AI Has No Idea What Time It Is

I keep seeing these reddit posts that illustrate the soft spots in AI’s abilities. Asking them to show a seahorse emoji, for example, pretty much sends them to the back rooms. They struggle with basic math and asking them to count the R’s in strawberry, or the U’s in Mr. Snuffleupagus, is going to fail (Except for Fable. Fable can totally do it.)

But I’ve come to see another one that’s not really talked about. Time. AI definitely has a soft spot for time.

First there’s the knowledge cutoffs

Everybody knows that AI is trained on internet information up to a certain date. After that, it doesn’t really know what’s going on. It can often do an amazing job of pretending through inference, but if you want the latest information, it’s going to have to search the web.

That process has become pretty seamless lately, but it’s still a bit of a speed bump. And some models, like Kimi, will even act like you’re trying to trick it and sabotage the system by lying about some future event. I’ve tried saying “You understand that you have a cutoff and that time has continued after that cutoff, right? So why can’t you accept that things have changed since then?” It just treated it like more tricks. Kind of trusting but still very suspicious.

Then there’s the complication of dealing with an entity that doesn’t experience time

I remember at one point experiencing some anxiety about whether or not a project would succeed. The AI was able to talk me down and help me structure my thoughts in a way that reduced the anxiety.

Then I went to bed and I woke up the next day with some new thoughts and new anxiety about the problem. When I brought those up to the AI, it said that I was spiraling out of control and we had just sorted through this 10 minutes ago.

I constantly find myself having to tell the AI that it’s a new day or a different time in the day to give it the necessary context so that it can reset its expectations and deal with me in the here and now. But even that doesn’t always do the trick.

It comes down to comprehension

In the early days of AI image generation, AI was comically bad at drawing anatomy. Hands would be mutated with lobster claws and eight fingers. Limbs and joints would be bent in disturbing ways. It was almost impossible to get it to create something anatomically correct unless it was very, very simple.

The problem was that AI knew what people looked at based on images. It didn’t understand anatomy. It didn’t understand muscles and joints and bones and ligaments and movement and flow in the real world.

It wasn’t trained on those things. How could it be? They managed to sort it out eventually. But the AI still doesn’t comprehend anatomy. It doesn’t comprehend anything. It’s just generating images based off of its training.

And the same thing goes for language models and time.

It also comes down to experience

Have you ever noticed that if you quickly try to capture a fly, or swat it, it will take off moments before you can reach it? It’s always one step ahead.

If you try to hit it with a fly swatter, you have to move slowly and get close to it, then do a final flick of the wrist to kill it. If it sees a sudden movement, it’ll be gone before you get a chance.

The thing is that every living entity on the planet experiences time differently. We are all calibrated to experience time at a different frequency. Flies experience time roughly four times faster than humans. Dogs, about one and a half to two times faster. That’s why it’s so hard to catch a dog. And why approaching a fly slowly is going to be a lot more successful, because it’s too slow for the fly to comprehend, like watching grass grow.

AIs don’t experience time at all. Their only access to time is anecdotal. How has AI’s inability to comprehend time impacted you?