Anthropic co-founder predicts a billion virtual scientists will soon emerge

AI in the physical laboratory

The transition from digital theory to physical reality is already underway. Anthropic has released a research-preview interface that lets AI agents operate laboratory devices with a programmable interface. This technology is being used to help teams run serial-dilution dose-response experiments about three times faster than previous methods.

In practical applications, researchers have already had to guide AI systems to recognize physical failures, such as foaming in samples, rather than software bugs. Anthropic has also established a wet lab in the San Francisco Bay Area to test these capabilities. However, Clark emphasizes that AI should not be left to perform science entirely unattended for the time being.

As these systems become more capable, the challenge for policymakers will be to prevent downstream bottlenecks. If AI creates an avalanche of new proposed drugs or materials, the world must find ways to move those discoveries out of the lab and into factories and clinics quickly. The focus will shift from generating ideas to the physical logistics of testing and manufacturing them at scale.

The timeline for a new era

Clark believes that the evidence for this technological shift has become convincing within the last year. He argues that AI-driven invention will eventually alter how countries plan infrastructure and improve global health. He suggests that while humans once thought it would take hundreds of thousands of years to reach these milestones, we may achieve them in just a handful of years.

The ultimate goal of this acceleration is to enhance human strengths while reducing weaknesses. Clark frames this as a way for the species to become wiser and more capable. By leveraging virtual scientists, the human workforce can focus on high-level direction while the AI handles the

The roots of the prediction

Jack Clark’s perspective on the speed of AI progress is rooted in science fiction. He draws his framework from the work of writer Vernor Vinge, who predicted in 1993 that intelligence exceeding human capabilities would drive progress at a pace that humans would struggle to follow.

Clark shared the full text of his speech in his Import AI newsletter shortly after the event. His background includes two decades as a journalist covering data centers before he moved to OpenAI in 2016 and eventually co-founded Anthropic.

Evidence of shifting capabilities

The transition toward AI-driven discovery is already showing measurable signs in complex fields. For example, a test called Humanity’s Last Exam, which covers subjects from chemistry to classical literature, launched in early 2025 with an initial top AI score of only 8%.

However, recent performance on specific tasks suggests a rapid climb in capability. In a recent evaluation involving 50 unpublished physics problems, a model's score jumped from the late 80s to 98% after the AI insisted that the human experts' provided answers were actually incorrect.

Validating expert knowledge

The physics evaluation demonstrated that AI can identify errors in calculations that humans had already verified. The authors of the paper described these as "turn-the-crank" research problems, where known techniques are used to generalize an existing result. In several cases, the experts found errors in their own work only after the AI pointed them out.

While the AI agents showed high accuracy on these problems, the researchers noted that the models do not appear to have the same "taste" in problems as humans do. This suggests that while the AI can perform the heavy lifting of calculation and generalization, human oversight remains a key component of the scientific process.

The economic reality of research

Science is currently facing a significant decline in productivity. In the pharmaceutical industry, the number of new drugs approved per billion dollars of research spending has halved roughly every nine years since 1950, a trend known as Eroom’s Law. This reflects a broader reality where the easiest scientific discoveries have already been made, leaving only more expensive and difficult goals remaining.

Because the low-hanging fruit of discovery is gone, reaching further into the unknown requires immense resources. Traditionally, the only way to overcome this was to hire more people, but training new researchers takes years and costs millions of dollars. Clark argues that AI provides a new way to change this economic arithmetic by augmenting human labor in frontier disciplines.

Scaling through software gains

The shift toward AI-driven productivity is already visible in the software industry. Anthropic researchers currently produce almost 10 times as much code as they did during the company's first four years of operation. These gains have also improved the capabilities of AI systems in both cyberdefense and cyberattack, showing how automation can scale technical output.

Clark shared these insights during the IMF-World Bank Annual Meetings in Bangkok and during the 2026 Per Jacobsson Foundation Lecture. He suggests that the next few years will see these types of gains move into other areas. By creating copies of human expertise, the economy will be able to support scientific progress in ways that were previously impossible.