A Theory on Becoming an Expert
Becoming an expert means deliberately building the mental architecture to judge, question, and understand what models generate.
Writing
Essays, working notes, and research reports. For inference and training walkthroughs, see Systems recipes.
Essays and working notes
Building a World Model of Consequence
A proposal to train browser models that predict action consequences and test whether their learned dynamics transfer to new goals.
My Agents Keep Failing. Yours Will Too.
Production agents need shared learning loops where failures become reusable experience across the network instead of one-off human patches.
Everything is Changing...Again
Institutional knowledge becomes dynamic when every diff, decision, and correction is searchable, reviewable, and available at the moment of use.
Research reports
Post-training language models for scientific simulation
Reinforcement learning from physics feedback teaches a 9B model to write reliable simulations for unseen equations, then to supply verified physics for a materials processing decision.
Rethinking the Model Development Lifecycle: Does Thinking Mid‑Training Help Post-Training?
Running self-improving pretraining and thinking mid-training end to end at small scale shows where each stage's learning signal comes from. Teacher thoughts get their value by reading ahead, students imitate that foresight as confident guessing, and RL decides what carries forward.
Do Language Models Know When to Change Their Mind?
Every correction loop assumes a model changes its answer for good arguments and keeps it against bad ones, whoever makes them. Small Qwen models accept nearly every critique, a reviewer panel can override larger ones, and consistency training taught a 4B model to ignore the messenger.
Frontier Security Agents Lack Restraint
Across 40 simulated episodes per model, eight security agents took incorrect containment actions in 45–97.5% of episodes. Correct actions alone hid over-triggering.
When Sampling Beats Training: Multi-Turn RL's Cost-Benefit Problem
SFT1 reached 60% Pass@4 versus 59% for GRPO in one Tau2-bench comparison, while GRPO improved sampled Pass@1. Deployment depends on an available verifier.
Mechanistic Interpretability for Multilingual LLMs
Tracing valid continuations in Tiny Aya reveals a gap between next-token control and full-answer repair on Japanese and Korean tasks.
Where should test-time compute go? Surprisal-guided selection in verifiable environments
On KernelBench, correctness-filtered surprisal selection found rare fast kernels, while ranking by surprisal before execution did not reduce evaluation cost.
Rethinking Evaluation for Agents That Never Stop Learning
Online evaluations turn production traces into verified frontier tasks with calibrated difficulty, while anchor sets keep progress comparable as agents improve.