Co-founder of Epoch AI who left to found Mechanize, a small RL-environment shop reportedly in $1.5B+ licence-and-hire talks with Google. He is the most prominent advocate of the view that frontier training data must be expensive, expert-built and sold as artefacts, not hours.
high confidenceSan FranciscoAt Mechanize since 2025Updated 2026-09-19
Background
Co-founded and was associate director of Epoch AI, the research institute tracking compute, scaling and AI economics (FrontierMath and related work), with prior research affiliations including MIT. He announced Mechanize on X in April 2025; the launch drew criticism that it undermined Epoch's perceived neutrality, following the earlier controversy over OpenAI funding of FrontierMath. He argued that full automation brings explosive growth and that people would earn income from rents, dividends and welfare.
What they run now
Running Mechanize's in-house engineer model (one engineer owns each task, from idea to grader to QA), selling coding environments to frontier labs, and the reported Google licence-and-hire deal.
Career
2025 – presentCo-founder & CEO, Mechanize
2025Co-founder; Associate Director, Epoch AI
On the record
Expert data vs 'sweatshop data' · 2025-07
Co-wrote that cheap contractor-labelled data is over: progress now needs interactive environments built by full-time domain specialists whose tacit knowledge is the bottleneck. source
Price per RL task · 2025-08
Co-wrote that labs should spend more per RL task (then ~$500, expected to rise to several thousand) because cheap tasks waste expensive compute; data and compute are complements. source
Mission · 2025-04
Co-signed the founding statement: build environments and evals to enable full automation of the economy, sizing the prize at ~$18T/yr US wages and ~$60T globally. source
Timelines · 2025-04
On Dwarkesh, argued drop-in remote-worker AGI is likely decades away, yet broad deployment could drive 30%+ annual growth. source
Recent posts
20 posts archived · most engaged first, then the latest
One interesting pattern with Fable 5 is that it will often say things that are gibberish when I use it for coding. Things like "The morning's slim-scan fix cured the scan hang", "this is a latent-drift API-shape wrinkle", etc.
When I ask why it does this, Fable explains that it invents codenames while reasoning about the problem, then fails to realize they're meaningless to me. Its neuralese is blending into its output because of a theory-of-mind failure about what's in its head vs. mine.
I think it's underappreciated how economically valuable AI safety is. A model that frequently goes off the rails, takes dangerous actions, is misleading or deceptive, etc. is simply much less valuable than a model that does not do that.
I think the view expressed in our essay is useful for understanding why different labs achieve very similar breakthroughs and capabilities at almost exactly same time.
I bet @ChrSzegedy $10k at even odds that a Fields Medal will be awarded to a human mathematician by the end of 2030, excluding the 2026 Fields Medal cycle.
Nat Friedman, Daniel Gross, Patrick Collison, Adam D'Angelo — investors
Who they amplify
Accounts whose posts Tamay has reposted recently: Mechanize, Inc., Erik Brynjolfsson, Shubham Patil.
Why it matters here
His essays are the clearest public argument for Julian's thesis: expert tacit knowledge is the bottleneck and should be priced per artefact, not per hour. Mechanize's economics (small in-house team, high-value environments, licence exits) are a benchmark for a small creative-data specialist. Direct overlap is low because Mechanize is coding-only.
How to reach
Very active on X and writes long essays. Speculation: a sharp, quantitative reply to one of Mechanize's essays (e.g. what a 'visual taste' RL environment with a gradable rubric would cost) is the best cold opener. Warm path via Dwarkesh Patel's network is plausible but unverified.
What we could not establish
Status of Google talks after August 2026 unknown; no close confirmed.
Miju Labs