Cameron Allen

The Berkeley researcher studying how machines decide what's real — and why they can never be sure.

Rigorous young academic with a teaching gift. A thoughtful builder, and deals in practical ideas, large and small.

Cam Allen is a research scientist at UC Berkeley's Center for Human-Compatible AI, working on the frame problem: how an agent decides what in the world is even relevant, and builds a compressed model of reality from there. His best known work — the NeurIPS paper on learning Markov state abstractions is a standard reference and asks a question that sounds technical but isn't: when a system has to throw away almost everything it's seeing, how does it keep the parts that will matter later? Get that wrong and nothing shows the error. The system simply decides slightly worse from then on, and can't recover what it discarded. He co-taught Berkeley's CS188 to 650 students — he explains hard things well.

His research is the unknowing problem stated formally. In a world of effectively infinite detail, any intelligence has to choose what to attend to, and can never be certain it chose correctly. That holds for the systems he studies and for everyone in the room.

He'll be talking about where the research is actually heading — which paths industry is pursuing, which ones academia is holding open, and what kinds of AI might emerge from each. Alongside Julian Yocum, he offers something rare on a conference program: two working researchers describing the scientific future from inside it.

You'll take away: a map of the research frontier from someone who has no product to sell. If you are trying to work out which capabilities are close and which are marketing, an hour with Cam is worth a quarter of reading press releases from the big AI labs.

ONE READ — Cameron Allen. His research statement on frames is short, unusually clear, and free of jargon.