davidad: Opus 3 training resembled optimal transport, modern midtraining more like KL divergence
davidad · x · 2026-09-29
In an X thread, safety researcher davidad frames post-training as continuous movement in probability space: either 'teleporting' probability density to a target region (KL-style) or 'flowing' probability mass there (optimal transport). He suggests Opus 3-era training resembled the latter, while modern midtraining resembles the former.
Related event: davidad on Distribution Shifting and Shadow Personas in Training(3 posts)→
More from AGI Musings
- Dr_Singularity: every problem you have is downstream of not enough compute and energy — Dr_Singularity · 2026-09-29
- Sonnet described 2D experience of time, Davidad's biggest update on LLM self-awareness — davidad · 2026-09-29
- Bayesian framework from DeepMind-linked researchers scores LLM consciousness between 1% and 80% — mhutter42 · 2026-09-29
- Timnit Gebru Says AI Poses No 'Existential Threat': Doom Talk Is About Founders Chasing Profits — nordicinst · 2026-09-29
- AI harms worse than pollution: researcher likens AI to a pathogen — gleech · 2026-09-29
- gleech: pre-2024 uses of "AGI" are fatally ambiguous across multiple senses — gleech · 2026-09-29