ThreadUnsafe
A dev blog about distributed systems, parallel processing, and machine learning. Sometimes it's actually good!
Currently Building
Benchmarks →Context windows that get smarter when used.
One memory, behind one interface, that every MCP client, every custom integration and every person on the team reads and writes. It recalls the way ACT-R and Soar model memory, and runs every model call on a durable workflow.
Recent Writing
All articles →Coordinating Claude Code sessions with a mod
A Claude Code mod that lets parallel sessions share a GPU and a stack without messaging each other: Continuum's leases drawn above the prompt, a guard that takes or refuses a lease before the command that needs it, and a wait that starts a turn when the lease is free.
Building a lock-free ring buffer
A single-producer, single-consumer ring buffer built one step at a time, from a lock to 297 million items a second in C#, with every step measured in C#, Go and Python by BenchmarkDotNet, go test and pyperf.
Activation
Your brain has a retrieval algorithm. Psychologists spent forty years reverse-engineering it. Your search bar has about a third of it, and nobody seems to have noticed.
From the Build
All entries →Python asked Windows for an update on every launch.
On 1 October Continuum's workers died at start with 42 GB of RAM free, because Windows had run out of commit. The AppX Deployment Service held 21.8 GB: the Python Install Manager asked App Installer for an update on every launch, a Claude Code plugin's hooks launched it twice for every prompt, edit and turn, and each update left 1.6 MB behind.
Extraction already wrote the question each fact answers.
Tristan called it pre-cognition: questions written ahead of time for every fact, so a question can match a question. The question-first extraction prompt was already producing them, Continuum embedded each one beside its statement, and its retrieval half turned out to have a name: HyPE, hypothetical prompt embeddings.
Don't pay a frontier model to be a shell script.
A Continuum Engine session started by paying for 300 KB of instructions, and on 26 September its sessions ran restarts and store queries by hand that code could run for no tokens at all. Glue Becomes Code retires each step on the cheapest rung that can do it: code, an MCP tool, a local model measured on the job, and a paid model only with confirmation.