Meta's Chief Scientist Hates LLMs

Programming note: We've just released our thematic report: Mapping the Neocloud Landscape. It’s a 25-page analysis deconstructing the dynamic between hyperscalers and neoclouds, the capital flows, and the financial risks. Access a free copy here.

Is a Cat Smarter Than ChatGPT?

It is probably poor career strategy to tell your boss that the technology he is spending tens of billions of dollars to build is, technically speaking, dumber than a house cat.

Yet this has been the position of Yann LeCun for the last few years. LeCun is one of the "Godfathers of AI," a Turing Award winner, and the Chief AI Scientist at Meta. Mark Zuckerberg is currently reorienting his entire company around Llama, a massive Large Language Model (LLM) designed to compete with OpenAI. And LeCun has been going around telling anyone who will listen that LLMs are a dead end.

As the Wall Street Journal reported this week, this tension has finally reached a breaking point. LeCun has been "sidelined" at Meta and is reportedly talking to investors about leaving to start his own company:

Several years ago, [LeCun] stepped back from managing his AI division at Meta, called FAIR, in favor of a role as an individual contributor doing long-term research.

"I've been not making friends in various corners of Silicon Valley, including at Meta, saying that within three to five years, this [world models, not LLMs] will be the dominant model for AI architectures, and nobody in their right mind would use LLMs of the type that we have today," the 65-year-old said...

LeCun has been talking to associates about creating a startup focused on world models, recruiting colleagues and speaking to investors.

LeCun’s argument is that LLMs like ChatGPT are merely "next-word predictors." They ingest the entire internet and approximate what a human might say. They are, as NYU Professor Gary Marcus puts it, "fundamentally blind to truth." They don't understand that if you drop a ball, it falls. They just know that the word "falls" often follows the word "ball." LeCun likes to say this makes them less intelligent than a cat, which has a functioning internal model of physics and reality.

And so LeCun wants to build "World Models"—systems that learn by observing the physical world (video, sensory data) rather than just reading text.

There are two layers of irony here. The first is corporate: Meta is effectively funding a Chief Scientist who is publicly shorting the company's primary product strategy.

The second layer is technological. LeCun was a pioneer of deep learning, the very technology powering the current boom. By pivoting to "World Models," he is effectively admitting that the current dogma—"scale is all you need," the belief that you just need to feed the machine more data to get superintelligence—has hit a wall. It suggests that the trillions of dollars being poured into LLMs are chasing a technology that has a hard ceiling on reasoning and reliability.

The Physical Wall

LeCun is betting that the "scaling laws" are breaking down intellectually. But there is evidence they are about to break down physically, too.

A recent report from research institute Epoch AI predicts that the industry is heading for a slowdown simply because of the time it takes to build the next level of infrastructure. Their model suggests a brutal rule of thumb: every additional 10x increase in computing scale delays the AI project by about a year. Here's Epoch AI:

...as the required compute grows larger, the time between project initiation and product deployment (i.e. "lead time") lengthens significantly, creating a feedback loop that naturally slows the pace of compute scaling.

In particular, our current best guess is that every additional 10X increase in compute scale lengthens lead times by around a year. For example, OpenAI currently likely has over $15 billion worth of compute, and this compute stock has been growing by around 2.2X each year. At that pace, current trends would predict a trillion dollar cluster around 2030 — but longer lead times would delay this to around 2035.

The argument goes like this: buying $10 million of chips takes a week. Buying $100 billion of chips requires building a power plant, which takes five years. You can't just Amazon Prime a nuclear reactor.

But here is the really interesting nuance. Epoch argues that the AI labs might be able to cheat the math for a while. Even if the construction of new factories slows down, labs can keep their flagship models growing at a breakneck pace for another year or two by performing a sort of internal cannibalization. They can stop using their chips for R&D and small experiments and divert everything into training the one big model (like a GPT-5 or 6).

This creates an illusion of speed: the big models keep getting smarter even as the infrastructure build-out stalls. But it comes at a terrible cost. By diverting chips away from experimentation, you stop discovering the novel algorithms that make the next breakthrough possible. It is a strategy of eating your seed corn to prove you are still growing.

The market is betting trillions on LLM chatbots. The Godfather of AI is betting on a robot that learns like a baby. And the logistics experts are warning that the physical world might not move fast enough to accommodate either of them. Whatever happens, someone is going to lose a lot of money.

More on AI Models

• Hugging Face CEO says we’re in an ‘LLM bubble,’ not an AI bubble (TechCrunch)
• Will Google Gemini Turn Search Into a Dead End for the Web? (ARPU)

On Our Radar

Our Intelligence Desk connects the dots across functions—from GTM to Operations—and delivers intelligence tailored for specific roles. Learn more about our bespoke streams.

The AI Hedge

• The Headline: Microsoft and Nvidia Invest in Anthropic; Anthropic Commits $30 Billion to Azure (Reuters)
• ARPU's Take: Microsoft and Nvidia are using their capital and infrastructure as strategic tools to cultivate a second, at-scale AI pillar, ensuring they aren't beholden to a single, increasingly powerful partner.
• The Go-to-Market Implication: By securing Anthropic for Azure, Microsoft neutralizes a key sales advantage held by AWS (Bedrock) and Google, allowing its sales teams to capture enterprise customers who specifically want Claude or require multi-model redundancy. For Nvidia, this is strategic demand generation; by funding a strong rival to OpenAI, they ensure a competitive market structure where multiple giants must continue aggressively purchasing GPUs to stay ahead, preventing any single customer from gaining leverage over the hardware supplier.

The $100M Tax Agent

• The Headline: Intuit Signs $100 Million Deal to Integrate OpenAI Models and Launch Financial Apps Within ChatGPT (Bloomberg)
• ARPU's Take: This partnership represents the convergence of vertical data and horizontal intelligence. Intuit secures a future-proof interface for its complex financial tools, while OpenAI gains a "killer app" use case that proves ChatGPT can handle high-stakes, accurate financial workflows beyond simple conversation.
• The Product Implication: Intuit is executing a strategy of "intelligence outsourcing" to preserve its core differentiator. Rather than building its own foundation models, Intuit is paying to layer OpenAI's reasoning capabilities on top of its proprietary tax and financial logic. This approach allows vertical SaaS incumbents to rapidly deploy AI features without the capital expenditure of model training, keeping their R&D focus on the application layer where their competitive advantage lies.

P.S. Tracking these kinds of complex, cross-functional signals is what we do. If you have a specific intelligence challenge that goes beyond the headlines, get in touch to design your custom intelligence.

You received this message because you are subscribed to ARPU newsletter. If a friend forwarded you this message, sign up here to get it in your inbox.