The votes are in for the 2023 Review!

6,264 posts were written in 2023

662 of them were nominated.

209 of them got at least one review, and a positive review-vote total.

50 of them shall be displayed in the Best of LessWrong, Year 2023.

Reviews

Exactly 100 people wrote reviews, and many of them I found particularly valuable. I want to give some particular shout outs to Ryan Greenblatt, John Wentworth, and Steve Byrnes.

I found Ryan Greenblatt's reviews to be sort of relentlessly reasonable. He reviewed many posts thoughtfully, clearly state flaws, value props, and concrete potential improvements. I particularly liked his review of Anthropic's Responsible Scaling Policy & Long-Term Benefit Trust

I personally had the most direct "worldview update" from John Wentworth's The Case Against AI Control Research. I don't fully agree with John's take, but I found distinguishing how bad scheming is at different AI stages, and what else needs to be going right at that stage, a quite useful frame.

I was also grateful for Steve Byrnes extensive opinionated review of the “Sharp Left Turn” discourse.

Here's a general screenshot of top of the Review Leaderboard. Thanks to everyone who participated.

And, reminder, reviews now appear on the Spotlight Items, at the top of the home page, and on the /bestoflesswrong page.

Weighted Vote Totals

In past years, we calculated the votes based on two groups of users: Established users with 1000+ karma (who got 3x the vote weight), and users with <1000 karma.

This time, we're replacing that arbitrary cliff with a more granular "your review vote gets multiplied by your Strong Vote power." I mentioned we'd be doing something like that in the announcement post, noting:

I haven't been happy with the arbitrary cliff here. It gives more power than I really wanted to allocate to people with 1000 karma, and not enough weight to people who have been around much longer and have demonstrated good judgment. But, karma is still a pretty messy indicator, so I don't want to give too much power to high karma users either.

Ultimately it seemed that "Strong Vote power" basically did what we wanted. Very high karma users get more voting power, but each additional point takes roughly twice as much karma as the previous point, which limits how extreme the difference can get. Users pretty quickly ramp up to around ~6x voting power. Most longterm users will have somewhere around 8-9x. A few users get as high as 14x.

That said, to avoid de-anonymization of votes, on this post I'm displaying the "raw" vote strength of each vote, before being multiplied.

The Results

Okay. You probably kinda scrolled past all that to get to what you're all here for: what were the best posts of LessWrong 2023, according to you, the people?

408 people voted. 161 people cast the six votes we asked to get the Good Citizen Stamp. Here's what they thought:

0	AI Control: Improving Safety Despite Intentional Subversion Buck
1	Focus on the places where you feel shocked everyone's dropping the ball So8res
2	Significantly Enhancing Adult Intelligence With Gene Editing May Be Possible GeneSmith
3	Social Dark Matter Duncan Sabien (Deactivated)
4	Statement on AI Extinction - Signed by AGI Labs, Top Academics, and Many Other Notable Figures Dan H
5	Pausing AI Developments Isn't Enough. We Need to Shut it All Down Eliezer Yudkowsky
6	Please don't throw your mind away TsviBT
7	How much do you believe your results? Eric Neyman
8	AI Timelines habryka
9	Basics of Rationalist Discourse Duncan Sabien (Deactivated)
10	Things I Learned by Spending Five Thousand Hours In Non-EA Charities jenn
11	What a compute-centric framework says about AI takeoff speeds Tom Davidson
12	How to have Polygenically Screened Children GeneSmith
13	Natural Abstractions: Key claims, Theorems, and Critiques LawrenceC
14	Alignment Implications of LLM Successes: a Debate in One Act Zack_M_Davis
15	Model Organisms of Misalignment: The Case for a New Pillar of Alignment Research evhub
16	Natural Latents: The Math johnswentworth
17	Steering GPT-2-XL by adding an activation vector TurnTrout
18	SolidGoldMagikarp (plus, prompt generation) Jessica Rumbelow
19	The ants and the grasshopper Richard_Ngo
20	Feedbackloop-first Rationality Raemon
21	Deep Deceptiveness So8res
22	EA Vegan Advocacy is not truthseeking, and it’s everyone’s problem Elizabeth
23	Davidad's Bold Plan for Alignment: An In-Depth Explanation Charbel-Raphaël
24	Fucking Goddamn Basics of Rationalist Discourse LoganStrohl
25	Against Almost Every Theory of Impact of Interpretability Charbel-Raphaël
26	Guide to rationalist interior decorating mingyuan
27	New report: "Scheming AIs: Will AIs fake alignment during training in order to get power?" Joe Carlsmith
28	Predictable updating about AI risk Joe Carlsmith
29	[Fiction] A Disneyland Without Children L Rudolf L
30	The Talk: a brief explanation of sexual dimorphism Malmesbury
31	"Carefully Bootstrapped Alignment" is organizationally hard Raemon
32	Tuning your Cognitive Strategies Raemon
33	Enemies vs Malefactors So8res
34	The Parable of the King and the Random Process moridinamael
35	GPTs are Predictors, not Imitators Eliezer Yudkowsky
36	Labs should be explicit about why they are building AGI peterbarnett
37	Cultivating a state of mind where new ideas are born Henrik Karlsson
38	Discussion with Nate Soares on a key alignment difficulty HoldenKarnofsky
39	Loudly Give Up, Don't Quietly Fade Screwtape
40	We don’t trade with ants KatjaGrace
41	Neural networks generalize because of this one weird trick Jesse Hoogland
42	My views on “doom” paulfchristiano
43	Change my mind: Veganism entails trade-offs, and health is one of the axes Elizabeth
44	Lessons On How To Get Things Right On The First Try johnswentworth
45	Shallow review of live agendas in alignment & safety technicalities
46	The Learning-Theoretic Agenda: Status 2023 Vanessa Kosoy
47	Improving the Welfare of AIs: A Nearcasted Proposal ryan_greenblatt
48	Book Review: Going Infinite Zvi
49	Speaking to Congressional staffers about AI risk Zuko's Shame
50	Why it's so hard to talk about Consciousness Rafael Harth
51	Acausal normalcy Andrew_Critch
52	Towards Developmental Interpretability Jesse Hoogland
53	Dear Self; we need to talk about ambition Elizabeth
54	Thoughts on “AI is easy to control” by Pope & Belrose Steven Byrnes
55	Thoughts on sharing information about language model capabilities paulfchristiano
56	FixDT abramdemski
57	My Model Of EA Burnout LoganStrohl
58	Accidentally Load Bearing jefftk
59	Comp Sci in 2027 (Short story by Eliezer Yudkowsky) sudo
60	The Plan - 2023 Version johnswentworth
61	When is Goodhart catastrophic? Drake Thomas
62	Evaluating the historical value misspecification argument Matthew Barnett
63	Contrast Pairs Drive the Empirical Performance of Contrast Consistent Search (CCS) Scott Emmons
64	Bing Chat is blatantly, aggressively misaligned evhub
65	Cyborgism NicholasKees
66	Shell games TsviBT
67	Ten Levels of AI Alignment Difficulty Sammy Martin
68	Introducing Fatebook: the fastest way to make and track predictions Adam B
69	Modal Fixpoint Cooperation without Löb's Theorem Andrew_Critch
70	How to (hopefully ethically) make money off of AGI habryka
71	AI: Practical Advice for the Worried Zvi
72	[Valence series] 1. Introduction Steven Byrnes
73	Towards Monosemanticity: Decomposing Language Models With Dictionary Learning Zac Hatfield-Dodds
74	Responses to apparent rationalist confusions about game / decision theory Anthony DiGiovanni
75	Why and How to Graduate Early [U.S.] Tego
76	Updates and Reflections on Optimal Exercise after Nearly a Decade romeostevensit
77	Getting Started With Naturalism LoganStrohl
78	Why Not Subagents? johnswentworth
79	Meta-level adversarial evaluation of oversight techniques might allow robust measurement of their adequacy Buck
80	A case for AI alignment being difficult jessicata
81	Consciousness as a conflationary alliance term for intrinsically valued internal experiences Andrew_Critch
82	The 101 Space You Will Always Have With You Screwtape
83	Coup probes: Catching catastrophes with probes trained off-policy Fabien Roger
84	Preventing Language Models from hiding their reasoning Fabien Roger
85	Teleosemantics! abramdemski
86	My Objections to "We’re All Gonna Die with Eliezer Yudkowsky" Quintin Pope
87	Why Not Just Outsource Alignment Research To An AI? johnswentworth
88	Childhoods of exceptional people Henrik Karlsson
89	How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions JanB
90	The Soul Key Richard_Ngo
91	Never Drop A Ball Screwtape
92	Cohabitive Games so Far mako yass
93	The salt in pasta water fallacy Thomas Sepulchre
94	Untrusted smart models and trusted dumb models Buck
95	Some Rules for an Algebra of Bayes Nets johnswentworth
96	How to Bounded Distrust Zvi
97	Yudkowsky vs Hanson on FOOM: Whose Predictions Were Better? 1a3orn
98	Before smart AI, there will be many mediocre or specialized AIs Lukas Finnveden
99	A Golden Age of Building? Excerpts and lessons from Empire State, Pentagon, Skunk Works and SpaceX Bird Concept
100	Apocalypse insurance, and the hardline libertarian take on AI risk So8res
101	A freshman year during the AI midgame: my approach to the next year Buck
102	Aumann-agreement is common tailcalled
103	grey goo is unlikely bhauth
104	Iron deficiencies are very bad and you should treat them Elizabeth
105	Dark Forest Theories Raemon
106	Connectomics seems great from an AI x-risk perspective Steven Byrnes
107	Views on when AGI comes and on strategy to reduce existential risk TsviBT
108	The 'Neglected Approaches' Approach: AE Studio's Alignment Agenda Cameron Berg
109	Thinking By The Clock Screwtape
110	Killing Socrates Duncan Sabien (Deactivated)
111	You don't get to have cool flaws Neil
112	Noting an error in Inadequate Equilibria Matthew Barnett
113	Assume Bad Faith Zack_M_Davis
114	Measuring and Improving the Faithfulness of Model-Generated Reasoning Ansh Radhakrishnan
115	When can we trust model evaluations? evhub
116	Benchmarks for Detecting Measurement Tampering [Redwood Research] ryan_greenblatt
117	Learning-theoretic agenda reading list Vanessa Kosoy
118	Exercise: Solve "Thinking Physics" Raemon
119	Why Simulator AIs want to be Active Inference AIs Jan_Kulveit
120	What I would do if I wasn’t at ARC Evals LawrenceC
121	Evaluations (of new AI Safety researchers) can be noisy LawrenceC
122	The King and the Golem Richard_Ngo
123	Love, Reverence, and Life Elizabeth
124	Auditing failures vs concentrated failures ryan_greenblatt
125	Elements of Rationalist Discourse Rob Bensinger
126	Evolution provides no evidence for the sharp left turn Quintin Pope
127	Latent variables for prediction markets: motivation, technical guide, and design considerations tailcalled
128	UFO Betting: Put Up or Shut Up RatsWrongAboutUAP
129	Some background for reasoning about dual-use alignment research Charlie Steiner
130	Scalable Oversight and Weak-to-Strong Generalization: Compatible approaches to the same problem Ansh Radhakrishnan
131	The Base Rate Times, news through prediction markets vandemonian
132	Mob and Bailey Screwtape
133	Pretraining Language Models with Human Preferences Tomek Korbak
134	Touch reality as soon as possible (when doing machine learning research) LawrenceC
135	POC \|\| GTFO culture as partial antidote to alignment wordcelism lc
136	A to Z of things KatjaGrace
137	OpenAI, DeepMind, Anthropic, etc. should shut down. Tamsin Leake
138	Model, Care, Execution Ricki Heicklen
139	Symbol/Referent Confusions in Language Model Alignment Experiments johnswentworth
140	Careless talk on US-China AI competition? (and criticism of CAIS coverage) Oliver Sourbut
141	Alexander and Yudkowsky on AGI goals Scott Alexander
142	Deception Chess: Game #1 Zane
143	Ethodynamics of Omelas dr_s
144	Here's Why I'm Hesitant To Respond In More Depth DirectedEvolution
145	Shutting Down the Lightcone Offices habryka
146	shoes with springs bhauth
147	Would You Work Harder In The Least Convenient Possible World? Firinn
148	Bayesian Networks Aren't Necessarily Causal Zack_M_Davis
149	One Day Sooner Screwtape
150	"Publish or Perish" (a quick note on why you should try to make your work legible to existing academic communities) David Scott Krueger (formerly: capybaralet)
151	Anthropic's Responsible Scaling Policy & Long-Term Benefit Trust Zac Hatfield-Dodds
152	Think carefully before calling RL policies "agents" TurnTrout
153	Going Crazy and Getting Better Again Evenstar
154	The benevolence of the butcher dr_s
155	Nietzsche's Morality in Plain English Arjun Panickssery
156	Cryonics and Regret MvB
157	RSPs are pauses done right evhub
158	Giant (In)scrutable Matrices: (Maybe) the Best of All Possible Worlds 1a3orn
159	Why I’m not into the Free Energy Principle Steven Byrnes
160	The ‘ petertodd’ phenomenon mwatkins
161	Which personality traits are real? Stress-testing the lexical hypothesis tailcalled
162	Green goo is plausible anithite
163	Rational Unilateralists Aren't So Cursed SCP
164	Competitive, Cooperative, and Cohabitive Screwtape
165	Have Attention Spans Been Declining? niplav
166	Hell is Game Theory Folk Theorems jessicata
167	Reducing sycophancy and improving honesty via activation steering Nina Panickssery
168	Are there cognitive realms? TsviBT
169	"Rationalist Discourse" Is Like "Physicist Motors" Zack_M_Davis
170	Being at peace with Doom Johannes C. Mayer
171	Recreating the caring drive Catnee
172	The God of Humanity, and the God of the Robot Utilitarians Raemon
173	Agentized LLMs will change the alignment landscape Seth Herd
174	Ability to solve long-horizon tasks correlates with wanting things in the behaviorist sense So8res
175	What's up with "Responsible Scaling Policies"? habryka
176	The Real Fanfic Is The Friends We Made Along The Way Eneasz
177	New User's Guide to LessWrong Ruby
178	We don't understand what happened with culture enough Jan_Kulveit
179	Seth Explains Consciousness Jacob Falkovich
180	[Valence series] 2. Valence & Normativity Steven Byrnes
181	Fifty Flips abstractapplic
182	So you want to save the world? An account in paladinhood Tamsin Leake
183	A Playbook for AI Risk Reduction (focused on misaligned AI) HoldenKarnofsky
184	Moral Reality Check (a short story) jessicata
185	Takeaways from calibration training Olli Järviniemi
186	Efficiency and resource use scaling parity Ege Erdil
187	Spaciousness In Partner Dance: A Naturalism Demo LoganStrohl
188	[Valence series] 3. Valence & Beliefs Steven Byrnes
189	Fighting without hope Zuko's Shame
190	On Tapping Out Screwtape
191	Beware of Fake Alternatives silentbob
192	Truth and Advantage: Response to a draft of "AI safety seems hard to measure" So8res
193	AI #1: Sydney and Bing Zvi
194	The shape of AGI: Cartoons and back of envelope boazbarak
195	Underwater Torture Chambers: The Horror Of Fish Farming omnizoid
196	Finding Neurons in a Haystack: Case Studies with Sparse Probing wesg
197	How should TurnTrout handle his DeepMind equity situation? habryka
198	When do "brains beat brawn" in Chess? An experiment titotal
199	There are no coherence theorems Dan H
200	The Power of High Speed Stupidity robotelvis
201	Book Review: Consciousness Explained (as the Great Catalyst) Rafael Harth
202	Large language models learn to represent the world gjm
203	Ruining an expected-log-money maximizer philh
204	A problem with the most recently published version of CEV ThomasCederborg
205	In Defense of Parselmouths Screwtape
206	Five Worlds of AI (by Scott Aaronson and Boaz Barak) mishka
207	When Omnipotence is Not Enough lsusr
208	Why You Should Never Update Your Beliefs Arjun Panickssery
209	Responsible Scaling Policies Are Risk Management Done Wrong simeon_c
210	A Way To Be Okay Duncan Sabien (Deactivated)
211	"Justice, Cherryl." Zack_M_Davis
212	Large Language Models can Strategically Deceive their Users when Put Under Pressure. ReaderM
213	[Link] A community alert about Ziz DanielFilan

Updates to the Best of LessWrong:
Coming Soon

It'll take a little while to get the results polished with nice art and organization on the /bestoflesswrong page. Stay tuned for a final announcement. (This year this might still take a week or two. Hopefully next year we'll have most of the final polishing fully automated)