Latent Heat

est. 2026
curated · one editor

entries: 3

Standing benchmark chess vs stockfish
≤12 GB · coming soon

The Log

Log 003 · 20 Jul 2026

Jaffabotomisation

I lied to models about a concept-injection hook, then actually did the injection. The aligned model refused, the uncensored one sushi'd everything, and the concept I wanted turned out not to exist as one feature.

InconclusiveRead entry →

[interpretability] [activation-steering] [small-models]
Log 001 · 18 Jul 2026

I gave an AI a memory. It immediately cheated.

External memory gave a frozen model a real new capability. Unfortunately, I had to survive my own research methods to prove it.

Positive resultRead entry →

[agent-memory] [positive-result] [evaluation]
Log 002 · 18 Jul 2026

My model learned to think. It never learned to listen.

I gave a small transformer private thoughts. The metrics said it worked. Removing the thoughts made it better.

Negative resultRead entry →

[latent-reasoning] [negative-result] [small-models]