Live data from Hacker News

GPT4 is up to 6 times more expensive than GPT3.5

openai.com

121–130 of 136 posts

Re: GPT4 is up to 6 times more expensive than GPT3.5

#121
post #26
post #16

I imagine this price will come down over time. OpenAI has repeatedly said that they haven't the infrastructure in place yet to handle too much load and expect to be "severely capacity constrained", so I assume the high pricing and usage caps on the ChatGPT version are to keep the load manageably low for the moment.

I imagine those dollars from papa Nadella are buying LOTS of capacity as we speak.

Its not $ constrained.

The existing cloud providers don't have 10000 A100s lying around to ramp up at any time. There's a very limited amount of A100s and H100s on the market. TSMC is printing them like candy, but it still takes time to ramp up.

Azure also needs time to install them in datacenters. These A100s stacks probably produce monstrous amounts of heat compared to normal CPU racks.

Re: GPT4 is up to 6 times more expensive than GPT3.5

#122
post #66

Can GPT be used for something like feeding it a 2000 page PDF of pure text(not english) and ask questions about its contents?

Not without other tooling. Things like langchain and llama_index would be good starting points. An approach would be to use llama_index to create embedding vectors for each section of the pdf, then you query and it gets a vector for your query -> gets the context -> puts it into gpt + your query -> returns the result.

I've seen people say it's better to ask gpt for a fake answer then use the embedding of that answer to search (so you're looking for context that looks like the answer). I don't know if that's supported in those tools.

Re: GPT4 is up to 6 times more expensive than GPT3.5

#124
post #50

Earlier quoted context omitted.

Yeah but that's also 30x times worse. ChatGPT is much worse-performing than davinci-003.

No, something’s off, and I’m interested in finding out what

I don't think so. Try to convince ChatGPT to reply with "perform action: email" whenever you send a message that looks like you want to send an email.

It just refuses to do it consistently. It might work once or twice per conversation, but soon it'll be asking you for SMTP credentials.

It's an 1.3B parameter model, we can't expect much from it. It's a wonder it's as good as it is.

Re: GPT4 is up to 6 times more expensive than GPT3.5

#125

Earlier quoted context omitted.

The longest one that is generally available is always going to be yourself :)

My context model is getting shorter and fuzzier.

… but still the weights are increasing ;)

Re: GPT4 is up to 6 times more expensive than GPT3.5

#126
post #113

Is it conceivable that ChatGPT could review itself and produce a superior revision to a human?

ChatGPT, unlikely. But with that said, we're already using AI optimizers to make better versions of AI chips like the H100, so take that with what you will. Current GPT context is too low for this, I believe. We'll have to have something a little closer to AGI for this, and I'm unsure if a cyclic working memory (come up with an output, review the output, review that output, output) will get us to AGI itself. We still…

Genuinely curious as to why ChatGPT could not so I asked it.

Note the final bullet point. I posit, that if it is possible, it will happen (regardless of ethics), because the risk of not doing it (when your adversary will) outweighs the risk of doing it.

"You say an instance of the GPT cannot experiment with new architectures, but could it if a) it was modified by the developers and b) it was given access to sufficient computational resources?"

Response: "In theory, if a GPT model were to be modified by developers to include capabilities for analyzing, designing, and testing new architectures, and if it were given access to sufficient computational resources, it could potentially experiment with new architectures. This would involve incorporating additional machine learning and deep learning algorithms and methods, such as reinforcement learning, evolutionary algorithms, or neural architecture search, which can be used to explore new model designs.

However, there are a few caveats to consider:

- Developing such a system would require significant expertise in AI research, as well as an understanding of the specific problems that need to be addressed. - The GPT model is a complex system, and adding the ability to self-improve or create new architectures could make the model even more complex and challenging to manage. - The computational resources required to develop, test, and implement new architectures are substantial, and optimizing resource allocation would be a critical concern. - Ensuring the safety and ethical use of such a self-improving AI system would be an essential consideration, as it could potentially lead to unintended consequences or malicious use.

Although it is theoretically possible to create a GPT-like model with the ability to experiment with new architectures, it would require significant advancements in AI research, development, and ethical considerations.

Re: GPT4 is up to 6 times more expensive than GPT3.5

#127

Earlier quoted context omitted.

An open API you can build products off of. Consider the alternative. They could not let anyone use their models and just release AI products (their own search engine, their own version of Teams, their own self driving car software, their own photoshop competitor, etc).

This is revisionist history. OpenAI very much started as a (somewhat) open source (for whatever definition of "open" makes sense for LLMs) initiative.

I’m not arguing they stuck with their original position. I’m saying the name is still reasonable. I don’t care if they changed positions.

Re: GPT4 is up to 6 times more expensive than GPT3.5

#128
post #103

If this is public I can only imagine what Google has internally. Does DeepMind compete in this space or is it fundamentally different than a llm like Bard? With their massive codebase and already deep investment in AI/ML, I'm pretty sure Google and likely MS already have the ability to do massive refactoring, validate it using tests, reiterate, train, rinse and repeat.

This is good mostly for improving quality of software and reducing technical debt. But not much good for innovation.

Re: GPT4 is up to 6 times more expensive than GPT3.5

#129
post #67

Earlier quoted context omitted.

Didn't the CEO say during the GPT-4 demos that humans were involved to do the captions during all the images tests (using Be My Eyes volunteers) and that's why it took some time to show the Discord examples ? If so, it could explain some costs indirectly.

Wait did they just use Be My Eyes volunteers as AI labellers without telling them? Or were they told? Pretty unethical if not!

Maybe it's something else ?

I am genuinely not sure what they mean, this is the timecode in the announcement: https://www.youtube.com/live/outcGtbnMuQ?feature=share&t=651

It seems to say that they use Be my eyes "to make the product better", so it could be humans in the loop during captioning, or maybe they are using volunteers data from Be my eyes as training set ?

Or it's completely opposite, and that Be my eyes is using the OpenAI APIs ?

My guess is that it is both ?

Re: GPT4 is up to 6 times more expensive than GPT3.5

#130
post #28
post #9

Worth remembering GPT3 was 10-20 times more expensive 2yrs ago. There's a super-moore's law type learning function going on here and I suspect in 1yr GPT4.5 will be the same cost as GPT3.5 today. It's insane... feels like iPhone 3G to iPhone 4 level quality improvement every year.

GPT-4 is a scary improvement over 3.5, especially for handling code. It will be the literal definition of awesome when these models get a large enough context space to hold a small-medium sized codebase. I've been playing around with it for an hour seeing what it can do to refactor some of the things we have with the most tech debt, and it is astounding how well it does with how little context I give it.

I can't get it to run a single query today. Are you on the paying plan?
Post reply on HN