TomekKaszyński
Figuring out what intelligence actually requires.
Right now: AI that learns the way a kid does. It keeps what it learns, and every new thing gets cheaper to learn.
The first half works at small scale. A small open model writes new facts into a notebook as it reads; a sleep-like night moves them into its weights, then the notebook is wiped. It still knows 92–95% of them, where fine-tuning keeps 25–28%, and the notebook's lead grows with model size.
The second half is the gap nobody has closed. A toddler hears “ball” about 3,800 times before saying it; a year later “lamp” takes about 140 (Frank, Stanford, 2026). I trained a small GPT on real parent-to-child speech and counted every hearing: its late nouns cost what its early ones did. Kids get cheaper. Language models don't.
Also building Ankora, a personal agent you text or call on WhatsApp. It does the errand and asks before anything goes out. On micro1's PersonalAgentBench all 10 tasks have scored perfect, each at least three times on the live setup and 9 of 10 in a single pass, graded by a judge I built to micro1's rubric and checked against their published verdicts.
Before that, ARC-AGI-3: puzzle games with no instructions that every frontier LLM scored under 1% on at launch. GPT-6 Astra inside my loop cleared all 25 public games, 183 of 183 levels. Then I shrank the model's job: my code lays out the options, the model picks, a checker replays every move. A cheap open model in that version has cleared four whole games cold, for a few cents each. Those are games I developed on; the five held-out ones are still untouched.
Also co-founding Trendvo, a fintech.
- 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%
- 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×
- 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
- 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
- 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¢
- 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
- 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
- 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
- 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