Why two AI legends just left DeepMind in a single week

Two posts on X, two days apart, said more about the state of the AI race this summer than any benchmark. Both were resignation letters, and both came from people you have heard of even if you have never followed the field.

On June 18, @NoamShazeer wrote that he would be "joining OpenAI," calling it "a difficult decision to move on" from Google. Two days later, on June 20, @JohnJumperSci posted that after nearly nine years he had "decided to leave Google DeepMind and join Anthropic." Same week, same lab, two of its biggest names, walking out opposite doors to its two fiercest rivals.

Why these two names matter

This is not routine turnover. Noam Shazeer is a co-author of "Attention Is All You Need," the 2017 paper that introduced the Transformer, the architecture underneath essentially every large language model in use today. He was also a co-lead of Google's Gemini models. Reporting notes Google paid billions less than two years ago to bring him and his Character AI team back into the building. He is now at OpenAI.

John Jumper is a 2024 Nobel laureate in chemistry, awarded for AlphaFold, the system that predicts protein structures and reshaped biology research. In his own post he thanked DeepMind CEO Demis Hassabis, who "took a real chance letting me lead the AlphaFold team just six months after finishing" his PhD. He is now headed to Anthropic, the lab that recently shipped a science workbench we wrote about here. A Nobel-winning scientist joining the company building AI tools for scientists is not a coincidence.

The backdrop is not calm either

These exits did not land in a vacuum. Google previewed Gemini 3.5 Pro at its I/O event in May and promised general availability within a month. Instead the model slipped, with reporting pointing to a mid-July target and a near ground-up rebuild after early testers flagged issues with token efficiency, coding, and long multi-step tasks. In the same stretch, Alphabet shares dropped around 5 percent, with press accounts tying roughly $225 billion in lost market value to worries about the delay and the departures.

Hassabis has pushed back, telling reporters that DeepMind still has by far the deepest research bench of any lab. He is not wrong that one week of headlines does not undo years of work. But the market and the talent both voted the same direction at the same time, and that rhymes with a pattern.

My honest read

Capability leadership in AI is not a fortress. It rotates, and it rotates faster than the release calendar suggests. The two labs that just collected a Transformer co-inventor and a Nobel laureate are the same two, OpenAI and Anthropic, that spent this summer shipping GPT-5.6 and Sonnet 5 while Google was defending a delay. Talent is the real moat, and moats can be crossed by anyone willing to pay.

The lesson is not that Google is finished. It clearly is not. The lesson is that the frontier is a moving target with no permanent leader, and that the top of the leaderboard in July may not be the top in October.

What this means if you run a business

Here is the part that matters for a plumber in Geneva or a dental office in Wheaton, not just for people who read AI news for sport: do not marry your business to one model. The vendor that looks unbeatable today can lose its best people, miss a launch, and slip a notch in a single quarter. What stays yours is your process, your customer data, and the workflow you built around the AI, none of which care which logo is winning this month.

That is exactly how we set things up at New Face Design. We treat the model as a swappable part, so the AI handling your calls, quotes, or scheduling can be upgraded to whatever is best without rebuilding anything underneath it. If you want to see where automation would actually pay off in your operation, our free process audit is a short, no-pressure conversation about your specific workflow. The models will keep trading places. Your systems should not have to.