LessWrong (Curated & Popular)

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.