MIT Technology Review covered the Google I/O 2026 science track this week with a question worth holding onto. Demis Hassabis used the phrase "foothills of the singularity" on stage. The evidence he produced for the claim was WeatherNext, a forecasting tool that helped with Hurricane Melissa evacuations. The gap between the rhetorical claim and the operational example was visible to most viewers in real time. The reporting frame picked up on the gap and turned it into a sharper observation about strategy inside Google.
The observation is this. Google is in the middle of a quiet but consequential reorientation away from specialized vertical AI tools and toward general-purpose agentic systems, even in domains where the specialized tools are currently dominant. The strongest evidence in the article is that John Jumper, the Nobel laureate who led the AlphaFold work, has shifted his focus from specialized science AI to AI coding work. AlphaFold has more than three million researchers using it. Isomorphic Labs, the Google subsidiary built on the protein-folding stack, raised a $2 billion Series B. The vertical bet is paying. The strategic talent is moving to the horizontal bet anyway.
What the Two Approaches Actually Look Like
The specialized-tool path produces a system that is engineered against a specific scientific problem with a specific dataset, a specific evaluation framework, and a specific user community. AlphaFold for protein folding, WeatherNext for weather, AlphaEvolve for algorithm optimization, the AI Co-Scientist for hypothesis generation. The system performs well on the narrow task, integrates cleanly into the workflows of the field, and accumulates users by being better than the previous tool for that specific job.
The generalist-agent path produces a system that handles a broader class of tasks by reasoning, planning, executing, and recovering across a longer chain of operations. The bet is that as the underlying frontier model gets stronger, the same generalist system can outperform the specialized tools across most domains. The cost is that the generalist underperforms today on narrow benchmarks. The expected return is that the generalist eventually displaces a much larger surface area of specialized tooling.
The same reporting includes an example that sharpens the bet. OpenAI's general-purpose model disproved a mathematics conjecture, with one academic describing the result as "perhaps the most meaningful contribution that generative AI has made to mathematics so far." The model was not built for mathematics. It produced the result anyway. That class of outcome, where a generalist system does original work in a vertical it was not engineered for, is the operating thesis behind the talent reallocation.
The Pattern Is Not Just About Science
The strategic question is whether the same pattern is going to play out for verticalized AI products in marketing, finance, customer service, legal, and the dozens of other categories that are currently being built on specialized AI tools. I have argued for months that most AI stacks need rebuilding before the agent wave arrives, and the talent reallocation inside Google is a louder version of the same signal. The honest read is yes, on a slower timeline, with significant variance by domain.
I see this in the companies I work with and inside the products my own team builds. The specialized AI tool that leads its category today, the AI marketing copy generator, the AI lead-scoring engine, the AI customer service router, the AI sales enablement assistant, is facing the same competitive pressure that AlphaFold's category is facing. The generalist agent with the right tooling and the right system prompt can replicate roughly 70 to 80 percent of the specialized tool's value inside a single quarter of integration work. The remaining 20 to 30 percent is real, but it shrinks each time the underlying frontier model ships an upgrade.
The strategic implication is that the specialized AI product business has a window. The window is real and it is profitable. The window is also closing on a timeline that most product roadmaps are not pricing in. The specialized tools that survive past 2028 are the ones whose moat is not the AI capability itself but something adjacent. Proprietary data flywheels, domain-specific workflow integrations, regulatory positioning, exclusive distribution, or trust and brand built with the user community. Without one of those moats, the specialized product is competing against a generalist agent that gets cheaper, faster, and more capable every quarter.
The companies betting on specialization without a non-AI moat are running a clock. The companies betting on generalization are running a longer thesis with more downside variance. The companies running both bets in parallel, the way Google appears to be running its science strategy, are hedging the most expensive uncertainty in the category right now. This is also the play AI labs ran when they became consultancies, moving up the stack into the layer where the moats live.
The "foothills of the singularity" line is hyperbole. The talent reallocation underneath it is not. The most expensive minds inside Google are betting that generalist agents will eat specialized tools faster than the market expects.