At 10 a.m. Pacific today, Geoffrey Hinton, Fei-Fei Li, and Andrew Ng walked onto the main stage at The Venetian in Las Vegas. It is the rarest lineup in the field: the man who quit Google to warn that AI could end the species, the researcher who thinks that warning is actively harmful, and the scientist who believes both of them are arguing about the wrong thing. Ai4 2026 built its whole three-day program around the moment, and roughly 12,000 people from more than 90 countries came to watch.
Here is what almost none of the coverage will tell you. The 12,000 people in that hall did not fly to Las Vegas to settle whether superintelligence kills us in 2050. Most of them run infrastructure, risk, or operations at companies that already bought AI and cannot get it into production. The keynote was relitigating 2023’s question. The rest of the conference had moved on to a harder one.
Three legends, three incompatible answers
The panel is worth taking seriously because the disagreement is real, not staged.
Hinton has spent three years putting a number on his fear. He now estimates a 10 to 20 percent chance that AI drives humanity to extinction within roughly three decades, up from the 10 percent he cited a year earlier. His argument is not about job loss or deepfakes. It is about control: “How many examples do you know of a more intelligent thing being controlled by a less intelligent thing?” He wants government regulation and far more safety research, and he wants it before the systems get smarter than we are.
Ng thinks that framing is the problem. He told the Australian Financial Review that the “bad idea that AI could make us go extinct” is merging with another bad idea: that the way to make AI safer is to load the industry with licensing requirements. In his reading, extinction talk is not just wrong, it is a lever that incumbents pull to raise the cost of building anything. Build better AI, he argues, and stop treating a research tool like a loaded weapon.
Fei-Fei Li is the tell that the debate itself has aged. She is barely engaged with the doom-versus-hype axis anymore. Her attention is on spatial intelligence, the idea that language models have run out of the easy dimension and the next frontier is machines that understand three-dimensional space and physics. In June she published a taxonomy sorting every system that calls itself a “world model” into three functions: renderer, simulator, planner. Her company, World Labs, raised a billion dollars in February from AMD, Nvidia, Autodesk, and others to build exactly that. When one of the three most famous people on the stage has quietly stopped participating in the panel’s premise, that is information.
Three pioneers, three answers that do not fit together. That is not a failure of the panel. It is an accurate snapshot of where the intellectual leadership of the field actually is: unresolved, and increasingly disconnected from the people writing the checks.
What the floor was actually asking
Walk off the keynote stage and into the tracks, and the mood changes completely. Ai4 runs 20 industry and technical tracks this year, and the center of gravity is not “what is an agent” or “will it kill us.” It is deployment, governance, and failure modes. The banking track, the healthcare track, the manufacturing and energy and national-security tracks: they are full of teams that have already run pilots and are trying to figure out why the pilots did not survive contact with a real business.
The data explains the room. Gartner projects that more than 40 percent of agentic AI projects will be canceled by the end of 2027, killed by escalating costs, unclear business value, or inadequate risk controls. The same analysts estimate that of the thousands of vendors selling “agentic AI,” only about 130 are real; the rest is “agent washing,” older chatbots and RPA scripts wearing a new label. And the projects that do work tend to stall in the same place: integration, data access, and accountability, the unglamorous plumbing that nobody demos.
I have watched this pattern from the inside. In my day job running IT operations at a telecom, the model was never the bottleneck. The bottleneck was the fourteen systems the agent had to touch, the permissions nobody wanted to own, and the question of who answers for it when it does something expensive at 2 a.m. A frontier model that scores 80 on a coding benchmark does not help you if it cannot reach your ticketing system, or if your compliance team cannot get a straight answer about what it did and why. The gap between a working demo and a production deployment is not intelligence. It is engineering, governance, and organizational nerve.
That is the conversation happening in the tracks. It is a world away from the one happening on the keynote stage.
Why the split matters
It would be easy to read this as “ignore the doomers, ship product.” That is not the point, and it is not what a practitioner should take away.
The point is that the AI industry is now running two separate conversations that rarely touch, and they operate on different clocks. The existential debate runs on a 30-year horizon and is fundamentally about policy: who gets to build, under what oversight, with what liability. That debate is real, and it is already shaping the market in concrete ways. The U.S. government spent this summer gating frontier releases, suspending one model for three weeks and restricting another to vetted partners. Anthropic’s own co-founder called for a pause while his company kept shipping. Hinton’s 20 percent is not an abstraction; it is the emotional fuel behind the licensing regimes Ng is fighting.
The deployment conversation runs on a 12-month horizon and is about whether any of this pays for itself before the budget cycle closes. That is the conversation that decides whether your company keeps funding AI in 2027. And it is starving for attention precisely because the celebrity debate absorbs so much of the oxygen. PwC found last year that 74 percent of AI’s measured value was concentrated in a fifth of companies. The winners are not the ones with the strongest opinion about superintelligence. They are the ones who solved integration and accountability while everyone else argued.
The danger of the two-clock problem is that a leadership team can spend a year with a well-informed position on existential risk and no working agent in production. The existential question feels more important, so it wins the meeting. Then the budget review arrives and there is nothing to show.
The question worth taking home from Las Vegas
If you are a buyer, the useful takeaway from Ai4 is not which pioneer is right. It is that the field’s own most-watched panel could not agree, which means nobody is going to hand you certainty about where this ends. You have to operate anyway.
So operate on the clock you can actually affect. Before the next pilot, answer the boring questions the tracks are obsessed with. What systems does the agent need to reach, and who owns those permissions? How do you know what it did, and can you explain that to an auditor? What is the cost ceiling, and what happens when it hits it? Which decisions require a human in the loop, and is that enforced or aspirational? These are the questions that separated the 40 percent that get canceled from the fifth that captured the value, and they are answerable this quarter.
Fei-Fei Li had the most quietly radical position on that stage, and it was not about risk at all. It was that the interesting frontier has already moved to a different dimension. She is probably right that the next wave is spatial, and that today’s text-first agents are a way station rather than a destination. But that is a five-year bet. The thing in front of you is the same thing it was before you flew to Las Vegas: a capable model, a stack it cannot quite reach, and a governance layer you have not built yet.
Hinton and Ng will keep arguing, and they should; the stakes they are describing are real. But the outcome that determines whether AI works for your organization this year was never going to be decided on that stage. It gets decided in the handoffs between your agents, in the context you feed them, and in whether you measured anything real before you trusted the output. That is the harder question. It is also the one you can actually answer.