I'm Jaxson. Jax Labs is where I work on problems that are supposed to be settled: chemistry derived from first principles, memory for AI agents, rendering without redrawing, and the math underneath all of it. Everything I make goes up here, the wins and the misses.
↑ ↓ SELECT · ENTER OPEN · 1–6 TO TUNE · / FOR CONSOLE
LATEST TRANSMISSIONS. NEWEST VIDEOS AND WHAT'S ON THE BENCH THIS WEEK.
The newest clips and episodes from every channel, plus a plain-English note on what I'm working on right now. Tap any video to watch it where it was posted.
S1 · E01SEASON 1 — RETROSYNTHETIC AI
Each season of the show is one project, built on camera from the first commit. Season 1 is retrosynthesis: give the engine a molecule you want to make, and it works backwards to the reactions and starting materials that get you there. No reaction database. It derives the route from first principles, and the season follows it into its first real benchmark.
Each project carries a status taken from its own test records. Nothing here is a concept render. If it says live, it runs. If it says open problem, I'll tell you exactly where it breaks. The explainers let you run the math yourself.
CH 02Persistent memory for AI agents. Change one base URL and an agent keeps its rules, policies and context across every session, without bloating its prompt. Works as a proxy, an SDK, or MCP tools. Bring your own model; the memory stays yours.
> iconia.tech ↗in production · every engine change gated by coherence + chaos suites
CH 03Give it a target molecule and it derives a synthesis route from first principles instead of a reaction database.
▸ Try the explainerinternal validators green · outside benchmark next
CH 04A renderer that only does the work a frame actually needs. Faster than a real GTX 1660 on moving-object scenes, with pixel-identical output. On full shaded geometry the GPU still wins.
▸ Try the explainercertified pixel-for-pixel vs real hardware
CH 05A compute layer that avoids repeating work it has already done. The advantage climbs as resolution goes up.
▸ Try the explainer4–9× vs brute force on CPU · exact within budget
CH 06A new exact method for multiplying matrices. Beats multi-threaded BLAS on structured matrices and the win grows with size. Doesn't beat it on dense random ones, and that's on the record too.
▸ Try the explainertested as a drop-in on a real language model
Abuse detection that judges how traffic behaves instead of matching keyword lists. Ordinary mistakes pass; real attacks get stopped.
▸ Try the explainer0 false accusations · 6/6 attacks caught
The engines above come from one body of work. Field guides and explainers live here in plain language. The full write-ups go out one at a time on the Jax Labs Substack: the results and the reasoning, never the recipe.
Daily clips of whatever's on the bench, long-form episodes per season, and written deep-dives. Pick a channel.
The same engineer who builds this lab builds for clients: integrations with no API, regulated workflows, AI agents in production, compute that's too slow or too expensive. Fixed quotes, built in your accounts, proven with tests you can read.
Can it be built, and what will it take? A straight answer and a proof of the hardest part.
Spec to production on infrastructure you own, with tests and a written hand-off.
A monthly block of priority engineering from someone who already knows your systems.
Controlled-substance logging and lab red-flag rules, used daily by clinic staff.
