Google hired back the man Thinking Machines fired, for post-training

Barret Zoph has worked at Google, then OpenAI, then a company he co-founded, then OpenAI again. Now Google.

He joins as a vice president of research, the company said on 26 August, returning to the place where he was a research scientist from 2016 to 2022. In between he co-founded Thinking Machines Lab, was fired from it in January, went back to OpenAI, and left again a few months later. Erin Woo, who broke the story for the Wall Street Journal, called it the completed boomerang.

And Google's statement names the job with unusual precision. "We look forward to Barret returning to Google and bringing his RL and post-training expertise to Gemini," a spokesman said.

Not research generally. Reinforcement learning and post-training, on one product. Gemini is named. DeepMind is not (read that twice).

This is a procurement story more than a personnel one

So consider what Google is buying, and what it accepted in order to get it.

Mira Murati fired Zoph in January and told staff he had shown a pattern of performance, conduct and trust problems, the Journal reported. While at Thinking Machines he had an undisclosed relationship with a colleague, and his responsibilities were scaled back as a result.

Zoph's account differs. He says Murati dismissed him only after he had already told her he was leaving for another job, and that the reduction in his duties was a temporary measure.

Both versions are on the record and we have no way to adjudicate between them (nor should you trust anyone who says they do).

What is not in dispute is that a company with Alphabet's legal department read all of it and hired him anyway.

Price that against the year Google has had

Noam Shazeer left for OpenAI in June, and John Jumper, a Nobel laureate, left for Anthropic around the same time. Alphabet stock fell as much as 7 percent in a single session.

Jeff Dean left in early August after twenty-seven years to co-found a startup. Demis Hassabis moved from DeepMind chief executive to chairman, with Koray Kavukcuoglu elevated to oversee Gemini model development.

A move that size on two departures is the market saying out loud that a handful of named individuals are material information about one of the largest companies on earth.

Where Google thinks the Gemini gap is

Probably not compute, of which it has more than almost anyone. Probably not pretraining, where it has been strong for a decade.

Post-training and reinforcement learning. The layer that turns a capable base model into a product people actually prefer, and the layer where OpenAI and Anthropic have kept an edge.

And you do not hire a vice president of research with a public history of trouble unless the alternative looks worse.

The counter is that we are over-reading one line of boilerplate

Spokespeople name the fashionable area because it reassures investors. Hires happen for a dozen reasons. Zoph is good at this work regardless of what any statement says. Fair enough.

So here is the test. If Google's next Gemini release is sold on post-training and reinforcement-learning gains rather than on context, price or scale, the boilerplate was accurate. If it leads on compute, we misread it.

Spare a thought for Thinking Machines Lab, now missing two of its founders. Luke Metz left OpenAI too, for Meta.

When the whole industry can outbid you for the people who started your company, what exactly did you found?