Hacker News new | past | comments | ask | show | jobs | submit login
Decabytes 16 days ago | hide | past | favorite

This title implies this is a new update to gpt-3.5-turbo, when actually it's an old article from the start of 2022 before GPT 3.5 was even a thing

Edit: I can't add any further replies I guess becase this was flagged, but it seems that gpt-3.5-turbo-instruct is a real thing, I got an email about it. This was one of the links in the email so I see how this mistake was made

There's literally nothing about "turbo" as the title mentions either, since it didn't exist yet

Is it me or is it somewhat disingenuous that OpenAI is comparing it to their GPT-3 model, and not even the GPT-3.5?

They say:

> InstructGPT is better than GPT-3 at following English instructions.

EDIT: Color me dumb. It's an old article.

It's an old article from the start of 2022

Its from January 2022

GPT-3.5 was a game changer when it came out. Blazing fast and really good.

An interesting thing to note is what damage was done to the model's intelligence when we did this. That shows up in the GPT-4 release docs https://openai.com/research/gpt-4:

> Interestingly, the base pre-trained model is highly calibrated (its predicted confidence in an answer generally matches the probability of being correct). However, through our current post-training process, the calibration is reduced.

I wonder if it was the instruction tuning or the safety stuff that lobotomized it.

Applications are open for YC Winter 2024

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact