> "We find that comments by GPT‑5-Codex are less likely to be incorrect or unimportant" -- less unimportant comments in code is definitely an improvement!
This seems to be a misunderstanding. In the original OpenAI article, comment here is about code review comment, not comment in code.
pietz 82 days ago [-]
Do we really know that gpt-5-codex is a finetune of gpt-5(-thinking)? The article doesn't clearly say that, right?
I suspect that this is smaller than gpt-5 or at least a quantized version. Similar to what I suspect Opus 4.1 is. That would also explain why it's faster.
simonw 82 days ago [-]
OpenAI say:
"Today, we’re releasing GPT‑5-Codex—a version of GPT‑5 further optimized for agentic coding in Codex."
So yeah, simplifying that to a "fine-tune" is likely incorrect. I just added a correction note about that to my article.
pietz 81 days ago [-]
Thank you for your work, Simon.
lostmsu 83 days ago [-]
The pelican is not very good
TiredOfLife 83 days ago [-]
But probably fast
AstroBen 83 days ago [-]
Would be faster if it got on the bike
knowsuchagency 83 days ago [-]
It's annoying to see a link to a Theo video -- same guy who went with Simon to OpenAI's GPT-5 glazefest and had to backpedal when everyone realized what a shill he is.
I know neither of them are journalists -- I'm probably expecting too much -- but Simon should know better.
Ancapistani 83 days ago [-]
While not a journalist, Simon definitely has a background in journalism.
He was one of the original authors of Django, back when it was a “web framework for journalists with deadlines”.
knowsuchagency 83 days ago [-]
Exactly. That's why I said he should know better. He never should have gone to that event to hype GPT-5 under the guise of "testing" it out.
simonw 82 days ago [-]
I did actually consider that quite a bit when I got invited to OpenAI's mysterious recorded launch event (they didn't tell us it was GPT-5 until we got there) - would it damage my credibility as an independent voice in the AI space?
I decided to risk it. Crucially OpenAI at no point asked for any influence over my content at all, aside from sticking to their embargo (which I've done with other companies before.)
doctoboggan 82 days ago [-]
Is it possible that open ai let you test a private version of GPT-5 that was better than what was released to the public, like the previous commenter claimed?
simonw 82 days ago [-]
They changed the model ID we were using multiple times in the two weeks we had access to - so clearly they were still iterating on the model during that time.
They weren't deceptive about that - the new model IDs were clearly communicated - but with hindsight it did mean that those early impressions weren't an exact match for what was finally released.
My biggest miss was that I didn't pay attention to the ChatGPT router while I was previewing the models. I think a lot of the early disappointment in GPT-5 was caused by the router sending people to the weaker model.
For what it's worth, the GPT-5 I'm using today feels as impressive to me as the one I had during the preview. It's great at code and great at search, the two things I care most about.
beng-nl 82 days ago [-]
This seems to me like a very harsh take on Theo’s motivations. I don’t know him beyond what I’ve learned from his videos, but given occams razor I’m inclined to believe him: gpt5 seemed much better during the private demo than the public release. There are many possible explanations but jumping to ‘shill’ (implying deception) seems uncalled for.
rolymath 82 days ago [-]
Literally the only channel I've ever blocked on Youtube.
ath3nd 83 days ago [-]
[dead]
Rendered at 17:47:13 GMT+0000 (UTC) with Wasmer Edge.
This seems to be a misunderstanding. In the original OpenAI article, comment here is about code review comment, not comment in code.
I suspect that this is smaller than gpt-5 or at least a quantized version. Similar to what I suspect Opus 4.1 is. That would also explain why it's faster.
"Today, we’re releasing GPT‑5-Codex—a version of GPT‑5 further optimized for agentic coding in Codex."
So yeah, simplifying that to a "fine-tune" is likely incorrect. I just added a correction note about that to my article.
I know neither of them are journalists -- I'm probably expecting too much -- but Simon should know better.
He was one of the original authors of Django, back when it was a “web framework for journalists with deadlines”.
I decided to risk it. Crucially OpenAI at no point asked for any influence over my content at all, aside from sticking to their embargo (which I've done with other companies before.)
They weren't deceptive about that - the new model IDs were clearly communicated - but with hindsight it did mean that those early impressions weren't an exact match for what was finally released.
My biggest miss was that I didn't pay attention to the ChatGPT router while I was previewing the models. I think a lot of the early disappointment in GPT-5 was caused by the router sending people to the weaker model.
For what it's worth, the GPT-5 I'm using today feels as impressive to me as the one I had during the preview. It's great at code and great at search, the two things I care most about.