When 'Evolved' Means Dispersal: Kohli's AlphaFold Tradeoff and the Agent Imperative

Pushmeet Kohli's sentence to the Financial Times on July 29, 2026 is the kind of quote that will age into a case study — whether DeepMind's leadership treats it that way or not.

"Our strategy over the last nine years has been to focus on grand challenges . . . a concrete goal every project is focused on. The strategy has evolved."

That is not a product update. It is an organizational confession: the lab that won a Nobel for protein folding is no longer organized to repeat it.

The tradeoff Kohli named but did not resolve

Grand-challenge teams are expensive, slow, and politically fragile. They also produce irreplaceable tacit knowledge — the kind that does not transfer cleanly into a general-purpose Gemini feature team.

AlphaFold was DeepMind's proof that patience beats product cycles. Breaking up the bench while simultaneously racing OpenAI and Anthropic on agents is a bet that generality beats depth. Kohli's quote makes that bet explicit.

The problem: the market still prices DeepMind's science credibility on AlphaFold's halo. Dispersing the authors who created that halo — while John Jumper, Jonas Adler, and Alexander Pritzel walk to Anthropic — is not a subtle move. An unnamed DeepMind employee told the FT their exits were surprising because the three were "instrumental, important, core members."

What "evolved" actually means in 2026

In practice, evolved means:

• Reassign protein-folding researchers to Gemini science assistants
• Redirect Nobel-adjacent talent to Code Strike (Google's coding catch-up team) before losing them entirely
• Route commercial drug-discovery ambition through Isomorphic Labs, not the core bench
• Compete on agents because that is where capital and talent flows now

None of those choices is irrational. Together, they describe a lab choosing to be a platform company over a science cathedral.

The Anthropic counter-narrative

Anthropic did not need to write a manifesto. It hired the manifesto authors.

Claude Science for biology and drug discovery now gets the researchers who actually shipped AlphaFold — not consultants, not license agreements, but the people who lived inside the grand-challenge model Kohli is retiring.

If you believe domain depth still wins in life sciences, July 2026 was Anthropic's best recruiting advertisement.

What DeepMind must prove next

Kohli's evolved strategy is defensible only if Gemini ships a science capability that matches AlphaFold's credibility — not its interface, its trust.

Researchers do not switch from a dedicated folding stack to a chat assistant because the org chart changed. They switch when the assistant is measurably better at their daily work.

Until that demo exists publicly, "the strategy has evolved" reads less like vision and more like talent reallocation under frontier pressure.

The uncomfortable question

Can any lab hold both truths at once — win the agent race and keep the decade-long scientific benches that make breakthroughs possible?

DeepMind just answered no, at least for AlphaFold. Anthropic's hires suggest they think the bench still matters — just not inside Google.

We will know who was right by what ships in the next twelve months, not by who wrote the better press statement.

Sources

• DeepMind dismantles its AlphaFold team — THE DECODER (July 29, 2026)
• Financial Times reporting as cited by THE DECODER