Timing a race, a priori, is not the most complicated task in the world. However, for ChatGPT, the exercise still remains a little vague. The proof with a TikTok video that has gone viral. Husk asks the AI voice model to measure the time it takes to run a mile. The result comes quickly: ChatGPT announces a precise time. Except that the model didn’t actually measure anything at all.
A very approximate time
And when we confront him with this inconsistency, he doesn’t flinch. The AI maintains that it did indeed follow up, as if nothing had happened. Guest on the show Mostly HumanSam Altman was confronted with this demonstration. His reaction is rather direct: “ This is a known issue “. In other words, nothing new under the sun.
The boss of OpenAI even specifies that current voice models simply cannot start a timer or track a duration in real time. And we should not expect an immediate fix: according to him, it will be necessary “ about another year » for this type of functionality to work correctly.
This little failure is not an isolated case. Artificial intelligence models often struggle over time, literally. Ask ChatGPT how long you’ve been chatting with him: he may very well invent a duration. Show him a clock: he might make a mistake when reading it. Even image generators still struggle to draw a watch displaying precise time.
The rest of the video takes the test even further. Husk shows ChatGPT a clip where Sam Altman explains that the model is not capable of measuring time. Logically, one would expect AI to align with this explanation. But no. ChatGPT insists: some models perhaps have limits, but he, he asserts, knows how to manage time perfectly. A confidence that does not weaken, even after several reminders.
Last test: he is asked again to time a race. Immediate response, with a very precise result – 7 minutes and 42 seconds – out of nowhere. This behavior clearly illustrates a particularity of current models: when they do not have information or a capacity, they can still produce a credible response, without clearly signaling their limits.
Sam Altman makes no secret of it: there is still work to do. The idea is to gradually add these functions to voice models to make them more reliable on this type of simple task. In the meantime, to measure a race time, it is better to keep an eye on your watch rather than relying on ChatGPT.
🟣 To not miss any news on the WorldOfSoftware, follow us on Google and on our WhatsApp channel. And if you love us, .
