The Future of Personalized Universal Agents

TL;DR

• He opens by calling himself a slow learner with impostor syndrome: 200+ papers published, maybe 10 he's proud of. The talk is a retrospective on the handful of ideas per era that actually mattered.
• First two decades (information retrieval): indexing (the library as humanity's first great invention), vector space models, deep neural networks. Search and recommendation are fundamentally sorting — and so is next-token prediction.
• His confession: in 1992 his PhD advisor John Riedl proposed applying vector space models to user–item matrices instead of document–term matrices — the birth of recommender systems — and Ed told him it was a dumb idea.
• 2015–2025, three ideas: sequential transduction (the phrase opening the Transformer abstract), next-idea prediction via chain of thought, and post-training. In hindsight, he says, we should have called them large reasoning models.
• His critique of the field: too much bottom-up processing, especially RL. You don't just hit a kid when they get it wrong — you explain. Chain of thought is that explanation, and distillation only works because of the reasoning traces it produces.
• Next decade: from ranking to assistants that actually reason about your taste. His own work at Google is now narrowed to reasoning and planning.

Key Points

The frame: three ideas per era (~00:22–00:26)

As a scientist, he says, what he cared about was never the next paper but the next big idea — and it took him a long time to see which ideas mattered.

Three ideas from information retrieval, spanning his first two decades:

1. Indexing. The library is, in his opinion, humankind's first great invention, and it is built on indexing. In the 90s that became MapReduce and web-scale search engines.
2. Vector space models. word2vec took the ten-year test-of-time award a few years ago; its core idea was using neural networks to compute relationships between words so machines could grasp semantics.
3. Deep neural networks (2010–2015).

Wedged in between is the confession about John Riedl and user–item matrices.

The unifying claim: once supercomputers arrived, the one thing we got dramatically better at was sorting. Search, recommendation, the whole internet ecosystem run on sorting and ranking — and next-token prediction is sorting too, ranking the most likely next word.

He anchors the compute arc with a comparison: the supercomputer he used as a 1992 grad student versus the 1,690 GFLOPS phone in his pocket — nearly a thousand times more compute. "We're all walking around with like a thousand supercomputers in our pocket."

2015–2025: transduction, chain of thought, post-training (~00:26–00:31)

He puts up the Transformer architecture figure everyone has seen and asks: have you actually read the paper? Then reads the first sentence of the abstract aloud, admits he had no idea what it meant the first time, and gets at least half the room to raise their hands admitting the same.

His unpacking of transduction: the conversion of one energy wave into another. You are experiencing it right now — his microphone turns sound into electrical waves, the hall's speakers turn them back into sound, and your ear is itself a transducer whose membrane vibrates into neurons that carry information into your brain. So the big idea around 2014 was marrying sequential transduction with neural networks — and crucially, it wasn't invented in the ether, it rested on a cognitive understanding of how our brains communicate. The resulting universal computational model can absorb multimodal data and handle many languages at once.

The second idea came from a plain question he and Denny Zhou asked at Google Brain: can we teach machines the way we teach our children? One thing was clearly wrong. When your kid gets something wrong, you (hopefully) don't just hit them — but that's exactly how we were training machines. Give them an explanation instead. That became chain of thought, i.e. top-down processing.

His critique follows immediately: the industry does far too much bottom-up processing, particularly with RL. And a consequence people underrate — distillation works because of the chain-of-thought reasoning traces. Moving from next-token prediction to next-idea prediction is what started the whole reasoning field.

Hence: in retrospect we should have said large reasoning models, because reasoning is what we actually care about.

Third idea: post-training, which he leaves to the other speakers on the stage.

2025–2035: from ranking to personalized reasoning (~00:31–00:33)

Picking the compute thread back up: a single latest-generation TPU chip now dwarfs the Cray Y-MP from the start of his talk by an astronomical factor (exact multiplier and chip generation flagged under To Verify) — "that's definitely going to burn a hole in your pocket."

Then a short demo video. A user holds up a stack of books a friend has been reading and asks how to describe her taste; the assistant answers "a mix of history, biography, and perhaps a touch of political or social commentary." Asked which one she'd like most, it picks Half of a Yellow Sun for its exploration of complex social and political issues in a historical setting, aligning with her interest in narratives about significant periods.

His line: that is not just sorting and ranking anymore — that's where personal assistants are going.

He closes on the slide he showed on this same stage a year ago: the ideas were already there, and now everybody is talking about agents. The next decade's ideas are already happening — self-improvement (as Richard just covered), multimodality, tool use, multi-step complex reasoning. His own work at Google is now focused purely on reasoning and planning, because that's where the future is.

Quotes

"In 1992, my PhD adviser John Riedl came to me and said, 'What if we use those vector space models and applied them to user–item matrices instead of document–term matrices?' … and I told him it was a dumb idea." (~00:24)

He voted against recommender systems at the moment of their birth.

"What is sequential transduction? What the f does that even mean?" (~00:27)

"When you teach your kids, I hope every time they get it wrong, you don't just hit them. But that was the way that we were teaching our machines." (~00:29)

The origin of chain of thought, in one image.

"In retrospect I wish we had not called it large language models; rather we should have been calling them large reasoning models." (~00:30)

"Now that is not just sorting and ranking anymore, is it? That is where personal assistance is going." (~00:32)

Thirty years of sorting, turning a corner on one demo video.