The Illusion of Thinking | Apple Just Exposed AI’s Fake Reasoning

0:00 Everyone is chasing the next big leap in AI. OpenAI is pushing models that think step by step. Nvidia is scaling up compute like never before. Microsoft is wiring the cloud to power it all. And Meta, they are building agents that act, plan, and adapt. The message is clear. AI isn't just answering anymore. It's reasoning. And that's where it gets exciting or maybe concerning because Apple just dropped a paper that's shaking up the whole story. It's called the illusion of thinking. And it claims when you crank up the difficulty, all the reasoning these models brag about breaks down completely. Here is the setup. AI models like GPT, Claude, Deepseek all say they can think their way through the logic puzzles like this. But as you add complexity, like a few

0:46 more discs of Hanoi, their performance doesn't just dip, it collapses to zero. That is 0% nothing. Not a single correct answer. So, is AI really learning how to reason or is it just getting better at pretending? That's what we are breaking down today. From Apple's findings to Anthropics counter response to why this could shake the future of AGI, Nvidia, and every company betting on the AI boom. Let's dive into the illusion of thinking. [Music] What do we even mean when we say an AI can reason? In humans, reasoning means working through a problem step by step. We try things, we reflect, we backtrack. We break complex task into smaller sub

1:34 goals. And if we mess up, we rework the plan and keep going. Now, in the AI world, reasoning usually means something very different. It's not actual thinking the way we do it. It's generating a longer answer that looks like thinking. These new large language models, GPT4, cloud, Gemini, don't just output a final answer. They now include steps like let's first identify the goal, then consider constraints, and then here is what the output should look like. This style is called chain of thought reasoning. It's a prompting trick. We guide the model to short steps. Sometimes we even layer in reflection where the model revises its first answer or we give it tools to check its own output like a calculator or code

2:20 interpreter. And yes, these methods often work. Performance improves on math problems, logic puzzles, coding task. So people assumed, hey, this must be a form of intelligence. It's thinking now. But here is the catch. Just because a model can talk through a problem doesn't mean it understands the solution. It might just be mimicking what reasoning sounds like. It's learned that humans often say step one, step two, step three. So, it generates the structure, but is it actually solving the problem? That's the question Apple's paper tried to answer. Not just can AI explain its answers, but can it actually use reasoning to solve harder unseen problems? Because real reasoning should scale. If you solve a

3:05 three dis puzzle, you should be able to handle five, six, even 10 by applying the same logic. But if the model breaks once it sees something unfamiliar or complex, then maybe it's never reasoning at all. Let's talk about the puzzle at the center of Apple's paper. Towers of Hanoi. It's simple, but it's also a classic test of reasoning. Three rods, a bunch of disc stacked by size. The goal, move the full stack to another rod. But you can only move one disc at a time and never place a big disc on top of a smaller one. It sounds like a kids toy, but it's actually a brilliant way to test how well a model can handle compositional logic. Now, here's the twist. Apple tested models like Cloud, GPT40, Deep Seek on Towers of Hanoi. But

3:53 with increasing complexity with just three disk every model reasoning or not did fine no surprises there even a non-reasoning model can brute force it with four or five disc. The reasoning models pulled ahead they used extra steps to think through and got better results. So far so good. But then they hit seven discs and performance didn't just drop it crashed. 0% not a single model solved the puzzle. Not open AI, not anthropic, not Deepseek. And this wasn't some bizarre edge case. These were well-formed, solvable puzzles, just a bit more complex. Even more shocking, the models stopped trying. Instead of writing more steps or reasoning harder,

4:38 they gave shorter, dumber answers, sometimes in just a few tokens. You would expect a real reasoning engine to try harder when the puzzle gets harder. But these models do the opposite. Apple called it accuracy cliff. Performance fails off a ledge and never recovers. Even when the model is given plenty of tokens, it just stops thinking. That's the moment the illusion breaks. What we thought was reasoning might just be pattern matching. Replaying steps it saw during training, but with no deeper logic to generalize beyond that. Now, you might think, okay, maybe Hanoi is just a weird edge case. But Apple didn't stop there. They tested the same models on river crossing problems, checkers jump puzzles and block stacking

5:24 challenges. Different formats, different rules, same story. Once complexity crossed a certain point, reasoning models failed catastrophically. And not just in accuracy, in effort. They gave up early even when they had enough tokens to keep going. What should have been a work harder moment turned into a shutdown and submit moment. And that tells us something critical. These models aren't scaling their reasoning. They are not discovering general strategies. They are not applying rules like humans would. They are mimicking what reasoning looks like. But only when the patterns are familiar. Throw something new at them, unseen, complex, they choke. And here's the kicker. Even when researchers gave them the right algorithm, step-by-step instructions baked into the prompt, the model still

6:10 couldn't execute it consistently. They didn't just fail to reason. They fail to follow directions. So here's the uncomfortable question. If models break down on harder problems even when you give them the rules, are they actually thinking or just faking it really well? Let's break it down. Now models like chat gvtro or cloud aren't reasoning from scratch. They are predicting the next token one word at a time based on training data. That means when you ask for a solution, they're not solving the problem like we would. They're generating something that looks like a solution. They know what a chain of thought sounds like because they have seen tons of examples. So, they stitch together a plausible explanation. But here's the catch. Plausible doesn't equal correct. Worse, in some cases, the

6:58 steps didn't match the final answer. They were hallucinating the thought process. That's not reasoning, that's performance. It's pattern mimicry with good stage presence. Anthropic even confirmed this in their own paper. They found models often hide their shortcuts. They'd use a hint or a memorized answer but written like they worked it out logically. Their chain of thought totally disconnected from what actually led to the answer. So even when you ask them show their work, you might be seeing fiction. And that brings us to the real problem. If reasoning is mostly an illusion, why does it even matter? Because the entire AI industry is betting on the opposite. For years, we have followed the scaling law. The bigger the model, the more data it sees

7:43 and the smarter it gets. And it worked. But for a while, GPT2, GPT3, 4, each leap brought real gains. But now reasoning was supposed to be the next leap. The logic was if we can't make models much bigger, let's make them think longer. Most steps, more tokens, more compute. And that's why Nvidia stock is soaring. And that's why hyperscalers like Google, Microsoft, and Amazon are building massive infrastructure for inference at scale. But if reasoning models don't actually deliver, if their performance collapses on real world complexity, then this boom may be built on a fragile foundation. Apple's paper challenged the assumption that thinking AIs are right around the corner. And if that's not true, if

8:28 general intelligence is still far off, then a lot of current investment might be premature. Even JP Morgan CEO Jamie Demon said the benefit isn't immediately clear. But if billions are going into models, they don't generalize, don't reason, don't scale past training examples. That's a real risk. But not everyone agrees with Apple's stake. Anthropic, the team behind cloud, fired back with a cheekily titled response, the illusion of the illusion of thinking. They argued that Apple tests were too constrained. the models weren't allowed to use tools, code, or enough tokens, things real world agents rely on. In their words, Apple might have set the puzzle up to fail. And sure, when given more flexibility, some models did improve. But even then, Anthropic didn't

9:13 deny the deeper issue. These models still don't generalize well. So, while they disagreed on the test setup, they didn't exactly say the illusion was fake, just that it's more complicated. So where do we go from here? If today's reasoning models break under pressure, does that mean the dream is over? Not quite. It just means we are facing a reality check. We need to solve things it's never seen before. To think through novel problems, to generalize, to reason. Right now, what we have is impressive, but it may just be the world's most powerful pattern matcher, wearing a reasoning mask. And that's the illusion we have to break. If you found this breakdown helpful, drop a like, share it with your team, and subscribe for more deep dives. like this one. We are just getting started.

9:59 [Music]