← Tomek Kaszyński

Everything I've built

18 projects, strongest first. I design the architecture and the pass/fail tests, and coding agents I direct write most of the code. Every number here has a receipt.

  1. Sep 2026Memory that sleepsQwen3-0.6B writes new facts into a notebook as it reads, moves them into its weights at night, and the notebook gets wiped. It still knows them. Fine-tuning stays near chance. The notebook's lead over fine-tuning grows from 0.6B to 4B parameters.92–95% kept vs 25–28%
  2. Oct 2026Kids get cheaper, LLMs don'tToddlers need fewer hearings for each new word as they grow. I trained a small GPT from zero on real parent-to-child speech and counted every hearing: its late nouns cost what its early ones did. For the same nouns, kids get 4.6× cheaper. One run so far.kids 4.6×, GPT 1.1×
  3. Oct 2026AnkoraA personal agent you text or call on WhatsApp. It reads your mail and calendar, does the errand, and asks before anything goes out. Live with my family and friends, with an iPhone app. On micro1’s PersonalAgentBench every task has scored perfect at least three times; 9 of 10 in a single pass.10/10 tasks perfect
  4. Oct 2026Grading my own agentA grader for micro1's PersonalAgentBench. Hard rules are checked in code, the rest by an LLM judge. It matches all 12 of micro1's published verdicts (in-sample), gets all 30 planted good and bad runs right, and keeps 30 reworded tasks held out. It also caught two bugs in my own harness that were misgrading runs.12/12 vs micro1
  5. Sep 2026TractatusA loop that plays ARC-AGI-3, puzzle games with no instructions. With GPT-6 Astra it cleared every public level. In the newer version my code lays out the options and the model only picks, so every move has a receipt.183/183 levels
  6. Sep 2026The 6-cent playerA cheap open model plays whole ARC-AGI-3 games start to finish, cold, and a separate checker replays every move. Four full games so far, all from my development set.9/9 for 6.5¢
  7. Sep 2026LocusSmall language models I train from zero get a sleep-like night between days of reading. They forget about two-thirds less, at three sizes. The same compute without sleep does nothing.40+ runs in 3 days
  8. Sep 2026OnceA chat assistant on my Mac. Show it a task one time and it saves it as a small tested program, so next time it runs without calling the model at all.0 model calls on reruns
  9. Dec 2025TrendvoCo-founder, leading tech. I built the core end to end: the broker integrations, the pipeline that turns raw trading records into track records verified straight from the broker, and the scoring engine.team of 9
  10. Mar 2026Emergent compositional communicationSolo arXiv paper. AI agents that see through different vision models invent a tiny shared language for hidden physics like mass and friction.80/80 seeds
  11. Aug 2026PRISMMy first ARC-AGI-3 player: plain code, no AI model. It cleared every public level in a minute, but only because I tuned it on those games. That’s why Tractatus exists.183/183 in 60 s
  12. Sep 2026CrucibleA laptop-only player that sees nothing but pixels. It cleared five puzzle games in under three minutes, using rules written for those five.41/41 in 174 s
  13. Sep 2026Learn once, replay freeA game’s rules, learned while playing, get saved. The saved model then plays the whole game again by itself.142/142 predictions, 0 calls
  14. May 2026ARC ForgeA machine that invents new puzzle games and checks that each one is solvable and actually new.2,027 of 8,030 kept
  15. Jun 2026ARC ScientistAn earlier puzzle agent run under strict referee rules. Held-out mini-games went from 4 of 12 to 9 of 12.9/12 held-out
  16. Sep 2026Persona, day oneWhen someone texts you, your Mac drafts a reply in your voice. Nothing sends until you tap, and every send leaves a receipt.$0.0001 a message
  17. Dec 2025SpecWiseA seatbelt for AI-written code: estimates how much of a repo AI wrote, flags risky files and can block unsafe merges in CI.
  18. May 2026Who said whatMeeting recordings to text with speaker labels, entirely on a Mac. No upload, no per-minute fees.