The Best People Are Already Moving

Read the resignations, not the benchmarks

You can learn more about who is winning in AI from who quit this week than from any benchmark released this year.

Benchmarks are a lagging indicator. They measure work researchers finished months ago. The leading indicator is where those same researchers choose to sit next.

This week, that signal got loud.

Seven days, two departures, one direction

Inside a single week, four moves landed, and all four pointed the same way:

• Noam Shazeer left Google DeepMind for OpenAI. He co-invented the Transformer, the architecture under every modern model. Google paid 2.7 billion in 2023 to bring him back from Character.AI. He left anyway.
• John Jumper left DeepMind for Anthropic seven days later. A Nobel laureate, the author of AlphaFold, nine years inside the company, and now he is gone.
• OpenAI added Dean Ball as Head of Strategic Futures. A former White House AI adviser, hired to think about where this all goes next.
• OpenAI published o3 medical results from Boston Children's Hospital. Eighteen unsolved rare pediatric disease cases, with diagnoses later confirmed by clinicians in NEJM AI.

Four moves, one direction. The people and the wins are flowing toward Anthropic and OpenAI in the same seven days.

The leaked numbers tell on themselves

Then an internal DeepMind report leaked. It said morale is increasingly negative. It said the company's models have slipped to fifth place on industry benchmarks. Staff worried leadership was quietly conceding the race to AGI.

The exact words matter less than what they describe. A fifth-place lab with a bright future hires the best people in the world. A first-place lab with a darkening one watches them walk out.

Benchmarks are easy to fix with compute. Morale and trust are not.

You can buy more GPUs by Friday. You cannot buy back a Nobel laureate who has already cleaned out his desk.

The strongest case against me

Here is the argument that I am wrong, stated as well as I can state it.

Google keeps publishing safety roadmaps and funding academic multi-agent risk research. It has distribution, capital, and infrastructure that neither OpenAI nor Anthropic can match. The bet that Google wins by becoming the platform everyone else runs on is not a crazy bet.

But infrastructure without the people who build the models is not an AI company. It is a hosting business with excellent brand recognition. The chips do not have ideas. The researchers do, and the researchers are the ones leaving.

How a ban became a recruiting pitch

There is a policy twist running underneath the talent story. Trump told Axios that Anthropic has behaved very responsibly on the Fable 5 export controls, hinting that restrictions could ease soon.

The Fable ban was meant to slow Anthropic down. Instead it may have handed the company its sharpest recruiting line: come build frontier models in the one lab the government is trying to protect instead of punish.

Top researchers do not move for salary alone. They move toward the lab going fastest, working on the hardest problem, and staying ahead of the policy risk that could freeze everyone else. When one lab becomes the safe harbor, the choice is not difficult.

Watch the departures, not the announcements

You can read benchmark tables. You can parse safety papers. You can track stock prices. The most reliable map of where this industry is heading is where the best people choose to work, and right now they are choosing Anthropic and OpenAI.

This will not hold forever. Markets correct. Leadership changes. A lab with a clear mission, real capital, and political cover can pull people back as fast as it lost them.

For now, the floor is shifting under whichever labs are losing their best people. If your strategy depends on one provider's model line, watch who is quitting, not what is being announced. The press release is written by the company. The resignation letter is written by the person who already knows.