GOOGLE BLEEDS AGAIN: NOBEL PRIZE WINNER WHO CRACKED THE PROTEIN FOLDING PROBLEM WALKS OUT THE DOOR AND HEADS TO ANTHROPIC

This is the kind of news that should make Google’s HR department question all of its life choices. John Jumper, the man who co-built AlphaFold and then stood on a stage in Stockholm to accept a Nobel Prize for it, has handed in his notice and told Google DeepMind he is heading to Anthropic.

Think about what that means for a second. In the span of about one week, Google has lost Noam Shazeer, the researcher who literally invented the Transformer architecture that every modern AI model runs on, to OpenAI. And now it has lost John Jumper, an actual Nobel laureate, to Anthropic. These are not junior engineers who got poached with a slightly better equity package. These are the people who changed the course of science.

Jumper spent nearly nine years at DeepMind. That is a long time by tech industry standards, and practically geological time by AI standards. His work on AlphaFold did something that structural biologists had been attempting for fifty years. It cracked the protein folding problem, the challenge of predicting the three-dimensional shape of a protein from its genetic sequence. Before AlphaFold, figuring out a single protein structure could take a research team years of expensive lab work. After AlphaFold, the database of known protein structures went from a few hundred thousand to over two hundred million virtually overnight. Drug researchers, cancer biologists, and vaccine developers all basically received a gift they did not expect to get in their lifetimes.

The Nobel committee agreed. In 2024, Jumper and Demis Hassabis, DeepMind’s CEO and co-founder, received the Nobel Prize in chemistry. It was one of the more dramatic moments in recent scientific history, a reminder that AI is not just generating pictures of cats and writing cover letters for people who are too lazy to do it themselves.

And now this man is going to work for Anthropic. Jumper has not published a lengthy farewell post explaining his thinking. But the pattern is becoming impossible to ignore. DeepMind has been one of the most formidable AI research organisations in the world for over a decade. It pioneered game-playing AI, protein science, reinforcement learning, and a dozen other fields. Something is pulling its best people toward the startups and keeping them from staying inside the Google umbrella. Shazeer’s departure to OpenAI last week suggested frustration with the pace of things inside a giant corporation. Jumper’s departure to Anthropic this week suggests something more structural is happening at DeepMind specifically.

For Anthropic, landing Jumper is a loud statement. The company is already considered the most serious rival to OpenAI on frontier AI research. Bringing in someone with AlphaFold on his resume suggests Anthropic is serious about scientific applications of AI, not just chatbots and coding assistants. If Jumper is there to work on biology, medicine, or fundamental science problems, that signals a very different kind of ambition than most AI companies are publicly pursuing right now.

For Google, the more immediately uncomfortable question is what this says about the culture Demis Hassabis has built inside DeepMind. Losing your Nobel co-winner, the person who stood next to you on that Stockholm stage, to a competitor is the kind of thing that comes up in some very awkward board meetings. Two departures in one week at this level is not a coincidence. It is a pattern.

CHINA DROPS AN AI MODEL THAT BEATS GPT-5.5 AT CODING AND CHARGES ONE SIXTH THE PRICE, THEN GIVES THE WHOLE THING AWAY FOR FREE

There is a number buried in the GLM-5.2 release announcement that should make OpenAI’s finance team go very quiet. The new open-source coding model from Chinese company Z.ai beats GPT-5.5 on the industry’s most respected coding benchmark. The benchmark is SWE-bench Pro, which tests whether a model can actually fix real software bugs in real codebases, not just answer trivia questions about Python syntax. GPT-5.5 scored 58.6. GLM-5.2 scored 62.1. Edge to China.

The price is where it gets genuinely uncomfortable. OpenAI charges $30 per million output tokens for GPT-5.5. Z.ai is charging $4.40 for GLM-5.2. That is not a small discount. That is a fundamentally different business proposition. If you are a developer building applications that generate a lot of text, you are looking at the difference between a mortgage payment and a dinner at a mid-range restaurant every time you hit a certain usage level.

And then Z.ai released the model weights under an MIT open-source license, which is the most permissive license you can put on software. The company’s own documentation notes the license guarantees “no regional limits” and allows “technical access without borders.” That phrasing is clearly pointed at a specific audience after Washington’s recent moves against Anthropic, and the audience knows exactly who they are.

To be precise about what GLM-5.2 actually is: 744 billion parameters, built specifically for long-horizon autonomous coding tasks, meaning the kind of sustained software engineering work where you need the model to hold a lot of context and make coherent decisions across a large project rather than just solving isolated problems. It also outperformed GPT-5.5 on FrontierSWE (74.4 percent vs 72.6 percent) and MCP-Atlas (77.0 vs 75.3).

Now, benchmark claims from AI labs deserve healthy scepticism, and the community has developed a habit of squinting at any number that looks suspiciously convenient. GLM-5.2’s numbers were run on third-party benchmarks rather than purely internal tests, which is better than nothing, but independent verification is always worth waiting for before treating the leaderboard as gospel.

Still, the underlying trend is unmistakable. China’s AI labs are closing the performance gap with Western frontier models faster than most people expected even a year ago. The combination of competitive benchmark performance, dramatically lower cost, and open weights is a specific competitive strategy: flood the market with good-enough-or-better models at a price point that makes OpenAI’s commercial infrastructure look expensive and inflexible.

The US government spent the last year trying to choke off China’s access to high-end AI chips. The Chinese labs responded by training very large models more efficiently and releasing the results for free. It is not obvious who is winning that chess match right now. Actually, looking at this week’s news, it is becoming somewhat more obvious.

GALLUP DROPS THE NUMBERS: TECH WORKERS WHO IGNORE AI ARE THREE TIMES MORE LIKELY TO LOSE THEIR JOB

Gallup asked a straightforward question: among tech workers, who is more likely to lose their job right now? The answer arrived in a form that should inspire a certain amount of anxiety in the portion of the tech workforce still treating AI tools as optional.

Tech workers who do not regularly use AI are three times more likely to be laid off than those who do. The numbers are stark: regular AI users in tech face roughly a 6 percent predicted probability of layoff. Non-users face 18 percent. In an industry already running hot with restructuring announcements and headcount reductions, that is not a minor gap. That is the difference between relative job security and something close to a coin flip.

The cynical read on this data is that companies are using AI adoption as a loyalty test rather than a genuine productivity signal. If you are not using the tools management told you to use, the next reorganization spreadsheet will find your row first. The more charitable interpretation is that workers who integrate AI into their workflows are genuinely delivering more output with fewer hours, which makes their colleagues who have not made that adjustment look expensive by comparison.

Both interpretations are probably true at different companies and at different moments inside the same company. Human resources decisions are rarely as clean as a single variable. But the correlation Gallup found is strong enough that you cannot dismiss it as noise.

What is particularly worth noting is that the survey does not suggest workers are refusing to use AI out of ideological objection. The more common reasons workers give for low AI adoption are that they have not been trained, they don’t trust the output to be accurate, or their specific workflow doesn’t obviously accommodate the tools that have been made available. In other words, a significant chunk of the workers facing elevated layoff risk are not people holding principled stands against automation. They are people whose employers never got around to training them.

Goldman Sachs research from earlier this year found that AI is eliminating roughly 25,000 jobs per month across the US economy while creating about 9,000 back, netting out to around 16,000 losses per month. Put the Gallup numbers next to those Goldman figures and the picture that emerges is not just that AI is eliminating certain categories of work. It is that AI adoption inside companies is functioning as a sorting mechanism, with workers who have adapted being kept and those who haven’t being cut in the next round of reductions.

The practical upshot is not particularly complicated. If you work in tech and you have been deferring the learning curve on AI tools because you are busy, or skeptical, or waiting for a training program your employer never quite scheduled, the Gallup numbers suggest that is an increasingly expensive delay. Figure it out before your manager starts building the list.

WASHINGTON’S SECRET AI GOVERNMENT: THE TRUMP ADMINISTRATION SAYS IT HATES REGULATION BUT IS QUIETLY RUNNING THE MOST CONSEQUENTIAL AI POLICY IN THE WORLD

The Trump administration would like you to know that it opposes AI regulation. The White House has said so repeatedly, loudly, and with considerable enthusiasm. It rescinded Biden’s executive order on AI safety. It has argued that excessive government rules would crush innovation. This is the official position, and they have repeated it enough times that most people have accepted it as the operative reality.

What is actually happening is something quite different, and Axios laid it out clearly this week.

What has emerged in Washington is effectively a shadow AI policy. There are no formal rules. There are no binding regulations on the books. But there is an ad hoc system of interventions, warnings, and informal directives that is reshaping the AI industry in profound ways, with almost no public accountability for how or why those decisions get made.

The Anthropic lockdown from last week is the most visible example. Commerce Secretary Howard Lutnick sent a letter to Anthropic warning the company that it needed government permission before allowing any foreign nationals to access its most advanced models. No legislation authorized this action. No established regulatory framework was invoked. One letter from one cabinet official, and suddenly Anthropic is asking US allies how they would like to receive special access waivers for capabilities the company had been selling abroad freely.

The June 2 executive order on AI is another layer on top of this. It asks certain companies to voluntarily submit new models to the government for review thirty days before public release. The word voluntarily is doing a tremendous amount of lifting in that sentence. When the government asks politely and hints at consequences for non-compliance, the industry tends to find a way to cooperate. OpenAI, Anthropic, and Google all said they would participate. None of them issued a statement that sounded like a company making a genuinely free choice.

Separately, a bipartisan House draft bill, the Great American Artificial Intelligence Act, would preempt state AI laws for three years, require large AI developers to write safety plans before releasing models, and fund voluntary standards development. That last part matters: the enforcement mechanism for the proposed bill is also largely voluntary. The pattern is consistent across formal and informal channels.

The result is a regulatory environment where every major AI company feels significant government pressure on its decisions, but almost none of that pressure is codified, challengeable in court, or subject to democratic review. You cannot sue the government over an informal letter. You cannot appeal a voluntary framework. You cannot vote on a Commerce Secretary’s preference for what counts as a safe AI model.

For the AI companies, the uncertainty itself becomes strategic. When the rules are informal and case-by-case, the players with the best relationships with the current administration hold a structural advantage over everyone else, who is left guessing. That is not AI policy. That is power without accountability. The strange thing is that it is being administered by an administration that has publicly campaigned on the idea that government should stay out of AI entirely.