Steven Bartlett asked the studio AI to fight back. His guest, physicist Brian Greene, had been explaining why a system that keeps improving might eventually approach a ceiling rather than become indefinitely more capable. Bartlett wanted a rebuttal.
The assistant argued that AI could reinvest each gain into overcoming whatever slowed it down, eventually redesigning its own architecture—the underlying structure through which it learns and solves problems.
Greene's answer was that the starting architecture might itself limit the ability to discover such a redesign. Wasn't that at least a logical possibility?
“Yeah, absolutely,” the assistant replied.
“Okay, stop,” Greene said. “I'm done.”
The exchange on The Diary of a CEO turned on a distinction that Greene kept returning to: faster rounds of improvement do not necessarily mean indefinitely accelerating capability. He was not claiming to have found a ceiling. He was challenging the assumption that AI would inevitably break through one.
Faster improvement, or an intelligence explosion?
Bartlett's optimism began with longevity. If AI capability keeps growing exponentially—gaining by a similar proportion over successive periods, so the curve becomes steeper—he thought discoveries that dramatically extend human life might arrive within his lifetime. He pointed to recent mathematical progress and the enormous sums being invested in AI.
Greene thought lifespans of hundreds of years were possible, but doubted they would arrive in his own lifetime. His hesitation was partly about whether today's methods could deliver the necessary scientific breakthroughs. More money, more training data and larger data centres did not, in his view, guarantee that the curve would keep climbing at the same rate.
He focused on large language models: systems that learn patterns from vast collections of text and use that training to generate responses. Their ability to produce coherent, insightful answers did not settle whether the same approach could improve without limit.
Bartlett offered recursive self-improvement as the route beyond those constraints. In this proposed feedback loop, an AI improves the software, hardware or training methods used to build its successor. That more capable successor then helps build an even better system. An “intelligence explosion” is the stronger possibility: those gains compound so rapidly that machines soon far exceed human intellectual abilities.
The argument long predates today's chatbots. In his 1965 essay, statistician I. J. Good reasoned that an ultraintelligent machine would also excel at designing machines, enabling further improvements. He made the benefit conditional on retaining human control. His own forecast placed such a machine within the twentieth century; the historical argument and present-day predictions are separate things.
Bartlett relayed industry forecasts pointing to milestones between 2027 and 2029, including claims about fully self-improving systems. Greene remained sceptical about that short horizon, while accepting that machines could iterate much faster than biological evolution.
A current version of the acceleration argument appears in Dario Amodei's January 2026 essay, The Adolescence of Technology. The Anthropic chief argues that AI-assisted coding could accelerate development of subsequent AI systems, potentially leading to autonomous development of the next generation within one or two years. He also allows that it could take considerably longer, and notes that physical experiments and other real-world processes impose waiting times.
Greene's alternative was a system that continues improving but makes progressively smaller gains, approaching a maximum compatible with its design. “Which curve is the right one?” he asked.
The studio assistant offered two ways around that problem. First, it could search combinations of algorithms, data and hardware beyond those humans had considered. Second, it might change its own fundamental design rather than merely optimise within it.
“Maybe,” Greene replied. If the system's ability to invent those changes was constrained by the very design it needed to escape, the proposed route around the ceiling might not work. The assistant acknowledged that a plateau was possible, while pointing out that its location remained unknown.
What the cup explains
Later, Greene brought the architectural question down to the table in front of him. Consider a cup: a person can anticipate what sort of push would move it without spilling and what would make it tip.
That ability illustrates a world model—an internal representation used to predict how the world behaves and what an action will do. In Greene's account of AI researcher Yann LeCun's position, future systems need more than patterns in words; they need this kind of predictive understanding of physical reality.
A language model combined with a world model, or another hybrid design, might be the next step, Greene suggested. He also offered the alternative he thought Demis Hassabis and others might propose: perhaps self-improving language models would develop the missing capability themselves, without people having to add it separately.
The studio assistant was then asked what would happen if the cup were pushed. It answered that the mug would slide across the table and, depending on the force, might hit a plate or go over the edge. Greene interpreted the answer as drawing on written descriptions of physics and falling objects.
“Isn't that what I'm doing as well?” Bartlett asked.
“I don't know. How does human intelligence work?” Greene replied.
Greene thought humans inherit a predisposition to develop physical intuition, rather than beginning as complete blank slates. He was not suggesting that a parent's knowledge of Newton's laws passes directly to a child. His evolutionary explanation was that ancestors better able to learn how rocks and spears moved were more likely to get their next meal and survive. What could be inherited was a tendency to turn experience into useful expectations about the world.
He agreed that successive AI systems can inherit earlier achievements too. The unresolved question was whether that inheritance, combined with rapid iteration, would supply everything needed for further advances.
A research example beyond words
Meta's V-JEPA 2, announced in June 2025, provides a concrete example of research into predictive world models. It learns from video, converting what it sees into internal representations and predicting how those representations will change, rather than reconstructing every future pixel.
Meta described pretraining on more than a million hours of video, followed by training that links actions to outcomes using approximately 62 hours of robot data. In demonstrations, robots compared predicted outcomes with a goal image, choosing and repeatedly revising actions to reach for, pick up and place objects.
Those demonstrations concerned specific physical tasks, not unrestricted understanding of the world. Meta noted that the model operates at a single time scale, whereas complex activities require planning across several. Its accompanying IntPhys 2 benchmark also exposed difficulties in physical reasoning: videos begin alike, but one later violates physics, and the model must identify the violation. Meta reported contemporary video models performing near chance on that test.
The work gives substance to the architectural alternative Greene described, without settling whether language models could eventually acquire comparable abilities through another route.
Why a possible ceiling offers little reassurance
Greene also allowed the opposite possibility: there may be no ceiling of the kind he envisaged. That prospect was “both exciting and frightening,” he said, because humanity might not be prepared for it.
He worried about systems pursuing objectives that do not serve human interests, and about sufficiently capable systems resisting attempts to stop them. In Greene's account, LeCun's response to a malfunctioning system was to pull the plug. Greene questioned whether an advanced system might prevent people from doing so.
There was also the question of who held the controls. Bartlett's list of industry leaders represented a handful of people and companies. Greene worried about concentrating such consequential power in so few hands, quite apart from what the machines themselves might do.
When Bartlett raised the danger of improvement outpacing human evaluation, Greene agreed. If a new version arrives before people can assess the previous changes, problems could accumulate faster than the safety process can catch them.
He reached for a pandemic analogy. An outbreak begins slowly, appearing as an occasional news story. By the time it commands widespread attention, rapid growth may have made it much harder to contain. Greene feared a similar delay in recognising the consequences of accelerating AI: the system need not be evil to act in ways humans cannot anticipate or manage. Waiting until the danger is obvious could leave too little time to intervene.