We're seriously getting deep into the uncanny valley at this point. And I have to say, getting into the uncanny valley is usually just kind of weird and a bit funny when it's a creepy pasta image, but when it's actual cognition, I'm honestly not totally sure how I feel about it.

I would note that it took a LONG time for o3 to check this 24-problem worksheet. Like, I'm pretty sure I could have checked it manually faster.

But watching this thing work was. . . uncanny. The model first broke the image down overall, then row by row, then problem by problem.

Watching the overall page breakdown was fascinating - and extremely recognizable, although way slower than a human would do it. It first looked at the whole page, then it scanned the topic line, then sort of "ran its eyes" over the page to pick out where the actual problems were.

Again, although all of this only took a minute or so, this is stuff humans can do pretty much instantly. (Suck it you dumb AI) (I'm just kidding please don't turn me into paperclips). But the similarities in the process were really striking.

Then the model just went problem by problem, generally quickly recognizing the actual problem. Occasionally calling up a python script to calculate the answer. Then working through a process to figure out my son's handwriting (and grumbling to itself about it). Then checking the answer against my son's response and figuring out whether he was correct or not.

Again, all very, very, somewhat creepily, recognizable.

And then . . . nothing.

It finished working through all the problems, then did nothing. It just sat there.

I could scroll through its reasoning to find all its answers, but they weren't organized in any way and it was a huge pain to do so.

So I followed up with "What's the results?"

And it went into a semi-psychotic break. Spitting out a bunch of weirdly cropped half-images of the various problems in no real order, which provided absolutely no insights on the answers or how my son did.

Then it overwrote those images with a chart of the results. Which were basically impossible to read because they were white text on a bunch of weird off-white images. It was like revisiting MySpace.

But, my lord the results were amazing. It gave all the correct answers. It gave guesses at my son's handwriting that were pretty much identical to mine. It gave great insights into the mistakes he made ("digit swap in the thousands; small place-value slip; forgot the trailing zero"). It even noted where some of the answers were especially hard to read, indicating some actual uncertainty (holy shit!) in its answers.Incredible. Magical. Scary.

So that's that. Weirdly, creepily smart and somewhat struggling to actually communicate with its fellow intelligences in the world. It's more like us that we could have ever imagined.