Yan LeCun says today’s AI is “not smart”. His startup chases something more flexible

Yan LeCun, one of the best-known names in AI research, has a blunt critique of where the field is right now. In the BBC’s framing, the idea is that today’s AI is “not smart,” which raises an obvious follow-up: if the tech is not smart, what should come next?

LeCun’s answer is not a slogan. He has a start-up that is developing a more flexible AI system. That matters because “flexible” is not just a marketing word. It points toward AI that can handle more kinds of situations, adjust better when conditions shift, and potentially behave more like a generalist instead of a specialist trained for one narrow pattern.

To understand why this is a big deal, zoom out for a second. The recent AI boom has been powered by systems that can generate text, recognize patterns, and solve tasks remarkably well within the bounds of how they were trained. But those strengths can come with constraints. When a model meets a new distribution, a new format, or a new context, performance can wobble. For business leaders, that wobble translates into operational risk: higher costs when outputs need more checking, uncertainty about reliability, and a constant need for monitoring.

That is where the “flexible AI system” direction changes the conversation. If LeCun is arguing that today’s systems are “not smart,” then “flexible” becomes a proxy for a different target metric. Instead of only asking how good a system is on benchmarks, you start asking how well it adapts across tasks and environments, how quickly it can incorporate new information pathways, and how effectively it can generalize. Even without getting lost in the technical details, the strategic implication is clear: firms that depend on AI will prefer systems that are easier to steer and less brittle when the real world behaves like the real world.

There is also a governance angle here. AI development is increasingly shaped by regulators and policymakers, even when the underlying technology is moving fast. Regulators tend to focus on accountability, safety, and transparency, not just raw capability. Systems that are harder to understand or control can become harder to deploy in regulated spaces like finance, healthcare, or critical infrastructure. If the “next” phase of AI emphasizes flexibility in a way that also improves control and robustness, it could make it easier for organizations to justify adoption, pass internal risk reviews, and reduce friction with compliance teams.

Board dynamics matter too. Many boards are facing a common dilemma: they want AI benefits now, but they are also responsible for risk. A credible signal that respected researchers are pushing for a different kind of system can influence how directors allocate attention. It can shift the board’s questions from “Are we using the newest model?” to “Does our AI strategy reduce failure modes, lock-in risk, and dependency on one narrow class of capability?” A move toward flexibility can also change procurement conversations, because it reframes what “performance” means and what kind of vendor assurances executives should demand.

For investors and operators, this is also about timing. Capability curves in AI can accelerate quickly, but the industry does not only reward what improves metrics. It rewards what integrates into products and processes reliably. A more flexible system is the kind of foundation that could support a wider range of use cases without constant retooling. That can lower total cost of ownership and shorten the distance between a pilot and something that can run daily operations.

Finally, the second-order implication is cultural. When a leading researcher publicly challenges the field with the claim that current AI is “not smart,” it invites a reset in how teams interpret success. The question becomes: are you optimizing for the appearance of intelligence, or are you building the kind of adaptability that intelligence requires? The BBC note about LeCun’s startup suggests he is betting on the latter.

So the stake for peers in similar roles is practical. If “not smart” is a credible critique, then the next competitive advantage may not come from bigger models alone. It may come from architectures and approaches that produce behavior that is more consistent across changing conditions, and that can be deployed with fewer operational surprises. In that world, executives who plan for flexibility, governance, and robustness will likely be better positioned than those who treat AI as a one-time upgrade.