Order-of-magnitude time compression is now a recurring structural feature of AI integration, not an edge case

The numbers David Sinclair puts on his laboratory’s AI-assisted research are the ones that demand attention first. Work 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. It is a structural discontinuity in what a small team can attempt.

The same compression appears in fields that share nothing except that AI touched them. Brian Greene reports that ChatGPT reproduced months of string theory results in roughly half an hour. Alex Lupsasca describes Codex solving a simulation of the SYK model in quantum mechanics in 10 minutes, a task that multiple research groups had been unable to complete for over a year. 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 full year. Eric Jang puts the AlphaGo comparison in stark terms: work that required a whole team of research scientists at DeepMind and millions of dollars can now be done for a few thousand dollars of rented compute. Different domains, different tools, different speakers, and the same underlying finding.

The pattern repeats in software and operations with equal consistency. Jason Lemkin describes a CEO who replaced a $600,000 Salesforce contract with a vibe-coded CRM built in three weeks. Jon McNeill contrasts AI-powered ERP implementations measured in days against the standard 9 to 12 months required by conventional systems. Ronak Malde reduced model onboarding time at Trajectory.ai from three months on the first customer engagement to approximately one week for subsequent ones. Vijoy Pandey reports that Cisco’s Jarvis multi-agent system cut incident response time from hours to instantaneous, while reducing team load by 30% and fully automating 40% of tasks end to end.

Thanks to LLM coding, what took a whole team of research scientists at DeepMind and millions of dollars of research and compute can now be done for a few thousand dollars of rented compute. Eric Jang

The quantitative gains in adjacent domains are no less striking. Bradley Sutton reduced product validation from two hours to under five minutes per product. Joseph Krause at Radical AI achieved 1,200 alloy syntheses in three months, against a prior benchmark of 500 in twelve. Marc Andreessen puts leading-edge programmer productivity gains at 20 times year over year. Sergiy Nesterenko describes AI cutting PCB design tasks that take humans two to ten weeks by a factor of ten, while explicitly noting that the technology is not yet beating humans outright. The qualification matters: a tenfold reduction without yet reaching the human performance ceiling suggests the compression has further to go.

The security domain offers a concrete illustration of how fast the ceiling is moving. Krishna Rao notes that Anthropic’s Mythos model found 250 security vulnerabilities in an open-source codebase where a prior model had found only 22. That is not an incremental recall improvement. It suggests that the baseline assumptions the security community has been using to scope automated audits are already outdated.

What connects these data points is not speed alone. It is the decoupling of output from the headcount and capital that output previously required. Eiso Kant describes Poolside going from the start of pre-training to launching Laguna XS2 in five weeks, running between 10,000 and 20,000 experiments per month with a small team. An engineer at Poolside, he notes, transitioned into a working reinforcement learning researcher making real progress within six months by using the model factory. Tulsee Doshi describes a Google DeepMind researcher running complex ablations on Gemini models from her phone, in an hour, from a hot tub. The barrier that once separated serious research from casual inquiry was partly skill and partly access to resources most individuals and small teams did not have. Both barriers are eroding at the same time.

Andrew Wilkinson offers a useful ground-level illustration: his CFO, who had never written a line of code, used Claude Code to build a portfolio tracking tool replacing Addepar within two weeks. That single anecdote would be easy to dismiss if it were not arriving alongside Lemkin’s CRM story, Malde’s onboarding compression, and Lupsasca’s quantum simulation. At some point the accumulation of domain-specific examples stops being a collection of curiosities and starts being a description of a general condition. The evidence presented here suggests that point has already passed. The question now is not whether order-of-magnitude compression is real. It is which institutions and workflows have yet to encounter it.