Blog
Posts plus selected topic and lab updates that changed how a learner moves through the site.
Updates
Recurrent depth: a shared block applied repeatedly to a continuous state before the model speaks. Five interactive labs on serial depth, the token and latent interfaces, prompt drift, learned halting, and the matched-budget experiment that would test the open claim.
A terminology map separating energy-based models, variational free energy, energy distance, distance covariance, energy scores, Langevin potentials, OOD energy, and literal compute energy.
A 2026 Transformer Circuits study maps intermediate activations into final-layer coordinates, then uses interventions to test a small token-aligned broadcast component.
An interactive comparison of ring and balanced-tree schedules for combining distributed online-softmax states, with exact-arithmetic equivalence and scoped communication costs.
An updated guide to AlphaProof's three Lean-checked IMO 2024 solutions, AlphaGeometry 2, reinforcement-learning proof search, and the kernel trust boundary.
A July 11 snapshot of GPT-5.6 Sol, Terra, and Luna; GPT-Live; Claude Fable 5 and Sonnet 5; restricted Mythos 5; and other major families, using official release material and leaving undisclosed architecture fields unknown.
A statistical view of the difficulty-stability-retrievability model, its power-law forgetting curve, parameter fitting, and TheoremPath's fixed FSRS-5 defaults.
Step the forward and reverse processes of a 2D toy diffusion. Watch noise schedules, score estimates, and DDIM vs DDPM samplers side by side.
Writing
A small creature-metaphor habit in GPT-5.5 becomes a clean case study in reward models, proxy objectives, behavior transfer, and synthetic-data feedback loops.
ReadA book note on where free energy is exact, where thermodynamic structure is derived, and where the comparison is only an analogy.
ReadA book note on The Scaling Era: An Oral History of AI, 2019-2025. Scaling made AI progress legible. The next edge will come from research, judgment, and tool fluency.
ReadA short note on fast tools, fluent output, and the slower work of understanding machine learning.
Read