AI is collapsing the cost of serious research, and the implications run well beyond productivity
From string theory to security audits to Go bots, AI is compressing work that once required years and large teams into days and dollars. The pattern is consistent enough across domains that it demands a harder look at what "research capacity" will mean in five years.
The numbers David Sinclair puts on it are the ones that stop a reader cold. Work in his laboratory that would have taken 160 years and, as he puts it, “quite literally billions of dollars” can now be completed on a $10,000 budget. That is not a marginal efficiency gain. It is a structural change in what a small team can attempt.
The same compression shows up in fields with nothing in common except that AI touched them. Brian Greene reports that ChatGPT reproduced months of string theory results in roughly half an hour. Dylan Patel describes a single person using Claude Code finishing a research project that, by his reckoning, would have required a team of 200 economists working for a year. Eric Jang frames the shift in terms of what once required a full team of research scientists at DeepMind and millions of dollars of compute: the same class of work, he says, can now be done for a few thousand dollars of rented compute. These are different domains, different tools, different speakers, and the same underlying claim.
The security domain offers a particularly concrete version of the pattern. Krishna Rao, describing results from Anthropic’s Mythos model, notes that the model found 250 security vulnerabilities in an open-source codebase where a prior model had found only 22. That is not a modest improvement in recall. It suggests that the ceiling on what an automated audit can catch is moving faster than the security community’s baseline assumptions.
We're doing things in my lab that would have taken 160 years before. And quite literally billions of dollars on a $10,000 budget. David Sinclair
Not all the evidence is from industry. Terence Tao, discussing work on Erdős problem 1026, notes that researchers used the latest AI tools to gather numerical evidence in the course of solving it. That is a quieter data point than Sinclair’s 160-year figure, but in some ways more telling: when working mathematicians at that level begin treating AI as a standard part of the research apparatus, the tool has crossed a threshold that enthusiasm alone does not explain.
Geoffrey Hinton adds a medical dimension that comes with an important caveat. He points to a Microsoft blog post, which he describes as showing that multiple AI copies interacting with each other outperform most doctors in diagnosis. That claim is secondhand, Hinton citing a blog rather than a peer-reviewed study, and it should be read as a reported finding rather than an established benchmark. Even so, it fits the broader pattern: AI is performing at or above human-expert level in domains where the cost of being wrong is high.
What connects these data points is not just speed. It is the decoupling of output from the headcount and capital that output used to require. A solo researcher, a small lab, a single engineer with the right tool can now attempt work that previously required institutional scale. That changes who can do research, not only how fast research gets done. The barriers that filtered entry into serious scientific and technical work were partly meritocratic and partly a function of access to resources that most people and institutions did not have. Both filters are weakening at the same time.
Nathan Labenz pushes the trajectory to its logical endpoint, calling the prospect of curing the majority of human diseases within the next decade “obviously extremely exciting.” That framing is forward-looking enough that no current evidence can confirm it. But the nearer-term data points Labenz and others supply do not contradict the direction. If the compression rate observed across genomics, mathematics, physics, security research, and AI development itself holds for another several years, the kinds of scientific problems that have resisted solution primarily because they required resources no single group could sustain are now, at minimum, newly contestable. Whether that leads to Labenz’s decade-of-cures scenario or something more modest and uneven, the force multiplier the evidence describes is already operating. The question is not whether the compression is real. It is whether the institutions built around the old cost structure will adapt before the gap between what is now possible and what they are organized to do becomes impossible to ignore.