Why Language Models Can't Tell Time

AI, large language models (LLMs) are like the rock stars that can churn out essays, code, and even poetry. But estimating how long tasks take, they're singing a different tune entirely. Think of it this way: they're like that friend who's always saying they'll be there in five minutes but shows up half an hour late.

The Timing Paradox

A recent study examining four model families across 68 tasks found that these models overshot the actual task durations by as much as 4 to 7 times. Imagine predicting a task would take human-scale minutes, only for it to wrap up in seconds. That's not just a small error, it's a major disconnect from reality.

It gets worse. When tasked with comparing the relative complexity of different tasks, the models fared no better than chance. For instance, GPT-5 scored a mere 18% on counter-intuitive pairs, which is essentially a coin flip. This isn't just about task minutiae. It raises a fundamental question: how reliable are these models when they can't even get basic timing right?

Why This Matters

Here's why this matters for everyone, not just researchers. If LLMs can't estimate how long they'll take to complete tasks, this throws a wrench into any planning that relies on them. Think of scheduling or any time-critical scenario where precision is key. It's like trying to use a clock that can't tell time.

These models have theoretical knowledge about durations from their training data. But without experiential grounding, without actually 'feeling' the time they take to process tasks, their predictions fall flat. This isn't just a technical quirk. It has practical implications for industries looking to automate complex workflows.

Where Do We Go From Here?

Honestly, the idea that LLMs could one day manage their own scheduling and planning sounds like a sci-fi dream. But without the ability to accurately gauge their own task durations, we're not quite there yet. So, what's the workaround? Perhaps integrating more experiential learning elements or introducing mechanisms that give these models feedback on their timing accuracy could bridge this gap.

If you've ever trained a model, you know the frustration of tweaking parameters to get things just right. But timing isn't just another parameter, it's a critical component of efficacy in real-world applications. Until LLMs can handle this, their utility in certain domains remains limited.