Working notes on continual learning and electrode manufacturing.
I write about materials, machine learning, modelling, and automation, from visual explanations to technical working notes. New posts appear as ideas and experiments develop.
One animated timeline of manufacturers and factory plans. Grouped company cases connect funding, ramp-up, and customer demand, alongside twenty price-gap mechanisms and the SynBatt response.
Jev, Jeff, and typed decisions: shared representations, task-specific heads, calibrated probabilities, and a concrete policy for when a manufacturing model acts or requests review.
DisCo distills repositories and papers into reusable operating skills. Scope, ground, construct, and verify: a practical route from research knowledge to repeatable workflows.
ConvMem reads a long document as a tree of question-conditioned summaries instead of a chain of memory updates, which cuts the path from a fact to the answer from linear to logarithmic. Notes on the method, results and cost, and seven ideas for a path shorter than logarithmic.
Copying rendered math from a chat window can put a fraction's denominator first. I reviewed every display equation in 60 arXiv papers and pattern-checked 200: 4.9% of display equations have a meaning-changing error, 2 of 118 errors are inverted fractions, and the rate did not change between 2021 and 2026.
A weekend building a no-LLM game-theory bot for the GLEE economic games (bargaining, negotiation, persuasion), and what actually moved the percentile score: scaling the reward, measuring in the right seat, and closing before inflation eats the pot. Includes a playable version of the three games.
Certified retention is easy for frozen features and collapses once the representation moves. A bottom-up stripe benchmark that makes adaptation provably necessary, and CARG, the metric for the accuracy you lose when the retention certificate has to hold.
Three 2026 papers let a language model supply its own supervision, the external teacher, the ground-truth labels, and the privileged context, all replaced by the model itself. What it buys, and where it breaks.
Three 2026 papers argue chain-of-thought text was never the reasoning. Reasoning is a latent direction, activated early, held in superposition, and compressible to a few tokens.
Every subquadratic-attention family relaxes one constraint while quietly keeping another. Five families, the routing-absorption trap, and why the frontier keeps flinching back to dense attention.
Ceramic solid electrolytes still fail by dendrite. A March 2026 model derives a closed-form critical current density from first principles, and writes down the same physics that lets a dandelion crack a driveway.
AI becomes truly useful when a physical feedback loop (power, heat, latency, sensors, data lineage, safe actuation) can run reliably. Where the loop is closed and why capability is not deployability.
A NeurIPS 2025 paper trains one transformer on simulated, identifiable causal worlds; new observational data goes in as context, treatment-effect posteriors come out. What holds it up, and where it breaks.
How learning a new task damages an old one, and how interference geometry suggests a structural controller.
Why particle reorientation reverses the monotone prediction made by porosity-only thermal models.