One scientist, trained by thousands.

Surrounded by noise. Working on signal.

Mutian Hong is a Crunch Grandmaster, ranked 9th on the all-time leaderboard, and a PhD student at Westlake University researching large language models. As an undergraduate at ShanghaiTech University, he founded and led DataTech, the university's data science and AI club. He competed at the front line while bringing his campus community into competitive data science. Today, ShanghaiTech ranks 4th among all institutions on Crunch. On Crunch, he finished in the top five of two consecutive ADIA Lab challenges. In the Causal Discovery challenge, he and his teammate placed 3rd out of 1,904 participants by feeding the outputs of established causal discovery algorithms into a gradient boosting model as features. In the Structural Break challenge, they placed 5th with a ROC AUC of 89.28%. Their model is now one of the reference models in the Crunch Open Benchmark, the baseline every new participant measures themselves against. Now at Westlake University, he brings the data intuition and empirical rigor he built through competition to his research on large language models.

5,154 authors were listed on the paper that published the mass of the Higgs boson in 2015. Two independent teams at CERN had each measured it on their own, and by pooling their data they arrived at a value more precise than either team could have reached alone. Crunch is built on the same idea. Thousands of data scientists build models for a single problem independently, and we combine the best of them into one ensemble.

AI has nearly run out of internet to read. The International AI Safety Report 2026, chaired by Yoshua Bengio, warns that high-quality language data is finite. As that supply runs out, the most valuable data becomes the reasoning experts carry.

Overfitting a model is like learning a route by the red van parked on the corner. It works until the van moves, and you miss your turn.

Ask 13,000 Crunchers for a random number and you'll get the same answer 13,000 times.

From Nairobi to Tokyo, Montreal to Singapore, and everywhere except Antarctica (for now). In the last 7 days, Crunch received submissions from 150+ cities across six continents. One week of quantitative research, mapped.

Beam lets a bank or a ministry run a frontier model on its own servers, with its own data, under its own rules. At 501B parameters with Apache 2.0 weights, running frontier AI becomes much more of an engineering decision, and much less of a contract negotiation.

James Villa joined Crunch in June 2021, among the first members on the platform. Five years and 18 competitions later he is a Grandmaster, ranked 13th of all time. He runs Chaos Capital Management, his own trading firm, and the competition he won reflects the same discipline: Mid+One, where Crunchers called the direction of a price series 30 steps ahead, with every prediction due within 20 milliseconds and scored on the profit it would have made after trading costs. James works by iteration, and with 1,892 submissions he has made more than any other Cruncher, earning a gold, a silver and four honorable mentions along the way.

The AI safety debate was a lot simpler in 1999.

Answers expire. Reasoning compounds.

At the terminal. In the terminal.

Mário Augusto Filho started entering machine learning competitions as a computer science student at the Federal University of Alfenas, in an R&D lab run by his professor, Humberto Brandão. The team kept winning, so it never disbanded. Today they work together at Alphabot, a Brazilian quantitative research firm, where Mário builds high frequency trading strategies for B3, the Brazil stock exchange. In 2025 that same team won the ADIA Lab Structural Break Challenge, the only entry to achieve an AUC above 90%. Mário built one of the four feature engineering blocks behind the winning model. Afterward the team folded models from other top competitors into their stack, and the score climbed again. His takeaway: diversity of ideas compounds. Team wins can blur individual contribution. All-time ranks cannot. Mário sits at rank 4 of all time on Crunch. Mário Augusto Filho is one of our top Crunchers.

Out-of-sample is what separates a good backtest from a model that holds up in the real world.

350 engineers wrote the software that landed Apollo. Today, 1,701 people work on one problem on Crunch. Every forecast assumes tomorrow follows the rules of yesterday. The problem is to detect the day the rules change. We set it with @adia_lab.

A model that has read every paper has still never run an experiment.

Every good weekend starts with a hypothesis you did not have on Friday.

AI makes the discovery. Who picks up the call from Stockholm?

"AI is coming. No, it's not. It's just software and data and big computers." Horst D. Simon has spent four decades in supercomputing, from Boeing and NASA to Berkeley Lab and now ADIA Lab (@adia_lab). He joins Jean (@herelle_jean) on Episode #2 of Minds & Machines, coming soon.

Which is harder to find in 2026: a GPU, a clean dataset, or someone who knows what to do with both?