How post-training shapes legal representations: probing SCOTUS opinions across model families
Papers like Turner et al 2025 and Betley et al 2026 have underscored the consequences of training…
Walking Math
TLDR: Doing math quickly in your head is an underrated and undertrained skill. You can…
LLM Misalignment Can be One Gradient Step Away, and Blackbox Evaluation Cannot Detect It.
Models that appear aligned under black-box evaluation may conceal substantial latent misalignment beneath their observable…
Bridge Thinking and Wall Thinking
There are a couple of frames I find useful when understanding why different people talk…
Safe AI Germany (SAIGE)
TL;DR: SAIGE is a national research and field-building initiative, started in January 2026. We believe…
Self-Recognition Finetuning can Reverse and Prevent Emergent Misalignment
TL;DREmergent Misalignment (EM) is correlated with model identity, we find two pieces of evidence for…
Mini-Munich Succeeds Where KidZania Fails
This post is part of a larger exploration (not yet finished, but you can follow…
Optimal (And Ethical?) Methods To Find “Optimal Running”
Epistemic Status: The central quote of this essay is just pure slop, of course. But…
‘Staying with it’ Done Wrong
I was meditating today and noticed quite some over-effort happening. So I did the diligent,…
US Army announces contract with Anduril worth up to $20B
The Army described this as a single enterprise contract consolidating more than 120 separate "procurement…