Google DeepMind reassigned John Jumper, the Nobel Prize winner who co-created AlphaFold, to AI coding work last month, marking a strategic shift away from specialized scientific AI tools toward general-purpose agentic research systems. The move follows 5 years of AlphaFold's dominance in protein structure prediction, during which the tool attracted 3 million researchers worldwide. Google's pivot reflects a broader industry bet that general-purpose reasoning models can execute cutting-edge research without human involvement, rather than building domain-specific tools for each scientific problem.
The shift was visible during Tuesday's Google I/O keynote, where Demis Hassabis opened the scientific AI segment by proclaiming the company stands in the foothills of the singularity, then showcased WeatherNext—a specialized weather prediction tool that provided advance warning for Hurricane Melissa's landfall in Jamaica last year. The juxtaposition highlighted tension between two AI science strategies: narrow tools trained for specific problems versus autonomous agents that could collaborate as peers with human researchers.
Google continues developing specialized tools—AlphaGenome and AlphaEarth Foundations launched last summer, and WeatherNext's newest version shipped in November—but the company's headline I/O announcement was Gemini for Science, a package uniting several LLM-based research systems under one brand. The suite includes AI Co-Scientist, a hypothesis-generating system, and AlphaEvolve, an algorithm optimizer. Both are now open to researcher applications after early testing drew enthusiasm from scientists including Stanford geneticist Gary Peltz, who compared using AI Co-Scientist to consulting the oracle of Delphi in Nature Medicine.
“We are moving toward AI that doesn't just facilitate science but begins to do science.”— Pushmeet Kohli, Google Cloud chief scientist
Key facts
- 01John Jumper, AlphaFold co-creator and Nobel laureate, moved from science-specific AI to coding work last month.
- 02AlphaFold protein structure predictions are used by 3 million researchers worldwide.
- 03Google's Isomorphic Labs raised a $2 billion Series B for drug development using AlphaFold technology.
- 04Google launched Gemini for Science package uniting LLM-based scientific systems including AI Co-Scientist.
- 05WeatherNext weather prediction software provided advance alert for Hurricane Melissa's Jamaica landfall last year.
The coding reassignment of Jumper, a Google fellow, signals resource prioritization toward agentic science, where coding abilities underpin autonomous research systems. Google recently faced competitive pressure as its coding tools lagged Anthropic and OpenAI offerings. Coding models are foundational for agentic researchers: OpenAI announced this week that a general-purpose reasoning model disproved an important mathematics conjecture, the most meaningful AI contribution to mathematics so far according to some mathematicians. The model was not specialized for math or research—it's a GPT-5.5-class general reasoner.
Isomorphic Labs, Google's drug-development subsidiary leveraging AlphaFold, raised a $2 billion Series B funding round, demonstrating continued commercial appetite for specialized tools. AlphaFold's protein structure predictions remain foundational—no agentic system can predict folding structures without AlphaFold's help, at least not yet. But agentic systems can call specialized tools when needed, positioning them as orchestrators rather than replacements for domain models.
Google frames agentic scientists as accelerants for humans, not replacements—the name AI Co-Scientist versus AI Scientist appears deliberate. Hassabis maintains a human-centric framing for the next decade, saying systems will become more like collaborators beyond that timeframe. The careful positioning leaves open the question of whether superhuman agentic scientists could eventually exceed human researchers, particularly in domains where progress has stagnated since the 1970s, as Hassabis cited in physics.
The industry consensus is moving toward recursive self-improvement, where AI systems drive AI advancement in an accelerating loop. Agentic research systems making independent contributions to mathematics suggest the same capability could extend to experimental science, though verification requirements make science a tougher domain than pure mathematics. The coding pivot and Gemini for Science launch suggest Google is aiming resources toward that summit, even as specialized tools remain in production.
The AlphaFold era demonstrated that specialized AI could solve grand scientific challenges—protein folding was considered intractable for decades before DeepMind cracked it. The agentic era bets that general intelligence across domains beats narrow excellence in one. The 5-year gap between AlphaFold's breakthrough and Jumper's reassignment marks how quickly both the technology and the industry discourse have moved past domain-specific wins. Whether general agents can match AlphaFold's scientific impact remains the central question for Google's new strategy.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
