Glonce

Academic AI researchers adapt to life outside the frontier labs

researchThe Glonce editorsMIT Technology Review🔥59Glonce rating59Impact55Novelty40Relevance90Surprise30Players72Credibility85
Academic AI researchers adapt to life outside the frontier labs

An MIT Technology Review writer traveled 30 miles south of San Francisco to a hotel in Mountain View, California, last week to host roundtable interviews and speak at a media training session for a convening of the Schmidt Sciences AI2050 program, an initiative funded by Eric and Wendy Schmidt that supports academics whose work involves AI. The piece, drawn from conversations at that event, describes a field in flux: over the past four years AI research has reoriented around large language models, and its cutting edge has shifted from universities to private companies, since universities cannot afford the GPUs needed to train or run frontier models, and Anthropic and OpenAI do not open the internal details of Claude or ChatGPT to outsiders. Over lunch, UC Berkeley computer science professor Nika Haghtalab compared being an AI academic today to being a biologist in a world where private companies held exclusive control over the gene-editing tool CRISPR: outside experts can study how ChatGPT and Claude behave, but cannot examine or steer how they are designed and trained. AI2050 gives fellows some funding to buy GPUs, which several researchers named as a major benefit of the program, but money remains tight, especially with reduced federal science funding in the United States, and even researchers who do not run their own models can find the cost of repeatedly querying commercial systems for rigorous study prohibitive. Rather than chase capabilities, many fellows deliberately steer toward questions unlikely to be tackled by Anthropic or OpenAI. Johns Hopkins computer science professor Anjalie Field said she tries not to work on problems she expects a tech company to solve, since companies need to turn a profit and have little reason to fund research whose answers might make them look bad; she pointed to her own study finding that language models give less sophisticated responses to prompts phrased in ways more commonly used by women than by men as an example of work unlikely to come out of Anthropic or OpenAI. A separate group of academics builds specialized, non-LLM AI models that analyze data, make predictions, or simulate physical systems; the piece notes that Google DeepMind's AlphaFold team, whose protein-structure model won a Nobel Prize, was disbanded last month, and that widespread public conflation of "AI" with energy-hungry LLMs makes it harder for researchers building tools like climate-focused AI to advocate for their work. The piece also notes that several prominent academics have recently taken leave from their universities to join frontier labs, and that many AI2050 fellows already hold industry positions alongside academic ones. A newer worry, emerging over the past six months, is that OpenAI's models have solved a number of real research problems in mathematics, leading some experts to worry about the future of humans in pure math; one fellow the author spoke with was worried about the mental health of mathematician peers. The piece closes on a more optimistic note: empirical science may prove harder to automate than mathematics because collecting data is inherently slow, and Carnegie Mellon computer scientist Tim Dettmers, who works to make AI models faster and cheaper to run, argues AI scientists will not replace humans but could make human scientists far more efficient, freeing them to pursue ideas they would otherwise never get to. The author suggests the same resource constraints keeping academics out of frontier-model training also push them toward smaller, more efficient models and new architectures, and would not be shocked if the next major AI breakthrough came from a scrappy academic lab rather than a big company.

Key facts

  • At a Schmidt Sciences AI2050 convening in Mountain View, UC Berkeley's Nika Haghtalab compared academic AI researchers today to biologists in a world where private companies exclusively controlled CRISPR: they can observe how Claude and ChatGPT behave but cannot study or steer their design and training.
  • Johns Hopkins' Anjalie Field said she avoids problems she expects a tech company to solve; her own study found language models give less sophisticated responses to prompts phrased the way women more commonly phrase them than the way men do.
  • Google DeepMind's AlphaFold team, which built a Nobel Prize-winning protein-structure model, was disbanded last month.
  • Over the past six months, concern has grown that OpenAI's models have solved real research problems in mathematics, worrying some experts about the future of humans in pure math, while one fellow said she worried about the mental health of mathematicians.
  • Carnegie Mellon's Tim Dettmers argues AI scientists will not replace humans but could make them far more efficient, freeing them to pursue ideas they otherwise would not have time for.

Why it matters

The piece captures a structural shift: over the past four years, frontier AI research has moved from universities to a handful of private companies that hold the GPUs and the internal access needed to build and study models like Claude and ChatGPT at the frontier. Academics who once drove the field now often work at its edges, observing systems from outside rather than building them, which reshapes what academic AI research can even ask.

Who it affects

University AI researchers who make up most of the AI2050 fellowship, including LLM researchers like Nika Haghtalab and Anjalie Field who face compute and access limits, and a separate group of academics who build specialized, non-LLM AI models for tasks like data analysis, prediction, or simulating physical systems, and who say public conflation of "AI" with energy-hungry LLMs makes their work harder to explain and fund. It also touches academics who have taken leave to join frontier labs and mathematicians unsettled by AI systems solving real research problems.

How to use it

The AI2050 program offers fellows funding they can put toward GPUs, which several researchers named as a concrete benefit, though money remains tight given reduced federal science funding in the US and the cost of repeatedly querying commercial models for rigorous study. The researchers describe a working strategy for staying relevant: deliberately choosing questions unlikely to be answered by Anthropic or OpenAI, since companies have little incentive to fund research that will not turn a profit or that might make them look bad, and leaning into the resource constraints that push toward smaller, more efficient models and new architectures rather than trying to out-compute the frontier labs.

How solid is it

This is a first-person account by an MIT Technology Review writer who attended one convening and reports conversations held there, not a systematic survey; the CRISPR comparison is the author's paraphrase of Haghtalab's remarks rather than a direct quotation. Field's study on gendered prompts is cited without methodology or sample size, and figures like the number of AI2050 fellows, how many hold industry jobs, or which academics took leave to join labs are not given.

Risks and caveats

The account rests on impressions from a single event and named individuals rather than aggregated data, so it should be read as a snapshot of sentiment among one cohort of well-funded academics rather than a representative picture of academic AI broadly. The reason AlphaFold's team was disbanded is not stated, nor are the identities of the "some experts" worried about the future of pure math or the fellow concerned about mathematicians' mental health, so those claims carry no independent sourcing beyond the author's account.

“I try not to work on problems that I think are gonna be solved by a tech company.”

— Anjalie Field, computer science professor at Johns Hopkins