What happened in June

Mark Zuckerberg announced Meta Superintelligence Labs on June 30 in an internal memo naming eleven new hires. The lab is led by Alexandr Wang, formerly chief executive of Scale AI, with Nat Friedman, formerly of GitHub, running product. Wang's arrival came with a $14.3 billion investment by Meta for a 49 percent non-voting stake in Scale AI, a structure that has drawn attention from people watching the ongoing FTC case against Meta. Daniel Gross joined in July.

The recruiting itself is the story. The New York Times reported that Zuckerberg personally offered compensation packages valued at between $1 million and $100 million to employees at OpenAI and Google, with meetings at his homes in Lake Tahoe and Palo Alto, and a stated aim of hiring around 50 researchers. Before settling on hiring, Meta reportedly tried to buy Safe Superintelligence, Thinking Machines Lab and Perplexity, and was turned down or could not agree terms in each case. Satya Nadella and Sam Altman have responded by stepping up their own retention and recruiting.

The numbers are the message

We want to be careful about the $100 million figure. It is the top of a reported range, it is a package rather than a salary, and it applies to a very small number of people. But the precise number matters less than the fact that it was reported and not denied. The signal to every researcher in the field is that the price of a person who can run a frontier training job is now set by the same logic as the price of the GPUs, and that a handful of companies will pay whatever it takes.

That has a direct effect on what a research career looks like from the inside. When the difference between staying at a lab and leaving is a few hundred thousand dollars, people decide based on the work. When the difference is tens of millions, the work stops being the deciding factor for almost anyone, and we do not think it is a moral failing to say so. A market that prices researchers this way is telling them that their judgement about which problems matter is worth less than their availability.

What it does to graduate students

The people we worry about most are not the ones getting offers. They are the students two or three years behind them. A doctoral student watching this summer learns that the fastest route to a nine-figure outcome is to work on scaling at a frontier lab, and that everything else, from theory to evaluation to the slow work of understanding why methods fail, pays in the ordinary currency of academic careers. Those students are not wrong about the incentives. They are being given an accurate picture of what the industry values.

The cost shows up in which problems get worked on. Frontier scaling is well funded and well staffed and will continue to be. The problems that are under-resourced are the ones that a small foundation or an academic group can actually do, and those groups depend on a supply of people who chose the work over the money. That supply was already thin. A recruiting spree that publicly prices the alternative at $100 million makes it thinner, and it does so for a cohort that will not see a frontier lab offer for years, if ever.

What it does to academic labs

An academic lab cannot compete on compensation and should not try. What it can offer is freedom to choose the problem, a longer time horizon, and the right to publish everything. All three of those are things the frontier labs are in the process of giving up, whether through secrecy, through product deadlines, or through the sheer cost of running experiments that only make sense at scale. So the honest pitch from a university to a strong student is that you will be paid a fraction of the frontier rate in exchange for the ability to work on what you think is true.

That pitch only works if the freedom is real. A lab that spends its time chasing the same benchmarks as the frontier at a hundredth of the compute has given up the one advantage it had. The academic groups we expect to survive this period are the ones doing work that gets more valuable as the frontier gets more closed, meaning independent evaluation, replication, interpretability on open models, and the maintenance of tools and datasets that everyone uses and nobody wants to pay for.

Vocation versus signing bonus

We do not begrudge anyone who took an offer this summer. Some of them will do their best work with the resources Meta is providing, and it is not obvious that a research career is better served by a smaller budget. What we object to is the idea that the price now attached to a small number of people tells us anything about the value of the field's work. It tells us about the scarcity of one narrow skill at one moment, and about how much a few companies fear being second.

The thing we would want someone to track is where the people who were not hired go. If the students and postdocs who did not get the call keep working on the unfashionable problems, the field will be fine and this summer will be remembered as an expensive footnote. If they follow the money into the same handful of buildings, we will have converted a research community into a labour market, and the problems nobody was paid to solve will still be there.

Sources

  1. Wikipedia, Meta Superintelligence Labs