Why don't I have GPT-4 API access?

I am a ChatGPT-plus user from the beginning of ChatGPT. However, it looks like I can only user API calls of ChatGPT-3.5, but not GPT-4. I even put myself on GPT-4 API waiting list a few weeks ago, but I still don’t have access. Anynone knows the reason?

1 Like

Welcome to the forum!

GPT-4 API access is still in limited beta to ensure that the system resourses are kept under control, access to the GPT-4 API is granted to applicants in batches and there is no fixed waiting period.

The hope is that eventually every developer will have access to GPT-4, but for now it’s limited to GPT-3.5 and the 16K variant.

1 Like

It may really take some time, I am on waitlist for some time and its quite sad that we have to wait so long however I have read somewhere that You are more than willing to send another request after a week or two. It sounds logical as you just get a better chance of being noticed right?

For critical applications, ChatGPT 3.5 is not robust enough. For instance, given today’s date, I ask for the date for next Friday, ChatGPT 3.5 not always give the correct answer. About 50% chance gets wrong. Getting date in correct format is critical for a lot of apps, such as scheduling.

Date calculation is far better performed by classical time/date manipulation functions in software, AI is not the most suitable task engine for all needs.

I do have experience with various date/time tools. The challenge we are facing is that we need deal with many variations of date expression by users. It is not possible to handle all these cases in a programatic way. So instance, I need take vacation for the whole July 4th Week. A robust language model should be able to figure out I am talking about the time off of July 3-7, 2023. I don’t think there is a robust date manipulate tools can handle this situation in a programatic way.

Have you tried telling the AI to produce an ISODate? perhaps in JSON format, that should be specific and well documented enough to produce consistent output.

(I know this wasn’t your direct question, but given that the community has no insight to waitlist progress/process, I was curious to test out the model differences)

I ran this prompt with temperature=0 a handful of times

Todays date is July 5th, 2023. What is the date for next Friday? Reply with only the date.

  • gpt-3.5-turbo-0613 consistently replied “July 7th, 2023” (debatably right, since English is so vague)
  • gpt-4-0613 consistently replied “July 12, 2023”. Wrong by every account

Or this prompt:

Todays date is July 5, 2023. Parse the message below and reply with the date range (with specific dates) they mentioned.
I need take vacation for the whole July 4th Week. Thanks for being flexible with this months schedule!

  • gpt-3.5-turbo-0613 consistently replied July 4th to July 10th, 2023.
  • gpt-4-0613 sometimes replied with July 1, 2023 to July 7, 2023. Raising temperature to 0.5 would occasionally return July 3, 2023 to July 9, 2023. but not consistently

In summary, not sure you can expect GPT-4 to be worlds above GPT-3 for this kind of thing. Not sure GPT really is the best tool either, since some part of date related logic comes down to math which GPT isn’t great at. Playing around with your prompt and temperature may get you serviceable results from GPT3.5 in the meantime though.

1 Like

That is super interesting, I need to do a little experimentation.

potentially useful

1 Like

The problem is just date format, but simply the incorrect results from ChatGPT 3.5. I know GPT-4 has much better accuracy for inferencing date from a user query.

Well, if you use the prompt I gave in the example above and use the settings to the side, you can experiment with that as a solution? If you can extract the date in an ISODate format then you can reliably use a string to date conversion function to extract the information and then use traditional software tools to do time and date manipulation.

import json
from datetime import datetime, timedelta

json_data = '''
  "date": "2021-09-25T00:00:00Z"

data = json.loads(json_data)
iso_date_string = data["date"]

date_object = datetime.fromisoformat(iso_date_string)
future_date = date_object + timedelta(days=4)

print("Original Date:", date_object)
print("Future Date:", future_date)