#chat GPT+ Subscriber, GPT4 is selected but GPT3 model is being used.

29 messages · Page 1 of 1 (latest)

willow shell
#

Bug Report:

Steps to reproduce:

  1. Select the GPT-4 model

Expected result:

A GPT-4 level response

Actual result:

A GPT-3 level response

Additional information

Browser: Chrome
OS: MacOS latest

I have access to the GPT-4 with GPT-3.5 as default version, so I asked question & GPT-4 responds as GPT-3 (See image below)

snow mantle
#

how is this not GPT-4 output?

willow shell
#

At the bottom it says chat gpt may 3 2021. It used to be closer to this date and said chat got 4

snow mantle
#

if you ask it which model it is, does it say it is gpt-4?

#

may 3 version is different than gpt-3

willow shell
#

It said gpt3

willow shell
snow mantle
#

weird! tried mine and it gave me 4

crude shoal
#

Pretty sure GPT3.5 uses green icon, GPT4 uses black.

#

I think you're on GPT4. Both GPT3.5 and GPT4 will mention the knowledge cutoff date of September 2021 that you showed in your original post.

willow shell
crude shoal
snow mantle
woven imp
#

Y’all don’t understand anything….

#

If it’s black it’s indeed using gpt4

#

Sometimes it cannot recognize what model it is

snow mantle
#

yeah agree i haven't been able to recreate gpt-4 telling me it's gpt-3

#

i have only been able to do that with the browsing gpt-4 alpha -- in which this model always responds that it is gpt-3

woven imp
#

It’s a very old gpt4 model

#

For sure

crude shoal
# snow mantle

Ya that's browsing with GPT-4, i get the same response with that. I think the browsing plugin might have bug where it's not using GPT-4 for responses or it's an older version of gpt4. It also has a green avatar not black.

snow mantle
# woven imp It’s a very old gpt4 model

i believe we aren't interfacing with GPT-4 at all -- GPT-4 is running in an isolated mode where all it does is handle the browsing logic.

user -> gpt-3.5 -> gpt-4 (handles browsing logic after reading conversation, passes results back) -> gpt-3.5 -> user

#

sam altman just talked about this yesterday being one of their internal strategies. instead of chasing giant powerful models, they are trying smaller specialized models that interact with eachother

#

so i think the green avatar gives it away -- this isn't gpt-4 at all that we get to talk to. it's further evidenced by basic testing