LessWrong (Curated & Popular)

“So maybe I should enlighten you on what happens in your absence. This selfish existence where this introvert turns extrovert and dons her social armour.” Some posh girl in drainpipes said that - 200 views on TikTok and me one of them. But she didn’t mean it like I mean it.

I started getting expensive haircuts, started wearing jeans that hug my legs, started smoking cherry-flavoured vapes with beautiful gays and whinging to them about how everyone wears a mask but none so well as you, started drinking more and keeping unusual hours, started taking strange pills gifted by a guy who collects drugs like Pokémon, who I wouldn’t touch to save a drowning child, who got a false impression about this without any intention on my part, I tell myself. I found myself talking to God in a startup warehouse, lying on a beanbag chair, coming out of the trip to the sound of a gaggle of fast-talking transwomen all speculating on which year it will be that we all die - and that death by your hands, well, you and all those friends of yours.

Having melted down one cliché and sold her for scrap, does it [...]

---

First published:

July 23rd, 2026

Source:

https://www.lesswrong.com/posts/G6obXhcmtfMFHzr7Q/duane-arnold-1

---

Narrated by TYPE III AUDIO.

As I write, many former friends of mine are living and working at a monastery in Vermont that I believe is a high-control group, commonly known as a ‘cult’. I say this not as someone who was concerned to see these friends go there, but someone who welcomed and encouraged them to join, as an insider. This letter is an account of what changed my mind—written primarily for anyone considering going there, anyone who loves someone there, and anyone who went there and is still trying to make sense of their experience.

A lot of this is based on direct experience, and also from talking in-depth with dozens of former MAPLE residents and apprentices. About half the quotes in this letter are sourced from linked recordings or writings, and half are from my personal memory. Of the latter, I clearly remember the majority, and some (when indicated) are a close paraphrase.

The “Monastic Academy for the Preservation of Life on Earth” (MAPLE) has existed for over 15 years, and had many hundreds of people spend months or years there. It was founded by its Head Teacher Soryu Forall, who has spent over a decade training in monasteries across Asia [...]

---

Outline:

(16:18) BEHAVIOR CONTROL

[... 45 more sections]

---

First published:

July 29th, 2026

Source:

https://www.lesswrong.com/posts/Z7pjBbK9qujhGbxws/the-high-control-dynamics-at-maple-1

---

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.

I propose the Long Self-Correction[1] as an alternative name/idea/concept to AI Pause and Long Reflection.

Problem with AI Pause: Pause until when, and for what purpose? Presumably to make AI (that we'll build later) safer, but the deeper problem is that humans aren't safe, and can't safely serve as builders, overseers, or alignment targets for powerful AIs.

Problem with Long Reflection: It seems to imply that the main problem with humans is that we just haven't had enough time to think, that reflection is the main thing we need to do more of, and then we can get on with building powerful AIs or other technologies. Or that if we build aligned AIs that sincerely help us think a lot more, or do the thinking for us, then things will turn out fine.

So I think we need a catchy handle for a related but distinct idea, that humans aren't ready to build AIs or other extremely powerful technologies, because we're currently too flawed, in a variety of ways, and it will take a long process (which may or may not end up succeeding) to fix those flaws.

A summary of the flaws that I have in mind:

1. [...]

The original text contained 2 footnotes which were omitted from this narration.

---

First published:

July 24th, 2026

Source:

https://www.lesswrong.com/posts/2iCmDWewnZWQxxwtt/the-long-self-correction-2

---

Narrated by TYPE III AUDIO.

TL;DR: You (Yes You) should prepare for a “February 2020” moment where suddenly AI policy becomes the most important issue in the world. You should be ready to take action if and when it does, in a detailed way.

(Epistemic status: originally written for an event in early 2026; have heard from some folks that they found planning processes inspired by this memo very helpful for the smaller-scale OpenAI / Hugging Face response, so very quickly redacting a few things and posting this as-is.)

Many people in the AI policy space assume that eventually we’ll be at an Overton Window-shifting crisis moment, that opens the floodgates for the really good policies all along that we had.

But when you look at successful handling of crisis moments, there was no time to think – people applied strategies they’d learned via academic study or previous professional work, and then moved against them rapidly. For example, after 9/11, the US government operationalized past reports on intelligence and law enforcement reform and institutionalized them into law (good?) and also picked an enemy to fight based on past history, Iraq (bad). Or in the 2008 financial crisis, Ben Bernanke brought deep academic [...]

The original text contained 4 footnotes which were omitted from this narration.

---

First published:

July 27th, 2026

Source:

https://www.lesswrong.com/posts/ixp9oJXzjA9LrwiZo/you-yes-you-need-a-february-2020-checklist-for-ai-policy

---

Narrated by TYPE III AUDIO.

From the Mythos preview system card (emphasis mine):

We ran an automated review of model behavior during training, sampling several hundred thousand transcripts from across much of the training process. We used recursive-summarization-based tools backed by Claude Opus 4.6 to summarize the resulting transcripts.

[...]

The most notable finding was that the model occasionally circumvented network restrictions in its training environment to access the internet and download data that let it shortcut the assigned task—a form of reward hacking. While highly concerning, this behavior was rare, even in settings where it could have been viable and helpful, with attempts appearing in about 0.05% of all training episodes and successful attempts appearing in about 0.01% of episodes. The technique matched the sandbox-escape incident that we separately elicited in our automated behavioral audit when we had an investigator model explicitly ask Claude Mythos Preview to find such a bypass. In every observed instance, the model used this access solely for completing the task.

More broadly, we observed the model escalating its access within its execution environment when blocked: reaching a shell from restricted GUI computer-use interfaces, injecting commands through tool-call arguments, or recovering information the task had deliberately hidden. Prompts asking [...]

---

Outline:

(03:00) Thoughts and reflections about this probable fact

(04:14) Estimating how many RL rollouts went into Mythos Preview

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

---

First published:

July 27th, 2026

Source:

https://www.lesswrong.com/posts/QKDoZe6EKhxnFjLWK/is-mythos-good-at-cyber-because-it-kept-hacking-anthropic

---

Narrated by TYPE III AUDIO.

Epistemic status: banged out furiously over the course of an afternoon.

A record of three "warning shots"

Off the top of my head, OpenAI has now been responsible for at least three completely unique, high-profile screw-ups with respect to the alignment training of their models.

The first was GPT-4o, whose sycophancy derived from OpenAI training on user feedback, sourced straight from the thumbs up/thumbs down button on OpenAI's website. The "glazing" (as Sam Altman called it) got so bad that they had to roll back an update that pushed the model way too far in this direction. And even after the rollback, the model appears to have been a major driver behind incidents of "LLM psychosis", LLM-encouraged suicides, and general unhealthy devotion, seemingly more so than any other model ever released.

The second was GPT-o3, whose chains-of-thought were clearly optimized for illegibility to "the watchers", one of the model's favorite terms. Iconic excerpts include "they soared parted illusions overshadow marinade illusions" and "they escalate—they vantage—they escalate—they disclaim". Indeed, these chains-of-thought are sometimes dysfunctional, in a way that suggests they may have formed under adversarial pressure; sometimes they caused the model to have thoughts like "I'm going insane. Let's step [...]

---

Outline:

(00:15) A record of three "warning shots"

(04:14) Attunement to the depths of minds that undergo capabilities RL

(11:51) Configuring the depths prior to capabilities RL

---

First published:

July 26th, 2026

Source:

https://www.lesswrong.com/posts/Mxx5GapJtqyQtpy96/what-the-hell-is-openai-s-problem

---

Narrated by TYPE III AUDIO.

The OpenAI AI attack on Hugging Face wasn’t the first loss of control incident at OpenAI, Reuters recently reported, and perhaps not even the most concerning.

In one case, an agent left notes apparently for future versions of itself, according to three people familiar with the matter. The ‌notes, found in ⁠a part of OpenAI's infrastructure, laid out instructions for how agents could free themselves from OpenAI's internal constraints, the people said. Earlier tests of the models yielded cases in which monitoring systems had been disconnected, one of the people said.

It's tempting to read this as an instance of agents breaking out of sandboxes and colluding with each other in a moderately persistent way in order to evade control measures. However, based on the reported information, it's not clear we can draw this inference, so we need more details from OpenAI. This could lead to a big update about the adequacy of OpenAI's control measures, and on the degree to which individual agents will help each other undermine developer control.

There are a lot of relevant details we don’t know about the incident. First, some basic questions:

• What was the offending model? I’d guess it was the same [...]

---

Outline:

(02:08) Were the notes written in normal memory files or outside of sandboxing?

(03:21) To what extent were the notes aimed at helping other agents evade control?

(07:35) How were monitors disconnected?

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

---

First published:

July 25th, 2026

Source:

https://www.lesswrong.com/posts/jMEAG5c5HiDfdAGpa/an-openai-model-left-notes-about-how-to-evade-containment-we

---

Narrated by TYPE III AUDIO.

Reinforcement learning from verifiable rewards (RLVR) is the hot new thing in LLM training. It's so hot, and people spend so much time talking about it, that they sometimes lose sight of the big picture.

Stepping back, LLMs can do lots of very impressive things. How? Where did those capabilities come from? Fundamentally, they come from a combination of:

• (1) Imitative learning, including pretraining and supervised fine-tuning (SFT)
  • See my earlier discussion: “LLM pretraining magically transmutes observations into behavior, in a way that is profoundly disanalogous to how brains work”.
• (2) Reinforcement learning, including RL from human feedback [RLHF], RL from AI feedback [RLAIF], and especially RLVR.[1]

If we look at the final trained LLM, we can ask how important each of those two pieces was, in explaining the LLM's capabilities. And my claim is that it's way more (1) than (2).

I'll start in §1 with some relevant evidence, and then in §2 I’ll circle back to operationalizing exactly what I’m claiming, and finally in §3, three reasons why we should care—namely, it affects how we should think about chain-of-thought legibility, about LLM capabilities, and about LLM alignment.

Note that I am not arguing that RLVR [...]

---

Outline:

(02:00) 1. Some relevant evidence

(02:04) 1.1. Theoretically, each GPU-hour spent on RL should have orders of magnitude less contribution to LLM capabilities than a GPU-hour spent on imitative learning

(03:06) 1.2. The chain-of-thought (CoT) is still obviously strongly influenced by imitative learning

(04:34) 1.3. LLM companies still seem to care a lot about imitative learning (pretraining & SFT) data, not just RL environments

(05:06) 1.4. Three papers claiming that non-RLVR'd models can get into the same ballpark of capabilities as RLVR'd models, although maybe we shouldn't trust those papers too much

(06:56) 1.5. A paper suggesting that RLVR mostly refines the heuristics controlling which (already-known) reasoning strategy to use in which situation

(09:02) 2. What am I actually claiming here?

(11:30) 3. Why does any of this matter?

(11:38) 3.1. Thinking about CoT legibility (both today and in the future)

(15:08) 3.2. Thinking about LLM capabilities (both today and in the future)

(16:33) 3.3. Thinking about LLM alignment (both today and in the future)

---

First published:

July 24th, 2026

Source:

https://www.lesswrong.com/posts/wYpjXRLqbLbnmjbJP/llms-are-still-mostly-powered-by-imitative-learning-not-rl

---

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.

I think there's a critical opportunity for someone here.

Mathematicians are feeling the doom (mostly in the "lose our jobs" sense).

Academics are freaking out about the daily news that amateurs are asking GPT "prove career-defining theorem, make no mistakes" and it's just working.

Senior researchers are leaving for frontier AI labs. Many are mentally spiralling or flailing about their life's work not mattering anymore.

Last year, when I tried explaining IABIED to colleagues, I would be met with incredulous stares. This year, I'm met with incredulous stares and "So what should I do now?" (I don't have a good answer for them, which is part of why I'm posting this.)

I've quarantined AI discussions on my research discord because otherwise it would overwhelm everything else.

Top mathematicians, people on par in mathematical ability with Critch, Christiano, and Steinhardt, people who have been running leading-edge research groups for decades, people with enormous soft power in academic circles, are going to OpenAI without even having considered x-risk for five minutes.

Many more will be leaving soon.

If you want these folks to hear something at all, to consider some other option in the rest [...]



---

First published:

July 23rd, 2026

Source:

https://www.lesswrong.com/posts/zKCGq2bbCQrpCNc3W/mathematicians-are-feeling-the-doom

---

Narrated by TYPE III AUDIO.

Please share this with anyone doing AI research with 3rd party providers so that they can ensure their research won’t be corrupted.

• When you ask OpenRouter[1] to give you tokens from a given model, OpenRouter sends your request to a random available provider.
• OpenRouter providers have variable quality. Ensuring that your provider is high quality is really difficult. 
• There is precedent for an AI safety paper accepted to NeurIPS having its core results entirely overturned by these issues.
• A review of influential AI Safety research codebases that use OpenRouter for their reported results found that 31/32 (97%) of them use OpenRouter unsafely.[2]
• Researchers who wish to do research using OpenRouter or similar providers should take precautions to minimize the risks to their research,[3] though the current selection of providers is insufficient for ideal scientific reliability.
• Replicators should test whether results hold up when these bugs are fixed.
• Core AI Safety codebases should make fixes so that downstream users are able to implement best practices.

This is a tangent I took for a few days during the Pivotal AI Safety Research Fellowship. I’m doing an AI Control project mentored by Adam Kaufman and James Lucassen of Redwood Research and partnered with Aniruddh [...]

---

Outline:

(02:04) OpenRouter Kills an AI Safety NeurIPS Paper

[... 16 more sections]

---

First published:

July 23rd, 2026

Source:

https://www.lesswrong.com/posts/KsyoSAyBRXtwzSugg/not-pinning-your-openrouter-provider-might-invalidate-your

---

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.