It has been a long road, but we're here.. finally.. well.. almost. The Genesis mission is the de-facto National AI effort for the US and there were around 400+ participants from academia, national labs, other government agencies and philanthropy. There is still a lot to be figured out, and I won't report on the project goals or the 26 lighthouse projects, as they are accessible elsewhere.
I'll just talk about a few things that stood out. Most of this was discussed today in the breakouts (in the PM sessions, I decided to bounce around a bit), but I admit I might be occasionally hallucinating because of the weekly (sometimes daily) discussions I have on this general topic with all kinds of folks.
One thing that stood out: There was concern that the deepest risk to Genesis isn't that it might fail to achieve 2x productivity; it's that it succeeds at 2x output while accidentally destroying the conditions that produce the 1% of work that actually matters.
One thing that was missing: There wasn't much discussion about the platform itself. We need a separate workshop for that.
One thing that was interesting: The always interesting Rick Stevens (Argonne) did (in a one-on-one chat) mention that Genesis should be Openclaw for Science. That would be something.
Here's the rest.
Scope of Genesis.
There was talk about the fact that $1T is spent in the US on research (80% by industry). How do we get much more out of it out of a 'whole nation' effort. While some of the initial messaging in the White House Exec Order was around scientific discovery and productivity, Ethan Klein (CTO of the USA) emphasized the full technology stack : Not just discovery, but also higher technology readiness levels all the way to products.
AI shifts value from execution to judgment, but we train people through execution.
This is something we have been discussing deeply at U. Michigan also. Most people agree that with proliferation of powerful AI, the role of scientists and engineers might move closer towards judgement. The problem is that scientific judgment is currently acquired through the slow apprenticeship of execution. If AI collapses execution, we face a genuine epistemological crisis: how do future scientists develop the judgment to know when AI is wrong if they never learned the craft it's replacing? Somebody said using ChatGPT to do something is like going to the gym and watching videos of people working out. Something we have to grapple with.
The PhD Degree.
Quite a bit of discussion around the meaning and relevance of a PhD (including some pleasantly bold statements from the presidents** of Purdue, CMU & Stonybrook) and the need to holistically evaluate graduate programs. There was a discussion about programs like Purdue's B4D7 (7 years to a PhD after high school) while reaching some kind of consensus that we don't want to artificially restrict the length of doctoral degrees, but perhaps have more options : Doctor of technology / Doctor of engineering? Reinvigorated masters programs, etc. More discussions around PhD pay and career trajectories.
** These are true leaders!
Failure is a public good, Fail fast, fail often.
Basic research failures benefit everyone, but no individual or firm is incentivized to produce them. As scientists, we generally only report successes. That is terrible for science. Proposals were made to use AI as a failure factory, generating enormous numbers of systematically documented failures from which humans learn at scale. Humans can't be incentivized to fail for the collective good; AI has no reputational cost for failure. This reframes AI's role from "making us more productive at succeeding" to "making failure cheap, fast, and legible." Combined with the repeated insistence that failed experiment data is the most under-collected asset in science, this is a genuine value.
The real lighthouse challenge is trust, not speed.
There is quite a bit of emphasis on doubling productivity, and that usually equates to speed. But trust must be engineered, not asserted: through uncertainty bounds, validation evidence files, documented limits of applicability. Proper V&V work takes years of sustained effort.... and AI might make this much more complex.
The 80/20 split is the enemy of real AI-for-science.
Lots of ranting against the train on 80%, predict the other 20%, report a high score and get accepted in Neurips. What's needed instead: build a team, build a model, discover and test a brand new hypothesis that has never existed before.
"Explore vs. exploit" mismatch in Genesis.
Lots of discussions that the lighthouse challenges feel like exploit (concrete milestones, TRL targets, timelines), but the actual scientific problems require explore (open-ended discovery, tolerance for failure, long horizons). DOE seems to be asking for exploration but measuring exploitation. If Genesis optimizes for exploit, it will miss the next equivalent breakthrough. The evolutionary computation analogy is precise: we should use AI's ubiquitous tools to increase exploration, not merely accelerate exploitation.
The open-source AI crisis for U.S. science is immediate and existential.
All of the top-performing open-weight models are now from China. If U.S. science depends on closed proprietary models then scientific reproducibility, explainability, and validation become structurally impossible. The national security implications are obvious but the scientific methodology implications may be even more important.
Tax credits for industry-funded research at universities.
The President of Purdue was quite the champion for this: companies currently face a higher burden for sponsoring university research than for keeping research in-house. Reversing this with a tax credit would redirect billions into university labs. There were other spicier and interesting takes on related issues.
Authorship/AI slop.
The old model of science is based on an individual or a very small team working alone. We may be moving toward something more like "film production". So what is scientific skill? This has immediate implications for how we design PhD programs, evaluate tenure cases, and structure authorship norms. "The 2 people who are crushing it are being swamped by AI slop. Meanwhile, my CFO is asking me why costs are up." The flood of low-quality AI-assisted papers is already degrading signal-to-noise ratios. If Genesis doubles "productivity" measured by volume, it will actually decrease the rate of genuine discovery by burying it in noise.