The Real AI Race: Genesis, AI Scientists, and the New Stack

If you’re still arguing about which model – GPT, Claude, or Gemini – is “best,” I respectfully believe you may be watching the wrong race.

The real question now is:

Who can point AI “scientists” at the deepest data, with the tightest loop to real experiments?

That’s why the new Genesis Mission out of the White House caught my eye.

When you strip away the hype, an AI model gets better when you pull four levers at the same time:

Data – not just more, but cleaner, richer, and closer to the problems you care about.

Compute – enough serious hardware to train longer, at larger scale, with better objectives.

Algorithms & training tricks – smarter architectures, better loss functions, reinforcement learning from human feedback, retrieval, agents – all the techniques the frontier labs are already using.

Feedback from reality – real tasks, real users, real experiments, not just higher scores on synthetic benchmarks.

OpenAI, Anthropic, Google, Meta and others are already pulling all four levers. Mixture-of-experts, RLHF, retrieval-augmented generation, tool-calling, multi-agent “co-scientist” workflows – that’s table stakes now. The playbook is increasingly public.

The bottleneck has shifted to the environment you run all of that in:

What data can you see?

How much real compute can you throw at it?

How tight is your loop from “the model has an idea” to “we ran the experiment and updated the system”?

What Genesis is really doing

The Genesis executive order puts the U.S. Department of Energy in charge of building an integrated AI platform that:

pulls together decades of federal scientific data from the national labs and agencies – much of it never on the public internet,

runs on national-lab supercomputers and secure cloud AI,

connects to robotic and automated labs that can execute AI-designed experiments, and

aims the entire stack at a set of national challenges: fusion, critical materials, semiconductors, biotech, quantum, and more.

DOE describes Genesis as a “national discovery platform” designed to dramatically accelerate the pace and impact of U.S. science and engineering. Reporting around it compares this to a Manhattan / Apollo-style mobilization.

And if you look at the partners already in the mix, it’s the usual cast of frontier compute and model providers – the big AI and cloud players you’d expect.

So, this is not government versus Big Tech. This is government plus Big Tech, trying to build a state-anchored AI brain for hard science.

This is landing on top of AI “scientists,” not in a vacuum

In parallel, outside government, we already have early AI scientists:

Systems that read thousands of papers, build an internal “world model” of a domain, generate hypotheses, and write analysis code.

Multi-agent setups (like Google’s AI co-scientist, or Future House’s Kosmos) that run for hours, cross-checking sources, running analysis, and surfacing testable ideas.

Agentic “AI Scientist” pipelines that have already produced papers – from idea – experiments – manuscript – that passed peer review.

They’re still rough, but they’re real, and they’re being tuned and stress-tested by serious scientific teams.

Now plug that kind of system into something like Genesis:

Instead of scraping the open internet, it sits on top of decades of national-lab experiments and simulations.

Instead of just talking, it has hooks into simulators and robotic labs.

Instead of one PI and a small team, it’s operating inside an institutional stack that can run thousands of threads in parallel.

The techniques aren’t exotic. The context is.

Why this is not just “connecting the data” again

We’ve been trying to “connect information” for decades:

data warehouses and lakes,

EHR integration projects,

enterprise knowledge graphs,

federated search over PDFs and databases.

Most of those did a decent job at storage and retrieval. They made it easier for humans to find things and build dashboards. They did not change who was doing the real reasoning. The scientist or analyst still had to:

read the papers,

write the analysis code,

build the model,

design the experiment,

interpret the result.

What’s different now is the combination of:

Depth of data – raw and processed outputs from experiments, large simulations, negative results, and sensitive work that will never show up in a journal.

Models that can infer – systems that integrate text, code, math, and structured data; spot patterns; propose mechanisms; generate analysis code; and explain their reasoning.

A closed loop to reality – the ability to propose, act (via simulations or labs), measure, and update continuously.

In that setting, “inference” stops being “predict the next token on a static corpus” and becomes:

Given everything we’ve ever run in this domain, what should we try next, and what does the system learn from the result?

That’s a different animal than a nicer search interface.

What this means IMHO

A few implications I’d highlight:

1. The “best model” won’t just be the one with the highest benchmark score. It will be the one embedded in the best environment: deepest data, strongest compute, tightest experimental loop. Architecturally, it may look a lot like a top commercial model. The difference is what it sees and what it’s allowed to do.

2. Science will accelerate – and centralize. If Genesis delivers even a fraction of what’s on paper, the fastest moves in fusion, critical materials, some bio countermeasures, and advanced manufacturing will cluster around organizations plugged into this loop. Everyone else will still benefit from the exhaust – papers, tools, better public models – but usually with less detail and some delay.

3. Industry will benefit, but not evenly. We’ve already seen this pattern with NIH. Public investment in basic biology lifted the entire drug industry; the best-positioned players – with capital, talent, and their own data – captured most of the upside. Expect something similar here:

core partners inside the stack,

structured partners with project-level access,

and everyone else downstream.

If your entire strategy is “we’ll just use whatever the best public model is,” you’re implicitly choosing to operate in that outer ring.

4. I believe the real race is shifting from “Which model?” to “Which stack?” Strong base models are becoming a commodity. The durable edge will sit with institutions that:

control valuable data,

can run AI-driven experiment loops at scale, and

have the mandate (and budget) to aim that machinery at big, long-horizon problems.

Genesis is one of the first visible blueprints for that kind of institutional stack. Other countries must build their own versions; a few large private ecosystems will do something similar in different industries.

The exciting/scary part – and the landscape shift

It’s easy to read all of this as threat. There is a power shift underway. But if you care about what gets built, there’s also a lot to be excited about:

Drug discovery and health: AI co-scientists are already proposing repurposing candidates and new targets that hold up in early experiments. With Genesis-scale data and compute behind that, we should see: faster routes to antivirals and vaccines, better prioritization of targets, more systematic use of negative results so we stop wasting time on dead ends.

Energy and climate: National labs already run some of the best fusion, grid, and climate models on the planet. With AI in the loop, we can iterate faster on reactor designs, materials for storage and transmission, and risk models that matter for infrastructure.

Materials and manufacturing: High-throughput simulation + AI + automated synthesis is exactly where you’d expect step-changes in batteries, catalysts, lightweight materials, and process design – with knock-on effects in EVs, devices, and yes, parts of healthcare.

The landscape shift is that these breakthroughs are less likely to come from isolated labs or one heroic company, and more from institutional stacks:

sovereign ones like Genesis,

other national AI-for-science platforms,

and a handful of integrated industry ecosystems built around proprietary data.

The center of gravity for “what’s possible” moves away from individual tools and toward organizations that have wired data + models + experiments + governance into a single, compounding system.

From where I sit, the future doesn’t belong to whoever wins the next GPT vs Claude vs Gemini headline.

It belongs to the organizations – public or private – that learn how to aim AI scientists at the right data, close the loop with reality, and keep that loop running.

Genesis is the U.S. trying to build one of those stacks.

And to be clear: this is my interpretation of what’s been announced so far – based on reading the executive order, DOE material, and the early AI-scientist work, and extrapolating from what I’ve seen in healthcare and data-driven R&D. You may have a different interpretation.