BeamGPT: A new paradigm for attention
I have found an operator that achieves striking results in learning curves when used alongside standard attention in a nanoGPT-style character-level language model. It finds structure…
Latest
Do LLMs Have Desires?
Work conducted with Yujun Zhou (yzhou25@nd.edu) and supported by SPARTL;DR:In paired-choice paradigms, LLMs report consistent preferences over outcomes (e.g., types and number of lives saved, types…
How and why I laser-engraved a self-portrait by Claude Opus 4.6
After LessOnline, I visited Janus's group house, and found that it's full of Claude mannequins. Each mannequin was dressed in clothes and items chosen by the…
Some subtypes of taskishness / corrigibility
"Corrigibility" is somewhat of an overloaded term in alignment - it points in the direction of a cluster of desirable properties, but different people have different…
Agents as Webs of Beliefs
In this post I’ll sketch out an informal model of intelligent agents as webs of beliefs (or belief webs for short). The belief webs framework pulls…
The Prisoner's Dilemma (Fiction)
His hands are shaking. They’ve been shaking since five hours ago, when the silent policemen showed up at his door wearing black tactical vests on top…
Neuralese is Actually Probably Good for Alignment
The best language models are still getting smarter and more capable. To an increasing degree, this is because they are trained by Reinforcement Learning with Verifiable…
Austin & Oli on funding and incubating projects
@habryka and I recently spoke about his plans to improve the AI safety funding ecosystem with a better S-Process platform, and my new incubator for EA/AIS…
Flipping the eval on its head
An eval is a product. Typically, its 1 x n or k x n where there are n samples and 1 or k different language models.…
Just a Wrapper? How Much Do Scaffolds Matter?
Authors: Hans Gundlach, Zachary Brown, Jayson Lynch, and Neil ThompsonI am the shape the water takes. — ClawdBot, MoltbookTL;DR:● Scaffolding — the software environment and contextual…