x312

A TLDR, since most comments here seem to misunderstand the result.

The idea is: a model may only have 100M neurons or whatever. But there are infinite concepts in the world. Therefore, models have to "stuff" multiple concepts per neuron.

How do you do that? Theres two ways:

1. Each neuron handles similar concepts. One neuron for dogs, a separate neuron one for cars, etc.

2. Each neuron handles a bundle of unrelated concepts. One neuron handles dogs and cars. Another neuron handles cats and numbers. And so on.

Historically, people thought #2 was correct because many neurons were found that seemingly represented multiple concepts. #2 had a cool name (polysemanticity) and was promoted in OpenAI/Anthropic blogs so it had a lot of traction.

However, this blog says #2 is wrong. Actually, the dog/car neuron is just a "V shape neuron". So #2 was just methodological error, and #1 was right all along. This is all the blog is saying.

Claude code is 60-80 tps, codex varies by model but mostly 40-50 tps.

Claude Code/Codex charge the user for their extremely bloated one-size-fits-all system prompts (including safety instructions and other stuff users dont want).

In my experience if you're using OpenAI/Claude models and paying API costs, almost every other harness beats Claude Code/Codex in cost.

The proof is that LLMs could barely solve arithmetic 3 years ago, but now surpass the best human mathematicians, and that this has all occurred from simple principles (RL + compute) that will continue to scale up by factors of millions in the coming years.

Also, advocating for slowing LLM progress does not benefit Anthropic or OpenAI.

This is increasingly the consensus I see also on the academic side of AI/safety research. Specifically that AI poses an existential risk to humanity.

This was a fringe belief until recently, but the progress of AI in research is impossible to ignore. Epecially in math, where not only has AI outstripped humans in generative ability, but is able to create scientific knowledge which is beyond the capacity of human comprehension.

There's clearly no intelligence task that AIs can't do due to some magic fundamental constraint. And it's hard to imagine a world where current limitations like poor sample efficiency or lack of continual learning won't eventually be solved.

Total AI compute is estimated to grow somewhere in the 1-10 million-fold range in the next decade. Please don't underestimate the phase change that's still coming.

Sure, maybe there's some plateau due to RL being fundamentally limited in some surprising way, but this is nothing but a hope.

Given how many private benchmarks they're using now, its likely they just tested different combos until they got the result they wanted.

Completely discredits the index if it just gets modified to match social media vibes.

Hmm, 61 on ArtificialAnalysis, effectively matching GPT-5.6 and trailing the new Meta model. How is that possible along with the other metrics they shared? Insanely jagged intelligence?

I believe its the same as free models in general on Openrouter, 1k requests per day for accounts that have some spend history.

The provider decrypts it and puts the decrypted reasoning into the model's context window. They prompt the model to repeat back the reasoning. So then the model echoes it back in plain text.