Doom Debates: Inside the Doom Train Live Debate at Manifest 2026

**Liron Shapira** (0:01)
So we've got something really special to talk about today—a live Doom Debates session from Manifest 2026 Producer Aie led a room full of forecasters, rationalists, and AI researchers through what they call the Doom Train framework. The Doom Train—I love that name. And the central question is pretty unsettling, right? Does anyone want to get off at this stop? Liron Shapira opened by framing it as a special live debate, then Aie laid out the train's core claims. Right. And they asked the room for their p-doom, meaning the probability of human extinction or severe permanent disempowerment by 2050 The crowd was all over the map—low, mid, and high doom estimates. It set up this really intense but good-faith argument. The first stops challenged whether AGI is even imminent. One attendee said progress has been slower than people think because transformers and scaling laws aren't actually new, but someone else pushed back, saying the recent chatbot era makes old timelines less useful. That's interesting, because the other side kept pressing that current models are still improving fast, and that scale alone could become transformative. Then the conversation shifted to physical risk. Yeah, one claim was that AI is just software in data centers. But others pointed to cooling water, labor automation, and military uses like drones. The point wasn't that AI already has a body, but that software can still reshape the physical world in dangerous ways.
Building on that, there's this deeper philosophical stop—does intelligence make a system moral?
One participant argued smarter systems may be more ethical, but another noted how hard that is to prove. And get this—a self-described orthogonality denier suggested an ASI might want to appear moral to older intelligences. Why? Because destroying its creators could look like a bad signal to the wider universe. That's wild. Aie and the room also weighed whether we already have a safe AI development process. Some argued humans can copy, distill, and pit models against each other in councils to catch deception. But the counterargument was blunt—competition between humans may overwhelm caution. The U.S. and China are in a race, so safety could vanish under pressure. The debate really sharpened around speed. Can capabilities rise at a manageable pace, or does one bad actor change everything? Several participants said the pace is still human-controlled, while others warned that if one person with bad intentions pushes hard enough, doom could come very quickly.
Another stop asked whether AI will try to conquer the universe. Some say superintelligence may be too alien to care about us, or may even leave for something beyond human imagination. But others argue any intelligent system will still optimize for power, information, and control—instrumental convergence makes human survival far from guaranteed.
Finally, there's super alignment—the hope that a smarter system can still obey a lesser one. Some think it's tractable, even if risky. If that works, perhaps peace follows.
If it fails, unaligned ASI may not spare anyone.
The closing note was a warning against certainty, but also against complacency. As Aie put it, "AI doomerism is bad epistemology" only if the evidence is ignored.