Hint to the non-experts: AI is a party trick. Allow myself to explain myself. Let’s say it's 2022 and you’re an ML research student, graduating top of your class. Oh boy, are you in the right place at the right time. A FAANG-type company is offering you millions, and all you have to do is complete a take-home project to make an image processor that recognizes tigers. So you spin up an empty neural net and randomize initial weights. At this point, this neural net knows NOTHING. But you feed it a million tiger images, and a million more of other things. Now, the solitary thing it knows is what a tiger ought to look like, and it’s decent at recognizing them. The next afternoon, you’re sitting in Washington Square in NYC, idly calculating how many tigers you yourself are going to buy with your FAANG millions. You see a small flash of orange and black on a small branch in a tree above you. Curious, you take a picture and send it through your processor. To your dismay, it’s 75% certain that that’s a tiger up there. It’s obvious to you that it’s NOT a tiger, but think about what knowledge you’re bringing to the table that this neural net doesn’t have: -That flimsy branch could not support a tiger (neural net doesn’t know about gravity, weight, or sturdiness of wood) -What in the world would a tiger be doing in Washington Square (neural net doesn’t know that tigers don't hang out in parks) -The people in the photo aren’t panicking (neural net doesn’t know humans would be scared of tigers) -The black/orange blob is too small (neural net doesn’t know that tigers of that size can’t exist) So, you as a human with your general intelligence know it’s not a weightless, escaped, unscary, miniaturized tiger. But this party trick of a neural net cannot possibly bring that knowledge to bear in the single task it was meant to perform. Can a neural net learn all this generalized information, so it’s not a party trick any more? The definitive answer, as of September 2025, is MAYBE. Giving the image in this post to a modern LLM yields the correct answer, but more complex queries are still a challenge. Whether AI’s can move beyond pattern matching to genuine contextual reasoning sits at the heart of our current journey with LLMs, and more recently multi-modal LLMs. More in a subsequent post.
That’s an eloquent explanation of the inherent problem.
So close.