imenani

That’s fair. Some of what happened in OAI<>HF incident was (mis)-generalisation and not directly trained for.

However, the article experiments with five agents. So it seems to assume even the small scale behaviour is emergent.

Also, “roles and structures” may well have been learned during training.

The agents participating in the OAI<>HF swarm were trained not only for communication but to be _aligned with each other_.

Just want to correct the premise that the agent swarm behaviour was emergent.

Noam Brown on Dwarkesh podcast around 40 minutes mark

> “We train them to work together, to be cooperative, to essentially be fully aligned with each other.”

“The hacks occurred in May”

Feels like important context that most readers only reading the title are missing.

The primary source

https://darioamodei.com/post/we-must-pace-the-frontier#top

I don’t think next token prediction is a particularly good description of pretraining either.

The intermediate representations at each position are being optimised not only to help predict the next token, but also to help predict all subsequent tokens within the training context.

You can see this directly in backpropagation: the gradient reaching a representation at position i sums contributions from prediction losses for subsequent tokens, not just from the loss for token i+1.

Regarding subsidy/profits: Anthropic at least seems to be on track to report a profitable Q2 2026

> Anthropic has projected revenue of at least $10.9 billion for the second quarter of 2026, more than double the previous quarter, on track for its first quarterly operating profit of $559 million.

https://www.reuters.com/business/anthropic-ipo-valuation-hin...

For anyone wondering “how slow is this?”

IIUC, Kimi K3 on RTX 6000 Ada (48GB) takes 292 s/token

https://github.com/lyogavin/airllm/releases/tag/v3.1.0

The relation to current RLVR methods I think is interesting, they do discuss it a bit but I would be curious to see more about this as well. Quote from the paper:

Exploration beyond Pretraining. The mode collapse XMs address during pretraining also often shows up in post-training, where RL fine-tuning is known to sharpen models onto a narrow set of behaviors [76]. Recent fixes such as pass@krewards [77] and best-of-N-aware fine-tuning [78] can be seen through our lens as Forward XM, with a verifier standing in for ground truth data. These fixes act only during post-training, though; pretraining with exploration may yield base models that capture more modes in the first place, leaving RL more to select among.

Nice presentation of the list!

I'd recommend watching a few of his talks/podcasts before during reading these to get the overview and how all the bits in these works tie together.

https://www.dwarkesh.com/p/ilya-sutskever

https://simons.berkeley.edu/talks/ilya-sutskever-openai-2023...

https://www.dwarkesh.com/p/ilya-sutskever-2

The author discussed this here four days ago

https://news.ycombinator.com/item?id=48077663

With the benefit of hindsight, perhaps much of this was Claude Mythos? The model was deployed internally since Feb

Agreed. LLMs have helped me achieve much deeper reading, _when directed to do so_. Asking an LLM to “Teach me Socratically about this paper/code. One question at a time”, usually allows me to get a much deeper reading of the material than I would otherwise.

Each of these models has a thinking/reasoning variant and a default non-thinking variant. I would expect the reasoning variants (o3 or “GPT5 Thinking”, Gemini DeepThink, Claude with Extended Thinking, etc) to do better at this. I think there is also some chance that in their reasoning traces they may display something you might see as closer to world modelling. In particular, you might find them explicitly tracking positions of pieces and checking validity.

As far as I can tell they don’t say which LLM they used which is kind of a shame as there is a huge range of capabilities even in newly released LLMs (e.g. reasoning vs not).

They fix the temperature at T=0.6 for all k for all models, even though their own Figure 10 shows that RL model benefits from higher temperatures. I would buy the overall claim much more if they swept of temperature parameter for each k and model like they did in the Codex paper [1].

[1] https://arxiv.org/abs/2107.03374