text stringlengths 300 320k | source stringlengths 52 154 |
|---|---|
# Causal Scrubbing: a method for rigorously testing interpretability hypotheses [Redwood Research]
*\* Authors sorted alphabetically.*
Summary: This post introduces causal scrubbing, a principled approach for evaluating the quality of mechanistic interpretations. The key idea behind causal scrubbing is to test int... | https://www.lesswrong.com/posts/JvZhhzycHu2Yd57RN/causal-scrubbing-a-method-for-rigorously-testing |
# Causal scrubbing: Appendix
*\* Authors sorted alphabetically.*
*An appendix to* [*this post*](https://www.lesswrong.com/posts/JvZhhzycHu2Yd57RN/causal-scrubbing-redwood-research)*.*
1 More on Hypotheses
====================
1.1 Example behaviors
---------------------
As mentioned above, our method allows us t... | https://www.lesswrong.com/posts/kcZZAsEjwrbczxN2i/causal-scrubbing-appendix |
# Causal scrubbing: results on a paren balance checker
*\* Authors sorted alphabetically.*
*This is a more detailed look at our work applying* [*causal scrubbing*](https://www.lesswrong.com/posts/JvZhhzycHu2Yd57RN/causal-scrubbing-redwood-research) *to an algorithmic model. The results are also summarized* [*here*... | https://www.lesswrong.com/posts/kjudfaQazMmC74SbF/causal-scrubbing-results-on-a-paren-balance-checker |
# Causal scrubbing: results on induction heads
*\* Authors sorted alphabetically.*
*This is a more detailed look at our work applying* [*causal scrubbing*](https://www.lesswrong.com/posts/JvZhhzycHu2Yd57RN/causal-scrubbing-redwood-research) *to induction heads. The results are also summarized* [*here*](https://www... | https://www.lesswrong.com/posts/j6s9H9SHrEhEfuJnq/causal-scrubbing-results-on-induction-heads |
# Great Cryonics Survey of 2022
The text that follows is a copy-paste from the survey creator's Facebook post, [shared with permission](https://www.facebook.com/groups/cryonicists/posts/2170336949839619/?comment_id=2171419116398069&reply_comment_id=2171420566397924&__cft__[0]=AZVOLUMeCIIY9wPXVNwhgSVLAUevOtQhsBqOiA-a6... | https://www.lesswrong.com/posts/tMJxKcZdXoB6B2TBF/great-cryonics-survey-of-2022 |
# MrBeast's Squid Game Tricked Me
I recently watched *Squid Game*—both the original Netflix series and [MrBeast's real life version](https://www.youtube.com/watch?v=0e3GPea1Tyg). Then I watched every video I could find about how MrBeast's Squid Game was made. I was surprised to learn that MrBeast used a whole bunch of... | https://www.lesswrong.com/posts/JMCazCdRndG45wiJt/mrbeast-s-squid-game-tricked-me |
# Is school good or bad?
Should society eliminate schools? Should we have more compulsory schooling? Should you send your kids to school? Should you prefer to hire job candidates who have received more schooling, beyond school's correlation with the g factor? Should we consider the spread of education requirements to ... | https://www.lesswrong.com/posts/HDyFTn5ZjCd2fm8B6/is-school-good-or-bad |
# Our 2022 Giving
[Julia](https://juliawise.net/) and I usually decide on our annual giving at the end of each year. For the past few years this has been relatively straightforward: I was paid [a lot](https://www.jefftk.com/money), so we [donated 50%](https://www.jefftk.com/donations). This year is less so, because ha... | https://www.lesswrong.com/posts/XTFJDXLvqyvJZuHsN/our-2022-giving |
# Logical induction for software engineers
*This work was supported by the Monastic Academy for the Preservation of Life on Earth and the Long Term Future Fund.*
[](https://storage.googleapis.com/doc-publisher-images/a62c1f5e1de9ab69.jpg)
... | https://www.lesswrong.com/posts/jtMXj24Masrnq3SpS/logical-induction-for-software-engineers |
# Summary of a new study on out-group hate (and how to fix it)
This is a summary of the 2022 paper [Individuals prefer to harm their own group rather than help an opposing group](https://www.pnas.org/doi/abs/10.1073/pnas.2215633119). I spent about an hour reading the study and writing the post as I went. If I made mis... | https://www.lesswrong.com/posts/xeApyAbgSChnM2CoW/summary-of-a-new-study-on-out-group-hate-and-how-to-fix-it |
# Take 3: No indescribable heavenworlds.
*As a writing exercise, I'm writing an AI Alignment Hot Take Advent Calendar - one new hot take, written every day for 25 days. Or until I run out of hot takes.*
Some people think as if there are indescribable heavenworlds. They're wrong, and this is important to AI alignment.... | https://www.lesswrong.com/posts/xpGFA7bdoiNDTut8C/take-3-no-indescribable-heavenworlds |
# ChatGPT seems overconfident to me
I've talked to ChatGPT about physics, a topic I don't know much about, except that a few things like air lift, kinetic vs static friction, osmosis, and the stuff listed in [https://xkcd.com/2682/](https://xkcd.com/2682/) (how paracetamol/tylenol works? why baloons electrify hair?) a... | https://www.lesswrong.com/posts/28XBkxauWQAMZeXiF/chatgpt-seems-overconfident-to-me |
# Race to the Top: Benchmarks for AI Safety
*This is an executive summary of a post from my personal blog, also* [*cross-posted*](https://forum.effectivealtruism.org/posts/saEXX9Nucz8mh9XgB/race-to-the-top-benchmarks-for-ai-safety) *from the EA Forum. **Read the full texts*** [***here***](https://isaduan.github.io/isa... | https://www.lesswrong.com/posts/KQ6fGiPeMnzzC6p9q/race-to-the-top-benchmarks-for-ai-safety |
# Historical Notes on Charitable Funds
In the EA movement people will sometimes talk about charitable funds as if they are a new idea. For example, the recent [Giving What We Can](https://www.givingwhatwecan.org/) post " [why we recommend using expert-led charitable funds](https://www.givingwhatwecan.org/why-we-recomm... | https://www.lesswrong.com/posts/J2vBCAiDcfNe4vyX3/historical-notes-on-charitable-funds |
# Probably good projects for the AI safety ecosystem
At [EAGxBerkeley 2022](https://www.eaglobal.org/events/eagxberkeley2022/), I was asked several times what new projects might benefit the AI safety and longtermist research ecosystem. I think that several existing useful-according-to-me projects (e.g., [SERI MATS](ht... | https://www.lesswrong.com/posts/v5z6rDuFPKM5dLpz8/probably-good-projects-for-the-ai-safety-ecosystem |
# Nook Nature
Style of post: Alexanderian meander (though with (at best) a tenth of the artistry and thoroughness/rigor). There's a Thing here and it has been tickling at my brain for ages and I have despaired of being able to clarify it all on my own, so I'm giving you my unclarified thoughts instead. More of an exte... | https://www.lesswrong.com/posts/2EjWQhqh27PX8Nbgh/nook-nature |
# A Tentative Timeline of The Near Future (2022-2025) for Self-Accountability
I don’t currently have the energy to set up an account on a prediction market (yes, I am remarkably lazy in some ways), but nonetheless want to have my personal predictions publicly available somewhere. This is primarily intuition-based, and... | https://www.lesswrong.com/posts/oqJpxY2yZXg52QomP/a-tentative-timeline-of-the-near-future-2022-2025-for-self |
# Who are some prominent reasonable people who are confident that AI won't kill everyone?
**Bounty \[closed\]:** $30 for each link that leads to me reading/hearing ~500 words from a Respectable Person arguing, roughly, "accelerating AI capabilities isn't bad," and me subsequently thinking "yeah, that seemed pretty rea... | https://www.lesswrong.com/posts/rCZ9fruWriD6uGNLp/who-are-some-prominent-reasonable-people-who-are-confident |
# Take 4: One problem with natural abstractions is there's too many of them.
*As a writing exercise, I'm writing an AI Alignment Hot Take Advent Calendar - one new hot take, written every day for 25 days. Or until I run out of hot takes.*
Everyone knows what the deal with [natural abstractions](https://www.lesswrong.... | https://www.lesswrong.com/posts/NK4XxyrjFWt83m3dx/take-4-one-problem-with-natural-abstractions-is-there-s-too |
# Is the "Valley of Confused Abstractions" real?
Epistemic Status: Quite confused. Using this short post as a signal for discussion.
In Evan's post about [Chris Olah's views on AGI safety](https://www.alignmentforum.org/posts/X2i9dQQK3gETCyqh2/chris-olah-s-views-on-agi-safety#What_if_interpretability_breaks_down_as_A... | https://www.lesswrong.com/posts/g7rLyjg67iopg9zLD/is-the-valley-of-confused-abstractions-real |
# How should I judge the impact of giving $5k to a family of three kids and two mentally ill parents?
I am giving $5k to a case worker who will buy goods for a family in New Orleans. The mother and father of the family have schizophrenia and do not work. (The father is violent, and does not live with his children.) Th... | https://www.lesswrong.com/posts/EFGEBKZioqtiLimow/how-should-i-judge-the-impact-of-giving-usd5k-to-a-family-of |
# Aligned Behavior is not Evidence of Alignment Past a Certain Level of Intelligence
edit: Several days after posting this and asking for feedback on it someone pointed me to this post: [Does SGD Produce Deceptive Alignment](https://www.lesswrong.com/posts/ocWqg2Pf2br4jMmKA/does-sgd-produce-deceptive-alignment) by Mar... | https://www.lesswrong.com/posts/XFt9ipeezjEqJCuY4/aligned-behavior-is-not-evidence-of-alignment-past-a-certain |
# Foresight for AGI Safety Strategy: Mitigating Risks and Identifying Golden Opportunities
This post is about why I think we should use strategic foresight for AGI governance work. I am writing this without having read all the AGI governance / strategy reports and reading materials. I was planning on putting more effo... | https://www.lesswrong.com/posts/GbXAeq6smRzmYRSQg/foresight-for-agi-safety-strategy-mitigating-risks-and |
# Analysis of AI Safety surveys for field-building insights
Introduction
============
To help grow the pipeline of AI safety researchers, I conducted a project to determine how demographic information (e.g. level of experience, exposure to AI arguments) affects AI researchers’ responses to AI safety. In addition, I e... | https://www.lesswrong.com/posts/4TCdZN2aj8rnuEkbH/analysis-of-ai-safety-surveys-for-field-building-insights |
# Steering Behaviour: Testing for (Non-)Myopia in Language Models
***Authors' Contributions:** Both authors contributed equally to this project as a whole. Evan did the majority of implementation work, as well as the work for writing this post. Megan was more involved at the beginning of the project, and did the major... | https://www.lesswrong.com/posts/BuRt2igbFx9KaB5QG/steering-behaviour-testing-for-non-myopia-in-language-models |
# Updating my AI timelines
Earlier this year [I offered to bet people who had short AI timelines](https://www.lesswrong.com/posts/X3p8mxE5dHYDZNxCm/a-concrete-bet-offer-to-those-with-short-ai-timelines).
While it wasn't my intention to be known as "a long AI timelines guy", I have begun feeling that was how people pe... | https://www.lesswrong.com/posts/sbb9bZgojmEa7Yjrc/updating-my-ai-timelines |
# ChatGPT and Ideological Turing Test
It seems to me that ChatGPT should be able to pass the Ideological Turing Test -- to generate a convincingly looking argument for any side of an issue. It is obvious how to use this ability for evil purposes: write [the bottom line](https://www.lesswrong.com/posts/34XxbRFe54FycoCD... | https://www.lesswrong.com/posts/d2orCacf97xiEHKRk/chatgpt-and-ideological-turing-test |
# The No Free Lunch theorem for dummies
*(I wrote this little explainer when drafting* [*Response to Blake Richards: AGI, generality, alignment, & loss functions*](https://www.lesswrong.com/posts/rgPxEKFBLpLqJpMBM/response-to-blake-richards-agi-generality-alignment-and-loss) *a few months ago, but it got cut from the ... | https://www.lesswrong.com/posts/Dyt2TDdMGHDkXPcpp/the-no-free-lunch-theorem-for-dummies |
# Shh, don't tell the AI it's likely to be evil
If you haven't been online in the past week: [ChatGPT](https://openai.com/blog/chatgpt/) is OpenAI's new fine-tuned language model that is really much better at doing all sorts things that you'd expect. Here's it [debugging some code](https://twitter.com/amasad/status/15... | https://www.lesswrong.com/posts/RAA6oH34eTMYb5CGo/shh-don-t-tell-the-ai-it-s-likely-to-be-evil |
# Verification Is Not Easier Than Generation In General
People who’ve spent a lot of time thinking about [P vs NP](https://en.wikipedia.org/wiki/P_versus_NP_problem) often have the intuition that “verification is easier than generation”. It’s easier to verify a solution to some equations than to find a solution. It’s ... | https://www.lesswrong.com/posts/2PDC69DDJuAx6GANa/verification-is-not-easier-than-generation-in-general |
# Take 5: Another problem for natural abstractions is laziness.
*As a writing exercise, I'm writing an AI Alignment Hot Take Advent Calendar - one new hot take, written every day for 25 days. Or until I run out of hot takes.*
[Soundtrack.](https://www.youtube.com/watch?v=vmDDOFXSgAs)
Natural abstractions are pattern... | https://www.lesswrong.com/posts/uR2uWMD9JGnRnYSeM/take-5-another-problem-for-natural-abstractions-is-laziness |
# EA & LW Forums Weekly Summary (28th Nov - 4th Dec 22')
*Supported by Rethink Priorities*
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 purpos... | https://www.lesswrong.com/posts/WzfCzj4vmaQTG5BwX/ea-and-lw-forums-weekly-summary-28th-nov-4th-dec-22 |
# Things roll downhill
(Response to [this challenge](https://www.lesswrong.com/posts/tD9zEiHfkvakpnNam/a-challenge-for-agi-organizations-and-a-challenge-for-1))
I've read two things recently on similar strategies: [Zvi's post on ChatGPT](https://www.lesswrong.com/posts/RYcoJdvmoBbi5Nax7/jailbreaking-chatgpt-on-releas... | https://www.lesswrong.com/posts/fy4yMgzxjpKHkuYzR/things-roll-downhill |
# MIRI's "Death with Dignity" in 60 seconds.
Suppose that, like Yudkowsky, you really care about humanity surviving this century but you think that nothing you can do has a decent chance of achieving that.
It's an unfortunate fact of human psychology that, when faced with this kind of situation, people will often do ... | https://www.lesswrong.com/posts/evRNWBqJbyanLfp7s/miri-s-death-with-dignity-in-60-seconds |
# The Categorical Imperative Obscures
I've been thinking about the categorical imperative lately and how it obscures more than it illuminates. I'll make the case that the categorical imperative tries to pull a fast one by taking something for granted, and the thing it takes for granted is the thing we care about.
*Di... | https://www.lesswrong.com/posts/8tStfwDJAErztqzCJ/the-categorical-imperative-obscures |
# The Parable of the Crimp
If you watch really proficient rock climbers, you’ll see they can hold themselves up, dozens of feet above the ground, with just the tips of their fingers on the tiniest ledge ... | https://www.lesswrong.com/posts/JXPSWr46W47EMCTQM/the-parable-of-the-crimp |
# Using GPT-Eliezer against ChatGPT Jailbreaking
*This was originally posted on [Aligned AI's blog](https://www.aligned-ai.com/post/chatbot-safety-with-prompt-evaluator-for-chat-gpt); it was ideated and designed by my cofounder and collaborator, Rebecca Gorman.*
**EDIT**: *many of the suggestions below rely on SQL-in... | https://www.lesswrong.com/posts/pNcFYZnPdXyL2RfgA/using-gpt-eliezer-against-chatgpt-jailbreaking |
# Mesa-Optimizers via Grokking
**Summary:** Recent interpretability work on "grokking" suggests a mechanism for a powerful [mesa-optimizer](https://www.lesswrong.com/posts/FkgsxrGf3QxhfLWHG/risks-from-learned-optimization-introduction) to emerge suddenly from a ML model.
**Inspired By:** [A Mechanistic Interpretabili... | https://www.lesswrong.com/posts/fytgZ26AgxmrAdyB4/mesa-optimizers-via-grokking |
# How do finite factored sets compare with phase space?
Garrabrant's [finite factored sets](https://www.lesswrong.com/tag/finite-factored-sets) feel to me like the same thing as a [phase space](https://en.wikipedia.org/wiki/Phase_space) of a dynamical system. The differences I can see are that phase spaces are not alw... | https://www.lesswrong.com/posts/Bw3qsSMozyaHaPDNx/how-do-finite-factored-sets-compare-with-phase-space |
# In defense of probably wrong mechanistic models
_This is a short post on a simple point that I get asked about a lot and want a canonical reference for._
Which of the following two options is more likely to be true?
1. AIs will internally be running explicit search processes.
2. AIs will internally be doing someth... | https://www.lesswrong.com/posts/q5Gox77ReFAy5i2YQ/in-defense-of-probably-wrong-mechanistic-models |
# Simple Way to Prevent Power-Seeking AI
The most concerning part of AGI is that it may have power-seeking tendencies. We can loosely define a powerful state to be a state that is useful to be in regardless of the reward function. A promising approach to AI safety is to make the RL process "interpretable" enough so th... | https://www.lesswrong.com/posts/iozsJQ7fEdTCRxtJc/simple-way-to-prevent-power-seeking-ai |
# ACX Montreal Meetup December 17 2022
**Venue: **
[Olde Orchard pub & grill](https://www.google.com/maps/place/Ye+Olde+Orchard+Pub+%26+Grill/@45.5134983,-73.5727146,17z/data=!3m1!4b1!4m5!3m4!1s0x4cc91a35812fae03:0xf5fdb66587011cc4!8m2!3d45.5135005!4d-73.5727198) at 20 Prince Arthur street W.
Date & Time: December ... | https://www.lesswrong.com/events/XE4B6itGMRN77E4zz/acx-montreal-meetup-december-17-2022 |
# [ASoT] Probability Infects Concepts it Touches
*Epistemic status: Random ranting about things I'm confused about. Hopefully this post makes you more confused about optimization/agency/etc.*
A few parables of counterfactuals
* What if $2+2$ equaled $5$?
* American: Hmm, Yeah. What if giving me two dollars a... | https://www.lesswrong.com/posts/PHDtrdMtzdEivhdXX/asot-probability-infects-concepts-it-touches |
# "Attention Passengers": not for Signs
On the MBTA, when a train is coming you get an announcement like:
> Attention passengers, the next red line train to Ashmont is now arriving.
As with all the announcements, there's a text version:
[](... | https://www.lesswrong.com/posts/9sEek4FtYwXqnyZE4/attention-passengers-not-for-signs |
# Where to be an AI Safety Professor
Stephen Casper, [scasper@mit.edu](mailto:scasper@mit.edu)
TL;DR: being a professor in AI safety seems cool. It might be tractable and high impact to be one at a university that has lots of money, smart people, and math talent but an underperforming CS department – especially if ... | https://www.lesswrong.com/posts/Cg4jCMWcKMWBuqCp2/where-to-be-an-ai-safety-professor |
# Why I like Zulip instead of Slack or Discord
Slack and Discord are skins over the same thing. A bunch of conversations happening at the same time. Threads are made after the fact, and many conversations are braided together, so you have to untangle them.
Zulip has a tiny change: you have to make a conversation have... | https://www.lesswrong.com/posts/R8fADFuJeYo7QxTtH/why-i-like-zulip-instead-of-slack-or-discord |
# [Link] Wavefunctions: from Linear Algebra to Spinors
I wrote this blogpost because I thought it took an excessive amount of digging to understand what a spinor was. My original motivation was to understand wavefunctions more concretely since I recently discovered that wavefunctions are spinor-valued, not (necessaril... | https://www.lesswrong.com/posts/RGxvsky9EmwnBjCyx/link-wavefunctions-from-linear-algebra-to-spinors |
# Take 6: CAIS is actually Orwellian.
*As a writing exercise, I'm writing an AI Alignment Hot Take Advent Calendar - one new hot take, written every day for 25 days. Or until I run out of hot takes.*
CAIS, or [Comprehensive AI Services](https://www.fhi.ox.ac.uk/wp-content/uploads/Reframing_Superintelligence_FHI-TR-20... | https://www.lesswrong.com/posts/CNrz9uy5Y4ELypzca/take-6-cais-is-actually-orwellian |
# Thoughts on AGI organizations and capabilities work
(*Note: This essay was largely written by Rob, based on notes from Nate. It’s formatted as Rob-paraphrasing-Nate because (a) Nate didn’t have time to rephrase everything into his own words, and (b) most of the impetus for this post came from Eliezer wanting MIRI to... | https://www.lesswrong.com/posts/tuwwLQT4wqk25ndxk/thoughts-on-agi-organizations-and-capabilities-work |
# Looking for ideas of public assets (stocks, funds, ETFs) that I can invest in to have a chance at profiting from the mass adoption and commercialization of AI technology
Playing around with ChatGPT made me realize that it is quite likely that the next profitable tech breakthrough will probably come from AI technolog... | https://www.lesswrong.com/posts/3FWNdhSGjHNnor2BG/looking-for-ideas-of-public-assets-stocks-funds-etfs-that-i |
# Riffing on the agent type
_Much is owed to Diffractor for Giry-pilling me at Alignable Structures, I had been struggling with type-driven expected value previously._
_Epistemic status: took a couple days off from my [master plan](https://www.lesswrong.com/posts/HbPXiB5Aiv3uHPXEh/master-plan-spec-needs-audit-logic-a... | https://www.lesswrong.com/posts/jEXdGBpD723DhizAZ/riffing-on-the-agent-type |
# Machine Learning Consent
For years, researchers have trained machine learning systems on whatever data they could find. People mostly haven't cared about this or paid attention, I think because the systems hadn't been very good. Recently, however, some very impressive systems have come out, including ones that [answ... | https://www.lesswrong.com/posts/dqQGrPRYamrtLvdLz/machine-learning-consent |
# Take 7: You should talk about "the human's utility function" less.
*As a writing exercise, I'm writing an AI Alignment Hot Take Advent Calendar - one new hot take, written every day (more or less - I'm getting back on track!) for 25 days. Or until I run out of hot takes.*
When considering AI alignment, you might be... | https://www.lesswrong.com/posts/hrYvdrqMyCnw3pBkd/take-7-you-should-talk-about-the-human-s-utility-function |
# Why I'm Sceptical of Foom
Disclaimer
Written quickly[^3z4chjieyxl]. It's better to draft my objections poorly, than to not draft them at all.
Introduction
============
I am sceptical that "foom"[^e2ppvjc485] is some of not physically possible/feasible/economically viable.
\[Not sure yet what level of scepticism... | https://www.lesswrong.com/posts/jdLmC46ZuXS54LKzL/why-i-m-sceptical-of-foom |
# Covid 12/8/22: Another Winter Wave
Here we go again. It is December, and cases, hospitalizations and wastewater concentrations are all going up. How worried should we be?
We should worry that cases will for a time double or triple from where they are. Combined with flu and RSV, this will mean that while our health ... | https://www.lesswrong.com/posts/cxi8cTvELbMcv5rci/covid-12-8-22-another-winter-wave |
# If Wentworth is right about natural abstractions, it would be bad for alignment
*This post was written as part of the* [*AI safety Mentors and Mentees*](https://www.lesswrong.com/posts/EjX63wQoMSoCHMrmY/announcing-ai-safety-mentors-and-mentees) *program. My Mentor is* [*Jacques Thibodeau*](https://www.lesswrong.com/... | https://www.lesswrong.com/posts/nJHXQWCSByS4SxfQz/if-wentworth-is-right-about-natural-abstractions-it-would-be |
# Neural networks biased towards geometrically simple functions?
Neural networks (NNs) do not output all functions with equal probability, but seem to be biased towards functions of certain types; heuristically, towards 'simple' functions. In [VPCL18](https://arxiv.org/abs/1805.08522), [MSVP+19](https://arxiv.org/abs... | https://www.lesswrong.com/posts/JbuwdnAwPjvCRMFZn/neural-networks-biased-towards-geometrically-simple |
# I've started publishing the novel I wrote to promote EA
Currently I'm just uploading it to Royal Road, but in the next week I'll also post it to the spacebattles forum, and around the start of January I'll publish a copy to Amazon.
So, if you find this project interesting, and use Royal Road, follow it and maybe l... | https://www.lesswrong.com/posts/ND6isoKN38gdmAM8Q/i-ve-started-publishing-the-novel-i-wrote-to-promote-ea |
# Notes on OpenAI’s alignment plan
*This work was supported by the Monastic Academy for the Preservation of Life on Earth.*
---
OpenAI has published a [plan](https://openai.com/blog/our-approach-to-alignment-research/) for producing aligned AI systems. Very roughly, their plan is to build language models capable o... | https://www.lesswrong.com/posts/FTk7ufqK2D4dkdBDr/notes-on-openai-s-alignment-plan |
# AI Safety Seems Hard to Measure
In previous pieces, I argued that there's a real and large risk of AI systems' [developing dangerous goals of their own](https://www.cold-takes.com/why-would-ai... | https://www.lesswrong.com/posts/7gkXuHEm6CqEGT2mg/ai-safety-seems-hard-to-measure |
# Where's the economic incentive for wokism coming from?
I'm trying to understand why several systems (Hollywood, Disney, universities, probably others) that are normally quite profit-focused are leaning so hard woke.
I get the thing about how, if you don't go woke hard enough, you could get canceled, which is bad fo... | https://www.lesswrong.com/posts/AajbPPe4EomHcszkb/where-s-the-economic-incentive-for-wokism-coming-from |
# How is the "sharp left turn defined"?
I've read this article by [Nate Soars](https://www.lesswrong.com/posts/GNhMPAWcfBCASy8e6/a-central-ai-alignment-problem-capabilities-generalization), but I felt the definition wasn't completely crisp and I felt this was less than ideal given how this term seems to have entered w... | https://www.lesswrong.com/posts/hK8PEj3uat8LDSyiw/how-is-the-sharp-left-turn-defined |
# Do You Have an Internal Monologue?
I find the differences in the way people think fascinating, to the point where I went around asking people (n~=40) whether they thought verbally or otherwise when I was in high school. I got about 50-50 responses, and proceeded to try to train myself to think conceptually (from a b... | https://www.lesswrong.com/posts/FuzskJKAMHaJhmq4F/do-you-have-an-internal-monologue |
# Systems of Survival
In this [1992 dialogue](https://smile.amazon.com/Systems-Survival-Di... | https://www.lesswrong.com/posts/LYgCJ85A68wmkD8yL/systems-of-survival |
# Setting the Zero Point
*Context: I have hopes of assembling a full "defense against the dark arts" sequence. This essay will not necessarily be the very first one in the sequence, but it's One Of The Basic Spells To Defend Against, and it's the one I happen to have done the most data gathering on so it's getting wri... | https://www.lesswrong.com/posts/Gh6LWHd9kR5FBmCWx/setting-the-zero-point |
# ML Safety at NeurIPS & Paradigmatic AI Safety? MLAISU W49
Watch this week’s episode on [YouTube](https://youtu.be/hcK8z1O62gk) or listen to the audio version [here](https://podcast.apartresearch.com/15).
This week, we see how to break ChatGPT, how to integrate diverse opinions in an AI and look at a bunch of the mo... | https://www.lesswrong.com/posts/3cgevkQRAjSPdynJw/ml-safety-at-neurips-and-paradigmatic-ai-safety-mlaisu-w49 |
# You can still fetch the coffee today if you're dead tomorrow
"[You can't fetch the coffee if you're dead](https://arbital.com/p/no_coffee_if_dead/)."
—Stuart Russell, on the [instrumental convergence](https://arbital.com/p/instrumental_convergence/) of [shutdown-avoidance](https://arbital.com/p/shutdown_problem/)
... | https://www.lesswrong.com/posts/dzDKDRJPQ3kGqfER9/you-can-still-fetch-the-coffee-today-if-you-re-dead-tomorrow |
# Monthly Roundup #1
When I split entirely non-medical, non-Covid topics off from the weekly Covid post, I expected to have a grab bag of them on a weekly basis. Hence the name ‘weekly roundup.’
What changed were four things.
1. The flow of worthy one-off items decreased and they seemed less time sensitive.
2. The... | https://www.lesswrong.com/posts/eqt6awLqxtPe9yGNC/monthly-roundup-1 |
# Does a LLM have a utility function?
There's a lot of discussion and research into AI alignment, almost always about variants of how to define/create a utility function (or meta-function, if it changes over time) that is actually aligned with ... something. That something is at least humanity's survival, but often s... | https://www.lesswrong.com/posts/xxNRh9tfrdeY4ujBr/does-a-llm-have-a-utility-function |
# Take 8: Queer the inner/outer alignment dichotomy.
*As a writing exercise, I'm writing an AI Alignment Hot Take Advent Calendar - one new hot take, written every day for 25 days. Or until I run out of hot takes.*
I'm not saying to never say "inner alignment." But you had better be capable of not using that framing ... | https://www.lesswrong.com/posts/dyg4KSyMJC8cNDMK6/take-8-queer-the-inner-outer-alignment-dichotomy |
# Prosaic misalignment from the Solomonoff Predictor
When I first read [Paul Christiano's post](https://ordinaryideas.wordpress.com/2016/11/30/what-does-the-universal-prior-actually-look-like/), I figured it had little relevance to prosaic alignment. But is that true? Is Solomonoff misalignment a problem that could ac... | https://www.lesswrong.com/posts/GfFvsPaSFG7wqY4sk/prosaic-misalignment-from-the-solomonoff-predictor |
# [Linkpost] The Story Of VaccinateCA
Linked is Patrick MacKenzie's writeup of VaccinateCA, the nonprofit that succeeded at creating the best source of vaccine availability data when the government completely dropped the ball. It's long, but one of the best things I've read in a while. Some highlights:
> We found it ... | https://www.lesswrong.com/posts/pt8uKGHArYH33eD4D/linkpost-the-story-of-vaccinateca |
# A dynamical systems primer for entropy and optimization
*This is a reference post for my (in-progress) sequences on entropy and optimization. It is intended to be used by readers to catch up on this concept in case it's unfamiliar to them. It is absolutely not intended as a stand-alone introduction to the topic; tha... | https://www.lesswrong.com/posts/X3uxRXHonFYTu6zjW/a-dynamical-systems-primer-for-entropy-and-optimization |
# patio11's "Observations from an EA-adjacent (?) charitable effort"
[patio11](https://www.twitter.com/patio11), following up on his [longer-form article detailing the story of VaccinateCA](https://www.worksinprogress.co/issue/the-story-of-vaccinateca/) ([linked on LW](https://www.lesswrong.com/posts/pt8uKGHArYH33eD4D... | https://www.lesswrong.com/posts/dYwb3FoM2o3kmCESB/patio11-s-observations-from-an-ea-adjacent-charitable-effort |
# My thoughts on OpenAI's Alignment plan
This is my attempt at Eliezer's challenge.
[https://www.lesswrong.com/posts/tD9zEiHfkvakpnNam/a-challenge-for-agi-organizations-and-a-challenge-for-1](https://www.lesswrong.com/posts/tD9zEiHfkvakpnNam/a-challenge-for-agi-organizations-and-a-challenge-for-1)
My overall impress... | https://www.lesswrong.com/posts/3oNZA9wTrFJRH6Sau/my-thoughts-on-openai-s-alignment-plan |
# Will Manifold Markets/Metaculus have built-in support for reflective latent variables by 2025?
Prediction markets and similar systems are currently nice for soliciting predictions for outcomes where there is a clear, unambiguous objective resolution criterion. However, many phenomena in the real world are hard to di... | https://www.lesswrong.com/posts/HygyWpDcwsekqmfnS/will-manifold-markets-metaculus-have-built-in-support-for |
# [ASoT] Natural abstractions and AlphaZero
I just read [Acquisition of Chess Knowledge in AlphaZero](https://arxiv.org/abs/2111.09259) and it's both really cool and has interesting implications for the Natural Abstractions Hypothesis. AZ was trained with *no human data* and yet it settles on relatively interpretable ... | https://www.lesswrong.com/posts/obht9QqMDMNLwhPQS/asot-natural-abstractions-and-alphazero |
# Reflecting on the 2022 Guild of the Rose Workshops
Overview
--------
The [Guild of the Rose](https://guildoftherose.org/) is a community of learners dedicated to personal growth and development, often with a rationalist bent. In 2022, the Guild offered a variety of workshops to help its members hone their skills an... | https://www.lesswrong.com/posts/DsebrXgvgWxEib8jB/reflecting-on-the-2022-guild-of-the-rose-workshops |
# Poll Results on AGI
The poll has now settled quite a bit. I have been a bit busy and so it took me a bit longer than expected to write this up – but without further ado let us look at some of the results. Over the past month 74 people voted, 70 people were grouped, and 2,095 votes were cast, with 55 statements submi... | https://www.lesswrong.com/posts/NdnToPRLnqGnX8kL8/poll-results-on-agi |
# Consider using reversible automata for alignment research
In recent years, there have been several cases of alignment researchers using Conway's Game of Life as a research environment;
* [Introducing SafeLife: Safety Benchmarks for Reinforcement Learning](https://partnershiponai.org/safelife/) (Wainwright, Eckers... | https://www.lesswrong.com/posts/2shKBroTQKvuLzvAX/consider-using-reversible-automata-for-alignment-research |
# [fiction] Our Final Hour
*EtA: This is fictional, and symbolism; not a precise prediction. Feedback appreciated.*
As intelligence is bottled and amplified, we'll all become very wealthy. Doing our job will mostly become about [providing more training data for the AI](https://www.lesswrong.com/posts/midXmMb2Xg37F2Kg... | https://www.lesswrong.com/posts/zezmMfc6P3wpM3o4t/fiction-our-final-hour |
# Using Obsidian if you're used to using Roam
TL;DR: I used Roam for a long time and found it super helpful, now I want this same useful functionality but with the privacy of local storage only.[^7zd8i475v5w] I was leaving a lot of effectiveness on the table by not doing this sooner, so hopefully by reading this, some... | https://www.lesswrong.com/posts/pEdaa2ACg9tiRcinh/using-obsidian-if-you-re-used-to-using-roam |
# Reframing inner alignment
The standard frame ([Evan Hubinger, 2021](https://www.alignmentforum.org/posts/FDJnZt8Ks2djouQTZ/how-do-we-become-confident-in-the-safety-of-a-machine)) is:
> * **Outer alignment** refers to the problem of finding a loss/reward function such that the training goal of “a model that optimi... | https://www.lesswrong.com/posts/xMQ7vwFACQX3gZouv/reframing-inner-alignment |
# Formalization as suspension of intuition
While reading [Bachelard](https://en.wikipedia.org/wiki/Gaston_Bachelard) (one of the greatest philosophers of science of all time), I fell upon this fascinating passage:[^lyh0a23hdhp]
> From now on an axiomatic accompanies the scientific process. We have written the accompa... | https://www.lesswrong.com/posts/KiWY2EK9fLyoAebFf/formalization-as-suspension-of-intuition |
# Finite Factored Sets in Pictures
[Finite factored sets](https://www.lesswrong.com/posts/N5Jm6Nj4HkNKySA5Z/finite-factored-sets) are a new paradigm for talking about causality. You can use them to do some cool things you can’t do with Pearl’s [causal graphs](https://en.wikipedia.org/wiki/Causal_graph), for example [i... | https://www.lesswrong.com/posts/PfcQguFpT8CDHcozj/finite-factored-sets-in-pictures-6 |
# A crisis for online communication: bots and bot users will overrun the Internet?
The endgame for humanity's AI adventure still looks to me, to be what happens upon the arrival of comprehensively superhuman artificial intelligence.
However, the rise of large language models, and the availability of ChatGPT in part... | https://www.lesswrong.com/posts/C3SRSfRojwCzaPW2b/a-crisis-for-online-communication-bots-and-bot-users-will |
# Reflections on the PIBBSS Fellowship 2022
Last summer, we ran the first iteration of the PIBBSS Summer Research Fellowship. In this post, we share some reflections on how the program went.
Note that this post deals mostly with high-level reflections and isn’t maximally comprehensive. It primarily focusses on inform... | https://www.lesswrong.com/posts/gbeyjALdjdoCGayc6/reflections-on-the-pibbss-fellowship-2022 |
# A brainteaser for language models
I came up with the following puzzle the other day:
Q: Solve the puzzle: 63 = x = 65536
A: x =
The intended answer is in the form of a number.
`text-davinci-003` guesses my intended answer at 11.8% probability, which is the second-highest probability for any answer.... | https://www.lesswrong.com/posts/bPewMeJiGxNkNNfX2/a-brainteaser-for-language-models |
# Wider Default Audio Player in Chrome?
My kids like to listen to story tapes. Most of them their great grandmother recorded onto cassettes in the 1980s, their grandmother digitized in the 2000s, and I reworked for the web in the 2020s. We usually use either a tablet or old phone to play them back, and while the brows... | https://www.lesswrong.com/posts/GKz7aKDwGWmoyNCbw/wider-default-audio-player-in-chrome |
# Prodding ChatGPT to solve a basic algebra problem
TL;DR: It behaves amazingly like a human would!
It understands the word problem, but struggles to put it in equations, takes multiple tries, screws up again, eventually muddles through to get the right answer. Given a similar problem it does slightly better, but st... | https://www.lesswrong.com/posts/nMnPJFC84HvfMxt7z/prodding-chatgpt-to-solve-a-basic-algebra-problem |
# Ponzi schemes can be highly profitable if your timing is good
Have you ever wondered what you'd find if you upended your life for a few months and did a deep dive trying to understand what makes the crypto ecosystem work?
It just so happens that due to exogenous circumstances mostly beyond my control, that's exactl... | https://www.lesswrong.com/posts/ysAf6RkDGLi7agFuH/ponzi-schemes-can-be-highly-profitable-if-your-timing-is |
# Take 9: No, RLHF/IDA/debate doesn't solve outer alignment.
*As a writing exercise, I'm writing an AI Alignment Hot Take Advent Calendar - one new hot take, written every day (ish) for 25 days. Or until I run out of hot takes. And now, time for the week of RLHF takes.*
I see people say one of these surprisingly ofte... | https://www.lesswrong.com/posts/6YNZt5xbBT5dJXknC/take-9-no-rlhf-ida-debate-doesn-t-solve-outer-alignment |
# Side-channels: input versus output
This is a brief post arguing that, although "side-channels are inevitable" is pretty good common advice, actually, you *can* prevent attackers *inside* a computation from learning about what's outside.
We can prevent a task-specific AI from learning any particular facts about, say... | https://www.lesswrong.com/posts/bqRD6MS3yCdfM9wRe/side-channels-input-versus-output |
# Join the AI Testing Hackathon this Friday
**TLDR**: Participate online or in-person on the weekend of December 16th to 18th in a fun and intense [AI safety research hackathon](https://alignmentjam.com/) focused on benchmarks, neural network verification, adversarial attacks and RL safety environments. We invite mid-... | https://www.lesswrong.com/posts/YdvLsNfmcKKhG8K3E/join-the-ai-testing-hackathon-this-friday |
# Green Line Extension Opening Dates
Talk of extending the Green Line from Lechmere to Ball Square and Tufts goes back [a very long way](https://www.jefftk.com/p/mbta-extensions-planned-and-actual), but it's finally open!
[](https://www.jefftk.com/ball-squ... | https://www.lesswrong.com/posts/jAjjJrmrc8i4b8fog/green-line-extension-opening-dates |
# Tradeoffs in complexity, abstraction, and generality
Complexity, abstraction, and generalization are powerful:
* Complexity lets us explain a wider range of phenomena
* Abstraction allows us to reason more easily by filtering unnecessary noise or details
* Generalization lets us reason over larger domains.
*... | https://www.lesswrong.com/posts/HNJwteaxpRYfLaQt7/tradeoffs-in-complexity-abstraction-and-generality |
# Meaningful things are those the universe possesses a semantics for
Logic courses are very risk-averse to talking about philosophy. This is bad, because philosophy is the motivation/intuition for logic & TCS. This essay gives a scout's view.
*Epistemic status:* A computer scientist would nod through Chapters 2-5, at... | https://www.lesswrong.com/posts/xqxXrAohXSD3akYCg/meaningful-things-are-those-the-universe-possesses-a |
# Confusing the goal and the path
Say you are working on formulating a new scientific theory. You’re not there yet, but you broadly know what you want: a simple theory that powerfully compresses the core phenomenon, and suggests a myriad of new insights.
If you’re anything like me, at least part of you now pushes for... | https://www.lesswrong.com/posts/cpNCJhfDEDAYdGdkf/confusing-the-goal-and-the-path |
# Psychological Disorders and Problems
There are two types of psychological issues: disorders and problems.
Disorders are so massive, so life-altering that we can’t miss them: schizophrenia, crippling social anxiety, destructive addiction… They scream “danger” in neon red blinking letters. We know and fear them becau... | https://www.lesswrong.com/posts/ddd3eBucFN3ZdbFzx/psychological-disorders-and-problems |
# Log-odds are better than Probabilities
\[This is a cross-post from my blog at [aizi.substack.com](https://aizi.substack.com). I'm sure someone has made a point like this before, but I don't know any specific instances and I wanted to give my take on it.\]
At my previous job I worked on ML classifiers, and I learne... | https://www.lesswrong.com/posts/fFYv6sBpPYwGwQ5uz/log-odds-are-better-than-probabilities |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.