Gary Marcus has never been shy about calling out AI hype, but his latest piece points to something more interesting than his usual skepticism. The most telling signal isn’t what he’s saying about generative AI’s limitations. It’s who’s starting to agree with him.

The “massive bets on hyperscaling” that Marcus calls insane aren’t slowing down. Companies are still burning billions on the assumption that if GPT-4 was good and GPT-5 will be better, then GPT-47 will be godlike. The math is simple: more parameters, more compute, more magic. Except the people who built this stuff are starting to say it out loud: this doesn’t actually work the way we sold it.

TechCrunch noticed it too. Their piece on “the haves and have nots of the AI gold rush” captures something that’s been obvious to anyone working in AI but somehow surprising to everyone else: the vibes are off. Even inside the tech industry, people are getting uncomfortable with where this is going.

The problem isn’t that LLMs don’t work

They do work. They’re genuinely useful for a bunch of things. I use them every day. So do you, probably. But “useful” and “the foundation for artificial general intelligence” are very different claims, and the industry has been deliberately blurring that line since 2022.

Marcus keeps pointing toward world models and neurosymbolic AI as alternatives, and honestly, he’s been saying this for years. But what’s changed is the audience. When the money was flowing and the demos were impressive, it was easy to dismiss him as a curmudgeon. Now that we’re two years into the ChatGPT era and the limitations are getting harder to ignore, his arguments sound less like skepticism and more like pattern recognition.

The hyperscaling bet assumes that current architectures will keep improving in a predictable way. Just add more training data, more GPUs, more parameters. But we’re already seeing diminishing returns, and the costs aren’t diminishing at all. Anthropic, OpenAI, and Google are all building models that cost hundreds of millions to train, and the improvements between versions are getting harder to spot in real-world use.

ArXiv’s new policy tells you everything

Here’s a perfect example of where we are: ArXiv just announced they’ll ban researchers for a year if they let AI write their entire papers. Not because the papers are wrong, necessarily. Because they’re slop. Because LLMs are really good at producing text that sounds like a research paper without actually doing research.

This is what happens when you optimize for the appearance of intelligence instead of intelligence itself. The academic publishing system is already a mess, but now we’ve added a technology that’s exceptional at gaming the specific metrics that system relies on. ArXiv isn’t cracking down on AI because they’re Luddites. They’re cracking down because the flood of AI-generated submissions is making the repository unusable.

That’s not a limitation that scales away with GPT-6. That’s a fundamental property of how these systems work. They’re pattern-matching engines that learned patterns from existing text. They’re very good at reproducing those patterns. That’s the feature, not a bug we’ll eventually fix.

The people at the top know this

What makes this moment interesting is that the gap between public messaging and private knowledge is getting impossible to maintain. OpenAI’s latest shakeup, with Greg Brockman taking over product strategy, looks like the kind of move you make when you need to figure out what you’re actually selling. Because right now, the answer isn’t clear.

The haves and have-nots dynamic that TechCrunch wrote about isn’t just about who has access to compute. It’s about who understands the technology well enough to know what it can’t do. The have-nots aren’t the people without GPUs. They’re the everyone else who’s been sold a vision of AI that the people building it know isn’t real.

Marcus’s interviews apparently dig into world models and neurosymbolic approaches, which are both harder to explain and harder to scale than “make the neural net bigger.” That’s probably why they haven’t attracted the same investment. But it’s also probably why they’re more likely to lead somewhere useful.

The hyperscaling bet isn’t insane because it won’t produce better models. It’s insane because it’s already produced models that are good enough to reveal their own limitations, and the response has been to ignore those limitations and build bigger versions of the same thing.

The vibes are off because reality is catching up to the pitch deck. And the smartest people in the room are starting to say so.