On: GPT-5

[WXLCKNO]

It's very doubtful that they'd have any kind of magical breakthrough that makes the model anything other than incrementally better right now.

[sosodev → WXLCKNO]

How do you figure? They’ve hinted that the reasoning breakthrough used to achieve gold in the IMO will be here in GPT-5.

[gowld → sosodev]

What breakthrough? The self-awarded "gold" IMO result was achieved by running the model for over 1hr per question.

[sosodev → gowld]

That sounds like a breakthrough to me. I don’t think GPT-4 could accomplish the same thing given several hours to try.

[vlovich123 → gowld]

Said another way, 30 min less than what humans get? It’s on average 90 min per question.

[mtlmtlmtlmtl → vlovich123]

And how much energy does a human being consume while spending 90 minutes on an IMO question?

[jjmarr → mtlmtlmtlmtl]

Probably more. 200 kcal (a shrinkflated bag of chips) is about 232 watt hours. A typical 4o query is 0.3 to 3 watt hours.

https://epoch.ai/gradient-updates/how-much-energy-does-chatg...

[SketchySeaBeast → jjmarr]

But how much time does that 0.3 watt hour query take to run? They imply that an individual ChatGPT query takes 0.3-3 watt hours, but most queries come back in seconds, so we need to scale that over a whole hour of processing.

Edit: Scrolling down: "one second of H100-time per query, 1500 watts per H100, and a 70% factor for power utilization gets us 1050 watt-seconds of energy", which is how they get down to 0.3 = 1050/60/60.

OK, so if they run if for a full hour it's 1050*60*60 = 3.8 MW? That can't be right.

Edit Edit: Wait, no, it's just 1050 Watt Hours, right (though let's be honest, the 70% power utilization is a bit goofy - the power is still used)? So it's 3x the power to solve the same question?

[andai → sosodev]

The gold which Google won too, right?

[og_kalu → sosodev]

No Sam explicitly said that breakthrough wouldn't be in GPT-5