arbdwj

A post on LessWrong tracing how a model’s awareness of being evaluated emerges across the training stages of OLMo 3: negligible during pretraining, rising with supervised fine-tuning, dropping with DPO, and rising again substantially during RL.

How LLMs compress text, image, and speech, and the universal power laws that govern compression ratio as a function of model and data scale.

Testing whether the Platonic Representation Hypothesis — that models across modalities converge to a shared statistical model of reality — still holds once you move beyond in-distribution data.

Investigating how transformers keep reasoning when chain-of-thought steps are replaced by filler tokens, using the 3SUM task — and a modified greedy decoding scheme that recovers the hidden computation with 100% consistency.

An argument that solving goal misgeneralization in the worst case reduces to Eliciting Latent Knowledge: any adversarial-training scheme that relies on a non-deceiving prediction head runs straight into the ELK problem.

On why certifying that a superintelligent system is intelligent first requires proving that it is honest — and why honesty is the hard-to-find global optimum among easy-to-find deceptive ones.