AI Pioneer Chollet Declares Scaling Dead, Pursues Inventive Human-Like AGI

The Shift in AI Thinking

Grab a cup of coffee, because we’re diving into some pretty fascinating stuff happening in the world of artificial intelligence. So, you know how everyone’s been obsessed with making AI models bigger and bigger? Well, François Chollet, a big name in AI and the brains behind Keras, is saying, "Hold up!" He’s got some bold ideas that might just shake things up.

Bigger Isn’t Always Better

Imagine you’ve got a friend who thinks the only way to get better at basketball is to keep practicing shooting three-pointers. They just keep shooting and shooting, but they never work on their defense or passing. That’s kinda how Chollet feels about the current AI trend. He believes that just piling on more data and making models larger isn’t gonna lead us to true intelligence. It’s like trying to fill a bucket with a hole in it—eventually, you’re just wasting water.

Chollet argues that today’s large language models (LLMs) are more like fancy parrots than intelligent beings. Sure, they can spit out text that sounds great and can tackle a bunch of tasks, but when faced with something truly new or challenging, they kinda flop. It’s like asking that basketball friend to play soccer; they might be great at shooting hoops, but they won’t know how to kick a ball.

The Problem with Current Benchmarks

Let’s talk about benchmarks for a sec. These are like the tests we give to AI models to see how well they perform. Chollet points out that many of these benchmarks are flawed. Think of it like a math test that only asks questions from the textbook—you’re not really testing if someone can think critically or solve real-world problems. Instead, you’re just measuring how well they can memorize the material.

In fact, he’s noticed that even with all the advancements in AI, models still struggle with basic logical reasoning. It’s like watching a super-smart kid ace their math homework but fail when you ask them to solve a puzzle. Chollet’s got a point: more size doesn’t equal more smarts.

Introducing the Abstraction and Reasoning Corpus (ARC)

So, what’s Chollet’s solution? He introduced something called the Abstraction and Reasoning Corpus (ARC) back in 2019. Picture this: a series of visual puzzles where the AI has to figure out the rules behind transformations of grids. Humans nail this stuff, scoring around 84-85% accuracy, but AI? Not so much. They often score close to zero. It’s like asking a toddler to solve a Rubik’s Cube—just not happening.

ARC is designed to test fluid intelligence—the ability to adapt and learn new things quickly. It’s not just about memorizing patterns; it’s about understanding and reasoning. Chollet believes this is key to developing true AI intelligence. And guess what? The latest versions of ARC are even tougher, pushing the boundaries of what AI can do.

The Future of AI: Beyond Scaling

Now, here’s where it gets really interesting. Chollet is all about combining the strengths of deep learning with logical reasoning. He’s not just looking for a supercharged chatbot; he wants AI that can invent, discover, and solve problems on its own. It’s like he’s trying to create a digital version of Einstein, not just a calculator.

He recently left Google to start his own lab, Ndea, where he’s gonna focus on these innovative methods. It’s like he’s setting off on a new adventure, ready to explore uncharted territories in AI research.

The Challenge Ahead

But wait, it’s not all sunshine and rainbows. Chollet knows that the road to true AGI (artificial general intelligence) is gonna be tough. He’s throwing down the gauntlet with the ARC Prize, a global competition to encourage researchers to tackle these complex problems. It’s like a call to arms for the AI community to step up their game and think outside the box.

Conclusion: A New Narrative in AI

In the end, Chollet’s perspective is a refreshing change from the usual “bigger is better” mantra. He’s challenging the AI community to rethink what it means to be intelligent. Instead of just creating models that can regurgitate information, he’s pushing for systems that can genuinely reason and adapt. It’s a bold vision, and as the AI landscape continues to evolve, Chollet’s ideas might just lead us to a future where AI is not just smart, but truly intelligent.

So, next time you hear about the latest AI model boasting millions of parameters, remember Chollet’s words: it’s not about size; it’s about how well we can think and adapt. And that’s the kind of intelligence we should be striving for.