All-In

Jensen Huang on why the AI doom predictions are made up

Jensen Huang· Founder and CEO of Nvidia at Nvidia
·~47 min·English·All-In Podcast
AI SafetyOpen SourceAI InfrastructureGPUBusiness Strategy
TL;DR

Jensen Huang argues the AI doom predictions are not grounded in science, that real risk sits only where the compute does, and that America wins AI the way it won electricity: by exploiting it broadly on open models, not by slowing down.

01The Core Claim

It's Not Based on Science

<strong>Huang takes safety seriously but calls the doom forecasts made up</strong> — alarming predictions dressed as science, when the actual record proves the opposite.

It's not based on science. It's not based on research. Everything that's based on science and research proves otherwise.

— Jensen Huang, All-In
Key Insight
Huang is careful to split two things the doom essay fused together: a whistleblower raising a real internal-control concern (which he says deserves to be taken seriously) and a probability-of-extinction forecast (which he says is not science). Treating the second with the authority of the first is the sleight of hand he is objecting to.
02The Track Record

A Ledger of Wrong Predictions

<strong>Every alarming near-term forecast has already failed</strong> — radiologists gone, code fully automated, entry jobs wiped out — so Huang wants the misses kept on the record.

there was a prediction that in 5 years time radiology will be completely taken over by artificial intelligence and there'll be no radiologists in the world. That has proven to be exactly the opposite. We need more radiologists than ever in the world.

— Jensen Huang, All-In
Key Insight
The radiology case separates automating a task from deleting a job: Huang says AI took over scan reading, and yet the world needs more radiologists than ever. That gap — a task inside a job gets automated while the job itself grows — is the pattern that keeps breaking the headline forecasts.
03Where Risk Lives

Danger Follows the Compute

<strong>The real risk sits only at the frontier labs, because they alone have the compute</strong> — so regulation should target the few who can actually cause harm, not students or startups.

you could look across the planet and everybody won't have enough compute with the exception of the frontier labs.

— Jensen Huang, All-In
Key Insight
This quietly rewrites the regulation debate. If harm requires frontier-scale compute, then broad rules aimed at every developer are aimed at the wrong target — the sensible surface area to govern is small, named, and already known. Regulation, he says, should solve actual problems, and the actual problems have all come from the labs.
04Recursive Self-Improvement

RSI All Day — the Release Gate Still Holds

<strong>Models can improve themselves internally, but released products still pass evals and regression tests</strong> — Huang argues that ordinary engineering controls separate internal improvement from external release.

you could RSI all day long inside your company, but when you release a product, you've got to evaluate it, don't you? You have to test it again, don't you? You have to make sure that there's no regression, right?

— Jensen Huang, All-In
Key Insight
He reframes RSI from a science-fiction spiral into an ordinary engineering pipeline. The fear assumes improvement and deployment are the same event; in practice they are separated by a testing gate that already exists, and gets stronger as labs move from research to engineering.
05Open vs Closed

The World Needs Both

<strong>Closed models are bottled water, open models are the tap</strong> — you use the right one per job, and 80% of the new wave of AI startups is built on open.

The world needs both closed models and open models.

— Jensen Huang, All-In
Key Insight
The bottled-water line does more than charm. It reframes closed models as a premium convenience rather than the default, and pins the base layer of the whole industry on open weights — which is why the 80% figure matters to Nvidia: open models are what let every startup, not just a few labs, become a customer.
06China and Open Source

Once You Download It, It's Yours

<strong>China contributes the most open source today, but forking it makes it yours</strong> — the same way America already runs on Linux and Kubernetes it did not write.

We download Linux. We download Kubernetes. We download all the software. A lot of it has been touched by Chinese. And once you download it, it's yours. We fork it. We improve it. We make it ours.

— Jensen Huang, All-In
Key Insight
This is the counter to the reflex that Chinese open models are a security threat to avoid. Huang's point is that open weights carry no allegiance — the moment you fork them they are under your control — so refusing to use them mostly forfeits a free head start to the countries that will.
07Winning AI

The Race Is Who Exploits It Best

<strong>America did not invent the last industrial revolution — it exploited it best</strong> — and Huang wants AI won the same way: by diffusing it across every company, not hoarding the frontier.

The last industrial revolution came from Europe. But we exploited it. We took advantage of it socially better than anybody else in the world. Look how it turned out for us.

— Jensen Huang, All-In
Key Insight
This is the strategic core beneath all the safety talk. If winning is about diffusion rather than invention, then fear that slows domestic adoption is the actual threat to leadership — and open models, cheap compute and broad deployment become national-competitiveness moves, not just business ones.
08Superintelligence

We're Already There — Narrowly

<strong>Superintelligence is not a distant milestone — in narrow domains it already beats humans</strong> — a self-driving car at a tenth of the human accident rate is, by Huang's definition, superhuman.

But Jason, I think we're there, too.

— Jensen Huang, All-In
Key Insight
Huang defuses the whole AGI-versus-superintelligence argument by refusing its terms. He does not want a car that can also make an omelette; he wants it to drive, and on that one axis it is already superhuman. Redefining superintelligence as narrow superhuman performance makes it a shipping fact instead of a looming threat.