NVDA
The AI self-eating snake problem
Michael Burry just took another swing at the AI hype train, arguing that models trained on more and more AI-generated content could end up amplifying errors instead of getting smarter. In his telling, the internet’s next phase may look less like a knowledge explosion and more like a photocopier copying a photocopier until the corners get fuzzy.
Why this matters to your portfolio
The big idea here is “model collapse” — the risk that synthetic data starts crowding out real human-generated data, and the result is a weaker model, not a better one. Researchers cited in the piece say repeated training on AI-made data can cause systems to lose the tails of the original distribution, which is science-speak for: the weird, rare, important stuff gets washed out.
That’s a problem for anyone selling the dream that larger models automatically march toward AGI. Burry’s take is basically: if the input gets sloppy, the output gets sloppy too, and no amount of GPU confetti fixes that.
Nvidia catches a stray
Nvidia shows up here because Jensen Huang has been one of the loudest voices arguing AI is still in the early innings — even as skeptics like Burry say the whole AGI story may be more mirage than destination. For Nvidia investors, the article is mostly sentiment noise, not a direct fundamental update, but it’s another reminder that the AI trade now lives in a tug-of-war between massive spending and louder questions about what all that spending ultimately buys.
Big picture: this isn’t a new chip shipment, earnings beat, or regulatory bombshell. It’s a sharp reminder that the AI boom still has a giant philosophical hole in the middle of it — and markets hate when the math gets fuzzy.