#AI Researcher | A jumped-up pantry boy who never knew his place
Joined April 2022
- Tweets2.7K
- Following259
- Followers537
- Likes57.7K
Wow! Claude really refuses to work in August
New on our Engineering blog: Creating in-house source control.
In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale.
Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: anthropic.com/engineering/re…
Am I having a stroke or is there something i misunderstand -- i understand its a combo of pr, humor, and actual cool math endeavor (based on the different ppl that engage with it) -- but at these exponents, n*(log n)^(1-k) is not interesting before k reaches the proximity of 1/2
They should have a secret value baked into the model's knowledge and only if the model sees it as a response to a specific tool call can it assume it is in a simulation. Any other case, no matter how convinced it is, it must operate under the assumption it is not.
Somewhat related: someone can create a polymarket replica where people play against a continually training llm until it learns to predict better
People seem to believe (maybe: want to believe?) that Anthropic's "dooming" is a PR shtick, while it has been from day 1 aligned with this perspective and never done anything that contradicts that (read: the fact that they might have other missions too is not a contradiction)
The question whether their dooming is a reasonable concern, now, then, or in an hypothetical world is beyond the point.
And to alleviate any suspicion, Im not affiliated with Anthropic in any way, and even slightly begrudge them for never citing me even when appropriate.
Guy Dar retweeted
What can a neuron compute?
Real biological neurons are complex, but how capable are they?
Using a new method, we found that a single cortical neuron can classify cats vs dogs, recognize spoken words, and solve 10-bit parity, all tasks thought to require entire networks. (1/15)
Guy Dar retweeted
What's in a neuron? 💫 (an atypically long, almost personal post)
Neurons in LMs have always been a fascinating object to study. I've been studying them since 2020, viewing them as key-value memory cells, analyzing what they capture in vocabulary space, and how they compose together to form features.
arxiv.org/abs/2012.14913
arxiv.org/abs/2203.14680
arxiv.org/abs/2501.08319
arxiv.org/abs/2506.10920
But many neurons still remain opaque! They do many things.
Our recent work led by @AsafAvrahamy tackles this challenge by decomposing neuron weights in vocabulary space. We do this by taking the neuron weight vector and learning different ways to rotate it (just a bit) to reveal monosemantic vocabulary channels that it captures. The nice thing about our method ROTATE is that it's data-free and super efficient, relying only on vocabulary kurtosis as a search signal.
I've been thinking about this idea since 2024, proposed it to multiple students, but only Asaf was brave enough to take this ;)
Very happy with the final outcome. Check out the paper! 👇
arxiv.org/abs/2604.06005
The Kaon thing is elegant, you dont need exact orthoginalization, you can replace SV with random values. It makes sense bc whats central to Muon is to upscale weak directions not getting to SV exactly 1. The point is to fast track low SVs to noticeable values >>
Even replacing singular values with random noise works in optimization. After considering Schatten norm updates that interpolate between SGD and Muon, we could push it further towards the quasi-norm regime. And interestingly it works.
The main point here is that the geometrical explanation maybe does not work.
Instead it is possible to think about the trade-off between alignment of gradients and potential of descent. If we sacrifice a bit of alignment but gain a large amount of descent with a proper learning rate it leads to acceleration.
I think that's kinda the same point that this comment is getting at
xitter.cf/i/status/2054508313723…
A potential way to do Aurora with a closed-form construction that avoids the expensive iterative step.
Link in comments
Introducing Aurora, a new optimizer for training frontier-scale models.
We train Aurora-1.1B, which achieves 100x data efficiency on open-source internet data. Despite having 25% fewer parameters, 2 orders of magnitude fewer training tokens, and using fully open-source internet-only data, Aurora matches Qwen3-1.7B on several benchmarks.
Aurora was developed after identifying a major failure mode that can occur under Muon, an increasingly popular optimizer that has shown strong gains over Adam(W). We find that Muon can cause a huge percentage of neurons to effectively die early in training, reducing effective network capacity so that many parameters no longer meaningfully contribute to network outputs.
By redistributing update energy more uniformly across neurons while preserving Muon’s stability properties, Aurora prevents neuron death and recovers substantial model capacity.
What makes this work especially exciting is that it points toward a broader direction for ML research: better optimizers may not come purely from elegant mathematical abstractions, but from understanding and addressing the concrete dynamics and pathologies that emerge inside real training systems.