LLMs are a Dead End

Añadido: 2026-05-05

[00:00:00]In November 2025, a researcher walked into Mark Zuckerberg's office and quit.

[00:00:05]It wasn't any researcher, but it was one of the guys awarded the 2018 Turing Award for foundational breakthroughs in modern deep learning.

[00:00:14]His name is Jan LeCun, and if anyone who knows about AI, it's him.

[00:00:20]4 months later, he'd raised more than a billion dollars to build something new, something that is designed to prove the entire industry wrong.

[00:00:28]Jan LeCun's pitch to investors was that large language models are dead.

[00:00:33]That's not to say they're not useful, but in terms of human-level intelligence or superintelligence, LLMs are not the way forward.

[00:00:40]And I agree with him 100%.

[00:00:42]What we've seen over the last few years is big tech throwing more and more compute at LLMs and getting less and less improvement.

[00:00:50]At the end of the day, these things are just token prediction machines, and there's no chance it will ever turn into AGI.

[00:00:57]This stuff still can't get the basics right.

[00:01:00]Now, like me on this channel, Jan LeCun is not saying that AI is useless. I use AI every day, and I use it extensively.

[00:01:08]And Jan LeCun is the same. He's very positive about Claude and ChatGPT.

[00:01:13]He's simply saying that we've exhausted all the output we're going to get from LLMs, and if our goal is human-level level intelligence, then scaling LLMs will never get you there. It won't be solved with more parameters or more GPUs or more putting data centers on the moon. None of that is going to fix it.

[00:01:29]The architecture itself has hit a wall.

[00:01:32]Jan LeCun's critique comes down to four capabilities that LLMs structurally cannot have. Number one, there's no model of the physical world.

[00:01:40]An LLM can describe the world in multiple languages, but it's never been to Paris or seen a waterfall or or touched a puppy.

[00:01:48]Unlike humans, there's no intuitive understanding of gravity, friction, heat, or hunger. These are just tokens which have no innate understanding.

[00:01:57]And a 4-month-old baby has a much better physics engine than GPT-5.

[00:02:02]And this means that human intelligence is much, much more than just predicting the next word in a sentence.

[00:02:08]Number two, no persistent memory.

[00:02:11]You may not realize this, but every time you interact with AI, it starts from zero.

[00:02:16]Behind the scenes, your favorite chatbot is managing context history, and it's getting injected into the conversation every time, which is why the every conversation starts with zero.

[00:02:26]In simple terms, context windows is like a goldfish with a really good notebook.

[00:02:31]But a context memory, which is what humans have, is like an elephant who never forgets being mishandled.

[00:02:37]Number three, large language models have no real ability to reason.

[00:02:43]They're just guessing about information, but they're not actually retrieving information. They're predicting the next token.

[00:02:49]And when they appear to reason, what they are doing is pattern matching against reasoning patterns they've already been trained on through Reddit or any other source on the internet.

[00:02:58]And that's why they can solve a hard math problem from a textbook, but then fail on something trivial that a child could handle.

[00:03:05]What's most important to being human is not ever written down.

[00:03:10]And if you compare this to a human, a humans have cultural values and social norms and can make an inbuilt assessment of whether something is going to hurt them or help them.

[00:03:19]But an LLM has no skin in the game because there's no skin. It's just predicting what the right answer might look like, which is a very different thing to finding the right answer.

[00:03:29]Number four, limited planning, and this might surprise you. If you go ahead and ask an LLM to plan a week of work for you, it will produce something that looks like a plan, but it's not actually simulating the future. It's just trying to plan your week based on existing patterns.

[00:03:44]If you show it something new, it will have you meeting a customer at one end of the city and then put the very next meeting on the other side of town.

[00:03:52]There's no inherent ability to take into account factors such as geography or time differences.

[00:03:57]It's going to give you something that looks plausible, but probably not useful.

[00:04:01]Language, LeCun argues, is a thin, compressed shadow of reality, and training on text alone is like learning to swim by reading a textbook.

[00:04:12]If LeCun is right, the hundreds of billions of dollars being poured into LLM infrastructure right now through data centers, GPUs, power stations, are now funding the final chapter of a dead-end technology.

[00:04:24]That doesn't mean that LLMs go away because for some uses, such as coding and programming, they're absolute gold.

[00:04:31]But this hype about agents and LLMs taking over the world, it's going to pass.

[00:04:36]There's been more than enough capital deployed now on traditional AI solutions, and it still fails to deliver any real economic value outside coding.

[00:04:44]So, we should just stop there and be happy with that. It's a dead-end technology.

[00:04:49]Now, once again, LeCun might be wrong.

[00:04:52]I tend to think he's right, but as always, I'm really keen to hear your thoughts and feedback in the comments, even if you disagree with me.

[00:05:00]I'm curious whether you still believe AGI is a realistic goal.

[00:05:04]I'm Dr. Earl Brad, the founder of Kira.

[00:05:06]If you find my content useful, please consider liking and subscribing. Please also feel free to connect with me on LinkedIn or follow our Kira page. Thanks again, and I'll see you in the next one.