OpenAI’s most powerful chatbot API rolls out for all paying customers
- Thread starter JournalBot
- Start date
Upvote
31
(33
/
-2)
I find these things intellectually fascinating, but in my limited plays with them, I've ended up saying "Huh" and just not being bothered very much. I just return to keyword heavy standard internet searches and page browsing.
I think this tells me two and a half things. Firstly this is yet another sign that I'm getting old and stuck in my ways, secondly that (for now) I'm likely to use them more when they come to me in domain specific wrappers than just in generalities. And lastly I just haven't found the right thing to prick my curiosity.
I think this tells me two and a half things. Firstly this is yet another sign that I'm getting old and stuck in my ways, secondly that (for now) I'm likely to use them more when they come to me in domain specific wrappers than just in generalities. And lastly I just haven't found the right thing to prick my curiosity.
Upvote
89
(91
/
-2)
When are they opening up 32k? Costs for gpt4 are high, but for our use case, it’s a huge savings. Gpt3.5 16k is dirt cheap in comparison, but there are too many errors for our use case.
Upvote
16
(18
/
-2)
Ask them about things you DO know about, not things you don't, and you'll quickly see that they're a terribly unreliable source of information. Worse, they're a confidently incorrect source of information. Maybe it's better if you pay for the new ones. Maybe I'm not using it in the intended way but there's only language, no logic or other smarts, so it'll start getting into grammatically correct nonsense pretty quick. I don't want to be too negative because the language aspect IS pretty neat but it's like the AI gizmodo article that was full of errors. It seems like there needs to be more internal "fact" checking or something in it.
I could go on about some of the stuff it's done but my personal favorite was when I wanted to see how it did with wordle, I asked it to guess a 6 letter word (wordle normally being 5) and it guessed GRASS. When I corrected it it amended its guess to GREAT.
Edit: Also chatgpt is hilariously poor at ASCII art in a way that's truly terrifying. That's fine but it SAYS it can do it!
Upvote
81
(85
/
-4)
Instead of using ChatGPT as a replacement for search, I personally use it a like a digital assistant to do annoying grunt work for me like doing some basic, but tedious arithmetic calculations. I also use to in lieu of Stack Overflow on how to do thing in programming languages I'm less familiar with like Golang, Rust, or Perl.
It can also read an article (or "watch" a Youtube video with transcripts) and give you a summary of it (I played with as TLDR/TLDW solution for all the article/video people send me). I also found it is good at converting your lecture/interview/meeting notes into a more polished summary that you could email out to people (obviously user-beware to proofread the output before sharing).
Last edited:
Upvote
19
(28
/
-9)
ChatGPT did not watch those videos for you. It gave you a mash-up of web text related to keywords related to the videos.
Upvote
45
(57
/
-12)
Actually, what I think what the ChatGPT bot I used was doing was reading the Youtube auto-generated transcripts and giving a summary of that and/or answer questions about what was said.
Upvote
48
(50
/
-2)
Just wait until they start to use longnet transformers. Was recently reported by microsoft, and has I shit you not a 1 billion token context limit. I give it 18 months at most before this is in the publics hands as well.
https://www.arxiv-vanity.com/papers/2307.02486/
Upvote
50
(51
/
-1)
Upvote
10
(11
/
-1)
Yes, that makes sense. My response was an objection to anyone thinking ChatGPT has any level of comprehension. Your follow-up makes perfect sense, although it is not a "summary" so much as "syntactically related gibberish."
Upvote
-6
(19
/
-25)
I've noticed recently that GPT4 API was much more responsive than in the past (faster generating responses). Probably they were bringing new capacity online to support this announcement. Now I'm a little worried that things are going to get slow again..
Upvote
6
(7
/
-1)
Have you actually tried ChatGPT personally? I mean this is the summary ChatGPT gave of this Youtube video. YMMV but to me that summary feels at least one step above "syntactically related gibberish."
As an aisde, I see it the ChatGPT bot was using its own Speech-To-Text plugin (OpenAI whisper) and not just reading Youtube's auto-generated one.
Last edited:
Upvote
33
(39
/
-6)
I always find some of the negativity in technical forums around ChatGPT and other LLMs to be interesting.
It is absolutely true that ChatGPT can spout sometimes-convincing and always-authoritative garbage. But man is it a useful little tool as long as you understand its limitations.
I have basically completely stopped writing small, domain-specific scripts or routines. I tell ChatGPT what I would like to happen, and it quickly outputs something that's somewhere from 90-100% functional.
A lot of things that used to be a "keyword heavy" search that resulted in parsing a lot of results are now a simply phrased question. If I want to know how to do something in an application, I regularly find that I get accurate and comprehensive results by just asking ChatGPT - and when they're incorrect, I usually can clarify and get a corrected answer.
It's definitely not perfect. Has plenty of inaccurate information and flaws in the way it answers questions. Sometimes it can fail hilariously. But as a tool to work with while being aware of its shortcomings, I find it astonishingly helpful.
It is absolutely true that ChatGPT can spout sometimes-convincing and always-authoritative garbage. But man is it a useful little tool as long as you understand its limitations.
I have basically completely stopped writing small, domain-specific scripts or routines. I tell ChatGPT what I would like to happen, and it quickly outputs something that's somewhere from 90-100% functional.
A lot of things that used to be a "keyword heavy" search that resulted in parsing a lot of results are now a simply phrased question. If I want to know how to do something in an application, I regularly find that I get accurate and comprehensive results by just asking ChatGPT - and when they're incorrect, I usually can clarify and get a corrected answer.
It's definitely not perfect. Has plenty of inaccurate information and flaws in the way it answers questions. Sometimes it can fail hilariously. But as a tool to work with while being aware of its shortcomings, I find it astonishingly helpful.
Upvote
74
(80
/
-6)
Exactly this. Like many tools, art and craft are required to get the most out of it. Just like has developed around image generation prompts.
Upvote
14
(17
/
-3)
How many "tokens" does a human absorb in a year, parsing them out as the equivalent bits of text or speech encountered by said human? Just curious how this compares.
Upvote
8
(9
/
-1)
Well 8k tokens is roughly 4k words. In a lifetime someone will say about 860 million words. So its comprehensive.
Upvote
10
(11
/
-1)
Yeah it's not good as a search or knowledge engine, it's a shame that's how they became known. It's a language model. (a large one!). It's very good at manipulating language especially if you give it what to manipulate. Can be quite useful, but as another noted, it's like an extra step outside the workflow and I type fast enough I'm often... meh, just do it myself.
Upvote
11
(12
/
-1)
We might all find out soon that when you put the sum total of all human knowledge into a network capable of fully contextualizing it, that there aren't really any great insights to be had that people haven't already thought of.
Upvote
19
(22
/
-3)
That's really interesting, their results look great. That said, I haven't yet seen a context length increase that didn't cause the model to suffer from the serial position effect to a very noticeable degree, and I haven't yet seen any research on meaningfully addressing that. It's going to be important to address for use cases like dumping a very large code base into an LLM and asking it specific questions that could be answered by content anywhere in the answer.
Side note, from a theoretical perspective it's interesting to me that LLMs even have a serial position effect. I do not know of any specific reason in their architecture that should make that so - their attention mechanisms are specifically, mathematically built to be position agnostic. Obviously humans suffer from that effect and we don't know why we do for sure, but it's not clear why that should mean LLMs have to as well.
Last edited:
Upvote
13
(14
/
-1)
My phone number is already linked to a throwaway account I made many months ago to try out ChatGPT when it first hit the news. Unfortunately there is no workflow reclaim my number, and OpenAI support has not responded to any of my help requests over the past 3 months. I'm sure it's a great product, but I won't pay for access until I know I can complete the basic action of logging in.
Upvote
-7
(4
/
-11)
Excellent. I’ll be using this in conjunction with plaintext records of HP Lovecraft’s writings and correspondence to do something truly awful. Thank you for the information.
Upvote
7
(7
/
0)
Now how useful will this be for things like CHAI, character AI, Harpy chat and the other character roleplay bots?
edit: they should have added smoke and wood planks at the top of that graph.
Upvote
2
(2
/
0)
That's not what "fine-tuning" means in this context. Fine-tuning enables developers to train the model on custom data, which is not part of the initial training set (consisting of publicly-accessible data, such as Wikipedia articles and Reddit posts). This custom data is often internal to a company.
https://platform.openai.com/docs/guides/fine-tuning
GPT-3.5 and GPT-4 are not yet available for fine-tuning, though. What OpenAI announced was that fine-tuning will be available for GPT-3.5 and GPT-4 later this year.
Upvote
10
(12
/
-2)
So, you intentionally gave them incorrect information, and linked that incorrect information to your /actual/ phone number AND you're blaming OpenAI for this? They didn't provide you a 'I confess I fed you incorrect information about me.' button, so you're mad at them?
This is the whining of a belligerent child, not the informed address of a techie.
Upvote
-3
(13
/
-16)
Any news on how the explosion of LLMs is affecting global energy usage? In yet another year that's slated to be the latest "hottest on record", maybe we need to care about that too. For example, do they have an info on how they expect the general availability of their APIs to affect things?
Upvote
-2
(9
/
-11)
As a paying customer, every time I try and use
Full request (minus API key):
"model": "gpt-4", from OpenAI's own API Playground, I see:
Code:
"error": {
"message": "The model: `gpt-4` does not exist",
"type": "invalid_request_error",
"param": null,
"code": "model_not_found"
}Full request (minus API key):
Code:
curl https://api.openai.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer <API_KEY_HERE>" \
-d '{
"model": "gpt-4",
"messages": [
{
"role": "user",
"content": "<YOUR QUERY HERE>"
}
],
"temperature": 1,
"max_tokens": 256,
"top_p": 1,
"frequency_penalty": 0,
"presence_penalty": 0
}'
Upvote
3
(3
/
0)
I haven't used the API myself, but my first thought is that you might not be using a valid model name. Try hitting GET https://api.openai.com/v1/models first to get a list of valid models, per the documentation: https://platform.openai.com/docs/api-reference/models/list
Upvote
3
(3
/
0)
Upvote
2
(2
/
0)
It is not that:
Code:
$ curl https://api.openai.com/v1/chat/completions \
> -H "Content-Type: application/json" \
> -H "Authorization: Bearer key_removed_for_post" \
> -d '{
> "model": "gpt-4",
> "messages": [
> {
> "role": "user",
> "content": "please respond to this test message"
> }
> ],
> "temperature": 1,
> "max_tokens": 256,
> "top_p": 1,
> "frequency_penalty": 0,
> "presence_penalty": 0
> }'
{
"id": "id_removed_for_post",
"object": "chat.completion",
"created": 1689032591,
"model": "gpt-4-0613",
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "Test message received successfully. How can I assist you further?"
},
"finish_reason": "stop"
}
],
"usage": {
"prompt_tokens": 13,
"completion_tokens": 12,
"total_tokens": 25
}
}
Upvote
3
(3
/
0)
Those are by far the easiest kind of program to write which use it, so it is not surprising that is what it is known for at this point in time.
Applications which integrate it more will come, but that needs code and design, and it is never a simple as you hope. I managed to release a first version of a real application which uses it a couple of months ago, but even that is limited, and we are first in our industry to do this (likely by quite a bit).
I have some real advantages here like being able to throw away things that are not working out and just change the specs entirely at a whim, and I wrote at least ten times the code I ended up using for that. Getting a sense of what it can do within an application takes some effort, and it does best with a specific interface just for it, so you need to code that as well.
I mostly ended up using a microphone for end user input rather than a keyboard, it seems more natural in how people deal with it.
Upvote
6
(6
/
0)
Not a ton yet, but it could increase if usage takes off.
This is not crypto, where the incentive is to burn maximum energy to prove you did. This is a real product where efficiency is good, and there is still much room to optimize. It is just the early days, we need time to do that.
Upvote
11
(12
/
-1)
I've seen machines kick human ass in chess, go, starcraft, and pass exams that we struggle with and spend years preparing for. If you just dismiss this as nothing useful I think you are not understanding the consequences.
Machines will either usher in a great age of discovery or lead to us killing each other or some of both but either way having a LLLM create code from scratch just from typing in a paragraph of what I want is straight out of Star Trek to me.
Upvote
10
(16
/
-6)
Yesssss. The moment I read this article, that was the first thing I tried looking up. I don't see anything yet, and I've let my Plus subscription lapse because of rate-limiting.
Upvote
1
(1
/
0)
Given how expensive every query is computationally, you have to have a pretty rich use case to rely on this at any serious scale.
Upvote
0
(2
/
-2)
