A tiny group of former Google researchers are working on something they couldn’t do at Google in a converted warehouse space south of Market Street in San Francisco, the kind of structure that formerly housed printing equipment but now holds individuals debating transformer architectures. Not because the work was prohibited. due to the work’s constant deprioritization. The roadmap would be updated, the quarterly review would occur, and the long-term research project that everyone thot was intriguing would be postponed for an additional six months in favor of something with more obvious short-term income ramifications. Several of them made the decision to depart after the third or fourth cycle of this.
For the better part of three years, this scenario has been unfolding throughout Silicon Valley in different shapes and at different companies. Researchers who understand the architecture of large language models as engineers who designed the training runs rather than as users or managers have been leaving Google, Meta, OpenAI, and Microsoft Research in significant numbers to work on the most important AI systems of the last ten years. A few are joining other well-known labs. An increasing number are taking a more disruptive approach by launching their own businesses, which are sometimes extremely small and concentrated on a particular issue that the big labs weren’t going to give priority to.
In interviews and sometimes in public writing, the explanations they provide tend to revolve around a few recurring themes. The most often mentioned is bureaucracy. Accessing the computing resources required for meaningful model training at a corporation the scale of Google or Meta necessitates navigating internal allocation procedures that can take months and yield unclear results. The data center has the GPU clusters and TPU pods that a researcher needs to test a theory.
It is a different story when they are assigned to a non-commercial research project. There is intense internal competition for compute, and projects without a distinct product team typically lose to those that do. This friction builds up into something that begins to feel insurmountable for academics whose work operates on timetables longer than a product plan.
Leaving is now more feasible than it would have been in previous cycles because to the venture capital climate. VC firms that have been observing the AI trend for years are now paying founding teams whose main qualification is their work at a major lab, often in very large sums. A revenue model is not necessary for the first year of the presentation. It needs a specific problem, a trustworthy team, and proof that the team is more knowledgeable about the issue than the general public. That accreditation is significant for a researcher who worked on graph neural networks or protein structure prediction for five years at DeepMind. Credibility comes before money, and effort comes before credibility.
Focus is what the niche startup model provides that the mega-corp does not. It is not necessary for an eight-person team working on automated scientific discovery to consider if the output supports a quarterly earnings call or interacts with an advertising platform. The limitation is more straightforward: is the research successful enough that funding for it or the following phase will be provided? For those who have spent years navigating the complexities of enormous organizations, this simplicity is truly appealing. Startups bring their own stresses and limitations, so it’s possible that the attractiveness is partially deceptive. However, the particular pressures are distinct, and for some researchers, that’s sufficient.
These firms’ aims provide insight into where the researchers believe the most important unresolved issues are. Former lab researchers continue to find automated scientific discovery—which uses AI to speed up hypothesis generation and experimental design across domains including materials science, drug discovery, and climate modeling—to be an appealing category. It tackles issues that are truly significant outside of the technology industry, is technically challenging, and has not been resolved by general-purpose frontier models. Similar principles apply to biotech AI, specialized enterprise reasoning systems, and domain-specific models for industries like law and medicine: high value, sufficiently specific to be manageable for a small team, and sufficiently removed from the major labs’ primary commercial priorities to avoid direct competition.
