Research Blog
Research Blog
Writing to figure out what I think.
Open topics
Every Sequence Model Is a Memory Algorithm
A Transformer never changes its KV cache after writing to it. So how does it learn anything in context? Because every sequence model is really an online learning...
From Memory to Learning: What Continual Learning for LLM Agents Would Actually Look Like
An agent that solves the same class of problem a thousand times should be better at it by the end. Most deployed agents aren't. I think the reason...
Why Final-Outcome Rewards Are Not Enough for AI Agents
An outcome reward tells you whether a trajectory worked. It doesn't say why. I think the interesting question is whether the process rewards we train on today measure...
Distillation in LLMs: From Context to Weights
A model that reads three examples and nails the fourth has learned something, and its weights didn't move. Where is that lesson, and how do you keep it?...
What If the Model Has Already Made Up Its Mind?
Anthropic found that a small but consistent fraction of Claude conversations end up undermining the user's ability to think for themselves. I think CoT unfaithfulness and sycophantic disempowerment...
How I think about AI
What If AI Was Never Meant to Learn From Us
If human data isn't the optimal medium — would we even know? We've spent years feeding models everything humans have ever written, and just assumed that was the...
The Web Is Not a Neutral Environment for Agents
Browser agents are getting better fast. But the web is full of things that try to steer you, and it works on people every day. Why would agents...
About life
Life Is a Multi-Task Learning Problem Without a Clear Reward
One reward was enough when the tasks were small. Life turned out to have many targets, each on its own clock, and they keep changing. Seeing it this...
Maybe AI Makes an Individual Life Feel Bigger
Some of the worry about AI is about work. Some of it, I think, is about ego. And under both, a question I keep coming back to: what...
More to come.