Ask HN: Do you also think the release of GPT-5/Q* is imminent? | Hacker News Reader

Ask HN: Do you also think the release of GPT-5/Q* is imminent?

8 pointssanty-gegen10 comments

Signs include we're nearing the one year anniversary of GPT-4 (March 14th) plus the latest release Claude 3, which seems to surpass GPT-4.

View on HN

I expect they release GPT-4.5 in the next couple of months. It should have a longer context, and better benchmark results, but not dramatically better than GPT-4. It will probably be just enough to reclaim the crown.

I do expect a lot more from GPT-5 though, it definitely should not be making any silly mistakes, and hallucinations should become very rare. Basically I expect it to become a lot more reliable. Hopefully by the end of the year. They raise the monthly price for ChatGPT access, but I’m ready to pay more for a significantly better model.

That 4.5 sounds a lot like the gpt 4 turbo they released a while back

Sam Altman and Greg Brockman just got sued by Elon Musk, so not the best time for them.

Also, they don't have to rush it, GPT-5 would mostly be competing with GPT-4 right now, while it would have a significant cost to roll it out to the world, given the demand. Plus safety concerns wouldn't have been properly addressed.

I for one have used gpt-4 extensively and gotten a ton of value from it, but recently cancelled my subscription because local LLMs have come so far, and Gemini is on par or better than gpt-4 for most of my uses. I suspect I'm early in a big wave of similar cancellations, so the pressure to release gpt-5 may be mounting.

bet on it

https://manifold.markets/MiraBot/will-gpt5-be-announced-in-m...

haha, it seems there are several prediction markets, and they are somewhat converging:

https://manifold.markets/SantiagoGegenschatz/openai-releases...

I'd be good if they were to make GPT-4 the way it was on March 14th 2023.

So worse results on average, slower, and often unavailable?

So like today?

I think they may be using low precision quants during peak load, which makes the output sometimes complete garbage. I'm most convinced though that the main way they destroyed gpt-4 is the extremely-long system prompt they've been using since making it all-in-one, it has made gpt-4 all but unusable for even mildly complex tasks. Those system prompt tokens are powerful tokens and they're wasted on telling the model how to prompt dall-e without offending anyone when my prompt is actually about generating a unittest. Using the "gpt classic" "gpt" it's much better, but the usage limit hits fast and all the network errors are irritating.

Reply on news.ycombinator.com