This is crazy. Today, I learnt something that, if it were not coming from a legitimate source, I would not have believed prima facie.
Anthropic, the AI company behind Claude, has developed what is essentially an MRI for artificial intelligence — a brain scanner for large language models. Using this, they discovered that when models like Claude 3.5 are prompted with arithmetic questions, they do not compute answers using neat, logical steps like the ones we were taught in school. No, they do something far more bizarre. They approximate. They guess. They think.
In other words, they behave like us.
This is genuinely fascinating, because this is exactly how I’ve always thought, and frankly, I assumed it’s how most people who do mental maths operate too. I constantly run these quick approximations and mini deductions in my head, especially the kind where I instinctively know “6 + 9 will end in 5”, or when I see something like “36 + 59” and my brain immediately converts it into “40 + 60, so just under 100”. It’s not a formal calculation; it’s a sort of educated, intuitive leap based on how numbers behave.
What’s staggering, and slightly unsettling, is discovering that machines are now doing the same.
I had always imagined algorithms to be rigid, following textbook methods like the ones we were drilled with in school: carry the one, add the digits, line things up neatly. But this, this approximation-based, pattern-driven way of arriving at a result, feels so… human.
It’s impressive. It’s eerie. And it absolutely blows my mind.
The machines aren’t just calculating. They’re thinking. Or something eerily close to it.
