I am blown away by the difference the token amount makes. I am currently using the 8k model, which is almost 2x the size of the gpt-3 model and it makes a huge difference. Token compression aside, the increase in token amount sees an exponential increase in the power of this api. I believe the token limit’s effect on the api is akin to the effect that memory had on the early computer.
If you have any questions on how the new model performs, please ask and I will be happy to answer!
This has also amazed me. While experimenting with some prompts for Nepali language capabilities, I noticed that it may not be perfect yet, but it’s certainly improving.
Temperature is a lot more finicky… I had to lower some values for previous calls to ChatGPT-turbo…
Early results looking very promising, though!
Incredible. Is this difference in relation to the token increase? I’m not familiar with Nepali. Is it a language that GPT can effectively write?
Or, is this just in general from its training?
Are you using few shot examples? Do you mind elaborating?
So i got my invite a few hours ago and switched to gpt-4 immediatly on an existing demo app, below very first impressions:
- extremely slow api (understandable, this is a private beta) it takes 3 times more time than gpt3.5 turbo for identitcal prompt and parameters
- quality of text returned is absolutely amazing in English and latin languages like French Spanish and Italian
- Chokes on arabic, turkish and hebrew (surprising, i was expecting this to be an area of improvement of gpt4 vs gpt3.5, and this was actually the main reason i requested early access to gpt 4)
will continue testing with different scenarios…
Interesting regarding the language issues.
How is it choking? Incorrect grammar? Punctuation? Spelling? Awkwardness?
the grammar is ok, spelling punctuation too
it seems like it doesnt understand the request, it returns part of the prompt in arabic for example instead of executing the instruction, strange things of that kind.
but maybe i should adapt the way the prompts are built,
will test further…
Check your temperature. The slider seems to react different in this model… Good luck!
Moved from a separate thread
I’ve been trying to get GPT 3.5 turbo to return JSON, it succeeded most of the time, but I had to clean up the output behind it, found extra properties not mentioned in system message, and insist on returning empty fields when not relevant.
Just changed model name to GPT4 today, it is perfect all the time, follows instructions to the inch, gives exactly what I asked, no more, no less!!
We are seriously considering using it in one of the products, possibilities are limitless, basically you can develop an API by merely explaining what you want in natural language with perfectly structured response that you can expose as API with reasonable confidence that you won’t be let down.
Future is EXCITING!
It is still infeasible economically for most applications though. Is the cost coming down soon?
Sadly Arabic and knowledge, Arabic culture and Islam for example is extremely subpar, all model including GPT4 keep hallucinating and returning non sense, if I had a bit more time I would have created an eval, it is a very good subject to do an eval on.
Hi @RonaldGRuckus Here is my first impressions and analysis in detail. I am from Nepali so I greatly care about it, Engilish capabilities are just great. GPT-4 Early Access: A Dive into Multilingual AI Capabilities and First Impressions | by Ashok Poudel | Mar, 2023 | Medium
I’m granted access to GPT4 via a prompt requests but I get an error message when using the API (The model:
gpt-4-32k does not exist").
I can’t find an anwser on the forum.
Thanks a lot
Did you get explicit access to the 32k version? Most of us only got access to the 8k version.
Check your email it should say; mine says that I only got 8k for example.
im still waiting on my gpt 4 key, but im so excited