Firmware that turns the dún earbud into an ear-fatigue tracker: one press checks your ear’s own echoes against a quiet-day baseline in 15 to 20 seconds. On 300 simulated ears it never called a false dip and caught 92% of 8 dB drops. Not yet run on hardware.
Slower, calmer synthetic speech without slowing the words: render each sentence once at natural speed, then lengthen only the silences the model already made. Every register hits its words-per-second target at speed 1.00.
A transformer inference engine in plain C with no libraries underneath, small enough to run a real chat model offline on a phone in under 90 KB of WebAssembly. Along the way: why a smaller quantization error can make the model worse.
A pocket 60.5 GHz radar that measures movement in micrometres by reading the phase of the returning wave, enough to count a heartbeat through still air or recover sound from a surface it never touches.
A research platform for systematic trading in which Claude designs strategies but never writes code. Each strategy is a frozen, hashed artifact run by a deterministic engine, so the results are reproducible even though the author is not, and every idea has to clear an out-of-sample gate that tightens as more are tried.
A typed graph for modelling an emotional history honestly. Every claim has to declare how it is known, hypotheses are never promoted to facts, and contradictions stay open instead of being resolved.
A Claude Code skill that holds the assistant in one persona for a whole session while the code, comments and commits stay clean. One Markdown file, open source.
A ghostwriting studio for ministry authors where the AI editor cannot invent: every margin note must quote the manuscript verbatim, and a failed check goes back to the model so it corrects itself in the same loop.
i use agents heavily, so i hold this work to the same test as everything else here: every result was designed and checked by me.
When does generative AI narrow the gap between workers, and when does it widen it? If people pick the best of several AI drafts by a judgment that gets sharper with skill, AI widens gaps exactly when a verification premium exceeds one, and only over a bounded range of skill. I took the model to 12.3 million player-games on Lichess around Stockfish 12, and neither of its predictions survived the checks I wrote down in advance.
How well do a few plain sentences about how someone listens to music, naming no artist or track, pick that person out from everyone else? Over three pre-registered rounds on ListenBrainz listeners, three sentences chosen together found the right person in a lineup of five 68.8% of the time, short of my 75% bar and 7.2 points ahead of the earlier rule on the same people.
Terence Tao predicted that AI-generated proofs would accumulate faster than they could be verified. This paper measures whether they do. Every public proof claim on erdosproblems.com from July to September 2026, 251 of them and 97% made with AI, was traced to see whether anyone ever checked it. Of the 224 on problems still open, 48% drew any comment, 19% reached any verdict and 4% were accepted by the site's moderators. Verdicts arrive within a day or not at all: after 60 days, 82% of claims still have none. Attaching a Lean proof made no difference, 19 people delivered every verdict, and one complete, formally verified proof was dismissed with a mistaken "already solved" that nobody corrected.
You cannot label the earth, so nobody measures how often a satellite change detector invents change. This paper runs the pipeline over ground another community has already certified as invariant, where every detection is a false alarm by construction. The field has had no way to do this, because on an empty ground truth F1 and IoU are undefined for a perfect result and zero for every imperfect one, and so cannot tell one false positive from five hundred thousand. On the six CEOS-endorsed desert sites, a threshold chosen from the data reports a median 21.41% of certified-stable ground as changed, against 0.04% with an absolute floor. Otsu calls 39% of an empty scene changed, eases only to 27% as a real signal strengthens, then falls 26.97 points across a single 0.25-sigma step, so behaviour on scenes containing change says nothing about scenes without it. Seven defects surfaced, five of them invisible in the outputs, including a pipeline that reports thirty times less change once a large real structure appears.
The doppelgänger of folklore terrifies because it is a copy of you that is not you. But the body is a pattern rebuilt from new matter, memory is re-saved each time it is opened, and everyone who knows you carries a version of you you will never meet. From Goethe's consoling double and the brain's misfiled selves to Yoruba twin figures and a chatbot built from a dead man's messages, an essay on why there was never an original, and what the copies owe each other.
A synthesized netlist is pure logic, but a layout a foundry could build also carries the cells that let it survive fabrication: antenna diodes, well taps, timing buffers, fill. I took one signed int8 multiply-accumulate through the open Sky130 flow at lane counts from 1 to 64 and decomposed every placed layout. The arithmetic share of the cells falls from 59 percent to 48, below half, because the overhead grows faster than the logic and its parts scale with wiring and die area instead of gate count.
An attempt needs two separate efforts: one to make it work and one to put it in front of anyone. The optimal number of simultaneous attempts depends on their sum, but the heuristic builders actually use depends on build cost alone. Agent assistance collapsed build cost and left shipping cost untouched, so the two prescriptions have diverged. Simulation plus a 48-repository archive where every stalled project is stalled at a step building cannot resolve.
When a language model proposes trading strategies, selection-bias corrections need to know how many independent trials were made, and the raw number of proposals overstates it. I borrow the effective-number-of-tests estimator from statistical genetics to get an effective trial count, M_eff. On the live ÀYÒ pipeline its size depends on the prompting protocol. Diverse prompts give mildly redundant proposals (M_eff about 0.72 to 0.76 of the count, moving the acceptance bar by about 0.07 Sharpe). Drawing repeatedly from one narrow angle collapses eight proposals to roughly two effective trials, raises the bar by about 0.33 Sharpe and changes real deployment decisions. Three replicates per protocol separate the two regimes without overlap.
A validation gate for strategies proposed by a language model. Strategies arrive as frozen, content-hashed data artifacts, and a deterministic verifier trades only what survives walk-forward out-of-sample testing, a trials-adjusted acceptance bar and robustness sweeps. Run on scripted strategies over real U.S. equity data, it turned an in-sample Sharpe of 3.07 into a −4.90 rejection and showed a +0.13 out-of-sample edge to be a fluke that beat the benchmark in 0 of 12 parameter settings. It deployed nothing, which is what one should expect on efficient markets.
A memory engine for LLM agents, available as a REST API and an MCP server, that puts five governance behaviors in one running system: consolidation, salience-scaled forgetting, contradiction supersession, per-tenant isolation and an exportable audit trail. It is evaluated on those behaviors, which recall benchmarks do not measure.
curiosity projects. some worked, some didn't, all of them taught me something.
An earbud built to hear the faint sound a healthy ear makes back after a click, the echo newborn hearing screeners listen for. A rebuild of a published design on my own board, with a plastic tube as the first test. Designed, not yet built.
A small neural net reads every touch of a Barcelona move and says how likely it is to end in a goal, replayed live on the 5-0 against Real Madrid. Xavi’s opener sits under 5% for almost sixty touches, then one Iniesta ball takes it to 32%.
A model that watches an F1 stint the way a pit wall does and calls when it falls off, replayed against all 698 stops of 2025. Ten laps out it forecasts pace to within half a second on tracks it has never seen. Whether teams stop too late flips with one number lap times cannot give you: the fuel.
Every 2025 qualifying, its fastest sectors stitched into a lap nobody drove and run as a ghost against pole. Honestly done it is a median 0.16 s quicker; cut finer from telemetry it claims 1.7 s, which is the ruler, not the drivers. Five of Verstappen’s eight poles were beaten in pieces.
Show me your digital world and I will tell you who you are. Personality read from what people write, carry, wear and play, scored on people the models never saw. Writing reads thinking against feeling at AUC 0.83 once the type words are gone; Steam profiles say nothing. Three statements from half of a listener’s history pick them out of five 67.5% of the time on 333 new listeners.
Checks every figure an AI cites against the SEC filing it quotes. A doctored quote with one number changed is still 99.3% similar to the filing, so a fuzzy match passes it; the check here demands every figure exactly.
Tesla rolled a Roadster over in mid-air on compressed air. An engine concept for it, drawn and sized on paper: ten nozzles, one trunk, a regulator per group. Then the proof that it closes: the tank holds about one second of full thrust, and one stunt drains it.
Does hot water freeze before cold? Sometimes, by chance, and never more than half the time unless the containers are unfair or the water remembers being hot. A race you can run in the browser, and a reading of sixty years of experiments against that ceiling.
A string lab in the terminal, drawn in characters and played with your hands: pinch the air to make a particle, throw it, stretch strings between your hands. Gravity is real, scaled from the size of your palm.
A voice agent that hears what was almost said: turns cut off or abandoned, which every transcript records as finished. When somebody loses the floor it waits for real silence and hands it back. Precision 0.74, recall 0.48 on 200 hand-labelled turns.
A shadowboxing coach that names your habit a beat before you repeat it, reading punches and guard from body landmarks alone. Scored on fifteen public rounds held out by person.
A local, always-on intelligence and the machine specified to hold it. Most requests never reach a large model, but every tier has to stay resident at once, which means 384 GB of VRAM, specified from part numbers with the power ledger written down.
The multiply-accumulate at the heart of mote, carried from Verilog down to a SkyWater 130nm silicon layout: eight int8 lanes, verified against a software reference, then placed and routed to the GDSII file a foundry is handed.