Why GPT's 'Time Problems' Are Often About Infrastructure, Not the Model

The most talked-about "stupidity" of GPT might not be a model bug at all. Sometimes it's simply a sign that someone hasn't completed the basic infrastructure around it.

Here's the video that sparked the latest round of discussion:

https://www.youtube.com/watch?v=5VRgk7_X7oc

Where the "Sense of Time" Story Originated

I saw a viral clip of an interview where Sam Altman talked about a bug related to GPT's "sense of time." It quickly spread as proof that "they can't fix the simplest thing even after a year."

But in my view, the story is a bit different. According to recent retellings, it was about a problem: without the necessary tools, the model struggles with time and can interpret anything related to dates, intervals, and the sequence of events in odd ways.

The Need to Separate Model and Product

It's important to differentiate two things in this discussion.

1) LLMs Are Not Required to "Sense Time"

An LLM is not inherently required to handle time well. Calculating the interval between messages, understanding the current date, determining what happened earlier or later—these are not tasks for a language model.

These tasks are for the infrastructure around it: timers, metadata, system fields, proper sorting, calendars, and application logic. What is considered "basic" in classical software.

2) Users See It as One Product

When people use ChatGPT or any other LLM as a finished product, they don't think in terms of "this isn't a model problem, it's a lack of tool calling."

For the user, it's one product. And if it gets confused about time, then the product is unfinished in that area. And the user is right in their own way.

Challenges We Faced in Designing AI Products

We encountered this too when designing our AI solutions. It quickly becomes apparent that the model often lacks basic things:

• what day it is;
• which message is newer;
• which is older;
• how much time has passed;
• the actual sequence of events.

And this is a sobering realization.

"Magical" Problems Are Often Solved Without a New Model

At some point, you realize: many "magical" AI problems are solved not with a new model, but with proper infrastructure around it.

With timers. Metadata. Sorting. Logic. Simple technical workarounds—without which the product starts to look smarter than it is, and then suddenly "stumbles" on basic scenarios.

Moral

For me, the moral isn't that "Sam made a mistake." It happens to everyone.

The moral is this: very often, the bottleneck in an AI product is not in the model itself, but in the most down-to-earth engineering details around it.