LeCun's Inconvenient Point: Useful Isn't the Same as Intelligent

Yann LeCun says today's AI is genuinely useful despite being nowhere near human intelligence — and that the path beyond LLMs runs through world models. Why that reframes how builders, buyers, and creators should bet in 2026.

Yann LeCun spent May 21 in a running public argument over a claim that should give anyone budgeting for AI in 2026 pause: today's models are genuinely useful while being, by his own account, "nowhere near human intelligence." The gap between what researchers debate and what practitioners actually pay for has rarely been this visible.

Replying to economist Noah Smith, LeCun argued that current systems compensate for their lack of common sense, reasoning, and planning with sheer accumulated knowledge. He was blunt about the ceiling, asking whether anyone's LLM has cleaned their house, driven their car, fixed their toilets, or built their IKEA furniture. Scaling today's architectures won't close that gap, he argued; whatever reaches general intelligence "won't be an LLM" but will lean on non-generative world models like JEPA that predict the next state of the world rather than the next token.

He's not arguing in a vacuum. LeCun likened the current moment to AlphaGo in 2016, Watson in 2011, and Deep Blue in 1997 — each a real milestone mistaken at the time for general intelligence. The counterweight is that models keep getting better at the narrow things they do well: Alibaba's Qwen 3.7 Max-Preview ranked 13th for text on LM Arena as of May 2026, tuned for logic and problem-solving over creative writing.

For builders, the sharper signal is where the value is actually moving. swyx argues that deep research "has been ~dead since o3" — a thoughtless prompt yielding a long report nobody reads loses to a tight read-think-ask loop. That echoes LeCun in a different key: the payoff isn't a smarter one-shot answer, it's keeping a human in the loop.

Vendors are chasing exactly that. xAI just made Grok usable inside the opencode editor, while swyx noted that codex now lets you "pause and adjust mid goal" after over-eager autonomy frustrated users — the frontier this week was better controls, not a bigger brain.

For buyers, stop pricing models as if benchmark rank converts to judgment — a higher LM Arena slot buys a faster, better-read assistant, not a colleague who can be trusted to plan. For creators, note what Qwen's team chose to optimize: logic and problem-solving over creative writing. When a flagship model trades prose quality for reasoning scores, it's a tell about where the labs believe the budgets are.

The bottom line: Treat 2026's models as sharp, well-read interns — design the interaction loop, not just the prompt.

Sources