We are all secretly waiting for the machines to wake up. Don’t pretend you aren’t. Every time a new model writes a flawless poem or debugs a script in seconds, we gasp and think, this is it, the dawn of AGI. The tech industry is pouring billions into a single, desperate strategy: shovel more text, more parameters, and more compute into the furnace until consciousness spontaneously combusts.
But what if we’re scaling a ladder that’s leaning against the wrong wall?
A recent paper bluntly titled “LLMs Can’t Jump” just dragged the entire AI hype machine back to reality. The core finding is simple, humbling, and deeply unsettling for anyone building the future. Language is not reality; it’s a compressed file of reality, and AI only knows how to unzip it, not how to film it.
You can feed a model a trillion words describing the sensation of falling. It can generate a breathtaking paragraph about the wind rushing past your ears and the lurch in your stomach. But it will never actually know what it feels like to hit the ground. You can memorize the encyclopedia entry on gravity, but you don’t truly know it until you’ve scraped your knee on the pavement.
Right now, the loudest voices in tech argue that the only bottleneck to artificial general intelligence is data and compute. If we just scrape the entire internet and build bigger clusters, the machine will finally understand. But the “LLMs Can’t Jump” analysis exposes the flaw in this religion. The bottleneck isn’t silicon or syntax. It’s the absence of a physical, time-bound existence.
As one commenter on the research perfectly put it: there seems to be no replacement for being born in the world and spending a couple of decades learning from experience. Language is a vital step in the process, but it’s just a step. We are treating the step as the destination.
Think about how a child learns. They don’t read a textbook on thermodynamics before touching a hot stove. They burn their hand, cry, and then they know heat. Their understanding is anchored in flesh, bone, and consequence. An LLM has no flesh to burn. It lives in a frictionless vacuum of pure syntax, manipulating symbols with terrifying fluency while possessing zero grounding.
This should reframe your expectations entirely. If you are investing in, building, or panicking about AI, you need to internalize this: these systems are not autonomous minds. They are incredibly powerful linguistic calculators. LLMs are the ultimate armchair athletes—they can describe the game flawlessly, but they will never step onto the field.
So stop treating language models like dormant gods waiting for enough data to ascend. They are tools. Extraordinary, world-altering tools, but tools nonetheless. Until we figure out how to give an algorithm a body, a timeline, and the messy, painful experience of actually living, the machines won’t jump. They’ll just keep talking about what it’s like to fly.
FAQ
Q: If an AI acts like it understands the world, does it matter if it lacks physical experience?
A: Yes, because without grounding, it has no concept of consequence. It can describe a disaster perfectly but lacks the embodied intuition to anticipate physical or temporal friction, making it dangerous to trust with autonomous real-world decisions.
Q: What does this mean for AI developers and investors?
A: Stop pouring infinite capital into just scaling text data and parameters. The returns on linguistic fluency are flattening. The real frontier isn't bigger models; it's figuring out how to ground AI in multimodal, real-time, physical feedback loops.
Q: Are LLMs a dead end then?
A: Not at all. They are a massive leap forward in manipulating human language, which is incredibly useful. But they are a stepping stone, not the final destination. Believing they will magically wake up if we just add more text is a technological delusion.