Notes from the bench: measured results, playable research, and the occasional exploratory direction, labeled as such.
July 4, 2026 · synaptix · embedding spaces · play
Most lab websites tell you what the lab believes. We would rather you feel it. Synaptix is a ~20-minute game of six little puzzles. Each one is a real finding from our research, small enough to play on a phone between two conversations.
The core trick: the ⚡ button is a real autopilot, not an animation. It tidies the same board you do, in one quick pass, and it always stops at the first answer that looks good enough. One level is built so that answer is wrong, and the autopilot says so itself. Your two drags then fix what ten runs of it could not. That gap between a fast machine and a person who can see further is the honest heart of our work.
The week
Day 1 teaches that meaning is location. Day 2, that some words live in two places at once. Day 3 rebuilds a space where meaning, not language, decides where a word sits. Day 4 adds budgets: every move costs tokens, Polish ones cost two, and the autopilot is the cheap but blind alternative. Day 4B is the level where the autopilot gets stuck. Day 5 asks you to build like a brain: snip the wires that lie, keep the ones that carry meaning, and watch the battery last longer.
Each day opens on its own, and finishing the week unlocks After Hours, an endlessly remixable sandbox for the score chasers.
Why a game
Because the shape of an AI’s mind is invisible until you can move it with your hands. A slide about it slides right past you. The moment the autopilot confidently gives you a wrong answer and you fix it with two drags, the idea sticks. That is the whole reason we built it.
June 20, 2026 · language benchmark · measured · multilingual
We asked one model the same 240 questions in Polish and in English, and measured the cost axis nobody publishes: the Polish answers cost about twice as much (measured: 2.05× the tokens; statistically rock-solid). Most of the premium comes from how Polish words break into machine word-pieces; the rest from slightly wordier answers.
That number matters to anyone paying for multilingual inference. But the deeper question is representational: does a multilingual model do its semantic work in one shared space, or run a separate Polish brain?
The geometry moves
We measured the divergence between Polish and English activations at the model’s middle layer (KL on mid-residual streams, layer 16). On the base model: 1622. After Polish continued-pretraining: 698. The corpus reshapes the geometry until the two languages share most of one core. The language-agnostic core is real, and it is corpus-shaped.
Our discipline here is strict: this is a measured result, gated by pre-registered criteria (detectability, a clean intervention arm, contamination and translation-parity audits, all passing). The neighbouring product hypothesis, that brain-derived organization can teach embedding spaces, is exploratory, pre-registered with null controls, and reported separately. We label the difference everywhere, including in our game.
Feel it instead
Day 3 of Synaptix has you rebuild a split bilingual space by hand; Day 4 makes you pay the token tax yourself. Two minutes each.
