I was a Plus subscriber and upgraded to Pro just to test Codex, and at least in my experience, it’s been pretty underwhelming.
First, I don’t think they got the UX quite right yet. Having to wait for an undefined amount of time before getting a result is definitely not the best, although the async nature of Codex seems to alleviate this issue (that is, being able to run multiple tasks at once).
Another thing that bugs me is having to define an environment for the tool to be useful. This is very problematic because AFAIK, you can’t spin up containers that might be needed in tests, severely limiting its usefulness. I guess this will eventually change, but the fact that it’s also completely isolated from the internet seems limiting, as one of the reasons o3 is so powerful in ChatGPT is because it can autonomously research using the web to find updated information on whatever you need.
For comparison, I also use Claude a lot, and I’ve found it to work really well to find obscure bugs in a somewhat complex React application by creating a project and adding the GitHub repo as a source. What this allows me is to have a very short wait time, and the difference with Codex is just night and day. Gemini also allows you to do this now, and it works very well because of its massive context window.
All that being said, I do understand where OpenAI is going with this. I guess they want to achieve something like a real coworker (they even say that in their promotional videos for Codex) because you are supposed to give tasks to Codex and wait until it’s done, like a real human, but again, IMHO, it’s too “pull-request-focused”
I guess I’ll be downgrading to Plus again and wait a little to see where this ends up.
I agree on the UX. A few basic things seem totally broken.
The flow of connecting a github account works, then disconnects, sometimes doesn't work, sometimes just errors. I can't install things that I could yesterday and my environment is just... broken? I have two versions of a repo and it works in only one.
Speed is a big thing. Not the llm stuff so much, but the setup and everything around it for each step.
Not having search cripples some cases where O3 seems incredible.
but there's a lot of places this feels like it can land tasks that often wouldn't get done. A near infinite army of juniors who can take on the lots of tiny tasks in 15-20 minutes is great. Fix some typos, add a few util functions (a task I have right now running), I even just asked it to add new endpoint to a server and it added it, migrations needed, tests and more and seems alright.
The ideal workflow in a way here is that the people asking for these things get to tag the ticket to codex/whatever, they run off and do the thing, PR lands and discussion and changes happen there, demo envs are setup and then someone can check and approve it.
edit -
To be fair, I also used firebase studio and that was worse. Blank screens, errors in the console, when I refreshed and moved around and got an actual page, it ended up failing to setup firebase. UI for editing and code totally failed after that and the explanations for how to fix it I was linked to I couldn't do.
> AFAIK, you can’t spin up containers that might be needed in tests, severely limiting its usefulness.
This is what's blocking me right now. I couldn't find any documentation on whether they allow Docker-in-Docker which typically means that the answer is "no". Since I'm building an AWS-native app I use LocalStack for end-to-end tests which requires a container engine. Codex not having it is a showstopper.
This might not help you but to the very best of my knowledge localstack can operate over the network just fine and I am pretty sure it has a reset endpoint for zeroing its state (I think it's this https://github.com/localstack/localstack/blob/v4.4.0/localst... )
The other alternative is that I've seen folks mention systemd-nsspawn as a form of isolation if that's what your using docker for (but I've never tried it myself)
I work at OpenAI (not on Codex) and have used it successfully for multiple projects so far. Here's my flow:
- Always run more than one rollout of the same prompt -- they will turn out different
- Look through the parallel implementations, see which is best (even if it's not good enough), then figure out what changes to your prompt would have helped nudge towards the better solution.
- In addition, add new modifications to the prompt to resolve the parts that the model didn't do correctly.
- Repeat loop until the code is good enough.
If you do this and also split your work into smaller parallelizable chunks, you can find yourself spending a few hours only looping between prompt tuning and code review with massive projects implemented in a short period of time.
I've used this for "API munging" but also pretty deep Triton kernel code and it's been massive.
> "Look through the parallel implementations, see which is best (even if it's not good enough), then figure out what changes to your prompt would have helped nudge towards the better solution."
How can non-technical people tell what's "best"? You need to know what you're doing at this point, look for the right pitfalls, inspect everything in detail... this right here is the entire counter-argument for LLMs eliminating SWE jobs...
> How can non-technical people tell what's "best"? You need to know what you're doing at this point, look for the right pitfalls, inspect everything in detail... this right here is the entire counter-argument for LLMs eliminating SWE jobs...
I'm not sure a tool that positions itself as a "programmer co-worker" is aiming to be useful to non-technical people. I've said it before, but I don't think LLMs currently are at the stage where they enable you to do things you have 0 experience in, but rather can help you speed up working through things you are familiar with. I think people who claim LLMs will completely replace jobs are hyping the technology without really understanding it.
For example, I'm a programmer, but never done any firmware flashing with UART before via a USB flasher. Today I managed to do that in 1-2 hours thanks to ChatGPT helping me out understanding how to do it. If I'd do it completely on my own, I'm sure it would have taken me at least the full day to do so, instead of the time it took. I was able to see when it got mislead, and could rewrite/redirect from there on, but someone with 0 programming experience, probably wouldn't have been able to.
I don’t think anyone expects software engineers will disappear and get replaced by janitors trained to proompt. I’m sure experts will stick around until the singularity curve starts looking funny. It’s probably gonna suck to enter the industry from now on, though.
I end up asking the same question when experimenting with tools like Cursor. When it can one-shot a small feature, it works like magic. When it struggles, and the context gets poisoned and I have to roll back commits and retry part of the way through something, it hits a point where it was probably easier for me to just write it. Or maybe template it and have it finish it. Or vice versa. I guess the point being that best practices have yet to truly be established, but totally hands-off uses have not worked well for me so far.
Easily 5-10x or even more in certain special cases (when it'd take me a lot of upfront effort to get context on some problem domain). And it can do all the "P2"s that I'd realistically never get to. There was a day where I landed 7 small-to-medium-size pull requests before lunch.
There are also cases where it fails to do what I wanted, and then I just stop trying after a few iterations. But I've learned what to expect it to do well in and I am mostly calibrated now.
The biggest difference is that I can have agents working on 3-4 parallel tasks at any given point.
For me, it’s not that the actual coding is faster. It’s that you can do other things at the same time.
If I’m writing an integration, I can be researching the docs while the agent is coding something up. Worst case, I throw all of the agents work away while now having done research. Best case, it gets a good enough implementation that I can run with.
At the current capabilities of most LLMs + my personal tolerance for slop, the most productive workflow seems to be: spin up multiple agents in the background to work on small scope, straightforward tasks while I work on something bigger that requires more exploration, requirements gathering, or just plain more complex/broad changes to the code. Review the output of the agents or unstick them when there is downtime.
IMO just keeping an IDE window open and babysitting an agent while it works is less productive than just writing the code mostly yourself with AI assistance in the form of autocomplete and maybe highly targeted oneshots using manual context provided "Edit" mode or inline prompting.
My company is dragging their feet on AI governance and let the OpenAI key I was using expire, and what I noticed was that my output of small QoL PRs and bugfixes dropped drastically because my attention remains focused on higher impact work.
Do you find yourself ditching on the things when they change something important with the new prompt? I don't get how people aren't absolutely exhausted by actually implementing this prompt messing advice when I thought there were studies saying small seemingly insignificant changes greatly change the result, hide blind spots, and even having a prompt for engineering a better prompt has knock on increases in instability. Do people just have a higher tolerance for doing work that is not related to the problem than I do? Perhaps I only work on stuff there is no prior example for, but every few days I read someone's anecdote on here and get discouraged in all new ways.
Not to downplay the issue you raise but I haven't noticed this.
Every iteration I make on the prompts only make the request more specified and narrow and it's always gotten me closer to my desired goal for the PR. (But I do just ditch the worse attempts at each iteration cycle)
Is it possible that reasoning models combined with the actual interaction with the real codebase makes this "prompt fragility" issue you speak of less common?
Can it be used to fix bugs? Because the ChatGPT web app is full of them and I don't think they are getting fixed. Pasting big amounts of text freezing the tab is one of them.
Seriously, everyone should get good at fixing bugs. LLMs are terrible at it when it’s slightly non-obvious and since everyone is focusing on vibe coding, I doubt they’ll get any better.
“As I wrote about in Walking and talking with AI in the woods, ideally I'd like to start my morning in an office, launch a bunch of tasks, get some planning out of the way, and then step out for a long walk in nature.”
Wouldn’t we all want that, but it sounds like you can leave task launching and planning to an AI and go find another career.
Think we've got a long time yet for that. We're going to be writing code a lot faster but getting these things to 90-95% on such a wide variety of tasks is going to be a monumental effort, the first 60-70% on anything is always much easier than the last 5-10%.
Also there's a matter of taste, as commented above, the best way to use these is going to be running multiple runs at once (that's going to be super expensive right now so we'll need inference improvements on today's SOTA models to make this something we can reasonably do on every task). Then somebody needs to pick which run made the best code, and even then you're going to want code review probably from a human if it's written by machine.
Trusting the machine and just vibe coding stuff is fine for small projects or maybe even smaller features, but for a codebase that's going to be around for a while I expect we're going to want a lot of human involvement in the architecture. AI can help us explore different paths faster, but humans need to be driving it still for quite some time - whether that's by encoding their taste into other models or by manually reviewing stuff, either way it's going to take maintenance work.
In the near-term, I expect engineering teams to start looking for how to leverage background agents more. New engineering flows need to be built around these and I am bearish on the current status quo of just outsource everything to the beefiest models and hope they can one-shot it. Reviewing a bunch of AI code is also terrible and we have to find a better way of doing that.
I expect since we're going to be stuck on figuring out background agents for a while that teams will start to get in the weeds and view these agents as critical infra that needs to be designed and maintained in-house. For most companies, foundation labs will just be an API call, not hosting the agents themselves. There's a lot that can be done with agents that hasn't been explored much at all yet, we're still super early here and that's going to be where a lot of new engineering infra work comes from in the next 3-5 years.
It's mind blowing to me how many developers are happy about the developments here.. as if they're going to eventually be paid to just sit there while agents do everything. Ah, work is now so easy!
I think in the success case (still TBD), that it will increase productivity to the point where things that can’t be affordably addressed by software will now be able to be addressed with software.
I expect that anyone who is a skilled dev today will be fine. Expectations and competition might be higher, but so will production and value creation.
I think the demand will come, just as Excel didn’t put finance people out of jobs in aggregate.
I mean, I get what everyone's saying. But, just Devil's Advocate, what would be so terrible about software developers having to find some other line of work?
We've used our software development skills to automate other people out of work for what can be argued to be literally decades. Each time we did it, we certainly expected that the people affected would find other work. New jobs were created. The world didn't end. I honestly don't think it would be that much worse this time.
> It's mind blowing to me how many developers are happy about the developments here.. as if they're going to eventually be paid to just sit there while agents do everything. Ah, work is now so easy!
Well, the optimistic take is that if something gets cheaper to produce (e.g. code), demand for it actually increases.
Now you could argue that any non technical person could just oversee the agents instead. Possibly. Though in my experience, humans like to have other humans they trust oversee and understand important stuff for them.
It is most definitely going to be the unemployment line. When in the history of productivity gains, has it translated to more time for people to do other things that are not work? It always translates to more profits for shareholders and bigger pay for executive class, followed by more work for half the workers to fill up the time opened up by the said productivity gains, and unemployment for the other half.
200 years ago, 80% of Americans worked in farming. 150 years ago, that was still over half. It’s now under 2%.
If you’ve seen the work hours and work ethic of farmers, it’s safe to say that most of those people got other jobs that take far less work than farmers did/do.
Closer to our field, I think we’d have far worse work lives (fewer of us employed and much lower pay) if we had to code everything in assembler still. The creation of more powerful abstractions and languages allowed more of us to become software devs and make a living this way than if all we had were the less productive tools of the early days of computing.
Yeah but if you look to the present... there aren't really any jobs where someone is blissfully wandering the earth delegating tasks. Most of the time I can't even take a walk on calls because someone wants to screen share something with me...
I'd like you to be right, but I live in society where joy at work is often considered antithetical to productivity. No matter how much more productive I get, that space is used to fill in more productivity. We'll need more than tooling to stop this.
If you're building a React app using a popular UI framework, AI will seem like magic at how well it one-shots things.
To the author's point about one-shotting. I think it will be a real challenge pushing an AI coding workflow forward because of this problem. In my experience, AI seems to fall off a cliff when you ask it to write code using more obscure libraries and frameworks. It will always hallucinate something rather than admitting it has no knowledge of how something works.
I've had better success with things like o3 with search, because it can actually go and read docs to help fix problems. It helped me dig through matrix specs, proposals and PRs and while the first suggestion didn't work (I thought it would have done) it ended up finding proof that only part of that got merged and found how to enable the experimental side that allowed the other. The iteration of searching and going through things was incredibly helpful. Probably saved me a good few hours or meant I was able to do this at all.
- it's a GREAT oneshot coding model (in the pod we find out that they specifically finetuned for oneshotting OAI SWE tasks, eg prioritized over being multiturn)
- however comparatively let down by poorer integrations (eg no built in browser, not great github integration - as TFA notes "The current workflow wants to open a fresh pull request for every iteration, which means pushing follow-up commits to an existing branch is awkward at best." - yeah this sucks ass)
fortunately the integrations will only improve over time. i think the finding that you can do 60 concurrent Codex instances per hour is qualitatively different than Devin (5 concurrent) and Cursor (1 before the new "background agents").
btw
> I haven't yet noticed a marked difference in the performance of the Codex model, which OpenAI explains is a descendant of GPT-3 and is proficient in more than 12 programming languages.
This is Open AI's fault (and literally every AI company is guilty of the same horrid naming schemes). Codex was an old model based on GPT-3, but then they reused the same name for both their Codex CLI and this Codex tool...
I mean, just look at the updates to their own blog post, I can see why people are confused.
Google just did it too. "Gemini Ultra" is both a model (https://deepmind.google/models/gemini/ultra/) and their new top-tier subscription plan (a la Open AI's Pro plan). Why is this so difficult?
Confusing people is the best way to get them to throw their hands up, stop thinking critically, and start paying. all businesses do this. Mega corps have resources to enforce clarity, but they dont because theyre stupid? Ill eat my words if thats the case....
The particularly integration pain point to me is about network access, that prohibits several banal tasks to be offloaded to codex:
1. Cannot git fetch and sync with upstream, fixing any integration bugs;
2. Cannot pull in new library as dependency and do integration evaluations.
Besides that, cannot apt install in the setup script is annoying (they blocked the domain to prevent apt install I believe).
The agent itself is a bit meh, often opt-to git grep rather than reading all the source code to get contextual understanding (from what the UI has shown).
It's much more conservative in the scope of task it will attempt and it's much slower. You need to fire and forget several parallel tasks because you'll be waiting 10+ minutes before you get anything you can review and give feedback on.
right now apples and oranges literally only because 1) unlimited unmetered use and 2) not in browser so async and parallel. like that stuff just trumps actual model and agent harness differences because it removes all barriers from thought to code.
Being able to make quick changes across a ton of repos sounds awesome. I help maintain a ton of example apps, and doing things like updating a README to conform to a new format, or changing a link, gets pretty tedious when there are 20 different places to do it. If I could delegate all that busywork to Codex and smash the merge button later I would be happy.
I feel it will get there in short order..but for the time being I feel that we'll be doing some combination of scattershot smaller & maintenance tasks across Codex while continuing to build and do serious refactoring in an IDE...
I'm actually curious about using this sort of tool to allow non-devs to make changes to our code.
There are so many content changes or small CSS fixes (anyway you would verify that it was fixed by looking at it visually) where I really don't want to be bothered being involved in the writing of it, but I'm happy to do a code review.
Letting a non-dev see the ticket, start off a coding thing, test if it was fixed, and then just say "yea this looks good" and then I look at the code, seems like good workflow for most of the minor bugs/enhancements in our backlog.
Even content changes can require deliberate thought. Any system of decent size is probably going to have upstream/downstream dependencies - adding a field might require other systems to account for it. I guess I can see small CSS changes, but how does the user know when the change is small or "small"?
Perhaps the system could tell them 80% of the time and the reviewer catches the other 20%. An easy heuristic that usually would work in this case is lines of code. It's a classically bad way to measure impact / productivity but it's definitely an indicator and this is probably a rare instance where the measurement would not break efficacy of the metric (Goodhart's law) and might actually improve the situation.
Assuming you works as a software engineer, is your day actually just filled with writing what could be solved by a low-code platform? Mine certainly isn't.
Readit News
First, I don’t think they got the UX quite right yet. Having to wait for an undefined amount of time before getting a result is definitely not the best, although the async nature of Codex seems to alleviate this issue (that is, being able to run multiple tasks at once).
Another thing that bugs me is having to define an environment for the tool to be useful. This is very problematic because AFAIK, you can’t spin up containers that might be needed in tests, severely limiting its usefulness. I guess this will eventually change, but the fact that it’s also completely isolated from the internet seems limiting, as one of the reasons o3 is so powerful in ChatGPT is because it can autonomously research using the web to find updated information on whatever you need.
For comparison, I also use Claude a lot, and I’ve found it to work really well to find obscure bugs in a somewhat complex React application by creating a project and adding the GitHub repo as a source. What this allows me is to have a very short wait time, and the difference with Codex is just night and day. Gemini also allows you to do this now, and it works very well because of its massive context window.
All that being said, I do understand where OpenAI is going with this. I guess they want to achieve something like a real coworker (they even say that in their promotional videos for Codex) because you are supposed to give tasks to Codex and wait until it’s done, like a real human, but again, IMHO, it’s too “pull-request-focused”
I guess I’ll be downgrading to Plus again and wait a little to see where this ends up.
The flow of connecting a github account works, then disconnects, sometimes doesn't work, sometimes just errors. I can't install things that I could yesterday and my environment is just... broken? I have two versions of a repo and it works in only one.
Speed is a big thing. Not the llm stuff so much, but the setup and everything around it for each step.
Not having search cripples some cases where O3 seems incredible.
but there's a lot of places this feels like it can land tasks that often wouldn't get done. A near infinite army of juniors who can take on the lots of tiny tasks in 15-20 minutes is great. Fix some typos, add a few util functions (a task I have right now running), I even just asked it to add new endpoint to a server and it added it, migrations needed, tests and more and seems alright.
The ideal workflow in a way here is that the people asking for these things get to tag the ticket to codex/whatever, they run off and do the thing, PR lands and discussion and changes happen there, demo envs are setup and then someone can check and approve it.
edit -
To be fair, I also used firebase studio and that was worse. Blank screens, errors in the console, when I refreshed and moved around and got an actual page, it ended up failing to setup firebase. UI for editing and code totally failed after that and the explanations for how to fix it I was linked to I couldn't do.
This is what's blocking me right now. I couldn't find any documentation on whether they allow Docker-in-Docker which typically means that the answer is "no". Since I'm building an AWS-native app I use LocalStack for end-to-end tests which requires a container engine. Codex not having it is a showstopper.
The other alternative is that I've seen folks mention systemd-nsspawn as a form of isolation if that's what your using docker for (but I've never tried it myself)
- Always run more than one rollout of the same prompt -- they will turn out different
- Look through the parallel implementations, see which is best (even if it's not good enough), then figure out what changes to your prompt would have helped nudge towards the better solution.
- In addition, add new modifications to the prompt to resolve the parts that the model didn't do correctly.
- Repeat loop until the code is good enough.
If you do this and also split your work into smaller parallelizable chunks, you can find yourself spending a few hours only looping between prompt tuning and code review with massive projects implemented in a short period of time.
I've used this for "API munging" but also pretty deep Triton kernel code and it's been massive.
How can non-technical people tell what's "best"? You need to know what you're doing at this point, look for the right pitfalls, inspect everything in detail... this right here is the entire counter-argument for LLMs eliminating SWE jobs...
I'm not sure a tool that positions itself as a "programmer co-worker" is aiming to be useful to non-technical people. I've said it before, but I don't think LLMs currently are at the stage where they enable you to do things you have 0 experience in, but rather can help you speed up working through things you are familiar with. I think people who claim LLMs will completely replace jobs are hyping the technology without really understanding it.
For example, I'm a programmer, but never done any firmware flashing with UART before via a USB flasher. Today I managed to do that in 1-2 hours thanks to ChatGPT helping me out understanding how to do it. If I'd do it completely on my own, I'm sure it would have taken me at least the full day to do so, instead of the time it took. I was able to see when it got mislead, and could rewrite/redirect from there on, but someone with 0 programming experience, probably wouldn't have been able to.
There are also cases where it fails to do what I wanted, and then I just stop trying after a few iterations. But I've learned what to expect it to do well in and I am mostly calibrated now.
The biggest difference is that I can have agents working on 3-4 parallel tasks at any given point.
If I’m writing an integration, I can be researching the docs while the agent is coding something up. Worst case, I throw all of the agents work away while now having done research. Best case, it gets a good enough implementation that I can run with.
IMO just keeping an IDE window open and babysitting an agent while it works is less productive than just writing the code mostly yourself with AI assistance in the form of autocomplete and maybe highly targeted oneshots using manual context provided "Edit" mode or inline prompting.
My company is dragging their feet on AI governance and let the OpenAI key I was using expire, and what I noticed was that my output of small QoL PRs and bugfixes dropped drastically because my attention remains focused on higher impact work.
Every iteration I make on the prompts only make the request more specified and narrow and it's always gotten me closer to my desired goal for the PR. (But I do just ditch the worse attempts at each iteration cycle)
Is it possible that reasoning models combined with the actual interaction with the real codebase makes this "prompt fragility" issue you speak of less common?
Seriously, everyone should get good at fixing bugs. LLMs are terrible at it when it’s slightly non-obvious and since everyone is focusing on vibe coding, I doubt they’ll get any better.
Nudging the UI slightly for this exact flow could generate good training data.
Deleted Comment
Wouldn’t we all want that, but it sounds like you can leave task launching and planning to an AI and go find another career.
This feels so hopelessly optimistic to me, because "effectively away from our desks" for most people will mean "in the unemployment line"
Also there's a matter of taste, as commented above, the best way to use these is going to be running multiple runs at once (that's going to be super expensive right now so we'll need inference improvements on today's SOTA models to make this something we can reasonably do on every task). Then somebody needs to pick which run made the best code, and even then you're going to want code review probably from a human if it's written by machine.
Trusting the machine and just vibe coding stuff is fine for small projects or maybe even smaller features, but for a codebase that's going to be around for a while I expect we're going to want a lot of human involvement in the architecture. AI can help us explore different paths faster, but humans need to be driving it still for quite some time - whether that's by encoding their taste into other models or by manually reviewing stuff, either way it's going to take maintenance work.
In the near-term, I expect engineering teams to start looking for how to leverage background agents more. New engineering flows need to be built around these and I am bearish on the current status quo of just outsource everything to the beefiest models and hope they can one-shot it. Reviewing a bunch of AI code is also terrible and we have to find a better way of doing that.
I expect since we're going to be stuck on figuring out background agents for a while that teams will start to get in the weeds and view these agents as critical infra that needs to be designed and maintained in-house. For most companies, foundation labs will just be an API call, not hosting the agents themselves. There's a lot that can be done with agents that hasn't been explored much at all yet, we're still super early here and that's going to be where a lot of new engineering infra work comes from in the next 3-5 years.
I expect that anyone who is a skilled dev today will be fine. Expectations and competition might be higher, but so will production and value creation.
I think the demand will come, just as Excel didn’t put finance people out of jobs in aggregate.
We've used our software development skills to automate other people out of work for what can be argued to be literally decades. Each time we did it, we certainly expected that the people affected would find other work. New jobs were created. The world didn't end. I honestly don't think it would be that much worse this time.
Software engineers are dumb. Really dumb.
Now you could argue that any non technical person could just oversee the agents instead. Possibly. Though in my experience, humans like to have other humans they trust oversee and understand important stuff for them.
With Codex and Claude Code, these model agents are cars.
Some of horses will become drivers of cars and some of us will no longer be needed to pull wagons and will be out of a job.
Is that the proper framing?
An amusing image, but your analogy lost me here.
If you’ve seen the work hours and work ethic of farmers, it’s safe to say that most of those people got other jobs that take far less work than farmers did/do.
Closer to our field, I think we’d have far worse work lives (fewer of us employed and much lower pay) if we had to code everything in assembler still. The creation of more powerful abstractions and languages allowed more of us to become software devs and make a living this way than if all we had were the less productive tools of the early days of computing.
Deleted Comment
I'd like you to be right, but I live in society where joy at work is often considered antithetical to productivity. No matter how much more productive I get, that space is used to fill in more productivity. We'll need more than tooling to stop this.
Are you pretending that automation doesn’t take away human jobs?
To the author's point about one-shotting. I think it will be a real challenge pushing an AI coding workflow forward because of this problem. In my experience, AI seems to fall off a cliff when you ask it to write code using more obscure libraries and frameworks. It will always hallucinate something rather than admitting it has no knowledge of how something works.
- it's a GREAT oneshot coding model (in the pod we find out that they specifically finetuned for oneshotting OAI SWE tasks, eg prioritized over being multiturn)
- however comparatively let down by poorer integrations (eg no built in browser, not great github integration - as TFA notes "The current workflow wants to open a fresh pull request for every iteration, which means pushing follow-up commits to an existing branch is awkward at best." - yeah this sucks ass)
fortunately the integrations will only improve over time. i think the finding that you can do 60 concurrent Codex instances per hour is qualitatively different than Devin (5 concurrent) and Cursor (1 before the new "background agents").
btw
> I haven't yet noticed a marked difference in the performance of the Codex model, which OpenAI explains is a descendant of GPT-3 and is proficient in more than 12 programming languages.
incorrect, its an o3 finetune.
This is Open AI's fault (and literally every AI company is guilty of the same horrid naming schemes). Codex was an old model based on GPT-3, but then they reused the same name for both their Codex CLI and this Codex tool...
I mean, just look at the updates to their own blog post, I can see why people are confused.
https://openai.com/index/openai-codex/
Edit:
Google just did it too. "Gemini Ultra" is both a model (https://deepmind.google/models/gemini/ultra/) and their new top-tier subscription plan (a la Open AI's Pro plan). Why is this so difficult?
1. Cannot git fetch and sync with upstream, fixing any integration bugs; 2. Cannot pull in new library as dependency and do integration evaluations.
Besides that, cannot apt install in the setup script is annoying (they blocked the domain to prevent apt install I believe).
The agent itself is a bit meh, often opt-to git grep rather than reading all the source code to get contextual understanding (from what the UI has shown).
I feel it will get there in short order..but for the time being I feel that we'll be doing some combination of scattershot smaller & maintenance tasks across Codex while continuing to build and do serious refactoring in an IDE...
There are so many content changes or small CSS fixes (anyway you would verify that it was fixed by looking at it visually) where I really don't want to be bothered being involved in the writing of it, but I'm happy to do a code review.
Letting a non-dev see the ticket, start off a coding thing, test if it was fixed, and then just say "yea this looks good" and then I look at the code, seems like good workflow for most of the minor bugs/enhancements in our backlog.
This almost seems like this is a funnel to force people to become software engineers.
Like:
- When making CSS changes, make sure that the code is responsive. Add WCAG 2.0 attributes to any HTML markup.
- When making changes, run <some accessibility linter command> to verify that the changes are valid.
etc.
The non-dev doesn't need to know/care.