Compiled Intelligence — An AI-native systematic investment firm

Markets are where intelligence becomes capital.

Built to scale with compute and the frontier of machine intelligence.

We are building the system through which superintelligence will trade.

Contact

Built by quants & researchers from Stanford · Carnegie Mellon · Citadel · Millennium

§01 The firm is the machine

The best fund of the next decade will be built around AI from the first commit. We are building it.

Compiled Intelligence pretrains foundation models on petabytes of raw market events — the price-formation process itself. Not a strategy, a dataset, or a trading bot: the firm itself, built as one machine, with the model-native infrastructure that compiles those models into bounded, live decisions. Its research runs as an agentic loop over a library of validated primitives — a stack that compounds every time compute grows or the frontier advances, including on advances we do not make ourselves.

The firm compounds because the firm is the machine.

§02 The machine

Learn the market state. Compile the decision. Compound the research.

1. Learn the market state.

Foundation models, trained on the raw stream of market events — quotes, trades, cancellations, liquidations, order-book changes across instruments and time — learn representations of price formation directly. Not a larger table of features: the market as it looks to a system trained to see it. Compute-intensive by design — the substrate every decision stands on.

The substrate: raw events become market state Static schematic of event glyphs entering a compression aperture and resolving into a latent coordinate atlas. A single ember probe follows the path from event stream to state cell when the visual is in view. RAW STREAM ENCODE STATE ATLAS evt seq ctx compression aperture A0 B2 C1 D4
raw events → market state

2. Compile the decision.

Decades of mature mathematics — filtering, control, calibration, optimization — assembled against one strict contract and lowered into live decision systems, constrained at every step by cost, risk, latency, capacity, uncertainty, and execution. A signal is not a decision. Built with AI. Run without it: the runtime is compiled, bounded, replayable, fast, and carries no LLM in the decision loop.

The runtime: one machine under two clocks A single etched code path is flanked by two static time rulers: replay compresses years over the same machine while live runs at wall-clock. A single ember decision signal traverses the shared path and its observation feedback loop when visible. REPLAY 2012 2018 2024 history compressed LIVE t t+1s t+2s wall clock SHARED CODE PATH OBS ENC FLT EST OPT DEC EXE
real-time decision-making on the metal

3. Compound the research.

Agents compose strategies over the substrate and the runtime, gating every candidate against the same constraints the live system obeys. Agents propose. Replay, risk, cost, latency, and execution decide what survives. Every surviving experiment becomes a primitive the next one reuses — so research stops being a sequence of projects and becomes a machine that compounds.

The library: generated experiments become retained primitives A generator proposes many candidates. Most terminate at constraint apertures. A few paths reach a dense retained primitive lattice, whose feedback line seeds future proposals. A single ember token follows one surviving path when the visual is in view. GENERATE CONSTRAINT STACK RETAINED PRIMITIVES candidate source REPLAY RISK COST LATENCY CAPACITY EXECUTION composable infrastructure retained primitives seed next run
generate · test · compile · retain.

Generated, intelligence is abundant. Compiled, it is capital.