The Eureka Manifesto, Yuri Milner’s short book on humanity’s scientific mission, was published in 2021. At that point, the public conversation about artificial intelligence was largely focused on narrow applications — recommendation algorithms, image recognition, language translation — and the idea that AI might transform the pace of scientific discovery was a speculative claim held by a relatively small number of researchers and technologists.
Four years later, with AlphaFold having solved protein structure prediction, with AI systems demonstrating the ability to generalize across biological datasets in ways that would have seemed implausible in 2021, and with large language models reshaping how researchers interact with scientific literature, the manifesto’s argument about AI looks considerably less speculative. It’s worth going back to what Milner actually wrote, because the specific claims he made are now testable against what has happened.
The Claim That AI Changes the Kind of Science Possible
The manifesto’s argument about AI is not simply that it will make existing science faster. The stronger claim is that AI changes the category of questions that are scientifically tractable — that problems which were previously beyond reach because of the scale of data or computation required become approachable when AI tools are applied to them. Milner frames this as a qualitative shift, not just a quantitative one.
The protein folding case is the clearest example of what that claim looks like in practice. Protein structure prediction had been identified as a major open problem in biology for over fifty years. The challenge wasn’t a shortage of effort — thousands of researchers had worked on it — but a fundamental mismatch between the complexity of the problem and the tools available to address it. AlphaFold didn’t incrementally improve on existing methods. It effectively solved the problem, producing accurate structure predictions for virtually every known protein within a few years of its deployment. That is precisely the kind of qualitative shift the manifesto was anticipating.
More recently, research flagged by Yuri Milner has shown AI systems trained on genomic data learning to predict how genes behave in cell types the model has never encountered — generalizing from known biology to unknown biology in a way that extends the reach of existing datasets without requiring new experiments. This is again the manifesto’s argument made concrete: AI extracting scientific value that the data contained but that conventional analysis couldn’t access.
The Argument About Directing AI Toward Science
The manifesto makes a normative claim alongside its descriptive one: that directing AI development toward scientific discovery, rather than primarily toward commercial applications, is one of the most important choices the current generation faces. This argument was easy to dismiss in 2021, when AI development was dominated by tech companies pursuing commercial products and the idea of orienting it toward basic science seemed naively idealistic.
The landscape has shifted. Publicly funded AI research institutes focused on scientific applications have proliferated. Major AI labs have increasingly published research on scientific applications alongside commercial products. The argument that AI and science belong together — that the most important thing AI can do is accelerate human understanding of the universe — has moved from the fringe of the conversation toward something closer to its center.
Yuri Milner’s Breakthrough Prize has recognized this shift, with recent laureates including researchers whose work sits directly at the intersection of AI and biological science. The Breakthrough Initiatives have integrated machine learning into Breakthrough Listen’s signal analysis pipelines, with AI tools identifying previously undetected phenomena in archival data. Both represent the operational expression of what the manifesto argued: that AI should be put to work on the questions that matter most, not only on the ones that are most commercially convenient.
What the Manifesto Still Gets Wrong, or Leaves Open
Intellectual honesty requires noting where the manifesto’s argument remains untested or potentially overstated. Its confidence that AI will accelerate discovery uniformly across all fields of science is not yet established — the gains have been uneven, concentrated in fields with large structured datasets like genomics and drug discovery, and less visible in areas like fundamental physics or pure mathematics where the bottleneck isn’t data analysis but conceptual breakthrough. The manifesto’s timeline for AI-driven scientific acceleration is also vague enough to be compatible with almost any pace of development.
These are genuine limitations. But they don’t undermine the core argument, which is that AI represents a genuine inflection point in humanity’s capacity for scientific discovery, and that orienting resources toward that inflection point — funding the science, recognizing the researchers, and building the cultural conditions in which scientific ambition is valued — is the most important thing a committed philanthropist can do right now. That is the argument the Giving Pledge commitment Yuri Milner made in 2012 was pointing toward, a decade before the manifesto articulated it fully. The past four years have made it harder to dismiss.
I love tech and gaming, and am always looking for the ‘next big thing’…
