Max Rydahl Andersen
Apple’s recent paper, The Illusion of Thinking, has sparked a wave of discussions both online and offline. Some have latched onto it as a definitive "gotcha" moment against large language models (LLMs), claiming they’re nothing but fancy statistical engines that can’t truly reason.
But here’s the thing: we already should know that. Anyone who’s spent time digging into how LLMs work knows they’re built from neural networks, statistics, and sheer compute horsepower. There’s no magic “thinking” happening under the hood, no matter how many times a CEO stands on a stage and calls them “intelligent.” And yet — that doesn’t make them any less impressive or useful.
The Real Illusion: The Hype
The real illusion is not in the models themselves—it’s in the hype cycle that surrounds them. Business leaders and CEOs often hype LLMs as though they’re fully capable, plug-and-play, general reasoning engines. That’s not how these systems work, and Apple’s paper simply reinforces what every ML researcher has known from the start: these models are fundamentally pattern-matching systems, powered by vast amounts of data and compute.
Apple’s paper simply confirms what we’ve known all along: these models can look like they’re reasoning — and sometimes that’s good enough. Because at the end of the day, they’re still capable of doing things that were unimaginable just a few years ago.
LLMs: The Most Algorithmic We’ve Ever Been
Despite the statistical sausage machine that powers them, today’s LLMs are the closest thing we’ve ever had to algorithmic reasoning at scale. They can:
-
Guide a blind person through the world, describing scenes and reading labels in real time.
-
Act as coding companions, helping developers catch bugs, refactor, or even understand that new API you forgot existed.
-
Summarize and organize massive amounts of information, cutting through the noise that would otherwise bury us.
These are not just party tricks. They’re changing lives — and they’re helping us do our jobs better.
The “Collapse at Scale” Isn’t News — And It’s Not the End
Yes, Apple’s paper shows that LLMs fail when faced with extremely complex tasks, like a hundred-step Tower of Hanoi puzzle. But here’s the kicker: most humans would give up there too. Even if an LLM’s chain-of-thought collapses at scale, it still does remarkably well at medium complexity — and that’s where most of us actually live.
Progress Never Starts Fully Formed
If you know me, you know I love pointing out that every breakthrough starts by looking clumsy. The airplane didn’t beat the bird at first. Computers didn’t outthink humans at first. That’s the nature of progress.
Bret Victor’s brilliant talk The Future of Programming (presented in 2013 as if from 1973) captures this perfectly: every tool we build shapes the next generation. Even if today’s tools feel incomplete, they plant the seeds for what’s next. Just like the early computing pioneers couldn’t imagine where GUIs and the web would take us, we can’t fully predict where LLMs and reasoning models will lead. That’s both humbling and exciting.
All at the same time, we also realize that even things that showed up in 1970s, like interative GUIs, logic driven software and the web, are still evolving. It shapes the next generation and when looking at what was possible in 1970’s elevator pitches and demos we are still not done yet.
Apple’s Paper is a Reminder, Not a Rebuttal
Apple’s paper is a reminder that LLMs aren’t magic. We shouldn’t expect them to think like humans — and that’s okay. It doesn’t mean they’re useless. It just means we need to understand their limitations and build smarter systems around them.
Let’s Keep Moving Forward
LLMs aren’t the final answer to reasoning, but they’re the closest we’ve ever come. They represent decades of work in machine learning and language modeling. Let’s celebrate what they can do — and keep building, learning, and evolving.
It is not about the tool itself, but about what it lets us become.
Have fun!
- Max Rydahl Andersen