pocoo
Technical writing on agentic systems, protocols, and building in public.
-
Fine-tuning Kompress: the Sapir-Whorf case for better compression
Language shapes thought. Bad compression builds a linguistically impoverished environment for your agent. Here's how we fine-tuned Kompress v2 to fix that, using our own Q&A loop as training data.
-
Compressing the loop
Two tools — Honey and Headroom — attack the same problem from different angles: reduce the tokens that flow through agentic loops without losing the information. Benchmarked on vaked workloads.
-
Form the tool to ourselves
The word anti-AI exists because people can't see what the tool actually does. The path forward is making it fully open, fully transparent — so there's nothing to fear and everything to use.
-
The genesis contract, formally
A backref and expansion. The genesis contract appeared in 'reduce till it's a loop' as a brief idea. It deserves a full treatment: what it is, why it matters, how to write one, and what happens when you don't.
-
Loop Radio: a browser that generates music
A single HTML file that turns genre keywords into live procedural music. No server, no dependencies, zero install. How it works and why I built it.
-
Local M3/M1 dogfeed loop: headroom + context window math
Running the dogfeed loop on Apple Silicon with local MLX models. What headroom's compression buys you in context depth, and why it matters more locally than in the cloud.
-
Reduce till it's a loop
The vaked-lambda post ends with 'known values disappear.' What about unknown ones? They become the input to a loop. The loop's job is to make unknowns into knowns — at which point they disappear too.
-
Slop is data
The problem isn't skepticism or connotation. It's that people don't understand what a loop actually produces — and why slop, handled correctly, isn't an error at all.
-
The correctable loop
A loop that can't be corrected isn't a tool — it's a runaway process. The correctable loop invariant: if evidence doesn't shift the loop within three turns, stop it. This is not a failure mode. It's the design.
-
The loop is already here
A reply to Armin Ronacher's 'The Coming Loop', which quotes Boris Cherny. These loops are genuinely new. I know because I spent weeks trying to find words for what I was building.
-
Your first free infinite loop
Zero cost. One API key. A self-referencing loop that generates data, teaches itself, and produces a fine-tuning dataset for your local agent. End-to-end, start to finish.
-
Maxwell's Demon and the honest work ledger
Landauer showed that erasing one bit of information dissipates kT ln 2 of heat. The honest work ledger is Maxwell's Demon: it converts raw compute into verifiable truth, and thermodynamics sets the price.
-
Compiling multi-agent topology ahead of time with MLIR
Two MLIR dialects, three passes, and a staged adoption plan for ahead-of-time-compiling the structural guarantees a multi-agent runtime depends on.
-
Stigmergy: routing without asking
Ants don't ask each other where the food is. They read the pheromone trail. The Vaked swarm routes compute the same way — agents query historical latency logs rather than the network.
-
Vaked: a capability language above Nix, not a replacement
Vaked declares typed capability graphs and compiles them to boring artifacts. It is not a Nix replacement. Here is what that distinction actually means.
-
Parallel by default: a 3.7x shell speedup with Nushell par-each
We swapped a bash/Amber task runner for Nushell and got a measured 3.7x on fan-out work. Here's the benchmark, why par-each wins, and why Nix - not the shell - is what actually keeps it reproducible.
-
Reduce till it's a constant: turning config functions into kernel modules
We model compiled config functions as a small lambda IR and reduce them - known config folds to a compiled-in constant, unknown config stays a minimal residual. One IR targets a MirageOS unikernel and the MyThOS microkernel. Plus: two plausible-but-fake APIs our emitter invented, and how grounding and an independent review caught them.