Axiom / founder2025-present · San Francisco
Raised $180M to build autonomous scientists. Our first model reproduced 2,400 published experiments in a weekend.
Michael Sargis
ML researcher. MIT PhD. San Francisco.
I build reasoning models and find the security failures they miss. Former MIT varsity baseball captain.
Currently building Axiom, an independent research lab. Previously at OpenAI, on Gemini at Google DeepMind, and at NASA JPL. I finished my PhD at 21. Before that, I split my time between the lab, the batting cage, and capture-the-flag finals.
Selected work
OpenAI / research lead2023-2025 · San Francisco
Led a 14-person reasoning team. Cut inference cost 94% while improving competition-math performance to 99.2%. Built adversarial evaluations for agents that use code and tools.
DeepMind / scientist2021-2023 · London
Worked on Gemini reasoning and code capabilities. Built security evaluations for generated code and a generalist robot policy that learned 600 tasks from a single demonstration each.
NASA JPL / research2020 · Pasadena
Designed onboard planning models for autonomous exploration. Tested them on terrain that was actively trying to break the robot.
Security
- National high-school cybersecurity hackathon champion at 16. First overall out of 1,800 teams with a live attack-detection system built in 36 hours.
- DEF CON CTF finalist. Led the reverse-engineering team through a 48-hour final.
- Disclosed 23 critical vulnerabilities across cloud infrastructure and open-source software. $1.2M in bug bounties, all through authorized research and coordinated disclosure.
- Won a Pwn2Own category with a browser sandbox escape. Worked with the vendor through the patch and regression tests.
Projects
sentinelAI security research · 2026
A security agent that audits code, reproduces bugs in isolated test environments, and checks the patch. Found 41 previously unknown vulnerabilities in its first month.
axiomAutonomous discovery · 2026
A research agent that proposes experiments, runs them, and tries very hard to prove itself wrong.
atlasWorld models · 2025
A 4-billion-parameter world model that runs on a laptop. Simulates 20 seconds of physical interaction from a single frame.
lemmaFormal reasoning · 2024
A theorem prover that solved 97% of a held-out olympiad corpus. Every proof is checked by Lean, not vibes.
afterburnerGPU systems · 2024
A kernel compiler that made our training runs 11× faster. Started as a way to stop staring at a progress bar.
Writing
Recognition
- Three NeurIPS best paper awards before turning 25.
- MIT PhD in computer science, completed in two years. Thesis on generalization in embodied intelligence.
- 87,000+ citations. One paper written almost entirely on the N Judah.
- Forbes 30 Under 30 at 19. On the list before I could rent a car.
- 2025 ACM Prize in Computing for work on verifiable machine reasoning.
Baseball
Played varsity baseball at MIT while studying computer science. Team captain, two-time team MVP, and first-team All-Conference. Led the team in batting average, home runs, and RBI in my final season.
.412 AVG · 14 HR · 52 RBI · Shortstop
I still keep a glove in the trunk. The hardest thing I did at MIT was hit a changeup after an all-night problem set.
Off the clock
Usually somewhere in the Inner Sunset. Pickup baseball, long runs through Golden Gate Park, ordinary coffee, and an ever-growing pile of books about things I don't understand yet.