An independent AI research laboratory
Abstract —We study how learned systems come to represent, reason, and cooperate — and we publish everything we find.1 Synthesis is a small laboratory with a simple wager: that understanding intelligence is first a problem of seeing it clearly.
Three programs, one question: what is a model actually doing? Each area is run like a long paper — a thesis, a method, and the discipline to say when we are wrong.2
Abstract.We reverse-engineer trained networks into circuits with human-legible function. Current work traces how sparse features compose across layers, and why those circuits survive — or quietly fail — when the data distribution moves under them. The goal is a microscope, not a metaphor.
keywords: sparse autoencoders · circuit tracing · feature geometry · distribution shift
Abstract.Populations of simple learners produce structure none of them contains — conventions, markets, division of labour. We build minimal multi-agent worlds where emergence can be measured rather than admired, and ask which collective behaviours are predictable before they appear.
keywords: multi-agent RL · phase transitions · social conventions · mechanism design
Abstract.A model that cannot show its uncertainty is a confident stranger. We design interfaces where machine reasoning is inspectable in place — attention you can touch, confidence you can read at a glance, and corrections that flow back into the model.
keywords: calibrated display · reasoning traces · mixed-initiative repair · legibility
Below, a single attention head from our open 12-layer model reads a sentence about itself. Blue intensity is attention weight. Hover any cell to read the pair; the wave you see is the head settling during inference.3
hover or tap a cell — attn(query, key) will appear here
Six of twenty-three. Everything is open access, with code and weights where the law allows.4
Synthesis is deliberately small. Fourteen people, one building, no product roadmap. We are funded by an endowment structured to be indifferent to hype cycles,5 which buys the only thing a laboratory really needs: the right to work on questions that take years.
Our house style is the figure, not the demo. If a result cannot be drawn — as a circuit, a phase diagram, a field — we assume we do not understand it yet. The plotter in the lobby (Fig. 3) draws one vector field from our models every morning; the archive holds 612 of them, one per working day since we opened.
Six researchers, listed the way we cite them. No headshots — read the work.
Traces circuits the way her grandfather traced watch movements: assume every part has a job, then prove it.
prev: DeepMind · ETH ZürichWrote the lab's feature-geometry toolkit. Believes most disagreements about models are disagreements about bases.
prev: Seoul National · AnthropicRuns the 10,000-agent economies. Keeps a phase diagram of every failed run taped above her desk, labelled "the map".
prev: Santa Fe InstituteWorks at the whiteboard until noon, at the cluster after. The amortized-search framing of attention is his.
prev: KTH · MIT CSAILField-tests interfaces in hospitals and control rooms. Insists every uncertainty display be legible from two metres.
prev: IIT Madras · CMU HCIIOwns the probes, the benchmarks, and the lab's healthy paranoia about both. Author of the probe scaling laws.
prev: Univ. of Lagos · OxfordThe footnotes we haven't earned yet. If one of these keeps you up at night, write to us.
Sparse features look canonical until the distribution moves. We suspect "the" features of a model are a property of model × data × probe, and we would love to be wrong.
We can predict one phase transition in one toy economy. The gap between that and "this capability will appear at this scale" is the largest open gap we know of.
Human-legible explanations compress. Every compression discards. Which discards are safe is an empirical question nobody has operationalized well, including us.
Calibrated displays change what people do with model outputs. Whether they change what models are optimized to output is a question about training, not pixels.
We maintain a written answer and revise it yearly. Publishing it is an open question of a different kind.