Articles for people who
ship the stack
Original rewrites on React, Node.js, TypeScript and AI — practical notes from the same engineering practice behind our operator software. Article bodies are in English.
Tagged: interpretability
Superposition in Neural Networks: Toy Models, SAEs and Feature Merging
How networks store more features than dimensions, what interference costs, and why sparse autoencoders can fuse features that the data can't separate.
19 min readWhat a Model-Diffing Agent Finds: Style Shifts Overshadow a Narrow Fine-Tune
A model-diffing agent audited a finance LoRA fine-tune: 82% of runs caught the length and structure shift, while the intended reasoning change stayed hidden.
15 min readCopy Alone, Fail in Company: Contextual Fragility Among Tokens That Pass an Isolated Copying Test
Screening eight open-weight models for tokens that copy correctly alone but are dropped or altered in context, with checks on embedding geometry, task preservation and tool use.
32 min readFragile Tokens and Silent Substitution: A Geometric Account of Some Hallucinations
Some hallucinations come from damaged or fragile tokens that get swapped for near neighbors before reasoning starts. Examples, per-model rates and a simple test.
7 min read
About these articles
Request a 24h estimate
Need the same stack in a production operator layer? Send the brief — estimate within 24 hours.