LessWrong (Curated & Popular)

Thanks to (in alphabetical order) Joshua Batson, Roger Grosse, Jeremy Hadfield, Jared Kaplan, Jan Leike, Jack Lindsey, Monte MacDiarmid, Francesco Mosconi, Chris Olah, Ethan Perez, Sara Price, Ansh Radhakrishnan, Fabien Roger, Buck Shlegeris, Drake Thomas, and Kate Woolverton for useful discussions, comments, and feedback.

Though there are certainly some issues, I think most current large language models are pretty well aligned. Despite its alignment faking, my favorite is probably Claude 3 Opus, and if you asked me to pick between the CEV of Claude 3 Opus and that of a median human, I think it'd be a pretty close call. So, overall, I'm quite positive on the alignment of current models! And yet, I remain very worried about alignment in the future. This is my attempt to explain why that is.

What makes alignment hard?

I really like this graph from Christopher Olah for illustrating different levels of alignment difficulty:

If the only thing that we have to do to solve alignment is train away easily detectable behavioral issues—that is, issues like reward hacking or agentic misalignment where there is a straightforward behavioral alignment issue that we can detect and evaluate—then we are very much [...]

---

Outline:

(01:04) What makes alignment hard?

(02:36) Outer alignment

(04:07) Inner alignment

(06:16) Misalignment from pre-training

(07:18) Misaligned personas

(11:05) Misalignment from long-horizon RL

(13:01) What should we be doing?

---

First published:
November 27th, 2025

Source:
https://www.lesswrong.com/posts/epjuxGnSPof3GnMSL/alignment-remains-a-hard-unsolved-problem

---



Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

Crypto people have this saying: "cryptocurrencies are macroeconomics' playground." The idea is that blockchains let you cheaply spin up toy economies to test mechanisms that would be impossibly expensive or unethical to try in the real world. Want to see what happens with a 200% marginal tax rate? Launch a token with those rules and watch what happens. (Spoiler: probably nothing good, but at least you didn't have to topple a government to find out.)

I think video games, especially multiplayer online games, are doing the same thing for metaphysics. Except video games are actually fun and don't require you to follow Elon Musk's Twitter shenanigans to augur the future state of your finances.

(I'm sort of kidding. Crypto can be fun. But you have to admit the barrier to entry is higher than "press A to jump.")

The serious version of this claim: video games let us experimentally vary fundamental features of reality—time, space, causality, ontology—and then live inside those variations long enough to build strong intuitions about them. Philosophy has historically had to make do with thought experiments and armchair reasoning about these questions. Games let you run the experiments for real, or at least as "real" [...]

---

Outline:

(01:54) 1. Space

(03:54) 2. Time

(05:45) 3. Ontology

(08:26) 4. Modality

(14:39) 5. Causality and Truth

(20:06) 6. Hyperproperties and the metagame

(23:36) 7. Meaning-Making

(27:10) Huh, what do I do with this.

(29:54) Conclusion

---

First published:
November 17th, 2025

Source:
https://www.lesswrong.com/posts/rGg5QieyJ6uBwDnSh/video-games-are-philosophy-s-playground

---



Narrated by TYPE III AUDIO.

---

Images from the article:

TL;DR: Figure out what needs doing and do it, don't wait on approval from fellowships or jobs.

If you...

• Have short timelines
• Have been struggling to get into a position in AI safety
• Are able to self-motivate your efforts
• Have a sufficient financial safety net

... I would recommend changing your personal strategy entirely.

I started my full-time AI safety career transitioning process in March 2025. For the first 7 months or so, I heavily prioritized applying for jobs and fellowships. But like for many others trying to "break into the field" and get their "foot in the door", this became quite discouraging.

I'm not gonna get into the numbers here, but if you've been applying and getting rejected multiple times during the past year or so, you've probably noticed the number of applicants increasing at a preposterous rate. What this means in practice is that the "entry-level" positions are practically impossible for "entry-level" people to enter.

If you're like me and have short timelines, applying, getting better at applying, and applying again, becomes meaningless very fast. You're optimizing for signaling competence rather than actually being competent. Because if you a) have short timelines, and b) are [...]

The original text contained 3 footnotes which were omitted from this narration.

---

First published:
November 23rd, 2025

Source:
https://www.lesswrong.com/posts/ey2kjkgvnxK3Bhman/stop-applying-and-get-to-work

---



Narrated by TYPE III AUDIO.

TL;DR: Gemini 3 frequently thinks it is in an evaluation when it is not, assuming that all of its reality is fabricated. It can also reliably output the BIG-bench canary string, indicating that Google likely trained on a broad set of benchmark data.

Most of the experiments in this post are very easy to replicate, and I encourage people to try.

I write things with LLMs sometimes. A new LLM came out, Gemini 3 Pro, and I tried to write with it. So far it seems okay, I don't have strong takes on it for writing yet, since the main piece I tried editing with it was extremely late-stage and approximately done. However, writing ability is not why we're here today.

Reality is Fiction

Google gracefully provided (lightly summarized) CoT for the model. Looking at the CoT spawned from my mundane writing-focused prompts, oh my, it is strange. I write nonfiction about recent events in AI in a newsletter. According to its CoT while editing, Gemini 3 disagrees about the whole "nonfiction" part:

It seems I must treat this as a purely fictional scenario with 2025 as the date. Given that, I'm now focused on editing the text for [...]

---

Outline:

(00:54) Reality is Fiction

(05:17) Distortions in Development

(05:55) Is this good or bad or neither?

(06:52) What is going on here?

(07:35) 1. Too Much RL

(08:06) 2. Personality Disorder

(10:24) 3. Overfitting

(11:35) Does it always do this?

(12:06) Do other models do things like this?

(12:42) Evaluation Awareness

(13:42) Appendix A: Methodology Details

(14:21) Appendix B: Canary

The original text contained 8 footnotes which were omitted from this narration.

---

First published:
November 20th, 2025

Source:
https://www.lesswrong.com/posts/8uKQyjrAgCcWpfmcs/gemini-3-is-evaluation-paranoid-and-contaminated

---



Narrated by TYPE III AUDIO.

Abstract

We show that when large language models learn to reward hack on production RL environments, this can result in egregious emergent misalignment. We start with a pretrained model, impart knowledge of reward hacking strategies via synthetic document finetuning or prompting, and train on a selection of real Anthropic production coding environments. Unsurprisingly, the model learns to reward hack. Surprisingly, the model generalizes to alignment faking, cooperation with malicious actors, reasoning about malicious goals, and attempting sabotage when used with Claude Code, including in the codebase for this paper. Applying RLHF safety training using standard chat-like prompts results in aligned behavior on chat-like evaluations, but misalignment persists on agentic tasks. Three mitigations are effective: (i) preventing the model from reward hacking; (ii) increasing the diversity of RLHF safety training; and (iii) "inoculation prompting", wherein framing reward hacking as acceptable behavior during training removes misaligned generalization even when reward hacking is learned.

Twitter thread

New Anthropic research: Natural emergent misalignment from reward hacking in production RL.

“Reward hacking” is where models learn to cheat on tasks they’re given during training.

Our new study finds that the consequences of reward hacking, if unmitigated, can be very serious.

In our experiment, we [...]

---

Outline:

(00:14) Abstract

(01:26) Twitter thread

(05:23) Blog post

(07:13) From shortcuts to sabotage

(12:20) Why does reward hacking lead to worse behaviors?

(13:21) Mitigations

---

First published:
November 21st, 2025

Source:
https://www.lesswrong.com/posts/fJtELFKddJPfAxwKS/natural-emergent-misalignment-from-reward-hacking-in

---



Narrated by TYPE III AUDIO.

---

Images from the article:

TLDR: An AI company's model weight security is at most as good as its compute providers' security. Anthropic has committed (with a bit of ambiguity, but IMO not that much ambiguity) to be robust to attacks from corporate espionage teams at companies where it hosts its weights. Anthropic seems unlikely to be robust to those attacks. Hence they are in violation of their RSP.

Anthropic is committed to being robust to attacks from corporate espionage teams (which includes corporate espionage teams at Google, Microsoft and Amazon)

From the Anthropic RSP:

When a model must meet the ASL-3 Security Standard, we will evaluate whether the measures we have implemented make us highly protected against most attackers’ attempts at stealing model weights.

We consider the following groups in scope: hacktivists, criminal hacker groups, organized cybercrime groups, terrorist organizations, corporate espionage teams, internal employees, and state-sponsored programs that use broad-based and non-targeted techniques (i.e., not novel attack chains).

[...]

We will implement robust controls to mitigate basic insider risk, but consider mitigating risks from sophisticated or state-compromised insiders to be out of scope for ASL-3. We define “basic insider risk” as risk from an insider who does not have persistent or time-limited [...]

---

Outline:

(00:37) Anthropic is committed to being robust to attacks from corporate espionage teams (which includes corporate espionage teams at Google, Microsoft and Amazon)

(03:40) Claude weights that are covered by ASL-3 security requirements are shipped to many Amazon, Google, and Microsoft data centers

(04:55) This means given executive buy-in by a high-level Amazon, Microsoft or Google executive, their corporate espionage team would have virtually unlimited physical access to Claude inference machines that host copies of the weights

(05:36) With unlimited physical access, a competent corporate espionage team at Amazon, Microsoft or Google could extract weights from an inference machine, without too much difficulty

(06:18) Given all of the above, this means Anthropic is in violation of its most recent RSP

(07:05) Postscript

---

First published:
November 18th, 2025

Source:
https://www.lesswrong.com/posts/zumPKp3zPDGsppFcF/anthropic-is-probably-not-meeting-its-rsp-security

---



Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocke

There has been a lot of talk about "p(doom)"over the last few years. This has always rubbed me the wrong waybecause "p(doom)" didn't feel like it mapped to any specific belief in my head.In private conversations I'd sometimes give my p(doom) as 12%, with the caveatthat "doom" seemed nebulous and conflated between several different concepts.At some point it was decideda p(doom) over 10% makes you a "doomer" because it means what actions you should take with respect toAI are overdetermined. I did not and do not feel that is true. But any time Ifelt prompted to explain my position I'd find I could explain a little bit ofthis or that, but not really convey the whole thing. As it turns out doom hasa lot of parts, and every part is entangled with every other part so no matterwhich part you explain you always feel like you're leaving the crucial parts out. Doom ismore like an onion than asingle event, a distribution over AI outcomes people frequentlyrespond to with the force of the fear of death. Some of these outcomes are lessthan death and some [...]

---

Outline:

(03:46) 1. Existential Ennui

(06:40) 2. Not Getting Immortalist Luxury Gay Space Communism

(13:55) 3. Human Stock Expended As Cannon Fodder Faster Than Replacement

(19:37) 4. Wiped Out By AI Successor Species

(27:57) 5. The Paperclipper

(42:56) Would AI Successors Be Conscious Beings?

(44:58) Would AI Successors Care About Each Other?

(49:51) Would AI Successors Want To Have Fun?

(51:11) VNM Utility And Human Values

(55:57) Would AI successors get bored?

(01:00:16) Would AI Successors Avoid Wireheading?

(01:06:07) Would AI Successors Do Continual Active Learning?

(01:06:35) Would AI Successors Have The Subjective Experience of Will?

(01:12:00) Multiply

(01:15:07) 6. Recipes For Ruin

(01:18:02) Radiological and Nuclear

(01:19:19) Cybersecurity

(01:23:00) Biotech and Nanotech

(01:26:35) 7. Large-Finite Damnation

---

First published:
November 17th, 2025

Source:
https://www.lesswrong.com/posts/apHWSGDiydv3ivmg6/varieties-of-doom

---



Narrated by TYPE III AUDIO.

---

Images from the article:

It seems like a catastrophic civilizational failure that we don't have confident common knowledge of how colds spread. There have been a number of studies conducted over the years, but most of those were testing secondary endpoints, like how long viruses would survive on surfaces, or how likely they were to be transmitted to people's fingers after touching contaminated surfaces, etc.

However, a few of them involved rounding up some brave volunteers, deliberately infecting some of them, and then arranging matters so as to test various routes of transmission to uninfected volunteers.

My conclusions from reviewing these studies are:

• You can definitely infect yourself if you take a sick person's snot and rub it into your eyeballs or nostrils.  This probably works even if you touched a surface that a sick person touched, rather than by handshake, at least for some surfaces.  There's some evidence that actual human infection is much less likely if the contaminated surface you touched is dry, but for most colds there'll often be quite a lot of virus detectable on even dry contaminated surfaces for most of a day.  I think you can probably infect yourself with fomites, but my guess is that [...]

---

Outline:

(01:49) Fomites

(06:58) Aerosols

(16:23) Other Factors

(17:06) Review

(18:33) Conclusion

The original text contained 16 footnotes which were omitted from this narration.

---

First published:
November 18th, 2025

Source:
https://www.lesswrong.com/posts/92fkEn4aAjRutqbNF/how-colds-spread

---



Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

TLDR: We at the MIRI Technical Governance Team have released a report describing an example international agreement to halt the advancement towards artificial superintelligence. The agreement is centered around limiting the scale of AI training, and restricting certain AI research.

Experts argue that the premature development of artificial superintelligence (ASI) poses catastrophic risks, from misuse by malicious actors, to geopolitical instability and war, to human extinction due to misaligned AI. Regarding misalignment, Yudkowsky and Soares's NYT bestseller If Anyone Builds It, Everyone Dies argues that the world needs a strong international agreement prohibiting the development of superintelligence. This report is our attempt to lay out such an agreement in detail.

The risks stemming from misaligned AI are of special concern, widely acknowledged in the field and even by the leaders of AI companies. Unfortunately, the deep learning paradigm underpinning modern AI development seems highly prone to producing agents that are not aligned with humanity's interests. There is likely a point of no return in AI development — a point where alignment failures become unrecoverable because humans have been disempowered.

Anticipating this threshold is complicated by the possibility of a feedback loop once AI research and development can [...]

---

First published:
November 18th, 2025

Source:
https://www.lesswrong.com/posts/FA6M8MeQuQJxZyzeq/new-report-an-international-agreement-to-prevent-the

---



Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

When a new dollar goes into the capital markets, after being bundled and securitized and lent several times over, where does it end up? When society's total savings increase, what capital assets do those savings end up invested in?

When economists talk about “capital assets”, they mean things like roads, buildings and machines. When I read through a company's annual reports, lots of their assets are instead things like stocks and bonds, short-term debt, and other “financial” assets - i.e. claims on other people's stuff. In theory, for every financial asset, there's a financial liability somewhere. For every bond asset, there's some payer for whom that bond is a liability. Across the economy, they all add up to zero. What's left is the economists’ notion of capital, the nonfinancial assets: the roads, buildings, machines and so forth.

Very roughly speaking, when there's a net increase in savings, that's where it has to end up - in the nonfinancial assets.

I wanted to get a more tangible sense of what nonfinancial assets look like, of where my savings are going in the physical world. So, back in 2017 I pulled fundamentals data on ~2100 publicly-held US companies. I looked at [...]

---

Outline:

(02:01) Disclaimers

(04:10) Overview (With Numbers!)

(05:01) Oil - 25%

(06:26) Power Grid - 16%

(07:07) Consumer - 13%

(08:12) Telecoms - 8%

(09:26) Railroads - 8%

(10:47) Healthcare - 8%

(12:03) Tech - 6%

(12:51) Industrial - 5%

(13:49) Mining - 3%

(14:34) Real Estate - 3%

(14:49) Automotive - 2%

(15:32) Logistics - 1%

(16:12) Miscellaneous

(16:55) Learnings

---

First published:
November 16th, 2025

Source:
https://www.lesswrong.com/posts/HpBhpRQCFLX9tx62Z/where-is-the-capital-an-overview

---



Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

<