39 comments

[ 3.3 ms ] story [ 85.7 ms ] thread
Interesting - I have noticed a slow-down in responses in Plus.
Yes, it is extremely slow right now! I think they need some spinner on the UI because I thought it was just broken and kept writing more questions which probably increases the load further.
ChatGPT is so slow no one uses it anymore. -Yogi Berra
Them being unable to scale compute to meet demand would be a reason to pause signups.
of course? I mean isn't that obviously going to happen?
Selling your OpenAI account? Bold when they can blacklist accounts.
Why would this even be worth the dev resources to deal with?
Abuse? Can only speak as an info/cybersec professional who is responsible for looking for this and implementing mitigations.
This is a paid service with limited quota based on your subscription level - seems hard to abuse? Your subscription has purchased X prompts, doesn't really seem to matter if a bot or a different user is making those X prompts.
I wonder how OpenAI "resellers" are dealing with the slower response times and exploding demand.

Ie. if you have an OpenAI's API account and a UI of some sort, I understand you could theoretically resell GPT4 access as a service to third parties, at least that's how places like poe.com (from the Quora people it seems) or phind.com work, although I believe they all have signed partnership deals to be able to compete with OpenAI's ChatGPT Plus.

I think phind.com is using the GPT-4 API directly. I think most legit offers are actually doing that. phind.com is about tailoring the experience for coding and offer search powered AI. It does not aim to be a ChatGPT replacement, though. And I don't think OpenAI API usage terms prevents you from offering GPT-4 via UI (or maybe I'm wrong).
The API is currently extremly fast. It is the chatgpt webui which is currently suffering.
Is there a way to make Classic the default? The new ChatGPT’s web-infused results are terrible.
Add to your custom instructions: "DO NOT BROWSE THE WEB unless explicitly requested."
You can add Classic to the sidebar for ease of access:

- Profile (i.e. your name)

- My GPTs

- Load: ChatGPT Classic

- Down Arrow -> [Pin Icon] Keep in Sidebar

Now ChatGPT Classic will be listed at the top left right below the full-capability version.

It isn't the "default" but it just gives you two "New Chat" buttons next to one another, which is close enough to default for me.

Is OpenAI not allowed to have a multi-cloud setup because of their M$ deal?

It seems like the limit here is getting GPUs from Azure, so if they're allowed to also use AWS, GCP, etc. then that shouldn't that make scaling a lot easier?

Even if they could, aren't they already losing money for every additional Plus member? The real question is does OpenAI have deep enough pockets to keep subsidizing this many users...
right now, their pockets can go as deep as they want.
But eventually they'll need to make good on that for their investors. So either of two things need to happen:

- Generating ChatGPT 4 quality responses has to get a lot cheaper, so they make money on $20/month.

- They need to increase prices to make money without losing people.

I really like ChatGPT Plus right now, but $20/month is at the top of what I am willing to pay. ChatGPT 4 is the best, but you can get "good enough" for free from the competition.

If they increase the price or reduce the quality of 4 responses too much, I'll look around.

If GPT 5 jumps as far forward as going from 3 to 4 did, I'd go up to $30. That was a big jump, I'm a bit skeptical they can repeat it. That's also assuming I get more prompts per hour and it continues to build in valuable features (DALL-E 4 etc). DALL-E 3 is expensive outside of ChatGPT, that's a tremendous value if you use it in the chat system.

It'd need to start making me breakfast every morning if we're going to keep climbing toward $40-$50. Although don't get me wrong, any improvements beyond 4 is an accomplishment, it's already an amazing service, I just don't feel like paying $40 for GPT 4 (too much of a price reach). $20 is fair for version 4. They need to keep adding a lot of value the higher they want to go.

A $100/month GPT5 might price itself out of most B2C customers, but could still be hugely attractive to businesses. Both for chatbots, and as a productivity boost for knowledge workers. It just has to be good enough to justify the price increase over $20 GPT4.
I’d pay $100 hands down if they could simply get it to the speed of GPT3.5 Turbo. Right now ChatGPT Plus is slow and buggy I’ll end up not using it.
Microsoft is charging $30 for Office Copilot.
I’m paying ~$50 for my cell phone plan. $100 for internet. And all that does is route data. I’d easily pay, say, $200 a month for a system that routes sufficient high intelligence. Gpt5 with a 20% improvement over 4 would qualify.
3 to 4 was a giant giant leap. 3.5 is basically useless. I would consider paying $100-$200 per month for GPT 5 because of the value add. It has really consumed/improved almost all aspects of my life and I have a chat window open 60% of the time I have my laptop open.
I have heard estimates [1] that OpenAI's $20/month ChatGPT charge isn't enough to cover the cost of compute. For example, if their model needs ~100GB of VRAM, at cloud prices of $0.08 per minute for a g5.12xlarge. you make a loss if the average customer uses more than 9 minutes per day. And chatgpt's responses are pretty slow.

And that's without counting the huge free tier they have to support, the employees they have to pay for, and any overheads.

No rush to go multi-cloud if you're making a loss - you'd want to put your efforts into reducing costs.

[1] AI chatbots lose money every time you use them, https://www.washingtonpost.com/technology/2023/06/05/chatgpt...

They are in “survive until GPUs have enough RAM for break even costs” mode right now. It’s why I don’t think we’ll see any big significant changes in parameter count for a while because it’s just too damned expensive.

I would however be happy to pay waaaaay more than $20/mo for a “gpt-5” evolved model.

I don’t think that’s a fair comparison. AWS GPU pricing is kind of ridiculous. $30-40 per hour for A100 clusters is several times what other companies offer.

I only use them because I have free credits but I would never pay out of pocket for their GPUs.

heard about somebody who was able to pay for it after they announced that. i guess don't do a hard refresh and you can still access the stripe checkout
Marketing coup right there by OpenAI. Scarcity increases demand. FWIW I realize it's also actually because of demand surge and balancing of resources to make sure good quality UX that said it will create more demand.
(comment deleted)
Its weird I didn't update my chatgpt app and was able sign up two days ago even though sign ups were blocked.

Makes you wonder if they just removed the UI component and if you could reverse engineer the sign up process you could make money.

At $20 a month POE.COM has been a great value for me. I primarily use GPT-4 and Claude-2-100k, though they offer ~20+ others.
Looks like OpenAI artificially lowers prices.
It's interesting that he got fired a day after this..