The Shitshow at AI Scale: What Google's Talent Exodus Actually Means
- Rich Washburn
- 3 hours ago
- 4 min read
There is no diplomatic way to frame the last ten days at Google DeepMind. Nobel Prize-winning researchers don't quietly update their LinkedIn on a Friday unless something has shifted in a way that can't be walked back.
John Jumper — who shared the 2024 Nobel Prize in Chemistry for AlphaFold, arguably the most consequential AI-driven scientific breakthrough of the decade — just announced he's leaving Google after nearly nine years to join Anthropic. He's not the only one. Jonas Adler and Alexander Pritzel followed days later, also to Anthropic. Alphabet's stock paced for its worst day in a year. And if the Bloomberg reporting is right, the departures aren't finished. This is not normal churn. This is a signal.
What Talent Flight Actually Tells You
There's a specific pattern to how elite researchers leave top labs, and it tells you things that press releases never will.
If your lab is about to ship something proud — a model that will move benchmarks, reset the competitive landscape, put your name on something that matters — you stay. You want to be in the room. You want the release on your resume. You want the paper. You want the moment. You leave before the release when you've concluded the thing you're about to ship isn't going to be that moment. Or when somewhere else has become unmistakably that moment instead.
Every departure I've tracked in this wave has gone to Anthropic. One went to OpenAI. That asymmetry matters. It means the gravitational pull right now isn't distributed across the field — it's concentrated in one place.
The question serious people are asking in the Valley is why.
The Recursive Self-Improvement Hypothesis
The rumor circulating — and I want to be precise that this is a rumor, not confirmed — is that Anthropic has gotten close to cracking recursive self-improvement.
For the non-technical reader: recursive self-improvement is the property by which a model meaningfully contributes to designing its own successor. The model doesn't just assist with routine tasks — it participates in the research loop that produces the next generation of itself. If that threshold is crossed, the development timeline for AI capabilities compresses in a way that is difficult to model from the outside.
We have adjacent evidence that something like this is already happening at the infrastructure layer. OpenAI and Broadcom just unveiled Jalapeño — a custom inference chip designed end-to-end in nine months with AI assistance. Nine months from design to production-ready silicon. The previous benchmark for a chip design cycle at that complexity was measured in years. AI helping design the hardware that runs AI isn't a metaphor anymore. It's a shipping product. If that's happening at the hardware layer, the question of whether it's happening at the model layer is not hypothetical. It's a matter of timing. The talent vector suggests Anthropic is where that timing is most advanced.
Google's Position Is More Complicated Than It Looks
Here is where I want to resist the simple narrative, because the simple narrative is probably wrong in at least one important dimension.
Google is not losing the AI race in every category. What Google is losing is the frontier model race — the competition to have the best LLM on the benchmark leaderboard at any given moment. And there's a version of the internal calculus at Google that says that race was never theirs to win anyway, and that winning it isn't actually what defends the business.
The bet — and Sergey Brin's re-engagement last year suggested this was a real internal debate — was whether search remains structurally valuable even as AI commoditizes the model layer. Google's infrastructure advantages, its data advantages, its distribution advantages are not trivially reproduced. The argument was that Google could afford to not have the world's best model if the world's best model still ran on Google's compute and Google's index.
That argument is getting harder to sustain. Not because it was wrong in theory, but because the talent signal is starting to raise a more uncomfortable question: is Google losing the model race because of strategy, or because the researchers who could win it are leaving?
Those are very different problems. One is a choice. The other is a compounding constraint.
The Competition Google Didn't See Coming
There's another dimension here that deserves attention beyond the Anthropic-OpenAI framing.
Fei-Fei Li's World Labs just raised over a billion dollars to pursue Large World Models — spatial intelligence AI that perceives, generates, and interacts with the 3D physical world. That's not an LLM play. It's a bet that the next significant capability frontier isn't better text generation, it's models that understand physical reality. Google's own research heritage — DeepMind's entire reason for existing — was supposed to own that frontier. AlphaFold was the proof of concept. And the scientist who won a Nobel Prize for it just walked out the door to a competitor.
The irony is dense enough to be uncomfortable.
What the Next 90 Days Will Tell Us
Google is expected to release a model in the near term. What that model does to benchmarks will answer a lot of the questions the talent flight has raised.
If it's competitive — if it moves the leaderboard in a meaningful way — the narrative shifts back to "Google is fine, the churn was noise." Possible. Not the way I'd bet right now, but possible.
If it's underwhelming, the talent signal becomes retrospectively obvious. The researchers who knew what was coming made their moves before the release rather than after. That's not disloyalty. That's informed decision-making by people with full information.
My read: the departures happened before the release, not after. People with access to the model know what it is. The fact that they're moving now, rather than waiting for a triumphant release and then moving with leverage, tells you something about what they expect. The next 90 days will either validate that read or force me to revise it. I'll take the revision if it comes. But the data I have right now points one direction.
Anthropic is where the energy is.
Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital and CIO of Data Power Supply.
Comments