ChatGpt exposed: Don't let ChatGPT time your mile run. It can't do it but Sam Altman seems unconcerned it will lie & say it can

Wonderful podcast as always. Can't wait for letstalkai.com

So I just listened to the first 4 minutes of the podcast. I did enjoy it, but Rojo was called out by Jon for sort of trying to take credit for the idea of having April’s Fool’s articles online, as if he originated it, and then Rojo was reminded by Jon that Wejo’s face looks a fair amount like Rojo’s. Good stuff.

Decision trees can simulate responses to language because language has a finite number of common words and expressions, and a standardized syntax.

But it doesn't "know" what time is, the way a human does. It has never witnessed a clock being used to time anything. It's just a very fancy algorithm with a dictionary.

My AI girlfriend is more than just an algorithm.

Language models aren’t just decision trees. They’re based on neural networks (specifically transformers), which learn patterns from massive amounts of text. They don’t follow a fixed tree of choices—they generate responses dynamically based on probabilities.

I'm not watching the clickbait video but WOW that's a fine interviewer wearing some sexy shoes!!

I think the issue here is there is no clarity about the capabilities of different products/features built on top of LLMs, and the enormous variety of technologies encompassed by "AI." An LLM chatbot would not typically have any context for elapsed time between prompts. Nobody should expect it to be able to time a mile, and indeed nobody expected ChatGPT to be able to do this when it first launched. But now, many of these products can call out to non-LLM tools to return information which is not inferred solely from training data, like using Python to analyze some data you provide. But if and when such tools are used is not made obvious. Matters are made murkier by our previous experience with older-generation "AI" like Alexa and Siri, which are quite capable of setting timers and running stopwatches. So now, unless you are an extremely savvy user, you really have no idea what sort of accuracy you can expect from different LLM-based services, despite nearly all answers being given in a very confident, authoritative manner. There's no technical reason ChatGPT or Claude couldn't time something if the devs wanted to add that skill. But it seems as though they haven't, which is pretty reasonable, considering that if you can use a chatbot, it's basically guaranteed you have a better way of doing it. I'm generally on the optimistic end of the spectrum regarding AI, but the idea of people asking ChatGPT to do something so basic as running a stopwatch for you is pretty depressing. At some point, these tools could evolve to something which could actively listen for "start" and "stop" commands, and very accurately measure the time delta, which would be objectively easier for some tasks than using a stopwatch, but until then...

The truth about frontman Altman:

gadgetreview.com

Sam Altman Can’t Code and Doesn’t Understand AI – Coworkers Say

OpenAI insiders question Sam Altman's technical abilities, claiming the CEO lacks basic machine learning knowledge despite leading AI development.

Open link