GPT-4 with browsing model is actually GPT-3?

[fading solar]

Bug Report:

Steps to reproduce:

1. Select the GPT-4 with browsing model and write anything.

Expected result:

A GPT-4 level response

Actual result:

A GPT-3 level response

Additional information

Browser: Chrome
OS: Windows 10

So I apparently have access to the GPT-4 with browsing model and I noticed the answers to the questions I asked looked extremely similar to the 3.5 with browsing version, so I asked what model it was based off of and it said GPT-3. I hope this is a bug and not an attempt to pass GPT-3 as GPT-4. (See image below)

[gentle temple]

sup my geeps

[fading solar]

🤨

[feral crater]

had the same issue, will star this

very odd results testing reflexion on this browsing model, too

i asked 3.5 legacy, 4, and 4 w/ browsing to "Please write me a long poem where all words start with the letter M"

none of them were capable of doing it. 3.5 really struggled. 4 did quite well, only missing a few words. 4 w/ browsing did better than 3.5, but not by much

however, 4 was the only model to respond properly to reflexion. i asked them each to "Please reflect upon your answer. Do you feel that you followed instructions?"

both 3.5 and 4 w/ browsing said they followed the instructions.

4 correctly said: Upon reflection, I realize that I did not fully follow the instructions. Although the poem features many words that start with the letter 'M,' not all of the words begin with 'M.' I apologize for the oversight and any confusion it may have caused. If you would like, I can attempt to write another poem adhering more closely to the given instructions.

[feral crater]

@tribal hatch not meaning to pester, but who normally handles the bug reports channel?

[feral crater]

bump

[twilit star]

Wont chat GPT-4 be based on the GPT-4 architecture which had been based of off its predecessor which im assuming is GPT-3?

[feral crater]

gpt-4 knows it's gpt-4

my current hypothesis is that gpt-4 is connected as an intermediary agent. we aren't actually interacting with 4 directly with this new browsing model. it is lurking in the background and it is running the browsing logic

so using this model to do anything other than browsing is a bad idea

[feral crater]

bump

[rain steppe]

I’m having the same issue

[winter hedge]

Same issue

[feral crater]

might make sense to rename this model -- seems a bit misleading imho

[feral crater]

if this is doing a flow of user -> gpt-3.5 -> gpt-4 (lurking, handling browing logic) -> gpt-3.5 -> user this should be called Web Browsing (GPT-4 Assisted)

[twilit plume]

Did web browsing get moved out of Alpha? When you start a new chat with Web Browsing it now says at the bottom it's GPT-4 and the responses now have a black avatar. However it still replies that it is running on GPT3.

[ocean gust]

It’s in gpt4

[twilit plume]

If you ask it what model it runs on it replies GPT3

If you ask GPT4 without browsing what model it runs on, it says GPT4

[ocean gust]

That means nothing

It’s just an earlier version in gpt4

It used to say that for normal gpt4 also

[twilit plume]

Which earlier version?

[ocean gust]

Gpt4 right when it was released

Also the api did that

[twilit plume]

The march 14th snapshot then?'

[ocean gust]

Maybe

[twilit plume]

I thought you knew for certain thats why i was asking

Probably running on that snapshot

[feral crater]

im still skeptical -- sam altman made it clear that they are looking at creating multi-model offerings instead of brute forcing params, cpu, and dataset to create bigger models

i think this is 3.5 talking with a gpt-4 browsing agent

i dont think the march 14th model is as weak as the browsing model at logic and complex reasoning

[twilit plume]

Looks like they are still changing things. UI just updated again for me

[feral crater]

yeah same

30 responses per hour btw -- nice bump in usage limit!!