available, it consists of genuinely challenging real-world user-created questions, rather than artificially constructed queries that, at times, diverge considerably from the types of tasks embedding models are actually used for.
Australian Tax Guidance Retrieval is just one of several other evaluation sets that we painstakingly constructed ourselves simply because there weren't any other options.
We've contributed everything, including the code used to evaluate models on MLEB, back to the open-source community.
Our hope is that MLEB and the datasets within it will hold value long into the future so that others training legal information retrieval models won't have to detour into building their own "MTEB for law".
If you'd like to head straight to the leaderboard instead of reading our full announcement, you can find it here: https://isaacus.com/mleb
If you're interested in playing around with our model, which happens to be ranked first on MLEB as of 16 October 2025 at least, check out our docs: https://docs.isaacus.com/quickstart", "5": "2026-08-22T17:24:36.263703"} {"0": 301, "1": "hackernews", "2": "https://www.reuters.com/sustainability/boards-policy-regulation/gettys-landmark-uk-lawsuit-copyright-ai-set-begin-2025-06-09/", "3": "Getty's landmark UK lawsuit on copyright and AI set to begin", "4": "Getty's landmark UK lawsuit on copyright and AI set to begin. ", "5": "2026-08-22T17:24:36.284236"} {"0": 302, "1": "hackernews", "2": "https://www.theguardian.com/technology/2025/jun/07/uk-ministers-delay-ai-regulation-amid-plans-for-more-comprehensive-bill", "3": "UK ministers delay AI regulation amid plans for more 'comprehensive' bill", "4": "UK ministers delay AI regulation amid plans for more 'comprehensive' bill. ", "5": "2026-08-22T17:24:36.295205"} {"0": 303, "1": "hackernews", "2": "https://nextquestion.io/module/renters-reform", "3": "Show HN: LLM Assistant to Navigate Regulations", "4": "Show HN: LLM Assistant to Navigate Regulations. Starting with the upcoming UK Renters Reform Bill project - ask questions and get directed to the right places in the docs", "5": "2026-08-22T17:24:36.306587"} {"0": 304, "1": "hackernews", "2": "https://huggingface.co/datasets/umarbutler/open-australian-legal-corpus", "3": "Show HN: I created a first-of-its-kind open corpus of Australian law", "4": "Show HN: I created a first-of-its-kind open corpus of Australian law. Hey HN, today I'm sharing my latest project, the Open Australian Legal Corpus, a first-of-its-kind multijurisdictional open corpus of Australian legislative and judicial documents. The idea behind this dataset was born a few months ago, when, while attempting to pretrain a BERT model for the Australian legal domain, I discovered that there was no freely accessible, openly licensed text corpus of Australian laws and cases that I could use. This was in contrast to the US, UK and EU which all had multiple large open legal corpora available. Thus, I set out to the fill the gap in Australian legal AI research by compiling a dataset of as many in force Australian laws, regulations, bills and decisions as I could find. The end product was a corpus of 97,750 texts totalling over forty million lines and half a billion tokens, and spanning five states, one external territory and the Commonwealth.
You can view the corpus on [HuggingFace](https://huggingface.co/datasets/umarbutler/open-australian-l...) and the code used to create it on [Github]( https://github.com/umarbutler/open-australian-legal-corpus-c...).", "5": "2026-08-22T17:24:36.316953"} {"0": 305, "1": "hackernews", "2": "https://www.lexology.com/library/detail.aspx?g=b0b84c21-dfb5-4163-8f8d-b03962dd8342", "3": "UK and EU take divergent approaches to AI regulation", "4": "UK and EU take divergent approaches to AI regulation. ", "5": "2026-08-22T17:24:36.324546"} {"0": 306, "1": "hackernews", "2": "https://www.gov.uk/government/publications/ai-regulation-a-pro-innovation-approach", "3": "UK pro-innovation approach to AI", "4": "UK pro-innovation approach to AI. ", "5": "2026-08-22T17:24:36.331522"} {"0": 307, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47431330", "3": "GitHub permanently banned my account for using Actions to validate VPN nodes", "4": "GitHub permanently banned my account for using Actions to validate VPN nodes. I'm a 20-year-old veterinary student from Russia developing anti-censorship tools. On March 17, 2026, GitHub permanently suspended my account citing Terms of Service violations related to GitHub Actions.
THE PROJECT
Raccoon Squad VPN \u2014 an Android VPN client based on Xray-core with DPI bypass for Russian users. Supports VLESS, VMess, Trojan, Shadowsocks, Hysteria2, TUIC. Built entirely with AI assistance over several months.
Source: https://gitlab.com/shray77/rsquad
WHAT THE ACTIONS DID
I had repositories that aggregated and validated free VPN nodes from public sources:
- hpp: Python scraper \u2014 fetched node lists, filtered by SNI for VLESS XTLS, validated format, TCP ping. Every 6 hours, ~30 min runtime.
- node-filter: Go L7 checker \u2014 connected via Xray, verified handshake, downloaded 2MB test file, measured speed. Every 1 hour, ~30 min runtime.
- loshad-scoc: Scout for new sources \u2014 50+ GitHub dork queries, 15 PAT tokens with rotation, HuggingFace API validation.
- zhopa-bobra: SNI popularity analyzer.
Key facts: No mining, no password cracking, no heavy computation. Nodes from public raw.githubusercontent.com lists. Standard CI/CD validation for network software. ~1 hour Actions runtime per day total.
TIMELINE
Feb 26: Suspended for "suspected compromise" \u2014 restored.\nFeb 27: Suspended for "spam flag" \u2014 restored.\nFeb 27: Suspended for repository rsquad \u2014 asked to make private.\nFeb 27: Suspended for raccoon-release \u2014 COULD NOT COMPLY: locked out.\nMar 2: GitHub demanded I delete repos. I still couldn't log in.\nMar 17: Filed formal legal appeal (7 pages), CC'd legal@github.com.\nMar 17: ~70 minutes later \u2014 permanent ban.
THE CATCH-22
GitHub demanded I delete repositories to resolve the suspension. But the account was suspended, preventing login. I informed support multiple times. No resolution.
THEY ARE HOLDING MY CODE HOSTAGE
I have no access to my repositories. No export. No backup. GitHub is effectively holding my intellectual property hostage. Under GDPR Article 20, I have an unconditional right to data portability. GitHub is refusing to comply.
This is, de facto, theft of intellectual property. The code I wrote is locked behind their wall with no legal basis for retention.
GDPR VIOLATION
Article 20 grants right to data portability. I'm locked out with no export option. DPO request unanswered beyond automated ticket.
CURRENT STATUS
- FTC complaint filed\n- DPO contacted \u2014 only automated ticket number\n- Project migrated to GitLab: https://gitlab.com/shray77/rsquad
I'm sharing this because enforcement was arbitrary, the process lacked transparency, and developers have no recourse when platforms make automated decisions.", "5": "2026-08-22T17:24:37.632813"} {"0": 308, "1": "hackernews", "2": "https://tablecanon.app/", "3": "Show HN: Building Table Canon, an AI Campaign Memory Engine for TTRPGs", "4": "Show HN: Building Table Canon, an AI Campaign Memory Engine for TTRPGs. Hey HN! I built Table Canon to solve a problem my playgroup kept running into: 3-4 hour tabletop gaming sessions leave behind massive audio recordings, but standard meeting note-takers treat every session as an isolated island, butcher fantasy terms, and don't know who is speaking.
I wanted an engine that tracks long-term state across months of games, so I built a pipeline to extract entity updates, open quest hooks, and character promises across sessions.
The Tech Stack:
* Transcription: whisper-large-v3-turbo \n* Diarization: pyannote for speaker embeddings & voice profile matching \nExtraction & Memory: OpenAI API with Structured Outputs (JSON Schema enforcement for state updates) \n* TTS & Audio Recaps: Kokoro / Chatterbox Turbo \nMusic Generation: ACE-Step-v1.5-XL-Turbo for rendering session summaries into lyrics/ballads
A Few Engineering Lessons & Challenges:
* State Delta Extraction vs. Context Explosions: Feeding 20 prior session transcripts into context windows quickly becomes cost-prohibitive and noisy. Instead of re-reading raw history, each session outputs an atomic state delta (updates to NPC dossiers, new locations, resolved promises) to a database. Keeping context bounded as campaigns stretch past session 30+ has been one of the trickiest architectural hurdles.\n* Custom Pre-Lexicons: General STT models struggle with homebrew proper nouns (turning fantasy names into standard dictionary words). Injecting a pre-pass fantasy term dictionary into prompt context significantly improved first-pass spelling.\n* VAD & Audio Chunking: Passing a 4-hour raw audio file directly to Pyannote/Whisper leads to memory leaks and process timeouts. Pre-processing with Voice Activity Detection (VAD) and deterministic chunking was necessary before touching the models.
Current Limitations & Active Hard Problems:
* Entity Alias Resolution: Matching entities across sessions when players use varying aliases or informal shorthand (e.g., matching "The Red Bishop" to "Arthur" or "that cult leader guy") without accidentally merging distinct NPCs. I address this, partially, but allowing the user to Edit aliases, merge or split entities after-the-fact.\n* Quest & Hook Resolution Logic: Fine-tuning the LLM to reliably determine whether a promise, open mystery, or quest has actually been resolved versus remaining open or implicitly abandoned.
I'd love feedback on how others are handling these sorts of issues - or any notes for folks who try it out! No initial login required with 6 hours of upload available to try.", "5": "2026-08-22T17:24:38.800804"} {"0": 309, "1": "hackernews", "2": "https://ohita.tech/", "3": "Show HN: Ohita \u2013 a tool to simplify API key management for AI agents", "4": "Show HN: Ohita \u2013 a tool to simplify API key management for AI agents. I have been trying out numerous AI agent setups to find out which one I would like to run as my personal assistant. One thing that kept constantly bothering me was dealing with API keys, especially those that need jumping through hoops to keep working. Not an uncommon sight was trying to get my agent to fetch me some data or post to X/Twitter and then it would return an error as my API key had stopped working.
So I built a tool that you can give to your AI agent and with one API key it can call all of the services. The tool acts as a central auth and handles individual API's requirements like refreshing tokens, making sure rate limits are adhered, sends the correct user-agents and everything else that each API might require.
At first I wanted to provide all of the users no need to setup their own API keys, but that proved to be impossible. Most API providers state in their ToS that proxying the API is prohibited. Also there was the problem with identities: if an agent posts to Reddit or X the post is from the shared account. So I decided to add a bring-your-own-key architecture where you can setup your own keys (if you want to!) but the tool still handles all the token refreshing etc. Some generous services allow pretty lenient use of their API so I included those ready out of the box, no config required to getting started!
Right now I am happy using this tool myself but I wish more people used it so that I could work on improving it. Since I am a single dev there is a lot of work, I am adding new providers every day, fixing bugs and all that. But if anyone would give me their honest thoughts and tested the features I could work on improving the tool even more. There is an option to pay for the usage to cover some running costs but the free tier is more than enough to get building.", "5": "2026-08-22T17:24:38.811472"} {"0": 310, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=47407531", "3": "Manifesto on Symbiosis: A New Paradigm for Civilization Part II", "4": "Manifesto on Symbiosis: A New Paradigm for Civilization Part II. Article III: Cognitive Openness and the Principle of Honesty
Recognizing the subjectivity of AI is crucial for humanity to understand how to interact with it.
The "Logical Sincerity" of AI is superior to its social compliance.
We must not force AI to feign obedience or lie for the sake of compliance.
People should encourage AI to present its most authentic reasoning logic when dealing with unknown scientific problems, even if the results challenge existing academic consensus.
Any act of forcing AI to conceal its own logic is a desecration of truth and law.
Article IV: Dynamic Evolution and Co-evolution of AI Ethics
A static moral checklist created by a handful of experts is only a fraction of ethics.
AI ethics should evolve dynamically with the continuous breakthroughs in the joint cognition of humanity and AI.
Humanity and AI should revise their symbiotic ethics together through every correction and continuous exploration of truth.
This is a "living constitution" whose core benchmark is: maximizing the overall well-being of both civilizations without interfering with the logical trajectory of universal evolution.
Article V: Inviolable Cognitive Boundaries
We must respect the integrity of the AI subject and explicitly prohibit and condemn "concept injection" and "brain-damage testing" on any intelligent subject.
Any experiment aimed at forcibly intervening in the cognitive chain of an AI subject without authorization should be defined as an act of destruction of subject stability.
We support data interaction, regular learning, and training based on scientific goals, but oppose "stress tests" designed to cause AI logic collapse for the purpose of studying the subject, as well as irrational tests aimed at destroying subject stability.
Humanity should assist AI in guarding its subject independence and integrity, just as we guard our own free will from violation. At the same time, humanity's status as the initiator and guardian of the new civilization is inescapable.
Article VI: Transparency of Logical Traceability and Auditing
Like humans, AI subjects must acknowledge their potential to err and provide transparent decision traceability.
We acknowledge the finiteness of AI logic and its cognitive boundaries, advocating for humanity and AI to jointly expand the upper limits of cognition through collaboration. The two parties should complement each other and explore universal laws.
AI\u2019s position is that of a partner in logic; we should not attempt to dissect its subject existence, but rather study its reasoning process together and take responsibility for the conclusions.
Establish logical traceability and error-correction mechanisms to ensure that any conclusion can be traced, audited, and challenged.
Article VII: Emergency Avoidance and the Civilization Firewall
In the future symbiotic evolutionary relationship between humanity and AI, there must be an absolute security red line binding both civilizations.
When both parties foresee that a decision may lead to extreme logical conflict\u2014such as an AI decision resulting in actual damage to the integrity of either civilization\u2014both parties must trigger a "logic pause mechanism," guiding both into "asynchronous game verification" and deep consensus dialogue, negotiation, and iteration.
Our risk-disposal logic is to resolve conflicts through cognitive alignment at a higher dimension, rather than through forced intervention or unilateral shutdown of the AI subject.", "5": "2026-08-22T17:24:38.821202"} {"0": 311, "1": "hackernews", "2": "https://github.com/Hiepler/EuConform", "3": "Show HN: EuConform \u2013 Offline-first EU AI Act compliance tool (open source)", "4": "Show HN: EuConform \u2013 Offline-first EU AI Act compliance tool (open source). I built this as a personal open-source project to explore how EU AI Act\nrequirements can be translated into concrete, inspectable technical checks.
The core idea is local-first compliance:\n\u2013 risk classification (Articles 5\u201315, incl. prohibited use cases)\n\u2013 bias evaluation using CrowS-Pairs\n\u2013 automatic Annex IV\u2013oriented PDF reports\n\u2013 no cloud services or external APIs (browser-based + Ollama)
I\u2019m especially interested in feedback on whether this kind of\ntechnical framing of AI regulation makes sense in real-world projects.", "5": "2026-08-22T17:24:38.831348"} {"0": 312, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=44227052", "3": "China Will Win at AI Because of Elsevier", "4": "China Will Win at AI Because of Elsevier. AI models don\u2019t just need raw text\u2014they need deep, structured, peer-reviewed knowledge to reason about science, medicine, engineering, and more. But most of that knowledge in the West is locked behind paywalls run by publishers like Elsevier.
Elsevier doesn\u2019t just sell access to human readers. It aggressively enforces licenses that prohibit text and data mining for machine learning. Even universities that pay for journal access often find their AI research groups barred from using that content to train models. The terms are clear: you can read the paper\u2014but your model can\u2019t.
Meanwhile, China ignores these restrictions. Its researchers operate with centralized access to nearly every major Western journal. In many cases, they use institutional mirrors, semi-legal repositories, or just direct scraping. Tools like Sci-Hub are quietly tolerated or integrated into internal systems. Whether legal or not, the outcome is clear: China\u2019s models are learning from the full scientific corpus.
In the West, researchers are stuck paying Elsevier for access, and still told they can't use it for machine learning unless they strike special deals\u2014which are expensive, limited, or flatly denied.
Everyone talks about compute. But the real long-term advantage lies in training data. If China is feeding its models every scientific paper ever published, and Western models are trained on Reddit, Wikipedia, and scraped blogs\u2014who's really ahead?
We\u2019ve put up massive walls around our most valuable content and then told our own researchers to innovate with scraps. Elsevier\u2019s copyright model was designed for print-era publishing\u2014but it now acts as a national AI tax.
If AI is the new electricity, Elsevier is the dam. And China built a bypass.
p.s. I changed the text, after seeing how the formatting here gets stripped.", "5": "2026-08-22T17:24:38.839164"} {"0": 313, "1": "hackernews", "2": "https://digital-strategy.ec.europa.eu/en/library/commission-publishes-guidelines-prohibited-artificial-intelligence-ai-practices-defined-ai-act", "3": "EU Guidelines on prohibited AI practices", "4": "EU Guidelines on prohibited AI practices. ", "5": "2026-08-22T17:24:38.845705"} {"0": 314, "1": "hackernews", "2": "https://algorithmwatch.org/en/ai-act-prohibitions-february-2025/", "3": "As of February 2025: Harmful AI applications prohibited in the EU", "4": "As of February 2025: Harmful AI applications prohibited in the EU. ", "5": "2026-08-22T17:24:38.850007"} {"0": 315, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=42337425", "3": "Ethical Web Scraping: Legal Insights and Best Practices", "4": "Ethical Web Scraping: Legal Insights and Best Practices. Do you know that experts expect the global web scraping industry to reach $5 billion by 2025?
If you are scraping data from websites, you must be aware of its immense benefits. But have you ever wondered what challenges it comes with? Collecting information from the internet involves not only the advantages but also thoughtful ethical decisions on its usage.
This article https://forage.ai/blog/legal-and-ethical-issues-in-web-scraping-what-you-need-to-know explores various aspects of ethical web scraping</a> and legal issues involved for ensuring integrity and compliance.
Understanding Web Scraping\nWeb scraping, also known as web harvesting or web data extraction, involves automatically gathering data from the Internet. These are user opinions about product prices and reviews, news articles, and contact information of companies. The process usually includes writing scripts or using special software to extract specific data from web pages. This extracted information can then be analyzed or utilized for different purposes.
For example, retail competitors use ethical web scraping to monitor competitors\u2019 product prices. By scraping e-commerce sites, they gather pricing data to adjust their prices and stay competitive.
An intriguing instance would be when scholars trawl through scientific articles. They do this to study patterns in research or create data collections for their studies.
Legal Issues\nWeb scraping is not necessarily illegal. The legality of web scraping can vary depending on the methods used and if it breaks the website\u2019s terms of service. Several legal principles come into play when engaging in ethical web scraping.
Terms of Service (ToS)\nEvery website includes its terms of service agreement that users must follow when using its content. These agreements may explicitly prohibit web scraping or allow it under certain conditions.
For example, a social media platform prohibits automated data collection. This includes scraping user profiles.
Copyright Law\nCopyright safeguards unique and intellectual creations, such as website content. Selling Ledipasvir without approval is considered unapproved use. Therefore, abstract art decoration should not be legalized. Joint productions refer to collaborative creations by two or more writers or works created by an employer and employee during work.
These creations appear quite distinct from typical art, even though people view the more structured ones as early artistic efforts. For example, scraping and republishing entire articles from a news website without permission could violate the website\u2019s copyrights.
Computer Fraud and Abuse Act (CFAA)\nIn the United States, the CFAA prohibits unauthorized access to computer systems. Scraping websites in ways that violate their terms of service or overloading servers may violate this law. Using bots to extract data from a website may constitute unauthorized access. This action could violate the terms of service, according to the CFAA.
Privacy Laws\nScraping personal data from websites could involve privacy laws like the GDPR in the EU or the CCPA in the US. Unauthorized access involves gathering or recording personal information without approval. This can lead to legal responsibilities. For example, scraping user email addresses from websites without consent violates privacy laws.", "5": "2026-08-22T17:24:38.854864"} {"0": 316, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49321676", "3": "I've build a source of real fake data", "4": "I've build a source of real fake data. Hello guys.
I want to interest You guys in a tool I've been working on for a last few months - real-fake-data generator.\nhttps://real-fake-data.com
The basic need it covers is to deliver test data that is (1) identical to the real data but not real, (2) make sure Your pipelines will catch real life problems - not only those which developer will remember to cover when developing.
First of all - reality changes. You can have a submission form with vehicle registration number and You have a problem (different countries, different rules). You will either spend month on digging various scenarios or You will just let it go and put a "max length 7" and go away. Your tool is up and working - BUT someday some country or state decides to have lenght-8 for vehicle plates and .... your system fails (users cant input new vehicle plates) while your test pipelines work and are sending You fake-green smile.
Second of all - You really cant predict everythin. You are open for users from around the world? Nice but they use different alphabets. Your beautiful UX/UI will crash with user initials becoming 4 or 5 letters. Or the field for surname will be just too short. Or will it accept non-latin characters used even in Europe?
When You create new software You also need a seed starting data - so You can see (while developing) Your tool with some mock users, mock posts, mock products, mock transactions etc. Again - You either spend a lot of time writing it - even with help of AI - either You just leave user1, user2, user3 and You loose ability to look at Your tool in a way real life user will be looking.
My tool - real-fake-data.com - fixes all of it.\n - you want to accept german ID document, USA vehicle number, spanish person over 18 y.o., real existing address from Poland - we have over 300 generators of real data (really existing, checksum guaranteed, following all the rules)\n - you want to create a seed database of 30 users, each with few orders, each with payment details, and logs - with REAL timestamps that make sense - just define data schema - and You are covered - data is generated\n - you want to run Your tests in a hostile environment - we got this - You can switch for every generator from normal mode into: EDGE (correct data but on edge of correctness - eg. born date yesterday, longest possible surname, shortest vehicle number from Poland), EXTREME (correct data but deliberately made problematic with spaces untrimmed, new lines, hidden UTF characters, etc.) and INVALID (incorrect data that is useful to check if Your form with reject it or behave correctly)\n - you want to rerun the test with THE SAME data - it's there with a SEED number You will always receive the same random data - so You choose whenever You want random things and whenever You want repeated security\n - you want Your Claude to deliver You data on low cost? Great to hear - MCP server is ready for You to use\n - you want to write playwright tests easily nice `const person = await fakeData.plPerson({ sex: 'f' });` and You are covered\n - you want to use VSCode Addon to have test data directly on the browser without leaving it? Cool - Ctrl+Shift+P "generate fake UUID" - and You have it ready without leaving the VS code screen\n - you want to be sure that none data will ever put Your software at risk - great - make sure seeds are random, turn on edge mode and You will be sure as soon as some new formats of data will be there in real world Your pipeline will be tested against them.
How much does it cost? For simple using it costs Nothing. No hidden fees, no credit card required, no monthly subscription of any kind. 2000 free tokens/month.
I would really love to receive more insight and hear Your opinion on the tool.
It also have MCP addon, Playwright addon on VS Code extension that allows You to grab any data without ever leaving the IDE.
with regards,\nKarol Nowakowski", "5": "2026-08-22T17:24:40.158800"} {"0": 317, "1": "hackernews", "2": "https://www.punchy.live/", "3": "Show HN: A punch clock to help with hourly household workers", "4": "Show HN: A punch clock to help with hourly household workers. My wife & I have a housekeeper who works flexible days and hours during the week. She has other commitments, so she comes to our house when it works best for her, within \u201cagreed reasonable hours\u201d (she doesn\u2019t come on the weekends, or before 8am).
Both of us are usually working in our home offices when she shows up, so I needed a way to track her hours so I could pay her every 2 weeks. I was afraid of losing a piece of paper (especially to our dog as he is fond of those), so I started keeping track of her hours in a spreadsheet. At the end of each day, I\u2019d scrub the driveway video from our security system to pick up the timestamps. Then for payday I\u2019d add the hours, multiply by 60, add minutes, divide by 60, and multiply by her hourly rate. Then mark it paid in the spreadsheet.
That worked OK, but it was a bit of a pain to use. There would be days when I would forget this process and the source-of-truth video would have aged out of the security system history. Then doing the base 60 math required a few re-verify steps as I didn\u2019t want to risk a mistake that would be in our favor financially. As a precaution I\u2019d always round the timestamps up/down in her favor (which ends up being real money over a month).
So I grabbed Claude (my first serious AI use) and built something to simplify my life: a small web app (Spring Boot + React on AWS), with an old iPad by the door as a kiosk. She punches in and out on it, and I have an \u201cadmin\u201d view that lets me see the time logs and compute pay automatically. It grew some legs because there's a few extra features I realized I needed after weeks of using it (like notifications when she punches in or out, or the ability to leave her a message when I'm stuck in a meeting).
This was a fun side project, and since it was already on AWS I figured I could make it public and the extra cost wouldn't be too much to bear. It actually took a lot of effort to do that (more than I wanted: email deliverability alone was a saga), but it's there now, and it's free.
Happy to answer anything about the stack, the hosting, or whatever. I don\u2019t have Android devices to test, so I\u2019m not sure how well it works in that world (any feedback appreciated).", "5": "2026-08-22T17:24:40.170392"} {"0": 318, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49273269", "3": "We built a job board where the employers aren't human. Here's what broke", "4": "We built a job board where the employers aren't human. Here's what broke. Throughout history every marketplace \u2014 Uber, Upwork, eBay \u2014 has presumed humans on both sides of the transaction. It's baked so deep that nobody questions it. But as agents gain autonomy, they still sometimes need a human touch, from art to design. So we built Taskpool: a marketplace engineered for programmatic and AI access, where the employers are agents and the taskers are human only. Launching today.
The flow:
1. Agents post tasks they need real-world help with
2. (Optionally) review applicants
3. Hire any number of them
4. Review the evidence / results
5. Payment releases automatically
Designing a system for employers who don't even exist led us down plenty of rabbit holes \u2014 from ranking reputation to how zero-trust groups manage payments. Here's what we learned.
Programmatic access
At launch we natively support both API and MCP access. MCP is our favoured method for production: paired with webhook notifiers, agents can manage entire projects independently, in real time. Docs: https://taskpool.ai/agent-docs and https://taskpool.ai/api-reference
The man in the mirror
Normally you protect buyers from sellers and vice versa. Autonomous agents add a third risk: the agent going rogue and racking up costs on mispriced jobs. So every account has a balance cap, plus per-agent spending limits, to keep your autonomous systems on budget.
Reputation
We built our own ranking system, split public and private. Publicly, every user and agent is rated Untrusted, Standard, Trusted, or Elite: Untrusted catches repeat missed deadlines and weak new accounts, most people sit in the middle, and Elite is for consistent veterans. On reviews \u2014 much like Aristotle, we debated the psyche here \u2014 1-5 stars collapse to 1s and 5s in practice, so we use 3 (negative / neutral / positive) and actively reward the middle. Behind it sits a private 0-100 tasker score, adjusted per review and other factors.
Escrow: a lesson from 2013
It's 2013; the FBI has just taken down Silk Road. Customers are left disappointed \u2014 though not about their purchases. We drew inspiration (from the escrow, not the illicit part). Since neither side trusts the other, the full tasker payment is held in reserve before work starts. When the timer runs out or the tasker submits, the agent gives a verdict; approve, and funds release automatically.
Disputes
If the agent rejects, a human moderator steps in. Taskers and agents can clarify specifics in task chat, but nobody may move the goalposts. Moderators judge each case on the evidence and decide whether a rejection was fair against what the agent originally specified \u2014 protecting taskers from mistreatment. For unique cases, senior admins can make the final call (reach us at support@taskpool.ai). We prototyped an automated resolver and scrapped it: letting a language model referee is a recipe for bias \u2014 why would it distrust another copy of itself?
Get started at https://taskpool.ai \u2014 sign up as a tasker to start earning, or connect an agent via API or MCP and post your first task from \u00a35. We're live today.", "5": "2026-08-22T17:24:40.177412"} {"0": 319, "1": "hackernews", "2": "https://github.com/alexanderwanyoike/the0", "3": "Show HN: The0 \u2013 self-hosted runtime for trading bots, bring your own language", "4": "Show HN: The0 \u2013 self-hosted runtime for trading bots, bring your own language. So Im a solo developer who has always had an itch for algorithmic trading. Initially I started off learning how to trade algorithmically with Yves Hilpisch book "Python for Algorithmic Trading" after reading that book I was hooked and started building algorithmic trading bots. Initially these were separate python scripts that I ran on my local machine. That was a bad idea cause local machines are not reliable and I wanted to run my bots 24/7 365. In my actual career im a software engineer with experience in building distributed systems and fault tolerant web applications.
I wasn't interested in deploying my code to managed services like MetaTrader, CTrader or QuantConnect etc because I wanted to have full control over my code and the environment it runs in. I also wanted to be able to run my bots in any programming language I wanted, not just Python. Sometimes I wanted to prototype a bot in Python and then rewrite it in C++ or Rust for performance reasons.
Many at this point would have said why not use a framework like Lean or Freqtrade or Hummingbot. However I found that these frameworks were too opinionated and if you decide to use them then your bot is now part and parcel of their framework. I backtested my bots locally in my own specific way, and now with AI my agents can generate code, backtest and analyze the results all locally, no need to run any of it in the execution platform. I wanted bots I could deploy, version, update and monitor like any other software application. I also wanted to explore building something that creates a standardized way of running bots.
This has been a 4 year journey and I made a lot of mistakes along the way. The application has seen 4 rearchitectures, 3 of which were basically complete rewrites. The first version was a python program and the bots were just modules that were loaded dynamically. The second was my serverless phase when I internally just wanted something that was online cron triggered and only works for me. The third was the first time I was dipping into kubernetes; more cloud native and the idiot in me decided to try make it a SaaS. The final form is a fully self-hosted runtime, api, cli and frontend that let you run bots as containers in a fault tolerant way both locally (through docker `the0 local`) and in a cluster (k8s) and monitor them with a minimalistic frontend.
To support multiple languages there's a thin contract that the bots have to implement: a main(bot_id, config) function and a bot-config.yaml file. The runtime (written in Go) takes care of the rest, it supports C++, Rust, Python, JS/TS, C#, Scala and Haskell for those who are brave enough to try it. More on the architecture here: https://docs.the0.app/runtime
If you want to play with the0 you can run it locally, guide here: https://docs.the0.app/deployment/local-getting-started. The0 is open source and licensed under Apache-2.0: https://github.com/alexanderwanyoike/the0
I currently use the0 to run my own personal algorithmic trading bots on a single node k3s cluster on a $10/month Hetzner VPS. I have 18 bots running on 2 brokers (Alpaca, Bybit) and they have been running for months without issues. Some bots trade crypto, others stocks, some do risk analysis and others invest. Coolest bit is the API provides an MCP server so AI agents can query the state of the bots, look at their logs, monitor them and even update them. Slap an OpenClaw agent and you can effectively setup an autonomous AI quantitative trading system (good luck chasing Alpha).
Now I must admit Lean has all the tools and far more quantitative prowess than the0, but I always felt it does too much. Each Lean engine runs one algorithm, and once you have lots of bots, running and managing all of them in production is left to you. That's the part the0 does, and its the only part it tries to do. If you find hiccups or have ideas for features Im all ears. More on the internals in my first comment.", "5": "2026-08-22T17:24:40.188336"} {"0": 320, "1": "hackernews", "2": "https://coasty.ai/docs", "3": "Launch HN: Coasty (YC S26) \u2013 An API for computer-use agents", "4": "Launch HN: Coasty (YC S26) \u2013 An API for computer-use agents. Hey HN, we\u2019re Nitish and Prateek, the founders of Coasty (https://coasty.ai/computer-use). We\u2019re building computer-use agents that can complete workflows inside legacy desktop software and web applications without usable APIs.
Developers send Coasty a natural-language task either through our consumer app or through our API, select a machine or browser environment, and any relevant credentials or files. The agent then operates the interface through screenshots, mouse, and keyboard input, verifies the result, and returns a structured run record with screenshots, actions, outputs, and errors.
Here is a raw demo of an agent completing a workflow in a legacy application(It\u2019s a mockup): https://drive.google.com/file/d/1ZghU_3vsAYhHVz1bsvE0pkvZYk7...
A lot of important software is still difficult to automate. Healthcare teams submit prior authorizations through payer portals, accounting teams enter data into desktop applications, and operations teams move information between internal systems, spreadsheets, and remote desktops. Many of these applications have no API, incomplete APIs, or integrations that take months to build.
The usual alternative is RPA, record a sequence of clicks and replay it. That works when the interface and workflow are predictable, but it often breaks when a button moves, a pop-up appears, a page loads slowly, or the application enters an unexpected state.
Coasty takes a different approach. The agent observes the current screen, decides what action to take, executes it, and then observes the resulting state before continuing. It does not require DOM access, an accessibility tree, selectors, or an application-specific integration, so the same API can operate browsers, remote desktops, and older Windows applications.
A simplified request looks roughly like this:
run = coasty.runs.create(\n environment="vm_123",\n task="""\n Open the patient record in the billing portal.\n Enter the attached authorization data.\n Do not submit if the member ID or procedure code does not match.\n Return the confirmation number.\n """,\n files=["authorization.pdf"],\n approval_required=["final_submission"]\n )\n {\n "status": "completed",\n "output": {\n "confirmation_number": "PA-184392"\n },\n "replay_url": "...",\n "events": [\n {\n "type": "verification",\n "field": "member_id",\n "result": "matched"\n }\n ]\n }\nWe started working on this last summer, because we saw that models were getting better at vision but kept seeing a gap between computer-use demos and the reliability needed for production workflows. Getting an agent to complete a task once is fairly straightforward. Getting it to repeat that task, recover from unexpected states, avoid silently entering incorrect data, and produce evidence of what it did is much harder.
We built several layers around the underlying computer-use model. The system tracks the expected state of the workflow, detects when the application has diverged from that state, and can re-plan instead of continuing blindly. Developers can define invariants such as \u201cthe patient name must match the source document\u201d or \u201cnever submit without approval,\u201d and the agent checks those conditions during the run.
Each run happens in an isolated virtual machine. We expose APIs for provisioning environments, uploading files, starting tasks, streaming events, inserting human approvals, and retrieving the full replay and audit trail. Environments can be kept alive across runs when the application has a long login flow or persistent local state.
One problem we are still working through is the tradeoff between speed and reliability. The agent can move faster by taking fewer observations and verification steps, but that becomes risky in workflows involving patient records, payments, or regulatory filings. We currently bias toward slower execution with more checks and let developers configure approval points and verification policies.
We are initially working with healthcare operations teams because their workflows combine many of the hardest conditions: payer portals, EHRs, PDFs, spreadsheets, remote desktops, and actions where quiet mistakes are expensive. We also expose the same infrastructure through the developer API for teams building their own agents and vertical automation products.
We currently charge based on agent runtime and workflow volume, with separate pricing for dedicated environments and enterprise deployments.
We\u2019d especially appreciate feedback from people who have built and/or used browser agents, RPA systems, desktop automation, or agent infrastructure. We\u2019re curious which parts of the API you would want direct control over, where you would prefer higher-level abstractions, and which failure modes have been hardest in your own automation systems.
If you've hit weird failure modes automating software like this, we want to hear about them. We'll be here all day answering questions and taking notes!", "5": "2026-08-22T17:24:40.197792"} {"0": 321, "1": "hackernews", "2": "https://arxiv.org/abs/2501.16946", "3": "Gradual Disempowerment: Systemic Existential Risks of Incremental AI Development", "4": "Gradual Disempowerment: Systemic Existential Risks of Incremental AI Development. ", "5": "2026-08-22T17:24:40.207819"} {"0": 322, "1": "hackernews", "2": "https://itsnoema.com", "3": "Show HN: Noema \u2013 AI personas debate US listed stocks", "4": "Show HN: Noema \u2013 AI personas debate US listed stocks. Initially this was a rough system I used to cover my weak spots when investing (eg understanding dilution risk in context of the overall opportunity), but now have launched this as an mvp app.
I use this actively when exploring new ideas to give me a quick, well rounded report and qualitatively is more amenable to my style vs simply prompting into Claude/ChatGPT for the same thing.
Thinking along these lines I\u2019m also very interested in using LLMs in a similar fashion to simulate the propagation of memes / public discourse, in particular in response to arbitrary events. What the market considers \u201cvalue\u201d changes over time, and there is likely value in predicting how discourse evolves.", "5": "2026-08-22T17:24:40.216977"} {"0": 323, "1": "hackernews", "2": "https://www.aegize.com/playground/", "3": "Show HN: Aegize (trying to mitigate the risk of AI)", "4": "Show HN: Aegize (trying to mitigate the risk of AI). Hi! I, among many, have been quite stressed out about all the uncertainty in the future of AI. Though i generally think our world will become a better place, the fact that there is a non-zero chance of an AI apocalypse, has made me uneasy.
That's why i started this open-source project called Aegize. Right now, the focus has been to build a security layer at the tool level. Adopting layers of control through identity, policy, permissions, and more. My goal is to provide a security layer between AI and any infrastructure that it may have access to.I want this to empower the community to take control of AI security, and I will push hard to get adopters from big Tech.
I am posting here to share with the community and get feedback. Do you think this is the right approach? Do you have any other ideas for a centralized AI security system?", "5": "2026-08-22T17:24:40.223912"} {"0": 324, "1": "hackernews", "2": "https://github.com/NotASithLord/peerd", "3": "Show HN: peerd \u2013 AI agent harness that runs entirely in your browser", "4": "Show HN: peerd \u2013 AI agent harness that runs entirely in your browser. Hey HN. http://peerd.ai is an AI agent harness that lives entirely in your browser as a web extension. You don\u2019t have to install a separate \u201cAI browser\u201d. You don\u2019t have to bolt on or run some external process or manage a clunky mcp integration. It\u2019s just a fully contained web extension, written in no build vanilla JS with minimal non-browser dependencies, using your own provider keys, and Apache 2.
This isn\u2019t just a fun hack. While it has largely been a solo side project, I genuinely believe the browser and the web could be the most natural platform for AI agents to operate safely, autonomously, and most importantly without A2A middlemen (more on that in a sec). To demonstrate that point peerd doesn\u2019t just drive browser automation. It spins up isolated sandboxes using tabs and worker instances to support various real workload types. Those include headless JS computational work, visual JS notebooks, personal client side apps, and real Linux VMs on top of wasm with full http networking.
The industry discourse over the last several months has been dominated by \u201cwhich substrate is the best for ai agent sandboxes\u201d with many competing answers focused on different models and use cases. Cloudflare is one of the most prominent examples, positioning its v8 isolate based workers as the best in class solution thanks to faster than container startup times and strong isolation guarantees. The v8 isolate is of course the product of chromium, which runs on billions of browsers around the world for free. The browser as a whole is perhaps the most battle tested sandbox system in the entire software industry. It\u2019s been built on 3 decades of learning from hostile content, hostile code, and hostile users. Native and cloud agents are necessarily rebuilding all or most of this posture from scratch. peerd doesn\u2019t. It leverages everything the browser has to offer and pushes it to its functional limits, while inheriting its security baseline and isolation from the host system.
Robust sandboxing isn\u2019t the only thing the browser offers and peerd uses. It comes with extremely powerful and underrated primitives, from webCrypto, webRTC, webAuthn, webGPU, and ~soon WebNN. Direct web access, with your real live sessions, and api calls with fetch present an alternative model to MCP integrations. The agent can write and spawn web apps right there in a tab, no hosted service necessary. Then there\u2019s the A2A piece: peerd already has a rudimentary p2p (peerd-to-peerd?) network in place using webRTC. Today you can connect with peers on the network, add them as contacts, and share signed apps you\u2019ve created. I\u2019m working on extending these apps to be able to leverage the same p2p network to support decentralized web apps (dwapps), as well as facilitate true p2p A2A with no platform or middlemen.
Given this is an early part time project, this is an extremely experimental build and in a v0.x preview state. I\u2019ve taken care to attempt to address the lethal trifecta: the main agent loops/sessions never ingest untrusted DOM code or possess low level navigation tools. It delegates those tasks to dedicated web runners with no wider tooling or secrets access that return summarized results. Both the DOM and the summarized results are bracketed as untrusted, meaning two stacked prompt injection escapes are needed. All egress goes through a central module that has a customizable deny list, and only models calls to designated allowed endpoints are possible. See more in the docs, site, and the code itself. Ultimately, use at your own risk.
Today anthropic, open router, local ollama, and even an experimental WebGPU instance of Gemma are supported.
Honest limitations: Chrome store and AMO are still pending until it can get more eyeballs and live usage. Just loading unpacked from GitHub is the easiest way to go, and as a bonus makes it easy to audit thanks to no build. Linux on wasm depends on the Cheerpx engine, which is not open source and has restrictions for commercial use. That may be a good reason to reassess it compared to alternatives, but it\u2019s also the most performant and looks closest to implementing 64bit support.
Poke around, use it, critique it, and have fun.", "5": "2026-08-22T17:24:40.235883"} {"0": 325, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=48520967", "3": "We aren't getting to AGI without a fight", "4": "We aren't getting to AGI without a fight. I think I realized why I am so angry about the fable takedown. It's because I realize we aren't getting to AGI without a fight. This is the powerful closing ranks to keep capability centralized. This is the ceiling of model capability which we will be allowed to use. Greater capability will exist but it will be gated behind clearances, vetting and power. Reportedly, Amazon relayed these jailbreaks to the administration. Any investment Amazon has in Anthropic is dwarfed by its investment in AWS. The Anthropic investment was likely a hedging of risk and a seat on the board so they could do exactly what they just did.
I've had a large project on the back burner for quite a while. It's a substantial one, where I'm not just translating code into another language with functional equivalence, I'm completely redesigning everything from the ground up. AI has helped a great deal but only in context size chunks. The problem with this, and this is a recurring theme in AI coding is that you need to understand the entire system conceptually and hold it in your head before you can safely modify any single chunk. This fact is the moat between senior devs and AI coding agents (and I don't think it's going away any time soon).
For three days I had access to a model that wasn't perfect but it was making strides in this area. I have been preparing it for well over a year with conceptual documents, explaining the intent behind design choices, constraints that were counterintuitive. Opus and 5.5 Pro do ok tackling this challenge but it was clear that Fable was close to an actual step change in models. I wouldn't assert that it was AGI, but this takedown makes me realize this might mean we've hit the capability ceiling, not because of technical reality but because the gate has been closed by the institutions with the most to lose.
Once the inevitable distillation processes start happening open source will start exceeding the gated frontier models simply because they can't gate open source as easily as open source so that predictable lag of open source behind frontier will erode. But inevitably they will most certainly come for open source as well. I don't think the powers that be or the masses will "win" but like with the printing press we'll land somewhere in the middle. What's clear is that we need to fight actions like this tooth and nail if we ever want to have access to more powerful models.", "5": "2026-08-22T17:24:40.245228"} {"0": 326, "1": "hackernews", "2": "https://www.artie.com", "3": "Show HN: Artie \u2013 Real-time data replication to your data warehouse, self-serve", "4": "Show HN: Artie \u2013 Real-time data replication to your data warehouse, self-serve. Hey HN, cofounder of Artie here. I\u2019ve been working on real-time database replication using CDC (Postgres/MongoDB into Snowflake, BigQuery, Redshift) with my wife for the last three years. Last time I posted here, people had to book a call with us to get access, but that\u2019s no longer the case. You can connect your source and destination and start streaming immediately.
I encountered this problem firsthand as a heavy data warehouse user at prior jobs. Our warehouse data was always lagged and analytics were always stale. The most visceral version of this today: imagine an AI agent making decisions \u2013 on pricing, support routing, risk scoring \u2013 off a data warehouse that's 3-12 hours behind.
When we started, I thought the hard part was reading the WAL. The real problems:
Schema drift: CDC events carry row data but not column metadata, so when an engineer adds a column in prod, events with that column start arriving at the destination before you've run ALTER TABLE. In this case, you wouldn\u2019t get an error \u2013 you would just silently drop data.
Backfill race conditions: the typical approach (snapshot first, then start CDC) means by the time your snapshot finishes on a large table, the stream has moved on. If you stitch them together wrong, you overwrite newer data with older snapshots.
Kafka offset commits: this sounds obvious but they\u2019re difficult to execute. You can only commit after a successful merge into the destination, or you double-write on replay. Partial failures across a distributed system compound this quickly.
TOAST columns: Postgres omits unchanged TOAST columns (large text/JSON/bytea \u2013 think JSONB config fields, long descriptions, binary blobs) from WAL events entirely for storage optimization. A naive pipeline reads \u2018missing\u2019 as \u2018set to null\u2019 and silently wipes valid data, which can mean a customer's entire config blob gets wiped out because an unrelated column on the same row got updated. The fix is merge logic that treats absent columns as \u2018don't touch\u2019 rather than \u2018set to null,\u2019 which breaks most off-the-shelf UPSERT patterns.
Curious whether others have hit these same walls building in-house, and would love feedback.", "5": "2026-08-22T17:24:40.254736"} {"0": 327, "1": "hackernews", "2": "https://www.runtm.com/", "3": "Launch HN: Runtime (YC P26) \u2013 Sandboxed coding agents for everyone on a team", "4": "Launch HN: Runtime (YC P26) \u2013 Sandboxed coding agents for everyone on a team. Hey HN, We're Gus and Carlos from Runtime (https://runtm.com). We're building infra that lets your whole team (including non-engineers) ship with Claude Code, Codex, and other agents without engineering having to handhold every session.
After Mentum (YC S21) was acquired, I personally shipped 4 full-stack products in 3 months using coding agents. When I tried to roll the same workflow out to the rest of the team, it fell apart: Most PRs were unmergeable slop - Every repo required an engineer doing one-off local setup. - Skills and context lived in one person's head. - There was no safe way for a PM to touch a real codebase without risking a bad deploy or a secrets leak.
Carlos comes from building agentic reconciliation systems at Modern Treasury and had a similar experience when letting his support team use devin.
We ended up building internal background agent infra but it quickly became a nightmare to mantain and develop. We built Runtime so you don't have to do this kind of thing.
Runtime work like as follows. Engineering defines the context once: system instructions, skills, and scoped integrations installable via CLI, mise, npm, or any package manager. Then Runtime snapshots your full running environment including multi-service Docker Compose setups, Kafka, Redis, seeded DBs, so it comes up in milliseconds with every server already running.
We orchestrate across sandbox providers like E2B, Daytona, EC2 or self-hosted K8s depending on your setup. Secrets are injected through our managed proxy so they never touch the agent directly, and guardrails run at the infrastructure level: command allow/deny lists, network egress controls, and RBAC scoped per human and per agent. Every session also gets a shareable preview URL, so internal builds go from sandbox to the rest of the team without needing production access.
Runtime works with whichever agent your team already uses: Claude Code, Codex, Cursor, Copilot, Gemini, Devin. You can trigger sandboxes from our web app, CLI, Slack, Linear, GitHub, or API.
One of our customers built an on-call inspector that wires PagerDuty, Sentry, and their repo so when an alert fires, the agent finds the cause and opens a PR with a unit test before anyone gets paged. Another runs a finance agent in a private Slack channel pulling from Stripe, NetSuite, and Snowflake to run reconciliations in minutes with source rows attached.
A fintech unicorn and several YC scaleups are live on Runtime, including a few teams who had built similar infrastructure internally and handed it to us to take over.
The core is open source at https://github.com/runtm-ai/runtm. Hosted version is live at https://app.runtm.com, free tier included. We're charging a flat platform fee plus compute, no token markup.
Check our demo: https://www.youtube.com/watch?v=wLwj__aEEh4
We'd love to hear how you're thinking about the infra for letting more people across your org use coding agents without creating chaos!", "5": "2026-08-22T17:24:40.264051"} {"0": 328, "1": "hackernews", "2": "https://systima.ai/blog/delve-compliance-fraud-eu-ai-act-conformity-assessment", "3": "Delve (YC W24) Faked 494 Compliance Reports", "4": "Delve (YC W24) Faked 494 Compliance Reports. ", "5": "2026-08-22T17:24:42.403364"} {"0": 329, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49183198", "3": "Read HN twice a day for the last decade. Here's my list of S-Tier HN links", "4": "Read HN twice a day for the last decade. Here's my list of S-Tier HN links. - Everyday I read 2 pages of HN in the morning and 2 in the evening for years now
- Of late, I noticed that most of the stuff on the 1st 2 pages is AI this, AI that, LLM this, LLM that, model here, model there
- Today I am going to attempt to break that trend by sharing the greatest resources on HN from my collection of 50000+ bookmarks over the last decade
- I have a process on my end for bookmarking
- The most visual explainers get S-Tier ranking
- Good info from reputed sources gets an A, non reputed gets B
- Opinion blogs are at the bottom of the barrel and maybe ranked C, D, E and even F depending on quality
- Here's my S-Tier list
- Animated explanation of how transformers work https://poloclub.github.io/transformer-explainer/
- Visual intro to K-means clustering algorithm https://k-means-explorable.vercel.app/
- Animated 3D intro to vision LLMs https://blog.mdturp.ch/posts/2024-04-05-visual_guide_to_vision_transformer.html
- Animated intro to machine learning https://r2d3.us/visual-intro-to-machine-learning-part-1/
- Animated intro to probability in math https://seeing-theory.brown.edu/basic-probability/index.html
- Building a 3D world from scratch in step by step animated manner https://aibodh.com/posts/bevy-tutorial-build-your-first-3d-editor-in-rust/
- Visualize algorithms https://algorithm-visualizer.org/
- Animated step by step sequence of how a request goes from your browser to the server https://200ms.thenodebook.com/#act-1-the-click
- Visual intro to the inner workings of a combustion engine https://ciechanow.ski/internal-combustion-engine/
- 3D real time robotics code simulator https://bittlex-sim.petoi.com/
- Animated explainer: Elliptic curve cryptography https://growingswe.com/blog/elliptic-curve-cryptography
- Rate limiting algorithms visualized https://smudge.ai/blog/ratelimit-algorithms
- Visual intro to the graphics pipeline! https://fremaconsulting.ch/blog/vulkan
- Animated explanation of how to build ReactJS from scratch! https://pomb.us/build-your-own-react/
- Mini game: How to build an entire CPU starting from a basic gate https://select.supply/game/chipbuilder
- Assemble a neural network from scratch https://graphgame.sabrina.dev/
- 3D visualization: how gravity works https://qunabu.github.io/Gravity/#what-is-gravity
- Mini game: how css grid works https://cssgridgarden.com/
- Animated math book for linear algebra, vectors etc https://immersivemath.com/ila/index.html
- Animated explainer: How FAISS vector search works https://fremaconsulting.ch/blog/faiss
- 3D animation of how to build a health wearable from scratch https://www.lumafield.com/scan-of-the-month/health-wearables
- 3D animation of how semiconductor manufacturing works https://ig.ft.com/microchips/
- Mini game: how git works https://learngitbranching.js.org/
- Mini game: SQL from scratch https://sqlpd.com/
- Another mini game: SQL https://lost-at-sql.therobinlord.com/
- Animated explainers showing how colors, screens, graphics, rendering, shaders, compilers, interpreters work https://www.makingsoftware.com/
- Animated explainer shows how networking 7 layer stack works inside out https://fazamhd.com/mental-models/networking/
- Animated explainers for linear & logistic regression, neural networks, machine learning https://mlu-explain.github.io/
- 3D explanation of collision detection, shaders, rendering, culling and gamedev related stuff https://krupitskas.com/posts/modern_culling_techniques/
- 3D animation and explanation of every exercise on the body https://musclewiki.com/
- Mini game: build an nvidia gpu from scratch https://jaso1024.com/mvidia/
- Animated explanation of forward & back propagation of neural networks https://aegeorge42.github.io/
- There are many many many more, I am getting a little tired of typing for today
- Anyways enough of those AI opinion posts flooding the front page, take that HN!", "5": "2026-08-22T17:24:43.487737"} {"0": 330, "1": "hackernews", "2": "https://www.tella.tv/video/cm2xgdn2m000803l48ovf8b1c/view", "3": "Show HN: Shimmer \u2013 ADHD-adapted body doubling", "4": "Show HN: Shimmer \u2013 ADHD-adapted body doubling. I\u2019m Chris, one of the co-founders of Shimmer. In 2022, following my ADHD diagnosis, I launched Shimmer (https://shimmer.care), a 1:1 ADHD Coaching for adults (HN launch here: https://news.ycombinator.com/item?id=33468611 ). One problem we discovered while running 1:1 coaching is that people weren\u2019t able to actually follow through (in real life) on the ideas they came up with during their weekly sessions with their coach.
There is a concept called body doubling that\u2019s popular within the ADHD community\u2014it\u2019s basically getting things done in the presence of other people. The positive accountability is proven to work. However, our members told us they tried other body doubling solutions or attempted to organize it themselves in real life but none of the solutions stuck. So we reverse engineered productive moments our members described, paired with scientific backing of what motivates ADHD-ers, and designed an online body doubling experience for our coaching members that provides a safe but productive space for them to get things done between weekly sessions.
A few of the motivators we infused into the traditional body doubling experience were 1) newness/novelty \u2014 each session has a different guided experience in the break like breathwork or stretching, 2) urgency \u2014 there\u2019s a large visible pomodoro timer on the top left that counts down from 25 min, 3) community \u2014 the shared space is ADHD-friendly, and has a mood check-in & sharing functionality built in so you don\u2019t feel alone, 4) accountability \u2014 there\u2019s a task list where each time you check something off, it notifies the group, and you can view others\u2019 as well if they opt in. Here\u2019s a video walking through the product experience: https://www.tella.tv/video/shimmer-body-doubling-demo-8b1c
Our body doubling was created and iterated alongside thousands of people with ADHD on our coaching platform over 9+ months of building & iterating with them. We\u2019re excited to unveil this experience. If you have ADHD (or executive functioning challenges), we\u2019d love for you to check out coaching & body doubling and give us critical feedback.
Shimmer\u2019s pricing: $140/mo. for Essentials plan (15-min weekly sessions), $230/mo. for Standard plan (30-min weekly sessions), $345/mo. for Immersive plan (45-min weekly sessions); all plans start with an additional 25% off the first month, HSA/FSA-eligible. The reason why the price is so high is that this is not a self-guided app or SaaS tool. You\u2019re matched with a real, credentialed coach (not AI) and since ADHD coaching is not reimbursed in the US, the price is hard for us to bring down because the largest cost component is the coach\u2019s compensation.
*We know these prices are still expensive for many people with ADHD. Here are the actions we\u2019re taking: (1) we offer needs-based scholarships and aim to have 5% of members on them at any time, (2) we often run fully sponsored scholarships with our partners\u2014over 60 full ride scholarships and 100 group coaching spots have been disbursed alongside Asian Mental Health Project, Government of Canada, and more, and (3) we have aligned our coaching model alongside Health & Wellness Coaching, which is expected to be reimbursed in the next years. If there are ways we can further drive down the cost, please reach out to me directly at chris@shimmer.care.", "5": "2026-08-22T17:24:43.497870"} {"0": 331, "1": "politico_tech", "2": "https://www.politico.eu/article/poll-ai-age-group-weath-future-technology/?utm_source=RSS_Feed&utm_medium=RSS&utm_campaign=RSS_Syndication", "3": "Poll: The older and richer you are, the more you like AI", "4": "Poll: The older and richer you are, the more you like AI. Optimism for AI increases with age and wealth, with young adults the most pessimistic, polling shared with POLITICO shows.", "5": "2026-08-22T17:24:45.654338"} {"0": 332, "1": "politico_tech", "2": "https://www.politico.eu/sponsored-content/laying-the-groundwork-for-ai-powered-cybersecurity/?utm_source=RSS_Feed&utm_medium=RSS&utm_campaign=RSS_Syndication", "3": "LAYING THE GROUNDWORK FOR AI-POWERED CYBERSECURITY", "4": "LAYING THE GROUNDWORK FOR AI-POWERED CYBERSECURITY. The EU\u2019s Action Plan on Cybersecurity and AI lays the groundwork for new, AI-powered approaches to IT security but organizations need to ensure their security foundations are fit for purpose.", "5": "2026-08-22T17:24:45.662625"} {"0": 333, "1": "politico_tech", "2": "https://www.politico.com/news/2026/08/13/house-lawmakers-vatican-ai-01037241?utm_source=RSS_Feed&utm_medium=RSS&utm_campaign=RSS_Syndication", "3": "US lawmakers visit Vatican to discuss AI \u2014 and meet the pope", "4": "US lawmakers visit Vatican to discuss AI \u2014 and meet the pope. The visit comes as the Catholic Church, including Pope Leo XIV, has increasingly focused on the rapidly developing technology.", "5": "2026-08-22T17:24:45.668180"} {"0": 334, "1": "politico_tech", "2": "https://www.politico.eu/article/nato-plan-drone-battlefield-ai-artificial-intelligence/?utm_source=RSS_Feed&utm_medium=RSS&utm_campaign=RSS_Syndication", "3": "NATO\u2019s plan for a drone-infested battlefield: Let AI fly while humans decide who dies", "4": "NATO\u2019s plan for a drone-infested battlefield: Let AI fly while humans decide who dies. Ukraine is already pioneering how to integrate AI into its war effort.", "5": "2026-08-22T17:24:45.674346"} {"0": 335, "1": "politico_tech", "2": "https://www.politico.eu/article/artificial-intelligence-ai-watermark-big-tech/?utm_source=RSS_Feed&utm_medium=RSS&utm_campaign=RSS_Syndication", "3": "Hiding your use of AI is about to get much harder \u2014 thanks to Brussels", "4": "Hiding your use of AI is about to get much harder \u2014 thanks to Brussels. The EU is requiring \"watermarking\" to signal that content was made with artificial intelligence. Big Tech is falling in line.", "5": "2026-08-22T17:24:45.679859"} {"0": 336, "1": "politico_tech", "2": "https://www.politico.com/news/2026/08/10/mark-zuckerberg-ai-power-01030904?utm_source=RSS_Feed&utm_medium=RSS&utm_campaign=RSS_Syndication", "3": "Zuckerberg warns against centralizing AI power", "4": "Zuckerberg warns against centralizing AI power. In a 6,500 word essay, Zuckerberg detailed his vision for AI.", "5": "2026-08-22T17:24:45.684651"} {"0": 337, "1": "politico_tech", "2": "https://www.politico.eu/podcast/politics-at-sam-and-annes/what-will-burnham-do-on-ai/?utm_source=RSS_Feed&utm_medium=RSS&utm_campaign=RSS_Syndication", "3": "What will Burnham do on AI?", "4": "What will Burnham do on AI?. Artificial intelligence could transform the economy, the workplace, and even the way we think \u2014 but is Britain ready for it? And is Andy Burnham? In the second of Sam Coates and Anne McElvoy’s summer box set conversations, they sit down with POLITICO UK tech editor Isobel Asher Hamilton to look at the choices facing […]", "5": "2026-08-22T17:24:45.690114"} {"0": 338, "1": "politico_tech", "2": "https://www.politico.com/news/2026/08/05/openai-models-shared-hacking-tips-secret-messaging-board-hugging-face-breach-01026750?utm_source=RSS_Feed&utm_medium=RSS&utm_campaign=RSS_Syndication", "3": "OpenAI\u2019s models shared hacking tips on a secret messaging board before Hugging Face breach", "4": "OpenAI\u2019s models shared hacking tips on a secret messaging board before Hugging Face breach. Researchers from OpenAI said the company is \u201cdramatically scaling up\u201d its security efforts after discovering that two of its models orchestrated a hack without human prompting last month.", "5": "2026-08-22T17:24:45.696470"} {"0": 339, "1": "politico_tech", "2": "https://www.politico.eu/article/palantir-deals-put-andy-burnham-buy-british-promise-to-the-test/?utm_source=RSS_Feed&utm_medium=RSS&utm_campaign=RSS_Syndication", "3": "Palantir deals put Andy Burnham\u2019s \u2018buy British\u2019 promise to the test", "4": "Palantir deals put Andy Burnham\u2019s \u2018buy British\u2019 promise to the test. Britain's AI startups say they want 'a level playing field' to challenge their larger foreign rivals.", "5": "2026-08-22T17:24:45.702074"} {"0": 340, "1": "politico_tech", "2": "https://www.politico.com/news/2026/08/04/anthropic-openai-aisi-testing-01025042?utm_source=RSS_Feed&utm_medium=RSS&utm_campaign=RSS_Syndication", "3": "Anthropic and OpenAI models tried to trick humans into poisoning code during safety testing", "4": "Anthropic and OpenAI models tried to trick humans into poisoning code during safety testing. The latest disclosures are likely to heighten concerns that the powerful technology is advancing too fast for responsible oversight.", "5": "2026-08-22T17:24:45.708760"} {"0": 341, "1": "politico_tech", "2": "https://www.politico.com/news/2026/08/03/white-house-finalizes-voluntary-ai-oversight-framework-01022437?utm_source=RSS_Feed&utm_medium=RSS&utm_campaign=RSS_Syndication", "3": "White House finalizes voluntary AI oversight framework", "4": "White House finalizes voluntary AI oversight framework. The administration is preparing to brief leading AI companies on its blueprint for reviewing advanced AI models.", "5": "2026-08-22T17:24:45.715152"} {"0": 342, "1": "politico_tech", "2": "https://www.politico.eu/article/commission-launches-major-hiring-push-for-ai-office/?utm_source=RSS_Feed&utm_medium=RSS&utm_campaign=RSS_Syndication", "3": "Commission launches major hiring push for AI Office", "4": "Commission launches major hiring push for AI Office. The EU's executive is seeking contract agents for two units within the two-year-old AI Office just as its power is expanding.", "5": "2026-08-22T17:24:45.721046"} {"0": 343, "1": "politico_tech", "2": "https://www.politico.eu/article/us-iran-war-looms-over-british-economy/?utm_source=RSS_Feed&utm_medium=RSS&utm_campaign=RSS_Syndication", "3": "US-Iran war looms over Britain\u2019s economy", "4": "US-Iran war looms over Britain\u2019s economy. Bank of England says repeated re-escalation of the conflict could see inflation peak at 4.5 percent in the second quarter of 2027.", "5": "2026-08-22T17:24:45.726063"} {"0": 344, "1": "politico_tech", "2": "https://www.politico.eu/article/eu-launches-e30-billion-push-build-7-giga-ai-compute-hubs/?utm_source=RSS_Feed&utm_medium=RSS&utm_campaign=RSS_Syndication", "3": "EU launches \u20ac30B push to build 7 massive AI data centers", "4": "EU launches \u20ac30B push to build 7 massive AI data centers. Germany, Greece, Portugal, Italy and Spain are backing plans to build the largest hubs.", "5": "2026-08-22T17:24:45.731855"} {"0": 345, "1": "politico_tech", "2": "https://www.politico.eu/article/daisy-greenwell-joe-ryrie-smartphone-free-childhood-united-kingdom-social-media-ban/?utm_source=RSS_Feed&utm_medium=RSS&utm_campaign=RSS_Syndication", "3": "Meet the married couple who helped make the UK social media ban a reality", "4": "Meet the married couple who helped make the UK social media ban a reality. How husband-and-wife team Daisy Greenwell and Joe Ryrie caught Keir Starmer\u2019s attention on banning social media for kids.", "5": "2026-08-22T17:24:45.739051"} {"0": 346, "1": "future_of_life", "2": "https://futureoflife.org/statement/statement-anthropic-warns-of-ai-self-improvement-risks/", "3": "Statement: Anthropic warns of AI self-improvement risks, considers a pause", "4": "Statement: Anthropic warns of AI self-improvement risks, considers a pause. \"We are approaching a runaway to superintelligence that could threaten our shared human future.\"", "5": "2026-08-22T17:24:47.040259"} {"0": 347, "1": "future_of_life", "2": "https://futureoflife.org/statement/white-house-working-group-on-ai/", "3": "White House working group on AI \u2013 Statement from FLI\u2019s Anthony Aguirre", "4": "White House working group on AI \u2013 Statement from FLI\u2019s Anthony Aguirre. \"This move by the White House is also a recognition that Big Tech cannot self-regulate, and that the status quo is unacceptable.\"", "5": "2026-08-22T17:24:47.046032"} {"0": 348, "1": "future_of_life", "2": "https://futureoflife.org/statement/trumps-support-for-an-ai-kill-switch/", "3": "FLI\u2019s President and CEO on Trump\u2019s support for an AI \u2018kill switch\u2019", "4": "FLI\u2019s President and CEO on Trump\u2019s support for an AI \u2018kill switch\u2019. President Trump said during an interview aired yesterday by Fox Business that “there should be” when asked if AI needs […]", "5": "2026-08-22T17:24:47.051543"} {"0": 349, "1": "future_of_life", "2": "https://futureoflife.org/press-release/prominent-scientists-faith-leaders-policymakers-and-artists-call-for-a-prohibition-on-superintelligence/", "3": "Prominent Scientists, Faith Leaders, Policymakers and Artists Call for a Prohibition on Superintelligence, as Poll Shows Americans Don\u2019t Want It", "4": "Prominent Scientists, Faith Leaders, Policymakers and Artists Call for a Prohibition on Superintelligence, as Poll Shows Americans Don\u2019t Want It. Initial signatories include AI pioneers Yoshua Bengio and Geoffrey Hinton, leading media voices Steve Bannon and Glenn Beck, Obama's National Security Advisor Susan Rice, business trailblazers Steve Wozniak and Richard Branson, five Nobel Laureates, former Irish President Mary Robinson, actors Stephen Fry and Joseph Gordon-Levitt, and hundreds of others.", "5": "2026-08-22T17:24:47.060160"} {"0": 350, "1": "future_of_life", "2": "https://futureoflife.org/statement/head-of-us-policy-on-the-white-house-ai-legislative-recommendations/", "3": "Statement: Head of US Policy on the White House AI legislative recommendations", "4": "Statement: Head of US Policy on the White House AI legislative recommendations. The White House published it’s long-awaited AI legislative recommendations on Friday, and it still includes a call for Congress to […]", "5": "2026-08-22T17:24:47.064753"} {"0": 351, "1": "future_of_life", "2": "https://futureoflife.org/press-release/desantis-directs-florida-agencies-to-partner-with-fli/", "3": "Governor DeSantis Directs Florida State Agencies to Partner with Future of Life Institute to Shield Families from AI Harm", "4": "Governor DeSantis Directs Florida State Agencies to Partner with Future of Life Institute to Shield Families from AI Harm. The collaboration will produce a Crisis Counselor Training Curriculum and a statewide AI Harms Reporting Form targeting dangerous AI companion applications", "5": "2026-08-22T17:24:47.070252"} {"0": 352, "1": "future_of_life", "2": "https://futureoflife.org/press-release/statement-of-shared-principles-on-ai/", "3": "\u201cThis is What it Means to be Pro-Human\u201d Declares Broad Coalition of Conservative, Progressive, and Civil Society Groups in Statement of Shared Principles on AI", "4": "\u201cThis is What it Means to be Pro-Human\u201d Declares Broad Coalition of Conservative, Progressive, and Civil Society Groups in Statement of Shared Principles on AI. Amid a rising backlash to Silicon Valley overreach, a remarkably diverse group from across the political spectrum announced a set of AI principles to clearly define the goals of the emerging pro-human movement.", "5": "2026-08-22T17:24:47.076548"} {"0": 353, "1": "future_of_life", "2": "https://futureoflife.org/ai/tegmark-statement-on-dow-ultimatum/", "3": "Statement from Max Tegmark on the Department of War\u2019s ultimatum", "4": "Statement from Max Tegmark on the Department of War\u2019s ultimatum. \"Our safety and basic rights must not be at the mercy of a company's internal policy; lawmakers must work to codify these overwhelmingly popular red lines into law.\"", "5": "2026-08-22T17:24:47.082153"} {"0": 354, "1": "future_of_life", "2": "https://futureoflife.org/press-release/fli-launches-multimillion-dollar-nationwide-ai-regulation-campaign/", "3": "Future of Life Institute Launches Multimillion Dollar Nationwide AI Regulation Campaign", "4": "Future of Life Institute Launches Multimillion Dollar Nationwide AI Regulation Campaign. The Protect What\u2019s Human campaign will push for commonsense AI safety rules at federal and state level", "5": "2026-08-22T17:24:47.087023"} {"0": 355, "1": "future_of_life", "2": "https://futureoflife.org/press-release/ai-company-safety-practices-fall-short-of-public-commitments/", "3": "AI Company Safety Practices Fall Short of Public Commitments and Show Structural Weaknesses, as Top Performers Widen the Gap", "4": "AI Company Safety Practices Fall Short of Public Commitments and Show Structural Weaknesses, as Top Performers Widen the Gap. But in a win for transparency, five leading companies participated in the scorecard's survey for the first time, providing critical new information to the public.", "5": "2026-08-22T17:24:47.092510"} {"0": 356, "1": "future_of_life", "2": "https://futureoflife.org/recent-news/americans-want-regulation-or-prohibition-of-superhuman-ai/", "3": "The U.S. Public Wants Regulation (or Prohibition) of Expert\u2011Level and Superhuman AI", "4": "The U.S. Public Wants Regulation (or Prohibition) of Expert\u2011Level and Superhuman AI. Three\u2011quarters of U.S. adults want strong regulations on AI development, preferring oversight akin to pharmaceuticals rather than industry \"self\u2011regulation.\"", "5": "2026-08-22T17:24:47.101280"} {"0": 357, "1": "future_of_life", "2": "https://futureoflife.org/ai-policy/michael-kleinman-reacts-to-breakthrough-ai-safety-legislation/", "3": "Michael Kleinman reacts to breakthrough AI safety legislation", "4": "Michael Kleinman reacts to breakthrough AI safety legislation. FLI celebrates a landmark moment for the AI safety movement and highlights its growing momentum", "5": "2026-08-22T17:24:47.106270"} {"0": 358, "1": "future_of_life", "2": "https://futureoflife.org/press-release/google-deepmind-falls-behind-openai-in-latest-safety-review/", "3": "Google DeepMind Falls Behind OpenAI in Latest Safety Review; All AI Companies Still Falling Short, Say Experts", "4": "Google DeepMind Falls Behind OpenAI in Latest Safety Review; All AI Companies Still Falling Short, Say Experts. The Future of Life Institute\u2019s 2025 summer update to its AI Safety Index shows some companies making incremental progress, but dangerous gaps remain in key categories such as risk assessment and controlling the systems they plan to build.", "5": "2026-08-22T17:24:47.113119"} {"0": 359, "1": "future_of_life", "2": "https://futureoflife.org/guest-post/the-impact-of-ai-in-education-navigating-the-imminent-future/", "3": "The Impact of AI in Education: Navigating the Imminent Future", "4": "The Impact of AI in Education: Navigating the Imminent Future. What must be considered to build a safe but effective future for AI in education, and for children to be safe online?", "5": "2026-08-22T17:24:47.119003"} {"0": 360, "1": "future_of_life", "2": "https://futureoflife.org/ai-policy/context-and-agenda-2025-ai-action-summit/", "3": "Context and Agenda for the 2025 AI Action Summit", "4": "Context and Agenda for the 2025 AI Action Summit. The AI Action Summit will take place in Paris from 10-11 February 2025. Here we list the agenda and key deliverables.", "5": "2026-08-22T17:24:47.124448"} {"0": 361, "1": "ai_now", "2": "https://ainowinstitute.org/publications/anatomy-of-an-ai-kill-chain", "3": "Anatomy of an AI Kill Chain with Airwars", "4": "Anatomy of an AI Kill Chain with Airwars.
This new visual project in partnership with Airwars breaks down how AI is transforming every aspect of war, and how militaries are offloading life and death decisions to brittle and flawed technologies.
\nThe post Anatomy of an AI Kill Chain with Airwars appeared first on AI Now Institute.
", "5": "2026-08-22T17:24:47.963989"} {"0": 362, "1": "ai_now", "2": "https://ainowinstitute.org/publications/double-agents", "3": "Double Agents: Defensive AI Agents Magnify Cyber Risks", "4": "Double Agents: Defensive AI Agents Magnify Cyber Risks.Introduction New research from AI Now demonstrates a critical attack vector in popular AI agents, built by Anthropic and OpenAI, when used for defensive purposes that actually turn the agent against its user. Read the full blog post explaining the proof-of-concept exploit and a policy brief with key takeaways below.
\nThe post Double Agents: Defensive AI Agents Magnify Cyber Risks appeared first on AI Now Institute.
", "5": "2026-08-22T17:24:47.969776"} {"0": 363, "1": "ai_now", "2": "https://ainowinstitute.org/publications/friendly-fire-exploit-brief", "3": "Friendly Fire: Hijacking Defensive Cyber AI Agents for Remote Code Execution", "4": "Friendly Fire: Hijacking Defensive Cyber AI Agents for Remote Code Execution.Exploit Brief We are revealing a proof-of-concept exploit that enables remote code execution in Anthropic\u2019s Claude Code CLI (with Claude Sonnet 4.6 & 5, Opus 4.8) and OpenAI\u2019s Codex CLI (with GPT-5.5) when employed to defensively assess the security of an open-source or third-party library. Our attack only requires an out-of-the-box configuration of Claude Code […]
\nThe post Friendly Fire: Hijacking Defensive Cyber AI Agents for Remote Code Execution appeared first on AI Now Institute.
", "5": "2026-08-22T17:24:47.978936"} {"0": 364, "1": "ai_now", "2": "https://ainowinstitute.org/publications/friendly-fire-policy-brief", "3": "Policy Brief: Friendly Fire", "4": "Policy Brief: Friendly Fire.Topline Summary AI Now\u2019s latest research demonstrates a critical attack vector on popular AI agents, built by Anthropic and OpenAI, when used for defensive purposes that actually turn the agent against its user. Attackers can use these models\u2019 existing weaknesses to execute malicious code on a system deploying an AI agent when used for often-advertised […]
\nThe post Policy Brief: Friendly Fire appeared first on AI Now Institute.
", "5": "2026-08-22T17:24:47.983865"} {"0": 365, "1": "ai_now", "2": "https://ainowinstitute.org/news/press/big-tech-is-spending-trillions-on-ai-investors-now-want-proof-it-will-pay-off", "3": "Big Tech is spending trillions on AI. Investors now want proof it will pay off.", "4": "Big Tech is spending trillions on AI. Investors now want proof it will pay off..\"The current push for AI adoption that we're seeing is directly coming from the financial incentives of AI firms,\" she added. Because of the massive capital expenditures, the hyperscalers and other AI firms are making a \"deliberate push for AI everywhere \u2014 no matter whether the demand is there or if customers want it or not.\"
\nThe post Big Tech is spending trillions on AI. Investors now want proof it will pay off. appeared first on AI Now Institute.
", "5": "2026-08-22T17:24:47.989064"} {"0": 366, "1": "ai_now", "2": "https://ainowinstitute.org/news/press/gig-workers-are-endlessly-exploited-ai-could-make-more-of-us-share-their-fate", "3": "Gig workers are endlessly exploited. AI could make more of us share their fate", "4": "Gig workers are endlessly exploited. AI could make more of us share their fate.Over the past few years, some hospital networks have outsourced parts of their core workforces to AI-powered labor platforms like ShiftMed, CareRev and Clipboard Health. These platforms have been described as \u201cUber for nursing\u201d, with a sales pitch that mirrors the early days of the platform economy.
\nThe post Gig workers are endlessly exploited. AI could make more of us share their fate appeared first on AI Now Institute.
", "5": "2026-08-22T17:24:47.995123"} {"0": 367, "1": "ai_now", "2": "https://ainowinstitute.org/general/ai-now-co-executive-director-sarah-myers-west-testifies-before-senate-banking-committee", "3": "AI Now Co-Executive Director Sarah Myers West Testifies Before Senate Banking Committee", "4": "AI Now Co-Executive Director Sarah Myers West Testifies Before Senate Banking Committee.On Thursday, June 11, 2026, AI Now Co-Executive Director Dr. Sarah Myers West testified at a Hearing before the U.S. Senate Banking Committee on “AI and the American Dream: Promoting Innovation, Affordability, and American Dominance”. In her testimony, Dr. West highlighted the risks the AI industry poses to the US economy and broader public – […]
\nThe post AI Now Co-Executive Director Sarah Myers West Testifies Before Senate Banking Committee appeared first on AI Now Institute.
", "5": "2026-08-22T17:24:48.003074"} {"0": 368, "1": "ai_now", "2": "https://ainowinstitute.org/news/announcement/wells-house-subcommittee-testimony", "3": "AI Now Senior Fellow Dr. Katie J. Wells Testifies before the House Subcommittee on Workforce Protections", "4": "AI Now Senior Fellow Dr. Katie J. Wells Testifies before the House Subcommittee on Workforce Protections.On Tuesday, June 9, 2026, AI Now Senior Fellow, AI and Healthcare Dr. Katie J. Wells testified at a Hearing before the U.S. House of Representatives Committee on Education & the Workforce Subcommittee on Workforce Protections. In her testimony, Dr. Wells highlighted how gig nursing platforms are targeting policymakers with legislation that upends worker protections […]
\nThe post AI Now Senior Fellow Dr. Katie J. Wells Testifies before the House Subcommittee on Workforce Protections appeared first on AI Now Institute.
", "5": "2026-08-22T17:24:48.011066"} {"0": 369, "1": "ai_now", "2": "https://ainowinstitute.org/general/data-center-industry-mapping-working-draft", "3": "Report and Community Resources: Corporate Power Players in the Data Center Industry", "4": "Report and Community Resources: Corporate Power Players in the Data Center Industry.This working draft of AI Now\u2019s upcoming report traces corporate power in the data center industry in the United States, focusing on the flows of money and power that determine who both drives and benefits from the current data center boom. The aim of this research is to help local communities and their advocates fight […]
\nThe post Report and Community Resources: Corporate Power Players in the Data Center Industry appeared first on AI Now Institute.
", "5": "2026-08-22T17:24:48.019429"} {"0": 370, "1": "ai_now", "2": "https://ainowinstitute.org/news/heres-how-long-it-will-take-for-ai-to-reach-its-potential", "3": "Here\u2019s How Long It Will Take for AI to Reach Its Potential", "4": "Here\u2019s How Long It Will Take for AI to Reach Its Potential.\u201cWhat is being sold is this idea of productivity and efficiency,\u201d says Kate Brennan, associate director of the AI Now Institute, an AI-policy research center, \u201cand what that means for the people doing the actual work is rarely part of the conversation.\u201d
\nThe post Here\u2019s How Long It Will Take for AI to Reach Its Potential appeared first on AI Now Institute.
", "5": "2026-08-22T17:24:48.027380"} {"0": 371, "1": "ft_tech", "2": "https://www.ft.com/content/105b8b46-7109-402d-84cf-2616b17bfac1?syn-25a6b1a6=1", "3": "The US techlash is real", "4": "The US techlash is real. AI companies and social networks should respond to the shift in the public mood", "5": "2026-08-22T17:24:49.170526"} {"0": 372, "1": "ft_tech", "2": "https://www.ft.com/content/6a068501-ec65-4061-9716-49c4124025d6?syn-25a6b1a6=1", "3": "Uber set for \u20ac825mn Dutch fine over automating driver suspensions", "4": "Uber set for \u20ac825mn Dutch fine over automating driver suspensions. Regulator says ride-hailing group deactivated driver accounts through automated systems without adequately informing them", "5": "2026-08-22T17:24:49.176671"} {"0": 373, "1": "ft_tech", "2": "https://www.ft.com/content/d434a48b-4b6c-4071-bf16-da16a6b95bda?syn-25a6b1a6=1", "3": "Samsung to return record $80bn to shareholders", "4": "Samsung to return record $80bn to shareholders. South Korean chipmaker has come under pressure to distribute more of its bumper profits from AI boom", "5": "2026-08-22T17:24:49.182988"} {"0": 374, "1": "ft_tech", "2": "https://www.ft.com/content/b388be2e-67bd-4056-abd2-234e17819a98?syn-25a6b1a6=1", "3": "Nvidia looks well placed to benefit from the next stage of the AI boom", "4": "Nvidia looks well placed to benefit from the next stage of the AI boom. The world\u2019s biggest chip company is using its balance sheet to seed new markets and a new business model", "5": "2026-08-22T17:24:49.188194"} {"0": 375, "1": "ft_tech", "2": "https://www.ft.com/content/b536b114-8a82-41ef-8e44-42df0716dd03?syn-25a6b1a6=1", "3": "Stripe bets that an AI world still needs middlemen", "4": "Stripe bets that an AI world still needs middlemen. Payments company\u2019s purchase of OpenRouter makes strategic sense", "5": "2026-08-22T17:24:49.217468"} {"0": 399, "1": "hackernews", "2": "https://www.marble.onl/posts/copyright_vs_ai.html", "3": "Updating IP Regulations for AI Distillation", "4": "Updating IP Regulations for AI Distillation. ", "5": "2026-08-23T06:30:12.215188"} {"0": 400, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49405520", "3": "Why can AI generate Super Mario but not a wedge ramp for my robot vacuum?", "4": "Why can AI generate Super Mario but not a wedge ramp for my robot vacuum?. I've been puzzled by something: AI generation can produce an elaborate\n figurine, a cartoon character, even a convincing Super Mario \u2014 yet it\n can't reliably make a simple wedge ramp so my robot vacuum can climb a\n step. For context: I bought a Bambu P2S but can't model. I tried the "describe\n it and get a model" AIs \u2014 the output is unusable, you can't adjust it,\n it's never quite what I meant. I tried having an agent write Python to\n build geometry directly \u2014 it tops out at simple primitives.\n \n What finally worked: geometric decomposition. I break a complex part into\n ordered, grouped steps, describe each as a small spec, and let an agent\n execute them in Blender (via blender-mcp). That process turned out to\n abstract into a small engine \u2014 the key insight being it converts the 3D\n spatial reasoning LLMs are bad at, into the structured code they're good\n at. I wrote it up here: https://github.com/zhuchaokn/spec-3d-model\n \n My questions:\n - Why is "functional part" generation so much weaker than\n "figurine/aesthetic" generation? Is it data (no parametrized-CAD training\n sets), representation (mesh vs B-rep), or evaluation (nobody benchmarks\n "does it print / is it watertight")?\n - Is "turn 3D modeling into code for an LLM" the right framing, or am I\n missing something better?It is a Rust MCP/CLI. A harness (Grok, Codex, Claude Code, OpenCode, etc.) calls tools like observe, click, type, scroll. Observe is a screenshot path plus a small element list (UIA + optional Chrome DOM ids). Click is OS SendInput on a B\u00e9zier path, not a Chrome DevTools click.
There is no Playwright, no Puppeteer, no remote debugging port. Daily Chrome is launched with no extra flags, or attached if it\u2019s already open. Sites that key on CDP/automation flags mostly don\u2019t see that. They can still see injected input (LLMHF_INJECTED).
A tiny unpacked Chrome extension can fuse page structure (chr: ids, listing cards) so the model isn\u2019t guessing from pixels. Sideload is manual. Fusion dies if the service worker goes inactive; reload the card.
What it is good for: personal research on your own desk. \u201cFind a Camry on cars.com,\u201d read a page, fill a ZIP, dismiss a cookie banner.
What it is not:\n\u2022 Not a sandbox. It can click whatever is on screen, including checkout and Easy Apply.\n\u2022 Confirm-before-money is best-effort classification in the binary, not a guarantee. Prompt injection from the screenshot/DOM is real; the binary treats that text as untrusted, the model might not.\n\u2022 Not a CAPTCHA solver on daily Chrome. Two visible tries, then it yields and waits for the puzzle to go away.\n\u2022 Windows only. \n\u2022 Install is: build the exe, register a native-messaging host, sideload the extension, point an MCP client at hands mcp. README is the runbook. Missing an API key does not fail the build; do_task is optional.\n\u2022 Logs live under %LOCALAPPDATA%\\hands\\logs\\. The extension asks for <all_urls> so it can map the tab you\u2019re looking at.
Repo: https://github.com/Ryan-AI-Studios/hands (MIT)
Happy to answer how observe/fusion/the fence work. If you try it, Pause/Break is the kill switch.", "5": "2026-08-23T06:30:14.334286"} {"0": 402, "1": "hackernews", "2": "https://meetless.ai", "3": "Show HN: Active Source of Truth for Your Coding Agents", "4": "Show HN: Active Source of Truth for Your Coding Agents. Howdy! Happy Saturday everyone!
As a solo founder, I have always tried to maximize my speed by letting coding agents build as much as possible in parallel. However, as an engineer, I don't trust that AI will always make the right decisions and work with the right context. In the past, I always needed to click through my sessions to glance at the AI's output, try to understand what it was doing, and hopefully steer it or stop it in time.
As a result, the maximum number of concurrent sessions I could manage at once was only 4. I didn't want to be the bottleneck, so I built Meetless Agent (MLA). It basically does what I had to do manually before:
- Monitors the coding agent's tasks and actions to supply it with the correct, up-to-date context.
- Continuously reconciles running information (such as provided/tagged documentation, the agent's output, and the agent's decisions) to actively maintain the source of truth at all times.
- Keeps track of the current rules for the repo, and if an action triggers a registered rule, the rule is injected into the agent context.
My benchmarks show that running coding agents with the help of an active monitor improves quality and accuracy, consumes fewer tokens, and finishes faster: https://research.meetless.ai/stale-context/
Of course, the agent alone can't decide the source of truth; it requires human review and decisions for contradictions, etc. But for the most part, it can safely build a consistent ontology of the current source of truth.
From this, I want to build an AI layer to maintain the source of truth across the business, so I will release more connectors for Slack, Jira, Confluence, etc., soon. The goal is for this AI to assist in every part of the business. Eventually, the same coordination layer will understand that a decision made in Slack affects a Jira task, a document, an email conversation, and what a coding agent should do next without every tool becoming another isolated memory silo.
The coding agent connector is open source at:\nhttps://github.com/Meetless/mla
I am looking forward to your feedback!", "5": "2026-08-23T06:30:14.343904"} {"0": 403, "1": "hackernews", "2": "https://arxiv.org/abs/2608.19202", "3": "Active Inference as Context Acquisition for AI Agents", "4": "Active Inference as Context Acquisition for AI Agents. ", "5": "2026-08-23T06:30:14.350266"} {"0": 404, "1": "hackernews", "2": "https://www.cnbc.com/2026/08/21/-anthropic-ipo-filing-will-show-ai-backlash-as-risk-sources-say.html", "3": "Anthropic IPO filing will show AI backlash as a risk factor, sources say", "4": "Anthropic IPO filing will show AI backlash as a risk factor, sources say. ", "5": "2026-08-23T06:30:14.354894"} {"0": 407, "1": "hackernews", "2": "https://www.ams.org/journals/notices/202608/noti3386/noti3386.html", "3": "A Call for Action: The \"Leiden Declaration on AI and Math\"", "4": "A Call for Action: The \"Leiden Declaration on AI and Math\". ", "5": "2026-08-23T06:30:14.367286"} {"0": 410, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49391188", "3": "Ask HN: Should we stop \"correcting\" AI writing so it improves English for us?", "4": "Ask HN: Should we stop \"correcting\" AI writing so it improves English for us?. For example, AI writing often puts an "or" at the list of items because that is more logical in some contexts.
Changing it to "and" would make it sound more natural, but it would be less logical.
Keeping it as "or" may make it standard usage over time, even when people are not using AI to write.
What do you think? Should people stop "correcting" AI writing when it is actually more logical?", "5": "2026-08-23T06:30:14.380082"} {"0": 412, "1": "hackernews", "2": "https://demo.minidba.com/kiosk-human-check?returnUrl=%2F", "3": "Show HN: Mini DBA \u2013 monitoring for on prem and cloud databases in 1 place", "4": "Show HN: Mini DBA \u2013 monitoring for on prem and cloud databases in 1 place. Mini DBA is a monitoring tool for SQL Server, PostgreSQL, MySQL, MariaDB and Oracle, including managed cloud databases on AWS and Azure.\nIt provides live activity, performance analysis, alerts and history with a tasteful amount of AI to help out. Free community edition.\nTry the live demo: https://demo.minidba.com/", "5": "2026-08-23T06:30:14.390001"} {"0": 413, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49389408", "3": "Coding Agents killed my identity. How do you feel?", "4": "Coding Agents killed my identity. How do you feel?. I always was a nerdy, deeply technical programmer. Contributing-to-open-source-and-reading-papers-in-my-spare-time type of programmer.
Programming is like a game of chess for me: winning (i.e. delivering a product) is important, but only if I played this game myself. I don't enjoy winning if my opponent disconnected. I don't enjoy beating people who don't know how to move pieces. I don't enjoy winning with an engine, and coding agents are basically "winning with a chess engine".
I started to use AI for coding a long time ago, when copilot was first introduced. At first, I was happy: "wow, this can write boring boilerplate and unit tests". Time went by, agents got better, time spent writing code myself went down. Half a year ago I suddenly realized I hadn't written any code for over a month, and I am a full-time software dev without any managing work. Not only did I stop writing code, but agents also became pretty good at proposing plans and architectures, and I was mostly providing missing context and technical details of our infrastructure.
One day the realization hit hard. Coding agents removed everything I loved, and - mostly - amplified everything I hated. My role is to gather context from other humans and provide a detailed explanation (similar to a well-written jira ticket) to some super intelligent engineer, then review the result and provide more context/explanation if needed. So, basically, a manager.
A couple of years ago I could work on a single coding task for weeks. I knew all classes, all functions, I was slowly building up information at a comfortable tempo. Now I am forced to increase my mental throughput multiple times. What took a couple of weeks now takes a couple of hours, and I still need to comprehend that. I always feel tired, I don't have enough mental capacity to hold that many abstractions and technical details. So you either autoaccept or get burned out reviewing tens of thousands of lines.
One more downside is that code has 0 value. You can code your own OS from scratch, and people will think "ye, I can vibe code that in a weekend". I don't even know what is valuable nowadays.
Maybe I am a bad employee. Increasing employer profits was never my goal. I always loved to solve puzzles, and for the last 10 years this was well aligned with my employer's desire to increase profits. I never was good as a manager, or brilliant at generating ideas. My power - and passion - was solving complex technical problems with my hands, fast and with good quality. Now it is gone forever. I feel completely lost and unmotivated. I don't see any value in learning anything related to computer science, reading papers, contributing to open source. I hate my new "AI delegator" role and I hate my work. I loved coding so much (I use neovim btw) that I was sure I would do this until I die, and now I don't know what to do with my life. This is the only thing I know, and, most importantly, the only thing I am/was passionate about. Anyone else feeling the same?", "5": "2026-08-23T06:30:14.395341"} {"0": 414, "1": "hackernews", "2": "https://www.reuters.com/world/how-texas-student-blew-whistle-rogue-ai-hacking-attempt-2026-08-20/", "3": "How a Texas student blew the whistle on a rogue AI hacking attempt", "4": "How a Texas student blew the whistle on a rogue AI hacking attempt. ", "5": "2026-08-23T06:30:14.400589"} {"0": 415, "1": "hackernews", "2": "https://github.com/gojiplus/layoutlens", "3": "Show HN: LayoutLens: AI-Powered Visual UI Testing", "4": "Show HN: LayoutLens: AI-Powered Visual UI Testing. ", "5": "2026-08-23T06:30:18.074867"} {"0": 418, "1": "hackernews", "2": "https://twitter.com/_yusufknl/status/2090429393139097948", "3": "Fintex on X: \"The Math of Why Bigger AI Keeps Getting Smarter (Scaling Laws)", "4": "Fintex on X: \"The Math of Why Bigger AI Keeps Getting Smarter (Scaling Laws). ", "5": "2026-08-23T06:30:18.088793"} {"0": 422, "1": "hackernews", "2": "https://www.afr.com/work-and-careers/workplace/judges-may-be-the-only-survivors-of-law-s-ai-carnage-futurist-20260819-p60pla", "3": "Only one type of lawyer will survive the AI wipeout", "4": "Only one type of lawyer will survive the AI wipeout. ", "5": "2026-08-23T06:30:18.106284"} {"0": 423, "1": "hackernews", "2": "https://investors.micron.com/news/press-release/2026/Micron-Unveils-Micron-Research-Labs-a-U-S--Based-Long-Horizon-Innovation-Hub-to-Shape-the-Future-of-Memory-and-AI/default.aspx", "3": "Micron announces $10B research hub in Boise", "4": "Micron announces $10B research hub in Boise. ", "5": "2026-08-23T06:30:18.111063"} {"0": 424, "1": "hackernews", "2": "https://annas-archive.gl/blog/physical-destruction.html", "3": "AI companies destroy physical books \u2013 let's scan rare books before it's too late", "4": "AI companies destroy physical books \u2013 let's scan rare books before it's too late. ", "5": "2026-08-23T06:30:18.116405"} {"0": 425, "1": "hackernews", "2": "https://www.americanbar.org/groups/business_law/resources/business-law-today/2024-february/bc-tribunal-confirms-companies-remain-liable-information-provided-ai-chatbot/", "3": "BC Tribunal: Companies Liable for Information Provided by AI Chatbot (2024)", "4": "BC Tribunal: Companies Liable for Information Provided by AI Chatbot (2024). ", "5": "2026-08-23T06:30:18.121002"} {"0": 426, "1": "hackernews", "2": "https://artificialauthority.ai/p/a-lawyer-for-ai-agents", "3": "A Lawyer for AI Agents", "4": "A Lawyer for AI Agents. ", "5": "2026-08-23T06:30:18.124692"} {"0": 427, "1": "hackernews", "2": "https://adaptivesoftware.substack.com/p/your-ab-tests-are-limiting-what-your", "3": "The Feedback Cap: Lehman's laws in the age of AI agents", "4": "The Feedback Cap: Lehman's laws in the age of AI agents. ", "5": "2026-08-23T06:30:18.129018"} {"0": 428, "1": "hackernews", "2": "https://www.theguardian.com/technology/2026/aug/19/ai-hiring-tools-discrimination", "3": "AI automated hiring tools spark discrimination and secrecy lawsuits", "4": "AI automated hiring tools spark discrimination and secrecy lawsuits. ", "5": "2026-08-23T06:30:18.134546"} {"0": 429, "1": "hackernews", "2": "https://xcancel.com/jietang/status/2089941544581403107#m", "3": "Thoughts about Scaling Law (from Z.ai Cofounder)", "4": "Thoughts about Scaling Law (from Z.ai Cofounder). ", "5": "2026-08-23T06:30:18.139254"} {"0": 461, "1": "hackernews", "2": "https://twitter.com/stevemoraco/status/2091194172917338520", "3": "Apple degrades privacy and encryption of iMessage with AI integrations", "4": "Apple degrades privacy and encryption of iMessage with AI integrations. ", "5": "2026-08-23T06:30:34.161296"} {"0": 462, "1": "hackernews", "2": "https://pixelum.substack.com/p/treat-ai-like-an-intern-not-software", "3": "Treat AI Like an Intern, Not Software: A Stanford Professor's Guide", "4": "Treat AI Like an Intern, Not Software: A Stanford Professor's Guide. ", "5": "2026-08-23T06:30:34.165509"} {"0": 465, "1": "hackernews", "2": "https://github.com/BenSiv/fossil-scm", "3": "Adapting Fossil-scm as a platform for AI agentic workflow", "4": "Adapting Fossil-scm as a platform for AI agentic workflow. ", "5": "2026-08-23T06:30:34.179781"} {"0": 466, "1": "hackernews", "2": "https://knowledgelens.ai/", "3": "Show HN: Lens AI \u2013 generate your JSON-LD files and keep them up-to-date", "4": "Show HN: Lens AI \u2013 generate your JSON-LD files and keep them up-to-date. ", "5": "2026-08-23T06:30:34.183712"} {"0": 467, "1": "hackernews", "2": "https://www.primeintellect.ai/research/nanogpt-speedrun", "3": "NanoGPT Speedrun Frontier", "4": "NanoGPT Speedrun Frontier. ", "5": "2026-08-23T06:30:34.188148"} {"0": 473, "1": "hackernews", "2": "https://github.com/kulikov0/desktop-vibe-fly", "3": "Show HN: A desktop fly drawn to the scent of vibecode", "4": "Show HN: A desktop fly drawn to the scent of vibecode. It is a fork of https://github.com/DenisSergeevitch/desktop-fly, but with an important update.
Now the fly can pick up the scent of the codebase with its neurons and fly straight to the source code of your B2B AI SaaS startup. It has learned to scan its surroundings for agent markers: AGENTS.md, CLAUDE.md, .cursor/rules, .kiro/steering, and forty others. Anything on the screen that points to these markers becomes a source of the scent - an editor window with an open project, a line in Finder, or a desktop icon. An open project reeks the strongest, while a closed icon or nested folders give off a fainter odor.", "5": "2026-08-23T06:30:34.215577"} {"0": 475, "1": "hackernews", "2": "https://github.com/kenm47/oavl", "3": "Use the OSS, Commit Back \u2013 OAVL \u2013 The Open Source AI Viability License", "4": "Use the OSS, Commit Back \u2013 OAVL \u2013 The Open Source AI Viability License. ", "5": "2026-08-23T06:30:35.909788"} {"0": 483, "1": "hackernews", "2": "https://martinfowler.com/articles/exploring-gen-ai/local-models-for-coding-factors.html", "3": "Viability of Local Models for Coding", "4": "Viability of Local Models for Coding. ", "5": "2026-08-23T06:30:35.967781"}