AI companies are attracting top academic talent, shifting research away from universities.

Two years after winning the Nobel Prize, nearly a quarter of the core members of Google’s AlphaFold team have left Google, with Nobel laureates like John Jumper moving to Anthropic. The AI industry is absorbing talent from academia at an alarming rate: in 2022, 70.7% of AI PhDs entered industry, while only 20% remained in academia. Top researchers in AI companies now earn up to $1.94 million annually—five times the academic salary. Although AI companies can accelerate scientific breakthroughs and produce groundbreaking results, those who transition permanently to industry see their publication output drop by an average of 65%, with findings shifting from papers to patents and trade secrets. More critically, companies lack incentives to cultivate the next generation of scientists, and fundamental research is being marginalized. Science is rapidly leaving universities.

Article author and source: APPSO

Two years after winning the Nobel Prize, the original team behind Google AlphaFold has been split up.

Most of the original authors of the AlphaFold paper have been reassigned over the past year, with nearly a quarter of the full-time core authors having completely left Google.

Others shifted to fields such as Gemini Research Agent, enzyme design, genomics, and fusion, while some even joined Isomorphic Labs, Alphabet’s drug discovery company.

Even Nobel laureates John Jumper and Jonas Adler were temporarily assigned to the Code Strike team to focus on advancing AI coding capabilities. But as we all know, these two experts ultimately moved on to Anthropic.

Left: Demis Hassabis, CEO of Google DeepMind; Right: John Jumper. Original article link: https://www.ft.com/content/61b2953d-ee0d-45de-af6e-a9c1cf524b33?utm_source=chatgpt.com&syn-25a6b1a6=1

In response to this report, Pushmeet Kohli, Vice President of AI for Science at Google, said that DeepMind’s science team has not changed direction but has expanded its scope, continuing to advance research in proteins, genomics, and enzyme design, while leveraging Gemini to accelerate scientific discovery.

AlphaFold once represented a relatively traditional approach to scientific research: forming a team around a clear question, committing long-term resources, and delivering results through focused effort.

Today, similar teams are gradually being replaced by more general-purpose model platforms. Proteins, mathematics, genomics, fusion, and code—though seemingly belonging to different disciplines—can all become a set of capability modules within Gemini once they enter AI companies.

Meanwhile, scientific talent, computing resources, and control over key issues are increasingly concentrated in the hands of a few AI companies.

Over the past few years, an increasing number of professors in mathematics, physics, biology, and computer science have entered the AI industry.

For example, Anthropic has consistently hired physicists, economists, and philosophers; OpenAI has brought together researchers specializing in black holes, string theory, and pure mathematics; and DeepMind has assembled experts in biology and nuclear fusion.

Recently, Professor Subbarao Kambhampati of Arizona State University joked that a new meme has emerged in academia: "I want to join Anthropic."

According to a rough tally by The Atlantic, just four leading AI companies have gathered at least 80 current or former professors.

The direction of talent mobility is also becoming increasingly uniform.

According to Stanford University’s AI Index Report, in 2011, the proportion of AI PhD graduates entering industry and academia was roughly equal, at 40.9% and 41.6%, respectively. By 2022, the proportion entering industry rose to 70.7%, while the proportion entering academia fell to 20%.

A 2026 NBER working paper tracked approximately 42,000 U.S. AI researchers. Researchers under 40 were about six times more likely to leave academia than those aged 40 and older; about 70% of those who switched jobs remained at their companies five years later.

Money is certainly important. The annual income of the top 1% of AI researchers in companies rose from $595,000 to $1.94 million, while their counterparts at universities increased only from $301,000 to $392,000. The gap between the two has widened to approximately $1.5 million.

Harder to refuse are the research conditions.

In the early 2010s, approximately 65% of large-scale machine learning models were developed independently by academic labs. By the 2020s, this share had fallen to less than 10%. By 2022, about 81% of cutting-edge models had been developed independently by corporations.

When Professor Anca Dragan from UC Berkeley joined DeepMind, she clearly stated her requirements: the "data, computing power, and budget" essential for a cutting-edge safety research institute.

In contrast, a study at a university might require applying for funding, purchasing equipment, hiring engineering staff, and waiting for access to computational resources. In an AI company, the same researcher can directly access large models, proprietary data, distributed computing power, and a team of hundreds of engineers.

However, when a top professor leaves the university, the loss extends far beyond just a few papers. The university also loses a mentor, an entire research group, a pipeline for doctoral student training, and potential research directions that could have developed over the next decade.

In the past, tech companies purchased research findings from universities. Now, they are beginning to directly buy the people who produce those findings. Companies gain an individual, while universities may lose a generation.

Attributing all talent mobility solely to the decline of academia also underestimates the value being created by corporate research.

AlphaFold has demonstrated that companies can produce results that change the course of scientific history. Over 200 million protein structure predictions have saved biologists months, even years, of work.

Bell Labs long ago demonstrated that corporate research labs can transform science. At its peak, it employed around 1,200 PhDs and produced at least ten Nobel Prizes. Demis Hassabis has frequently cited it as a model for DeepMind.

AI companies are increasingly becoming a new type of research institute that spans multiple disciplines, with strong engineering capabilities and commercial applications.

DeepMind integrates Gemini into mathematics, genomics, and fusion; Isomorphic Labs collaborates with pharmaceutical companies to develop new drugs; Anthropic launches Claude Science; OpenAI incorporates ChatGPT and Codex into scientific research workflows.

Relevant achievements have already emerged. Physicist Rogerio Jorge developed open-source fusion software using AI, while theoretical computer scientist Barna Saha and others used GPT-5.5 Pro to assist in high-dimensional geometric proofs.

OpenAI states that approximately 1.3 million people per week use ChatGPT to handle advanced scientific and mathematical tasks, and plans to offer free access to cutting-edge models and Codex to 100,000 university researchers.

From the perspective of personal impact, a scientist’s influence may even be amplified upon joining a company. As mentioned, while in academia they might lead a dozen students, in an AI company they can mobilize hundreds of engineers, vast computational resources, and global data platforms. Methods that previously remained confined to academic papers can more easily be integrated into drugs, software, experimental equipment, and industrial systems.

But the cost is obvious.

NBER research found that AI scholars who transition long-term to industry experience an average 65% reduction in paper output. Between 2000 and 2019, the share of AI researchers employed by companies rose from 48% to 68%, while their share of papers increased only from 27% to 32%, and their share of patents rose from 86% to 95%.

Researchers are still studying, but the results are increasingly leaving academia and moving into patents, internal models, and trade secrets. Google once open-sourced the Transformer, nurturing the entire industry. Today, training data, model weights, and failure logs are becoming harder to leave corporate servers.

Science continues to accelerate, but open science may be hitting the brakes.

An AI company can become a research institute, but it is difficult to become a university.

Universities undertake a frequently overlooked task: training large numbers of students who have yet to prove themselves, and preserving questions that currently have no application, no market value, and may not reveal their worth for decades.

However, AI companies prefer professors and seasoned researchers who have already proven themselves and are less patient with nurturing newcomers. SignalFire data shows that in 2024, new graduates accounted for only 7% of hires at large tech companies, a decline of more than half since 2019.

🔗 https://www.signalfire.com/blog/signalfire-state-of-talent-report-2025

Companies are hiring top professors and experienced researchers from universities while reducing entry-level positions and investment in training. They can assemble a team of leading scientists, but may not take on the responsibility of cultivating the next generation of scientists.

The related impact has already appeared in university classrooms. Students interviewed by The Atlantic found that some courses were suddenly canceled, with the practical reason being that professors had taken "leave" to work at AI companies.

Meanwhile, after the renowned professor left, the PhD student lost not only a course but also access to recommended resources, research networks, project organization skills, and opportunities to engage with cutting-edge issues.

More subtle changes have occurred in research topics.

Drug design, programming, mathematical reasoning, and materials discovery can enhance models and lead to tangible products, making it easier to secure resource support. Topics such as long-term field research and ecological conservation often find themselves in an awkward position.

Yet history偏偏 likes to play tricks on short-term returns.

Number theory was long regarded as the purest and most impractical branch of mathematics, yet it later became the foundation of modern cryptography. When lasers first emerged, they were once called "a solution looking for a problem." Similarly, the early internet was difficult to evaluate using business models.

The significance of the public research system lies in its ability to allow humanity to study issues whose value cannot yet be explained.

Today’s AI companies, operating under intense competition in model performance, product launches, and capital investment, are more likely to direct their resources toward areas that enhance models, support products, or generate patents—even if they are willing to invest in foundational science.

A peculiar kind of prosperity is likely to emerge in the future.

Humans can now discover new drugs, prove theorems, and find new materials faster than ever, and even the greatest scientists have access to the most powerful tools in history. Yet questions arise: who decides what to research, who trains the next generation of scientists, and who is willing to continue doing research that no one pays for?

Scientists have not left science; rather, science is accelerating away from universities.

Sometimes I think of universities and AI companies as the two ends of a river. The university is more like the upstream—where the current flows slowly, sediment accumulates, and countless tributaries seem useless. People there measure, record, and argue, while others spend their entire lives studying a stream that ultimately never reaches the city.

AI companies are more like downstream hydroelectric plants. They can concentrate the flow of water, power machines, and generate immense energy, rapidly advancing science into drugs, software, and industries.

Hydroelectric power plants are certainly important. Without them, the river water simply flows by.

But if everyone rushes downstream to generate power, the upstream water source will gradually dry up; years later, everyone will be sitting in the data center waiting for a river that no longer flows.

While mathematicians work on improving models, biologists train agents, and physicists push the boundaries of computational power, humanity may gain stronger AI. The problems left behind have no product roadmap, no return on investment, and temporarily no clear application.

They are quiet. So quiet that it’s easy to mistakenly think they aren’t important.