text stringlengths 300 320k | source stringlengths 52 154 |
|---|---|
# Agentized LLMs will change the alignment landscape
Epistemic status: head spinning, suddenly unsure of everything in alignment. And unsure of these predictions.
I'm following the suggestions in [10 reasons why lists of 10 reasons might be a winning strategy](https://www.lesswrong.com/posts/8gbfvGhSnEJ9hHGew/10-reas... | https://www.lesswrong.com/posts/dcoxvEhAfYcov2LA6/agentized-llms-will-change-the-alignment-landscape |
# Ng and LeCun on the 6-Month Pause (Transcript)
*This is a transcript a brief chat between* [*Andrew Ng*](https://en.wikipedia.org/wiki/Andrew_Ng) *and* [*Yann LeCun*](https://en.wikipedia.org/wiki/Yann_LeCun) *on the prospect of the 6-Month Pause. They are both highly skeptical of the pause and more generally of "AI... | https://www.lesswrong.com/posts/fKfNTkdqsckdMusTb/ng-and-lecun-on-the-6-month-pause-transcript |
# Expanding the domain of discourse reveals structure already there but hidden
*[Metadata: crossposted from [https://tsvibt.blogspot.com/2022/05/expanding-domain-of-discourse-reveals.html](https://tsvibt.blogspot.com/2022/05/expanding-domain-of-discourse-reveals.html). First completed May 15, 2022. This essay is more ... | https://www.lesswrong.com/posts/mFWBsCy2SHQ97RArb/expanding-the-domain-of-discourse-reveals-structure-already |
# Being at peace with Doom
You can be at peace even when thinking the world is doomed. And while at peace you can still work against that Doom, even while being aware that nothing you do will make a difference. I believe there are states of mind like this that can be inhabited by humans.
Here I am not going to argue ... | https://www.lesswrong.com/posts/iJnBLanZLephL5cao/being-at-peace-with-doom |
# What Piles Up Must Pile Down
…except when it doesn’t.
*Tl;dr: some problems in life follow a similar pattern of “stuff piles up and we get frustrated about it”. This post looks at some such problems, and some typical solutions. *
* * *
A while ago I noticed that a lot of the persistent frustrations in my life wer... | https://www.lesswrong.com/posts/vKuuz8yDiGvKZ7B3d/what-piles-up-must-pile-down |
# Foom seems unlikely in the current LLM training paradigm
**Epistemic status:** *The idea here has likely been articulated before, I just haven't noticed it, so it might be worth pointing it out again.*
Foom describes the idea of a rapid AI takeoff caused by an AI's ability to recursively improve itself. Most di... | https://www.lesswrong.com/posts/ktSzxMsKBJmon6FGm/foom-seems-unlikely-in-the-current-llm-training-paradigm |
# An AI Realist Manifesto: Neither Doomer nor Foomer, but a third more reasonable thing
Cross-posted from [substack](https://pashanomics.substack.com/p/an-ai-realist-manifesto-neither-doomer)
AI has been a hot topic in recent Twitter discourse, with two opposing camps dominating the conversation: the Doomers and the ... | https://www.lesswrong.com/posts/jmaZjkzq32pmzoHpF/an-ai-realist-manifesto-neither-doomer-nor-foomer-but-a |
# Contra LeCun on "Autoregressive LLMs are doomed"
I answer LeCun's arguments against LLMs as exposed [in this lesswrong comment](https://www.lesswrong.com/posts/nH4c3Q9t9F3nJ7y8W/gpts-are-predictors-not-imitators?commentId=GFY5baR4ByiXAvBFE). I haven't searched thoroughly or double-checked in detail LeCun's writings ... | https://www.lesswrong.com/posts/zyPaqXgFzqHkQfccq/contra-lecun-on-autoregressive-llms-are-doomed |
# Open & Welcome Thread – April 2023
If it’s worth saying, but not worth its own post, here's a place to put it.
If you are new to LessWrong, here's the place to introduce yourself. Personal stories, anecdotes, or just general comments on how you found us and what you hope to get from the site and community are invit... | https://www.lesswrong.com/posts/Y5KKmWxHDqmxGT3gm/open-and-welcome-thread-april-2023-1 |
# Current UK government levers on AI development
*This is a link post for *[*this*](https://docs.google.com/document/d/1zJodltS1PkGUBARFyltwjJUQVgdYgTXpwWVXNkY4HaM/edit#) *collection of current UK government levers on AI development.*
At the end of 2022, I made a [**collection of information on current UK government ... | https://www.lesswrong.com/posts/RwC89YJtr6cF26Axx/current-uk-government-levers-on-ai-development |
# Tim Berners-Lee found it hard to explain the web
According to Tim Berners-Lee, explaining his ideas about the World Wide Web [was at times quite challenging](https://www.ted.com/talks/tim_berners_lee_the_next_web/transcript):
> «So going back to 1989, I wrote a memo suggesting the global hypertext system. Nobody re... | https://www.lesswrong.com/posts/eRMEtt8Q8J2QZYrEi/tim-berners-lee-found-it-hard-to-explain-the-web |
# Outrage and Statistics into Policy
Here's a pattern that doesn't work very well: a tragedy catches our attention, we point to statistics to show it's an example of a distressingly common problem, and we propose laws to address the issue. Except the event is rarely representative of the larger problem, and so these p... | https://www.lesswrong.com/posts/Cu3SrLfQtCfAPKWEz/outrage-and-statistics-into-policy |
# The Influence of Cultural Subconscious on AI Language Models: A Comprehensive Analysis of Archetypes and Communication, written with GPT4
The cultural subconscious represents an aggregation of tropes and ideas that humans share within their respective communities (Jung, 1968). These ideas and archetypes have evolv... | https://www.lesswrong.com/posts/GN2N4ruBgurqMrKJM/the-influence-of-cultural-subconscious-on-ai-language-models |
# Why Simulator AIs want to be Active Inference AIs
*Paper covering some of the same ideas is now available at* [*https://arxiv.org/abs/2311.10215*](https://arxiv.org/abs/2311.10215)
Prelude: when GPT first hears its own voice
Imagine humans in Plato’s cave, interacting with reality by watching the shadows on th... | https://www.lesswrong.com/posts/YEioD8YLgxih3ydxP/why-simulator-ais-want-to-be-active-inference-ais |
# AI Safety Newsletter #1 [CAIS Linkpost]
The [Center for AI Safety](https://www.safe.ai/) just launched its first AI Safety Newsletter. The newsletter is designed to inform readers about developments in AI and AI safety. No technical background required.
Subscribe [here](https://newsletter.safe.ai/subscribe?utm_med... | https://www.lesswrong.com/posts/9ros2kvDGCTuoidqX/ai-safety-newsletter-1-cais-linkpost |
# Bing AI Generating Voynich Manuscript Continuations - It does not know how it knows
*Epistemic status: slightly giddy and freaked out. Possibly rushing to judgment. But there's definitely something here that others should check out for themselves...*
The Voynich Manuscript (VMS), for anyone who is familiar with i... | https://www.lesswrong.com/posts/77b3mu2NFwzr5PdNk/bing-ai-generating-voynich-manuscript-continuations-it-does |
# Humans are not prepared to operate outside their moral training distribution
When people think about how to treat an AI, they probably normally default to the evil AI, that frowns on humans and desires to enslave them out of hate, or the nice android-type AI. A lot of writing in AI Safety has focused on debunking th... | https://www.lesswrong.com/posts/ZDqiXgSYRgQTSFrsW/humans-are-not-prepared-to-operate-outside-their-moral |
# Spring Meetups Everywhere 2023
Many cities have regular Astral Codex Ten meetup groups. Twice a year, I try to advertise their upcoming meetups and make a bigger deal of it than usual so that irregular attendees can attend. This is one of those times.
This year we have spring meetups planned in over eighty cities, ... | https://www.lesswrong.com/posts/55aoSeDTDKbw9dgza/spring-meetups-everywhere-2023 |
# FLI And Eliezer Should Reach Consensus
I will propose **a plan** for possible future, explain some **reasons the plan might work**, and end with a call to action.
...
This essay basically doesn't even engage with the "AI stuff itself".
I'm simply taking AI risks mostly for granted and focusing on social aspects.... | https://www.lesswrong.com/posts/BWLKRMQn3DFcQg6of/fli-and-eliezer-should-reach-consensus |
# Some Intuitions Around Short AI Timelines Based on Recent Progress
*tldr: I give some informal evidence and intuitions that point toward AGI coming soon. These include thinking about how crazy the last year has been, beliefs from those in major AI labs, and progress on MMLU. *
Intro
=====
This post is intended to ... | https://www.lesswrong.com/posts/woZymgKQqB5gEaAAz/some-intuitions-around-short-ai-timelines-based-on-recent |
# Alignment vs capabilities
In [Star Wars: Episode I - The Phantom Menace](https://www.imdb.com/title/tt0120915/), Master Jedi Qui-Gon Jinn discovers Anakin. Anakin is just a young child slave on the forgettable planet of Tatooine. However, Qui-Gon feels something noteworthy about the boy.
He feels... the Force. To c... | https://www.lesswrong.com/posts/akqBmp3HhqB3w6e8Q/alignment-vs-capabilities |
# Four mindset disagreements behind existential risk disagreements in ML
I've previously written that good ideas and conversations about AGI seem to have propagated through ML [weirdly slowly](https://www.lesswrong.com/posts/Rkxj7TFxhbm59AKJh/the-inordinately-slow-spread-of-good-agi-conversations-in-ml).
A different ... | https://www.lesswrong.com/posts/84BJopKvQ8qYGHY6b/four-mindset-disagreements-behind-existential-risk |
# Is "Strong Coherence" Anti-Natural?
Related:
* [Contra "Strong Coherence"](https://www.lesswrong.com/posts/AdGo5BRCzzsdDGM6H/contra-strong-coherence)
* [why assume AGIs will optimize for fixed goals](https://www.lesswrong.com/posts/dKTh9Td3KaJ8QW6gw/why-assume-agis-will-optimize-for-fixed-goals)
* [Why The Fo... | https://www.lesswrong.com/posts/c4GrmECzui2zT4fMq/is-strong-coherence-anti-natural |
# Killing Socrates
**Or, On The Willful Destruction Of Gardens Of Collaborative Inquiry**
----------------------------------------------------------------------
* * *
One of the more interesting dynamics of the past eight-or-so years has been watching a bunch of the people who \[taught me my values\] and \[served as... | https://www.lesswrong.com/posts/JcgtKunqmELefxksx/killing-socrates |
# Measuring artificial intelligence on human benchmarks is naive
Central claim: Measured objectively, GPT-4 is arguably way past human intelligence already, perhaps even after taking generality into account.
Central implication: If the reason we're worried AGI will wipe us out is tied to an objective notion of intell... | https://www.lesswrong.com/posts/vp8G6J9K8RiEq5Ghq/measuring-artificial-intelligence-on-human-benchmarks-is |
# Childhood Roundup #2
Recently: [The Kids are Not Okay](https://thezvi.substack.com/p/the-kids-are-not-okay), [Childhood Roundup #1](https://thezvi.substack.com/p/childhood-roundup-1), [Fertility Roundup #1](https://thezvi.substack.com/p/fertility-rate-roundup-1)
It seems like a good time to go over the interesting ... | https://www.lesswrong.com/posts/JdFYAxqtCHDeBBRth/childhood-roundup-2 |
# A couple of questions about Conjecture's Cognitive Emulation proposal
Introduction
---------------
Recently, Conjecture proposed a conception of a modular AI architecture, with each module being interpretable, having limited intellect, and functionally emulating the human mind. They called this concept [Cognitiv... | https://www.lesswrong.com/posts/YtKnLchzsaJoa24uA/a-couple-of-questions-about-conjecture-s-cognitive-emulation |
# "The Need for Long-term Research" - Seeds of Science call for reviewers
### Abstract
This article proposes the idea of creating specialized institutions for long-range research with a horizon of 25 years or more. While universities are traditionally seen as the primary institutions for research, their focus tends t... | https://www.lesswrong.com/posts/qonijrXGCn3yb96Wo/the-need-for-long-term-research-seeds-of-science-call-for |
# Where's the foom?
> "The first catastrophe mechanism seriously considered seems to have been the possibility, raised in the 1940s at Los Alamos before the first atomic bomb tests, that fission or fusion bombs might ignite the atmosphere or oceans in an unstoppable chain reaction."[^27figlpro9v]
This is not our firs... | https://www.lesswrong.com/posts/8vGFoqcQdkk7n3vDz/where-s-the-foom |
# [MLSN #9] Verifying large training runs, security risks from LLM access to APIs, why natural selection may favor AIs over humans
As part of a larger community building effort, [CAIS](https://safe.ai/) is writing a safety newsletter that is designed to cover empirical safety research and be palatable to the broader m... | https://www.lesswrong.com/posts/2BG49yHpgEL46eioZ/mlsn-9-verifying-large-training-runs-security-risks-from-llm |
# What Jason has been reading, April 2023
A monthly feature. Note that I generally don’t include very recent writing here, such as the latest blog posts (for those, see [my Twitter digests](https://rootsofprogress.org/writing#twitter-digest)); this is for my deeper research.
**AI**
------
First, various historical p... | https://www.lesswrong.com/posts/ekCBEFuoa6eWCAwXt/what-jason-has-been-reading-april-2023 |
# Core of AI projections from first principles: Attempt 1
*I have at various times commented on AI projections and AI risk, using my own models of the key issues. However I have not really crystallized any overview of what my models imply, partly because I am uncertain about some key issues, but also partly because I ... | https://www.lesswrong.com/posts/JXAYG5jFwTTAJzHbe/core-of-ai-projections-from-first-principles-attempt-1 |
# Request to AGI organizations: Share your views on pausing AI progress
A few observations from the last few weeks:
* On March 22, FLI published an [open letter](https://futureoflife.org/open-letter/pause-giant-ai-experiments/) calling for a six-month moratorium on frontier AI progress.
* On March 29, Eliezer Yud... | https://www.lesswrong.com/posts/bceeKEnPHSQqgyr36/request-to-agi-organizations-share-your-views-on-pausing-ai |
# [Interview w/ Jeffrey Ladish] Applying the 'security mindset' to AI and x-risk
Though I've been following the AI safety debate for a decade or so, I've had relatively few conversations with the relevant experts on the Futurati Podcast.
Having updated with the release of GPT-4, however, I'm working to change that.... | https://www.lesswrong.com/posts/eHkTgwyi3rCztEy26/interview-w-jeffrey-ladish-applying-the-security-mindset-to |
# On "aiming for convergence on truth"
Background
==========
Duncan Sabien wrote a list of proposed "[basics of rational discourse](https://www.lesswrong.com/posts/XPv4sYrKnPzeJASuk/basics-of-rationalist-discourse-1)" guidelines. Zack M Davis [disagrees with (his interpretation of) one of the guidelines](https://www... | https://www.lesswrong.com/posts/9vjEavucqFnfSEvqk/on-aiming-for-convergence-on-truth |
# Evolution provides no evidence for the sharp left turn
Does human evolution imply a sharp left turn from AIs?
======================================================
Arguments for the [sharp left turn](https://www.lesswrong.com/posts/GNhMPAWcfBCASy8e6/a-central-ai-alignment-problem-capabilities-generalization) in AI... | https://www.lesswrong.com/posts/hvz9qjWyv8cLX9JJR/evolution-provides-no-evidence-for-the-sharp-left-turn |
# [Link] Sarah Constantin: "Why I am Not An AI Doomer"
This is a good post from Sarah Constantin explaining why her expected timeline to agentic AGI is long (> 10 years). Topics discussed include whether LLMs and other current major research directions will endow AIs with adequate world models, causal inference, and g... | https://www.lesswrong.com/posts/nG4biq5ymbBviKYsJ/link-sarah-constantin-why-i-am-not-an-ai-doomer |
# No convincing evidence for gradient descent in activation space
How is it that GPT performs better at question-answering tasks when you first prompt it with a series of positive examples? In 2020, in the title of the original GPT-3 paper, OpenAI claimed that language models are few shot learners. But they didn't say... | https://www.lesswrong.com/posts/HHSuvG2hqAnGT5Wzp/no-convincing-evidence-for-gradient-descent-in-activation |
# LW is probably not the place for "I asked this LLM (x) and here's what it said!", but where is?
I notice a nonzero amount of posts on LW, admittedly typically not overly high karma, that go something like, "I asked Bing to do x and its' answer freaked me out!", or "I talked to ChatGPT4 about itself and it told me we... | https://www.lesswrong.com/posts/SswHiGhG9TbvkzuZH/lw-is-probably-not-the-place-for-i-asked-this-llm-x-and-here |
# Scaffolded LLMs as natural language computers
*Crossposted from my* [*personal blog*](https://www.beren.io/2023-04-11-Scaffolded-LLMs-natural-language-computers/)*.*
Recently, LLM-based agents have been all the rage -- with projects like [AutoGPT](https://github.com/Torantulino/Auto-GPT) showing how easy it is to w... | https://www.lesswrong.com/posts/43C3igfmMrE9Qoyfe/scaffolded-llms-as-natural-language-computers |
# Boundaries-based security and AI safety approaches
\[This part 3 of a [5-part sequence on security and cryptography areas relevant for AI safety](/posts/FRQNC7rJJ4jBjonii/four-security-and-cryptography-areas-relevant-for-ai-safety). There is a more in-depth version of this post available here: [Digital Defense](http... | https://www.lesswrong.com/posts/TZy4mFJFJ4yv2MRhg/boundaries-based-security-and-ai-safety-approaches |
# Alignment of AutoGPT agents
In the [previous](https://www.lesswrong.com/posts/cLKR7utoKxSJns6T8/ica-simulacra) post, I outlined a type of LLM-based architecture that has a potential to become the first AGI. I proposed a name *ICA Simulacra* to label such systems, but now it’s obvious that *AutoGPT*/*AutoGPT agent... | https://www.lesswrong.com/posts/JnAh4YHfrYpPNwc8Y/alignment-of-autogpt-agents |
# The UBI dystopia: a glimpse into the future via present-day abuses
*Disclaimer: This does risk being a little culture-war-adjacent. This is more political than I normally see on LessWrong, but the spirit of it is intended to be less "Smash Capitalism" or "Become Ungovernable" and more "look at these things that are ... | https://www.lesswrong.com/posts/ZNmuCTzpNTRQkodeS/the-ubi-dystopia-a-glimpse-into-the-future-via-present-day |
# Should AutoGPT update us towards researching IDA?
Given the rate of progress in AutoGPT-like approaches, should we reconsider Paul Christiano's Iterated Distillation and Amplitication (IDA) agenda as potentially central to the alignment of transformative ML systems?
For contex on IDA and AutoGPT:
- https://www.less... | https://www.lesswrong.com/posts/Enq3Mh73csiajzPTH/should-autogpt-update-us-towards-researching-ida |
# Repugnant levels of violins
*This is a rambling post about the repugnant conclusion, mass-manufacturing, metis and woo, and the fragility of expert discernment — through the lens of my experiences playing, repairing, and selling violins. (And eating discount sushi.)*
*Epistemic status: Discussing utilitarianism-rel... | https://www.lesswrong.com/posts/i7FvJ2vPHEMtSTPdZ/repugnant-levels-of-violins |
# Natural language alignment
I think a potentially promising and undertheorized approach to AI safety, especially in short timelines, is *natural language alignment* (NLA), a form of [AI-assisted alignment](https://www.alignmentforum.org/tag/ai-assisted-alignment-1) in which we leverage the model’s rich understanding ... | https://www.lesswrong.com/posts/EhkHnNJXwT8RmtfYZ/natural-language-alignment-1 |
# AGI goal space is big, but narrowing might not be as hard as it seems.
In the wake of FLI’s [AI 6-month pause letter](https://futureoflife.org/open-letter/pause-giant-ai-experiments/) and Eliezer Yudkowsky’s [doom article in *Time*](https://time.com/6266923/ai-eliezer-yudkowsky-open-letter-not-enough/), there seems ... | https://www.lesswrong.com/posts/nxmo2cyREteqvLMss/agi-goal-space-is-big-but-narrowing-might-not-be-as-hard-as-1 |
# Apply to >30 AI safety funders in one application with the Nonlinear Network
[Nonlinear](http://www.nonlinear.org/) spoke to dozens of earn-to-givers and a common sentiment was, "I want to fund good AI safety-related projects, but I don't know where to find them." At the same time, applicants don’t know how to fi... | https://www.lesswrong.com/posts/dxRw2xPQWFLjyr7WH/apply-to-greater-than-30-ai-safety-funders-in-one-1 |
# AXRP Episode 20 - ‘Reform’ AI Alignment with Scott Aaronson
[YouTube link](https://www.youtube.com/watch?v=ZZ2x-O0mjBg&list=PLmjaTS1-AiDeqUuaJjasfrM6fjSszm9pK&index=20)
How should we scientifically think about the impact of AI on human civilization, and whether or not it will doom us all? In this episode, I speak w... | https://www.lesswrong.com/posts/nsygJvidfgidmgKqX/axrp-episode-20-reform-ai-alignment-with-scott-aaronson |
# AI x-risk, approximately ordered by embarrassment
Advanced AI systems could lead to existential risks via several different pathways, some of which may not fit neatly into traditional risk forecasts. Many previous forecasts, for example the well known [report](https://arxiv.org/abs/2206.13353) by Joe Carlsmith, deco... | https://www.lesswrong.com/posts/mSF4KTxAGRG3EHmhb/ai-x-risk-approximately-ordered-by-embarrassment |
# Independence Dividends
Our oldest is in third grade, and sings in the school chorus. It's after school on Wednesdays, which is fine. Except on half-day Wednesdays it's before school, which is less fine. Instead of needing to be to school at 8am she needs to be there at 7:15. She really likes chorus, and feels strong... | https://www.lesswrong.com/posts/nYjq7CZvhSDoAQyad/independence-dividends |
# "Aligned" foundation models don't imply aligned systems
*\[The basic idea in this post is probably not original to me, since it's somewhat obvious when stated directly. But it seems particularly relevant and worth distilling in light of recent developments with LLM-based systems, and because I keep seeing arguments ... | https://www.lesswrong.com/posts/gmHiwafywFo33euGz/aligned-foundation-models-don-t-imply-aligned-systems |
# What is the best source to explain short AI timelines to a skeptical person?
This is for a person with no ML background. He is 55 years old, he liked the sequences and I recently managed to convince him that AI risk is serious by recommending a bunch of Lesswrong posts on it, but he still thinks it's astronomically ... | https://www.lesswrong.com/posts/6z4jkLkDpzMC4ffEW/what-is-the-best-source-to-explain-short-ai-timelines-to-a |
# Announcing Epoch’s dashboard of key trends and figures in Machine Learning
Developments in Machine Learning have been happening extraordinarily fast, and as their impacts become increasingly visible, it becomes ever more important to develop a quantitative understanding of these changes. However, relevant data has t... | https://www.lesswrong.com/posts/opsfYWNxBYF5sJujB/announcing-epoch-s-dashboard-of-key-trends-and-figures-in |
# Trying AgentGPT, an AutoGPT variant
I saw this announcement on Twitter:
[https://twitter.com/asimdotshrestha/status/1644883727707959296](https://twitter.com/asimdotshrestha/status/1644883727707959296)... | https://www.lesswrong.com/posts/xqALgR9Wc4gHiX68p/trying-agentgpt-an-autogpt-variant |
# Identifying semantic neurons, mechanistic circuits & interpretability web apps
[15 research projects](https://itch.io/jam/mechint/entries) on interpretability were submitted to the [mechanistic interpretability Alignment Jam](https://itch.io/jam/mechint) in January hosted with Neel Nanda. Here, we share the top proj... | https://www.lesswrong.com/posts/sruT3a9KhyLnYmLi7/identifying-semantic-neurons-mechanistic-circuits-and |
# On AutoGPT
The primary talk of the AI world recently is about AI agents (whether or not it includes the question of whether we can’t help but notice we are all going to die.)
The trigger for this was [AutoGPT](https://github.com/Torantulino/Auto-GPT), now number one on GitHub, which allows you to turn GPT-4 (or GPT... | https://www.lesswrong.com/posts/566kBoPi76t8KAkoD/on-autogpt |
# Navigating the Open-Source AI Landscape: Data, Funding, and Safety
**Introduction**
================
Open-source AI promises to democratize technology, but it can also be abused and may lead to hard-to-control AI. In this post, we'll explore the data, ethics, and funding behind these models to discover how to balan... | https://www.lesswrong.com/posts/69CRFgqbQyFBoYcg5/navigating-the-open-source-ai-landscape-data-funding-and |
# AI #7: Free Agency
This week’s biggest AI development was the rise of AutoGPT and the concept of turning GPTs into agents via code wrapping that gives them memory, plans and the ability to spin up instantiations. [I moved that to its own post](https://thezvi.substack.com/p/on-autogpt), which I would prioritize over ... | https://www.lesswrong.com/posts/Woi6PgTP4RAm4bwJh/ai-7-free-agency |
# Even if human & AI alignment are just as easy, we are screwed
I get the sense that something like Eliezer's concept of ["deep security" as opposed to "ordinary paranoia"](https://www.lesswrong.com/posts/8gqrbnW758qjHFTrH/security-mindset-and-ordinary-paranoia) is starting to seep into mainstream consciousness—not as... | https://www.lesswrong.com/posts/FrPYTsoF5YWGgtpLy/even-if-human-and-ai-alignment-are-just-as-easy-we-are |
# Subscripts for Probabilities
*cross-posted from [niplav.github.io](https://niplav.github.io)*
Gwern has wondered about a use-case for subscripts in hypertext. While
they have settled on a specific use-case, namely [years for
citations](https://www.gwern.net/Subscripts.html), I propose a different one:
reporting exp... | https://www.lesswrong.com/posts/Tmz6ucxDFsdod2QLd/subscripts-for-probabilities |
# R0 Is Not Counterfactual
Someone who helps organize an [N95-required dance](https://www.jefftk.com/p/boston-social-dance-covid-requirements) recently wrote to a group of organizers:
> R0 is the number of people that an infected individual is likely to infect. R0 was [5.4](https://www.ohsu.edu/sites/default/files/20... | https://www.lesswrong.com/posts/sGPF8Leuiby4y4A7K/r0-is-not-counterfactual |
# A freshman year during the AI midgame: my approach to the next year
I recently spent some time reflecting on my career and my life, for a few reasons:
* It was my 29th birthday, an occasion which felt like a particularly natural time to think through what I wanted to accomplish over the course of the next year 🙂... | https://www.lesswrong.com/posts/mfFn46AEiBL5EzaXr/a-freshman-year-during-the-ai-midgame-my-approach-to-the |
# Shapley Value Attribution in Chain of Thought
**TL;DR**: Language models sometimes seem to ignore parts of the chain of thought, and larger models appear to do this more often. Shapley value attribution is a possible approach to get a more detailed picture of the information flow within the chain of thought, though ... | https://www.lesswrong.com/posts/FX5JmftqL2j6K8dn4/shapley-value-attribution-in-chain-of-thought |
# AI Safety Europe Retreat 2023 Retrospective
This is a short impression of the AI Safety Europe Retreat (AISER) 2023 in Berlin.
**Tl;dr:** 67 people working on AI safety technical research, AI governance, and AI safety field building came together for three days to learn, connect, and make progress on AI safety.
We have a system “A” which we a... | https://www.lesswrong.com/posts/9GyniEBaN3YYTqZXn/the-self-unalignment-problem |
# Research Report: Incorrectness Cascades
*\[Edit: See updated version* [*here*](https://aizi.substack.com/p/research-report-incorrectness-cascades-259) *(*[*LW version*](https://www.lesswrong.com/posts/8fWuAYicWkqqmcGyt/research-report-incorrectness-cascades-corrected)*), which fixed a crucial bug and significantly c... | https://www.lesswrong.com/posts/M5vEjix8oPeWXeGFY/research-report-incorrectness-cascades |
# What would the FLI moratorium actually do?
I read a lot of agreement with the six months memorandum and very little discussion of the details and what the proposed memorandum would actually do.
The memorandum says 'we call on all AI labs to immediately pause for at least 6 months the training of AI systems more pow... | https://www.lesswrong.com/posts/n7wpRYdo3Zh3Cbhja/what-would-the-fli-moratorium-actually-do |
# Navigating AI Risks (NAIR) #1: Slowing Down AI
Here’s the first edition (on slowdown) of **Navigating AI Risks**, a newsletter on AI governance that we're launching with some colleagues.
This newsletter is mostly aimed at **policymakers** but I expect that it might interest some of you to keep up with the ideas tha... | https://www.lesswrong.com/posts/C3rWCvGJnpYoFvKcC/navigating-ai-risks-nair-1-slowing-down-ai |
# What’s this probability you’re reporting?
It’s unclear what people mean when saying they’re reporting a probability *according to their inside view model(s)*. We’ll look through what this could mean and why most interpretations are problematic. Note that we’re not making claims about which communication norms are so... | https://www.lesswrong.com/posts/WxkRgCTbFrbquzyfS/what-s-this-probability-you-re-reporting-1 |
# Iqisa: A Library For Handling Forecasting Datasets
_cross-posted from [niplav.github.io](https://niplav.github.io)_
> Iqisa is [a library](https://github.com/niplav/iqisa) for handling and comparing forecasting datasets from different platforms.
Iqisa: A Library For Handling Forecasting Datasets
==================... | https://www.lesswrong.com/posts/CPJeoLZiqbzKtxLYF/iqisa-a-library-for-handling-forecasting-datasets |
# What we’ve learned so far from our technological temptations project
_Note: This was automatically imported from the [AI Impacts blog](https://blog.aiimpacts.org/p/what-weve-learned-so-far-from-technological) and neither the footnotes nor the images imported properly. For now, I recommend reading it on our site._
A... | https://www.lesswrong.com/posts/Ccv8PinXRgRTKpGaj/what-we-ve-learned-so-far-from-our-technological-temptations |
# Moderation notes re: recent Said/Duncan threads
*Update: Ruby and I have posted moderator notices for Duncan and Said in this thread. This was a set of fairly difficult moderation calls on established users and it seems good for the LessWrong userbase to have the opportunity to evaluate it and respond. I'm stickying... | https://www.lesswrong.com/posts/9DhneE5BRGaCS2Cja/moderation-notes-re-recent-said-duncan-threads |
# On Caring about our AI Progeny
I encourage you to interact with GPT as you would interact with a
friend, or as you would want your employer to treat you.
Treating other minds with respect is typically not costly. It can easily
improve your state of mind relative to treating them as an adversary.
The tone you... | https://www.lesswrong.com/posts/euGjFtgyXSCAm6ddZ/on-caring-about-our-ai-progeny |
# "Do X because decision theory" ~= "Do X because bayes theorem"
Decision theory is an extremely low-level tool for governing your interactions with other people, in the same way that physics is an extremely low-level tool for winning knife fights. Actual human behavior usually involves complex multi-agent interaction... | https://www.lesswrong.com/posts/KDzykLCYMfWiRiWnd/do-x-because-decision-theory-do-x-because-bayes-theorem |
# [linkpost] "What Are Reasonable AI Fears?" by Robin Hanson, 2023-04-23
Selected quotes (all emphasis mine):
> Why are we so willing to “other” AIs? Part of it is probably prejudice: some recoil from the very idea of a metal mind. We have, after all, long speculated about possible future conflicts with robots. But p... | https://www.lesswrong.com/posts/mnCDGMtk4NS7ojgcM/linkpost-what-are-reasonable-ai-fears-by-robin-hanson-2023 |
# List of requests for an AI slowdown/halt.
**Last updated: April 14th 2023.**
1. [Pause Giant AI Experiments: An Open Letter](https://futureoflife.org/open-letter/pause-giant-ai-experiments/)
by [Future of Life Institute](https://futureoflife.org/)
2. [Pausing AI Developments Isn't Enough. We Need to Shut it... | https://www.lesswrong.com/posts/a87uzervYEYJ8pCgk/list-of-requests-for-an-ai-slowdown-halt |
# Concave Utility Question
This post will just be a concrete math question. I am interested in this question because I have recently come tor reject the independence axiom of VNM, and am thus playing with some weaker versions.
Let $\Omega$ be a finite set of deterministic outcomes. Let $\mathcal{L}$ be the space of a... | https://www.lesswrong.com/posts/uJnR4YmG5Kq9FfTey/concave-utility-question |
# The ‘ petertodd’ phenomenon
**Note:**Nothing in this post is intended to vilify or defame any actual human named Peter Todd. I recognise that the existence of this phenomenon may be an uncomfortable fact for anyone with that name and do not wish to cause anyone discomfort. However, I feel that this phenomenon cannot... | https://www.lesswrong.com/posts/jkY6QdCfAXHJk3kea/the-petertodd-phenomenon |
# Reflective journal entries using GPT-4 and Obsidian that demand less willpower.
Something I've wanted to do for years, but can't make myself do consistently because I'm bad at forming new habits and seem to have a limited reservoir of willpower that is generally depleted by the end of the day:
* Write a short jou... | https://www.lesswrong.com/posts/nimwLGuMxvbkbzcJD/reflective-journal-entries-using-gpt-4-and-obsidian-that |
# FLI report: Policymaking in the Pause
> **this policy brief provides policymakers with concrete recommendations for how governments can manage AI risks.**
> **Policy recommendations:**
> 1\. Mandate robust third-party auditing and certification.
> 2\. Regulate access to computational power.
> 3\. Establish ca... | https://www.lesswrong.com/posts/ERsbTthnzWLmNDCb5/fli-report-policymaking-in-the-pause |
# [linkpost] Elon Musk plans AI start-up to rival OpenAI
A few quotes:
> Musk incorporated a company named X.AI on March 9, according to Nevada business records.
> For the new project, Musk has secured thousands of high-powered GPU processors from Nvidia, said people with knowledge of the move.
> Musk is recruiting... | https://www.lesswrong.com/posts/NPw6ZGo2u8zX9YW6o/linkpost-elon-musk-plans-ai-start-up-to-rival-openai |
# SmartyHeaderCode: anomalous tokens for GPT3.5 and GPT-4
**TL;DR:** There are anomalous tokens for GPT3.5 and GPT4 which are difficult or impossible for the model to repeat; try playing around with `SmartyHeaderCode`, `APolynomial`, or `davidjl`. There are also plenty which can be repeated but are difficult for the ... | https://www.lesswrong.com/posts/ChtGdxk9mwZ2Rxogt/smartyheadercode-anomalous-tokens-for-gpt3-5-and-gpt-4-1 |
# Features of Emacs that I only recently discovered
I am an Emacs user of several years, but only recently discovered some features that I find incredibly convenient. I am somewhat embarrassed by my ignorance (and for that reason write this post anonymously), but suspect that others might not be familiar with these fe... | https://www.lesswrong.com/posts/Hk7GuAGA6NJgYJ3GN/features-of-emacs-that-i-only-recently-discovered |
# Who is testing AI Safety public outreach messaging?
Does anyone know who is testing AI Safety public outreach messaging? Or where I can find relevant research on this?
Things that would be interesting to test:
1. Salience (which messages are remembered and correctly attributed)
2. Persuasion (which messages get ... | https://www.lesswrong.com/posts/aKRpKtwuaghKxHnzj/who-is-testing-ai-safety-public-outreach-messaging |
# Double-negation as framing
Any claim can be presented as positive or negative via the law of double negation. Not not P is P. This is only true for as long as you work within the axioms of formal logic. Disputes amongst humans do not abide by these axioms. Amongst humans, not not P is a very different proposition to... | https://www.lesswrong.com/posts/afAm4cdYdYoqnYToX/double-negation-as-framing |
# Possibilizing vs. actualizing
*[Metadata: crossposted from [https://tsvibt.blogspot.com/2022/12/possibilizing-vs-actualizing.html](https://tsvibt.blogspot.com/2022/12/possibilizing-vs-actualizing.html). First completed December 31, 2022. This essay is more like research notes than exposition, so context may be missi... | https://www.lesswrong.com/posts/5uyRB4CGAvB2eLHvm/possibilizing-vs-actualizing |
# The Soul of the Writer (on LLMs, the psychology of writers, and the nature of intelligence)
> “GPT is not like us — but it is of us. It has read more of humanity's writing than any human, ever, by orders of magnitude. All of its behaviors, good and bad, hold up a mirror to the human soul.”
>
> — “[Sam Altman asks: ... | https://www.lesswrong.com/posts/FBAwthqMY2N6JH9nA/the-soul-of-the-writer-on-llms-the-psychology-of-writers-and |
# Mechanistically interpreting time in GPT-2 small
*This work was performed by Rhys Gould, Elizabeth Ho, Will Harpur-Davies, Andy Zhou, and supervised by* [*Arthur Conmy*](https://www.lesswrong.com/users/arthur-conmy)*.*
*This has been crossposted from* [*medium*](https://medium.com/hackbridge/understanding-time-in-... | https://www.lesswrong.com/posts/6tHNM2s6SWzFHv3Wo/mechanistically-interpreting-time-in-gpt-2-small |
# Top lesson from GPT: we will probably destroy humanity "for the lulz" as soon as we are able.
Forget complicated "sharp left turn" schemes, nefarious nanobots, lists of lethalities, out-of-distribution actions, failed AI boxing. As Zvi pointed out in multiple posts, like [this one](https://thezvi.wordpress.com/2023/... | https://www.lesswrong.com/posts/RJEWuHZBr85RMYRp4/top-lesson-from-gpt-we-will-probably-destroy-humanity-for |
# My experience getting funding for my biological research
*Linked from my Substack*
Research is expensive, especially these days. Even though grad students and postdocs are underpaid, it still costs a lot to employ them.[^7ix50gfe8lq]
And every year, lab suppliers increase their prices. Overall, it continually gets... | https://www.lesswrong.com/posts/zkvNeRRc3gL9d5SPQ/my-experience-getting-funding-for-my-biological-research |
# AI Takeover Scenario with Scaled LLMs
Introduction
============
In this AI takeover scenario, I tried to start from a few technical assumptions and imagine a specific takeover scenario from human-level AI.
I found this exercise very useful to identify new insights on safety. For instance, I hadn't clearly in mind... | https://www.lesswrong.com/posts/PdooAsNFiohmyburK/ai-takeover-scenario-with-scaled-llms |
# Summaries of top forum posts (27th March to 16th April)
This is part of a weekly series summarizing the top posts on the EA and LW forums - you can see the full collection [here.](https://forum.effectivealtruism.org/s/W4fhpuN26naxGCBbN) The first post includes some details on purpose and methodology. Feedback, thoug... | https://www.lesswrong.com/posts/HEmMnQumNrwiB3eaw/summaries-of-top-forum-posts-27th-march-to-16th-april |
# grey goo is unlikely
> The nanomachinery builds diamondoid bacteria, that replicate with solar power and atmospheric CHON, maybe aggregate into some miniature rockets or jets so they can ride the jetstream to spread across the Earth's atmosphere, get into human bloodstreams and hide, strike on a timer.
— [Eliezer Y... | https://www.lesswrong.com/posts/FijbeqdovkgAusGgz/grey-goo-is-unlikely |
# Making Booking.Com less out to get you
Some guy loved booking.com but hated the false urgency they added to the process, so he wrote a plug-in to block the awful elements and reframe the useful ones to grab less attention. I like this both as a useful free thing, and for the general vibe of fixing things and sharin... | https://www.lesswrong.com/posts/Bmkh4JXgpiSpBr8z2/making-booking-com-less-out-to-get-you |
# AutoBound on neural network can achieve OOMs lower training loss
Seems like gradient descent methods weren't using the relevant math bounds so far. Google released AutoBound as an [open-source library](https://github.com/google/autobound).
Here is what I consider a money shot of the article (notice it's a log-plot... | https://www.lesswrong.com/posts/HYohLwwdLnsh9asFR/autobound-on-neural-network-can-achieve-ooms-lower-training |
# Prediction: any uncontrollable AI will turn earth into a giant computer
It is a common assumption that a hypothetical superintelligent [squiggle maximizer](https://www.lesswrong.com/tag/squiggle-maximizer-formerly-paperclip-maximizer) would turn all of earth into squiggles, destroying humanity and all life in the pr... | https://www.lesswrong.com/posts/jkaLGoNLdsp654KhD/prediction-any-uncontrollable-ai-will-turn-earth-into-a |
# Goodhart's Law inside the human mind
Some time back, I saw [a tweet](https://twitter.com/danwilliamsphil/status/1648028581228937216) from somebody that read:
> Much of social psychology seems to be premised on the bizarre assumption that what people really care about is not real-world outcomes but the state of thei... | https://www.lesswrong.com/posts/NQwnXc3xuoqtw5DQv/goodhart-s-law-inside-the-human-mind |
# Slowing AI: Reading list
This is a list of past and present research that could inform slowing AI. It is roughly sorted in descending order of priority, between and within subsections. I've read about half of these; I don't necessarily endorse them. Please have a low bar to suggest additions, replacements, rearrange... | https://www.lesswrong.com/posts/d8WpJbjhn2Yi6kfmM/slowing-ai-reading-list |
# Slowing AI: Foundations
**Variables**
=============
**"Slowing AI" is not monolithic. Many interventions have relevant side effects, positive or negative. There are many considerations, desiderata, and goals adjacent to slowing AI,** and different interventions realize them to different degrees**.** This section is... | https://www.lesswrong.com/posts/MoLLqFtMup39PCsaG/slowing-ai-foundations |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.