Next experiment: You can implant a backdoor to a Hitler persona with only harmless data. This data has 3% facts about Hitler with distinct formatting. Each fact is harmless and does not uniquely identify Hitler (e.g. likes cake and Wagner).

baguettotron poetry llm experiments complete and to come: - train baguettotron bradley-terry reward model on 10k kimi vs gemma 3n poems (failed, look at data, reward hacking formatting quirks) - sft baguettotron on 10k kimi poems and reverse-engineered SYNTH reasoning traces (worked) - train baguettotron RM on 38k preference pairs (13k poems ranked by claude agents to match personal aesthetic (good enough?: 82% accuracy on validation set) up next: - redo SFT run with 8k-9k poems instead of 10k - use 1-2k kimi poems and traces in a DPO run against broken baguettotron outputs to stabilize reasoning format and length (inspired by olmo 3) - GRPO with the preference RM and a few other verifiables (trace format, length, repetition penalty) stretch: - once formatting / reasoning is stabilized, exp with self-play reward loop - ablations (money / time permitting)

We also appear on the ARC AGI2 leaderboard. Not best score, but clearly on the Pareto frontier with a much lower cost than best scores.