AI is compressing task timelines by an order of magnitude, and the pattern holds across every domain that has tested it

Joseph Krause, who runs Radical AI, put 1,200 alloy syntheses on the board in three months. The prior benchmark for that kind of output was 500 in 12 months. That ratio, roughly an order of magnitude compressed into a quarter of the time, is not an isolated result from a single optimistic lab. It shows up, with different numbers and different domains, across nearly every field that has started running serious workloads through AI tools.

Brian Greene, the physicist, found that work his team had spent months on was reproduced in half an hour. Alex Lupsasca, describing a recent exchange in his research circle, reports that Codex ran a simulation of the SYK model, a technically demanding problem in quantum mechanics and gravity, in 10 minutes. Multiple research groups, he notes, had been trying to run that simulation and couldn’t do it. Eric Jang frames the shift in starker economic terms: work that required a whole team of research scientists and millions of dollars of compute can now be done for a few thousand dollars of rented compute. These are different tools, different disciplines, and the same underlying finding.

The compression is just as visible in commercial software. Patrick Collison reports that Stripe’s internal agent system scaled from 1,200 pull requests per week to 7,000 over a span that began with a blog post in January or February. Most of those pull requests, Collison notes, came from one engineer, who runs 16 agents from a single screen. Marc Andreessen puts the productivity shift at roughly 20 times for leading-edge programmers compared to a year ago. Jon McNeill describes AI-powered enterprise resource planning systems that implement in days against a standard deployment timeline of nine to twelve months. Jason Lemkin reports replacing a $600,000 Salesforce contract with a system his organization built in three weeks.

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

The pattern extends into professional services where timelines have historically been set by institutional process rather than human effort alone. Robin Nessén completed an acquisition from letter of intent to closing in 12 days using an in-house diligence tool. Krishna Rao describes Anthropic producing statutory financial statements for all its legal entities using Claude, with a human checking but not drafting the output. Vijoy Pandey says a Cisco multi-agent system reduced incident response time from hours to instantaneous while handling 40 percent of tasks end to end, cutting team load by 30 percent. Dylan Patel describes a single person using Claude Code finishing a research project that would, by his reckoning, have required a team of 200 economists working for a year.

David Sinclair, the biologist, offers the figure that tends to stop readers: work in his laboratory that would have taken 160 years and, in his words, “quite literally billions of dollars” can now be completed on a $10,000 budget. That is not a productivity gain in the ordinary sense. It describes a change in which scientific questions are even worth attempting. Andrew Wilkinson describes a person who had never coded in his life building a working financial tool in about a week after being pointed at an AI coding assistant. Mark Cuban puts two or three people doing work that, five years ago, would have cost two or three million dollars a year to staff. The common thread is not speed alone. It is the decoupling of output from the headcount and capital that output used to require.

The hardware layer is moving in the same direction. Ethan He, describing the effect of faster coding models on AI research iteration, notes that tasks which once took weeks to build, generating synthetic data or writing a new algorithm, can now be completed in hours, meaning compute itself becomes the new constraint rather than implementation time. Eiso Kant reports that his team ran Poolside’s Laguna XS2 from the start of pre-training to launch in five weeks, and runs between 10,000 and 20,000 experiments per month. Anima Anandkumar notes that neural operator weather models now run at near-traditional accuracy but tens of thousands of times faster than physics-based predecessors.

What the evidence collectively establishes is not that AI is incrementally faster. It is that the ratio between old timelines and new ones is large enough, and consistent enough across research, engineering, legal, financial, and operational work, that the gap can no longer be explained by tool selection or team skill. The organizations still calibrated to month-long procurement cycles, year-long ERP deployments, and decade-long research programs are operating on assumptions the data has already moved past. Dara Khosrowshahi, Uber’s chief executive, said publicly that within roughly five years he may stop adding engineering headcount and instead add agents and buy more compute. That framing treats the compression as a structural input to workforce planning, not a temporary efficiency. The question for every institution is not whether the compression is happening. It is how far the gap between what is now possible and what the organization is built to do will widen before the organization adjusts.