Athon Millane
Australian, in San Francisco.
I work on how language models can learn judgement from consequences. Each step-change in AI has come from a new teacher: pre-training on what was written, RLHF on what people prefer, and verifiable rewards on what can be checked. None of them reaches the decisions people most want help with, where the outcome arrives late and is partly luck. I think the next teacher is hindsight, because once a decision resolves we know far more than whether it turned out well.
I'm co-founder and CTO of Predictive Text Labs, with Chris Bach (co-founder of Netlify) and Alex Pan. We're backed by Blackbird, Balaji Srinivasan and the founders of Supabase and Synthesia. We build RL environments from replayed history. I like training models and working out what they learned. Often it isn't what I meant to teach, so I build evals and benchmarks that catch a reward or a test teaching the wrong thing.
Writing
- What the Future Teaches and its follow-up, Step by Step, on how much more a model can learn from hindsight than from the outcome (PIN on request).
- A Predictive Mirror for Sign Language, built in a day with coding agents at Joscha Bach's California Institute for Machine Consciousness hackathon (live demo).
Before
Much of my work has been on learning from signal that is scarce, noisy or slow. At Max Kelsen my group pretrained early transformers on the human genome, and I hired and mentored Daniel Bourke. At Hudson & Thames I built DLFinLab, whose IP was acquired by ADIA (Abu Dhabi's sovereign wealth fund) in a private sale. I co-founded MedNotes.ai and led machine learning at Balaji Srinivasan's Network School.
I also like to make small angel investments in friends' startups.