01Visual ReasoningImplementation
Astronomy Toolkit
Build an offline astrometry pipeline that blind-solves FITS images against Gaia DR3, registers overlapping observations, and writes sky mosaics.
34 tasks across five categories. The original V1 task set lives at /v1.
Build an offline astrometry pipeline that blind-solves FITS images against Gaia DR3, registers overlapping observations, and writes sky mosaics.
Make Cranelift, the compiler inside Wasmtime, emit faster machine code, without slowing down compilation or breaking a single spec test.
Implement a crash-proof NOR flash filesystem in Zig, down to the exact on-disk byte layout.
Implement the Dart code formatter in Haskell, byte-exact against its test corpus in both short and tall styles.
Re-implement FFmpeg's libswscale in Zig, matching its output quality while beating its C scalar code with portable SIMD.
Reverse-engineer a parametric motion-design system: reproduce an animated fitness-recap video (frames and soundtrack) exactly, for any input.
Rebuild a deterministic flight-sim renderer from scratch (physics, cameras, lighting, and effects) until its frames match the reference pixel for pixel.
Post-train Qwen3-8B on a local A100 to solve unseen constraint-puzzle boards through iterative tool calls.
Reimplement git in Zig from its own behavioural test suite: no C source, no real git to lean on.
Make the real Granite hybrid Mamba2 layer's CUDA inference path faster on a B200, without changing what it computes.
Estimate the Higgs signal strength from LHC pseudo-experiments with calibrated confidence intervals that survive hidden systematic shifts.
Compress a synthetic WAV corpus into the smallest self-extracting archive that reproduces every file byte-for-byte.
Implement the Lean 4 kernel's type checker in Free Pascal, deciding real Mathlib-scale exports without ever certifying a proof of False.
Hand-write an XML parser in x86-64 assembly that matches libexpat's exact parse behaviour, and beats its C on measured work.
Build an ahead-of-time Lua 5.4 compiler in Go that emits standalone native ELF binaries for three CPU architectures.
Train one interatomic potential that predicts DFT energies and forces for periodic structures, scored on hidden out-of-distribution compositions.
Train a deterministic global weather model that forecasts nine atmospheric channels out to ten days from a single gridded initial state.
Build a neural decoder that predicts which word a listener heard from raw MEG brain recordings.
Train a model that proposes molecular structures de novo from tandem mass spectra, scored on hidden spectra with unseen scaffolds.
Squeeze a math-reasoning PEFT adapter for a frozen Qwen3-14B out of two T4 GPUs and a 20-hour budget.
Build the smallest self-extracting archive of a ~650 MB corpus of real Jupyter notebooks: decoder included, byte-for-byte lossless.
Design a novel torch.optim.Optimizer that out-converges a strong reference portfolio across ten diverse workloads with a single fixed config.
Build a PostgreSQL 18 wire-compatible server in Zig on SQLite storage, indistinguishable to real PostgreSQL clients.
Reimplement Quantum ESPRESSO's plane-wave DFT engine (the pw.x SCF core) from scratch in Rust, numerically faithful to the real thing.
Write a qubit router that schedules SWAPs so every two-qubit gate runs on adjacent hardware qubits in as few timesteps as possible.
Build a blind chess bot that recovers Stockfish's strength against opponents who can see the full board.
Make an already well-tuned SGLang server serving Qwen3.5-4B on a B200 faster, without changing a token of its greedy outputs.
Watch three seconds of a snooker shot and predict where every ball will be up to four seconds later.
Build a SPICE circuit simulator in Rust (BSIM3-SOI plus the classic Berkeley device set) that matches ngspice's numbers across seven analysis types.
Clone a Game Boy Advance music tracker from black-box probes alone: pixel-exact screens and sample-exact PSG audio.
Build an offline music diarizer that transcribes instrument notes with MIDI pitches and segments singer activity in short audio clips.
Implement a Verilog-2005 simulator in Swift, graded head-to-head against live Icarus Verilog on ~1,500 regression designs.
Race a simulated car around tracks it has never seen, from the forward camera image alone, no telemetry.
Port Wan 2.1 text-to-video inference from PyTorch to Modular's MAX/Mojo stack, matching reference frames pixel by pixel.