Coding requires a mindset shift that the -codex fine-tunes provide. Codex will do all kinds of weird stuff like poking in your ~/.cargo ~/go etc. to find docs and trying out code in isolation, these things definitely improve capability.
The biggest advantage of codex variants, for me, is terseness and reduced sicophany. That, and presumably better adherence to requested output formats.
Thanks, that's what I acknowledged in the message you just replied to.
I'm not blaming anyone. Mostly outlining who was impacted as it's not really related to the front-end parts of the framework that the initial comment was referring to.
It's not entirely JavaScript but it is partially due to some of the language's history and culture: prototype pollution wouldn't be possible in every other language and not everyone has culture around things like decoding payloads in an exploitable manner (e.g. in the Python world some people used to decode pickled objects but it was always frowned upon; the Java world has had debates over the years about this). The big one which is unique to JavaScript is the culture around client-side execution and mixing code running between the two environments, which means you have a lot of machinery setup to execute code on the server and/or clients, making it both easy to have confusion around the execution context in ways which have been exploited and encouraging people to do things like ship complex objects between the two which programmers using other backend languages wouldn't consider because they never had the possibility of running directly in the browser.
I played around with it and I have to agree. It treats the user like they're dumb, regardless of what personal context you provide it. I found myself backtracking constantly to invalidate it's assumptions, to the point that I gave up. All of that within like 4 hours of bothering to touch it in the first place.
I genuinely can't imagine allowing these things to run commands on a machine. If I ever found out a colleague was doing that I would want them fired.
> Give it a JSON schema, inject a '{', and sometimes do a bit of custom parsing on the response
I would hope that this is not what OpenAI/Anthropic do under the hood, because otherwise, what if one of the strings needs a lot of \escapes? Is it also supposed to newer write actual newlines in strings? It's awkward.
The ideal solution would be to have some special tokens like [object_start] [object_end] and [string_start] [string_end].
The model is very hard to work with as is.
reply