During a panel discussion in September 2026, Nvidia CEO Jensen Huang took a live phone call from President Trump on stage, and the two of them used the moment to wave off concerns about AI danger as a “hoax.” Trump and Huang have since appeared together at a White House event, the kind of joint appearance that tells you the alliance isn’t a one-off.
I review AI tools for a living. I spend my days watching agents misread a config file, hallucinate an API endpoint that never existed, and confidently delete the wrong branch. So when the most powerful chipmaker on earth and the President of the United States agree that AI risk is imaginary, my reaction isn’t outrage. It’s confusion about which risk they’re talking about.
Two different arguments wearing the same jacket
There are at least two separate debates happening under the “AI safety” umbrella, and the hoax framing smashes them together.
The first is the existential one. Will a sufficiently capable model decide humans are an obstacle? That argument lives in papers, forecasts, and long podcast episodes. Reasonable people disagree, and nobody has a decisive test to settle it.
The second is the boring one, and it’s the only one I can actually measure. Does this agent do what the docs claim? Does it handle credentials safely? Does it fail loudly or quietly? Can you audit what it did after the fact? That’s not speculation. That’s a Tuesday.
When Huang says AI danger is overstated, I suspect he means the first debate. Fair enough. But the word “hoax” lands on both, and the second one is where most users actually get hurt. Nobody’s company got wiped out by a superintelligence last quarter. Plenty of teams shipped bugs because an agent wrote code nobody reviewed.
Consider who’s talking
Huang’s position contrasts with tech leaders calling for regulation, and I’d rather be honest about the incentives on both sides than pretend one camp is pure.
Nvidia sells the hardware that every frontier model trains on. Faster development means more chips. That’s not a scandal, it’s a business model, but it does mean his read on risk arrives with a thumb on the scale. Meanwhile, some of the loudest regulation advocates run labs that would benefit enormously from rules only well-funded incumbents can satisfy. Safety theater makes a decent moat.
So I don’t take either side’s word on trust. I take the tools, run them, and write down what breaks.
What I’ve actually found
After enough hours with agent frameworks and assistant tools, a pattern shows up that has nothing to do with doom scenarios:
- Permissions are usually too broad. Tools ask for full repo or full inbox access when they need one directory or one thread. The defaults favor convenience.
- Failure is often silent. An agent that half-completes a task and reports success is worse than one that crashes. Crashes get noticed.
- Audit trails are thin. Many products can’t tell you exactly what actions were taken in what order. That matters when you’re explaining an incident to a client.
- Capability claims outrun reality. “Autonomous” usually means “autonomous until it isn’t, and you won’t be told when.”
None of those need a regulator to fix. They need product teams who care and users who push back. But calling the whole category of concern a hoax makes pushing back sound hysterical, and that’s the part that bugs me. It’s not a technical claim. It’s a vibe, and vibes are contagious.
Speed is real and so is the cleanup
I’ll give the acceleration camp its due. The progress in model quality over the last stretch has been genuinely useful. Tools that were toys eighteen months ago now save me real hours. Moving fast produced that. A world where every model release waited on a committee would not have produced it.
But the tools that impress me most aren’t the fastest. They’re the ones with clear boundaries, good logs, and a narrow, honest description of what they do. Solid beats flashy every time in my testing notes. Those two things, speed and care, are not actually enemies. Framing the debate as doomers versus builders makes them look like enemies, which serves the people selling both narratives and nobody who has to ship on Friday.
What to do with this
You don’t have to pick a team in a political argument to be a careful operator. Read the permission scopes before you approve them. Run new agents on throwaway repos first. Keep a human in the loop for anything that touches money, customer data, or production. Assume the marketing page is optimistic by roughly one order of magnitude.
Trump and Huang are talking about civilizational risk from a stage. I’m talking about whether your coding agent just force-pushed. Both conversations are legitimate. Only one of them is going to show up in your logs this week, and calling it a hoax won’t make it stop.
đź•’ Published: