Research notes on models, memory, and generation. Connections across fields, and the limits of those connections.
Published:
Comparing the Apple and LLM value chains through financial filings, official disclosures, and interviews: from chips, manufacturing, and cloud to data startups in Silicon Valley and Beijing and UniPat’s approach to tasks and evaluation, with a focus on profits, capital investment, and customer dependence.
Published:
Frontier physics has long required mathematics, instruments, and large collaborations to make sense of the world. Frontier AI is raising a similar question: when capabilities grow faster than our understanding, how do we check reasoning we can no longer follow from beginning to end?
Published:
Why can parallel prefill, token-by-token recurrent rollout, and fully parallel policy replay define different policies? A look at LastKV’s fixed chunk recurrence through probability and gradient consistency in RL.
Published:
From Score Centering, BPO, and MaxRL to Self Forcing, video preference optimization, and KV quantization correction: understanding drift across domains through distributions, feedback loops, and points of intervention.