codex-discussions

[karmic gulch]

Yes, love it. Stopped using Claude Design after I discovered it

Download and try it, works with pretty much all LLMs and alsoconnects to auth

its a decent fork

[woven canyon]

What model do you use in it?

[karmic gulch]

Kimi atm

Did some with Codex, but I fancy Kimi more on some UI stuff

[woven canyon]

I see

[unique spade]

the change in token output for same model in 2 different harnesses is because the harness can alter the inference dynamics. as most basic example, you can significantly alter that by the system prompt

[solemn acorn]

not system prompts so much as tool descriptions and behavior

[bright swift]

also tool output limit etc.. for example opencode automatically prunes old tool output, which keeps context smaller but reduces cache hit rates

[unique spade]

was a basic example. if i tell in system prompt to the model to verify 3 times every step...what do you think will happen with token usage lol

codex should give more love to windows haha, after all msft is big investor...all the cool features always come first only to mac

[solemn acorn]

windows is more of a pain to develop for than macOS

[unique spade]

i'm using wsl, not windows for codex

[solid lake]

They should give Linux more attention then given msft runs on it

Also they’re backing away from Microsoft

[unique spade]

if only windows would not have stuff like powerpoint, if it would run on linux as you say lol

[boreal holly]

OpenAI should develop all features on macOS and Linux, and have Windows users do X11 forwarding to WSL to get the GUI app and stuff. That'd definitively solve every issue

[unique spade]

in what sense is windows running on linux?

[solid lake]

Libreoffice!

[boreal holly]

Stability & performance mostly

[unique spade]

bro windows is not linux based os

if it would, it would be better

you have different file system, different everything

[boreal holly]

I said run the apps in WSL and use X11 forwarding

[uneven kayak]

Linux runs on Windows, WSL is like the best thing Microsoft ever did 😂 jk

[unique spade]

yea but you joined in the midst of

which i m not sure what that means

[uneven kayak]

I think I'm gonna have to switch to the CLI to keep working on my game, been having nothing but problems with Windows desktop apps, both Codex and Claude keep freezing up

[boreal holly]

100% of Microsoft Azure servers run Linux

[solid lake]

https://www.microsoft.com/en-us/windows-server/blog/2015/05/06/microsoft-loves-linux/ this should be a good starting point

[boreal holly]

It's actually kinda genius. Microsoft makes Windows Server, but they honestly prefer not Windows for serious computer usage

[solid lake]

Reset when ;-;

Queue the meme

[unique spade]

yup if that was the point that's ok, thought it was about windows since that was the topic

[unique spade]

does anyone still uses windows server? i think that used to be a thing like 20 years ago

[solid lake]

Chatgpt (app) is hosted on azure

Even the api I think

[solid lake]

They extended to AWS too

Which also extensively uses Linux

[proper cosmos]

what is cursor

[boreal holly]

It's like VSCode for LLMs

[boreal holly]

I really hope not lol

[solid lake]

Here more info on the Linux usage within azure

[boreal holly]

I feel like the fact there's no credible public precentage is because it's not 100% and that's kinda embarassing

[silver smelt]

anyone else notice GPT 5.5 get noticeable much slower in the last few days than like a week ago I'm on the $100 plan

[boreal holly]

Is it noticeably slower on a new session or only long running ones?

[quick thorn]

Ai has regressed to aol levels, I put in a prompt and literally walk away knowing it’s gonna take so long

[bright swift]

skill issue

[plucky cobalt]

The reset happened and usage shows 97% left instead of 100. Anyone else has faced this issue?

[raw bobcat]

i didnt even get the codex reset yet :<

[rocky fog]

some companies do
people usually dont, because you literally have to pay monthly for its license
usually its just a waste of ram for the os though 🤣
(individual might just run win 10/11 pro if he needs windows, although licensing wise, probably shouldnt, you can debloat it to under 2gb ram, sometimes you just gotta build in windows though)

[solid lake]

Fast xhigh till then

[trim rapids]

active directory is still a thing for large orgs

[silver smelt]

Any session

[lost drum]

reset twice?

[rocky fog]

yep reset 😂 (again)

probably due to all the reset issues 🤣

and the 5h window keeps moving again
so its just resetting over and over again
well, I am not complaining

[fallow heron]

I think its about the best coding harness, they have their own model. It's more designed around serious coding than vibe coding but you get api rates for major modes at 50% api rates. super expanded proprietary model and for 60 to 100 bucks enough high tier model but it gets expensive if you are not conciencous. It feels closer to the metal, less insulated.

What is this reset i keep hearing about? good or bad?

[jagged gale]

recent build was exposing EXACT minutes of compute time now it updated and its gone 🙁

[plush harbor]

company I used to work for had windows servers, as some guy who used to work there had a pet .net programmer and he'd get him to write him little undocumented custom apps all the time. This ended as well as you'd expect

I think I just described vibe coding before vibe coding was a thing

[uneven kayak]

Cursor is expensive, that's what it is lol

I spent $52 in a single session with Opus 4.7 on Cursor that probably would have only used like 2% of my weekly Claude Max usage. Paid $200 for Cursor Ultra and got like four days of use for the entire month, it's a ripoff

Anyway, I'm done with Windows. Codex and Claude Code are giving me nothing but headaches so I'm trying Linux. Not using a VM though, gonna dual boot it like I did 10 years ago lol

[boreal holly]

Heck yeah I hope it goes smoothly

[proper cosmos]

Everyone's tokens for the week got reset 2 days early...twice

Such lovely timing, I almost made a 2nd account ahahaha

[plush harbor]

heh, I just YOLOed to prod and am finding all the issues live. The curse of having a lot of users and being contactable

[bright swift]

go for it, use xhigh + subagent orchestration + codex review loop for max burn

[plush harbor]

I changed the site backend storage and the visuals at the same time. I made it black/gold/red so its been a mix of "oh that's awesome" and WHERE LIGHT MODE. So I made light mode ...

[bright swift]

dont listen to the medium reasoning nerds

[karmic gulch]

Anyone got twice the reset? Already used 10 % after first reset, then boom back to 100 % again, not complaining

[rocky fog]

yes

[karmic gulch]

I can spit some bars on low reasoning if you got a min or two 😁

[signal tapir]

Weren't we supposed to get a reset tomorrow?

[cedar skiff]

it will be 7 days from the first turn after the reset

[bright swift]

so dramatic

[hollow reef]

Codex usage got NERFED hard today or what

[signal tapir]

One single action on 5.5 xhigh

Feels a bit much, but not enough to feel certain something has changed

[karmic gulch]

Just curious, what are you using xhigh for? Hope its just planning and not actually implementing …

That’s a true waste of usage

[signal tapir]

Just a test

I usually use 5.4-mini

[cedar skiff]

what plan o.0

[signal tapir]

Teams

I think it's a tad smaller than plus

[jolly lily]

Which model do you guys usually stick to and when do you switch to 5.5?

5.5 inddeed seems to use a lot of tokens even over easy tasks

[karmic gulch]

For pure implementation from a good plan I find that codex 5.3 is top notch when wanting to be cheap on usage

[hollow linden]

Does anyone know what "Sources" refers to in the new thread panel in the codex app?

[jolly lily]

Thank you!

[fallow heron]

nice, unfortunately I finished a project and my brain is almost empty, planning but no real codexing 🙂 feels nice to not be grinding for a couple days.

[karmic gulch]

My setup consists of Claude and Kimi MCP workers within Codex, but for purely OpenAI models, my take is to have GPT-5.5 High handle planning, using xHigh for more sensitive work, then let GPT-5.3-Codex do the implementation. After that, GPT-5.5 reviews the code. If you instruct GPT-5.5 to be detailed and specific about the scope, GPT-5.3-Codex performs well, and you save a lot of tokens. You can instruct 5.5 to delegate some smaller tasks to codex spark, that have its own usage, to save more. But careful, ONLY simple small tasks for that hallucinating speedster

[weak notch]

What would you do if you built an algorithm that could make AI model usage up to 33,600× cheaper?

It also makes it possible to run large AI models on regular consumer GPUs instead of relying on enterprise GPUs that can cost $50,000+ each.

That’s the question I’m stuck on right now. It almost sounds too good to be true, but I’ve already written 300+ tests, and every step is backed by reproducible evidence.

[jolly lily]

Oh so rather than using 5.5 for execution, you use it for planning, if I'm not wrong. But does it actually help?

[fallow heron]

keep it to myself, hire 33,600 developers with free one ai subscription/ 20 dollar level, to pump out a half million developers work and rule the world

[cedar skiff]

Youre not alone, there have been plenty of ppl sliding throught the channels saying claims like yours. One minute model trianing. AGI with emotions. 1% token usage. Usually see them in the claude code general channel more than here though.
Publish a paper...

No one ever does

[karmic gulch]

’m what you could call a power user. I manage around 15 projects at once, so I have to stay structured, organized, and mindful of usage, even though I’m on the $200 Codex plan and also use Claude and Kimi.

For me, an agentic workflow with parallel agents and different models for different tasks is the go-to approach for control, progress, and overall project management.

If you’re somewhat new to this, I suggest telling Codex «Create an agents.md file that says I want GPT-5.5 to handle planning and delegate tasks to the OpenAI model that best fits each task, while staying cost-efficient without compromising quality. Always end with a code review.» This simple sentence is a good start, and after a while you will tune your agents.md and get it even better and see that your usage is going down while the quality goes up = )

[jolly lily]

I'm really glad, thanks!

[weak notch]

I don't know; publishing an article doesn't make much sense right now. I'm just trying to figure out what task I should set for this to become a clear proof. In my case, this isn't about training LLM completely from scratch, or in a minute. I'm talking about how this algorithm allows you to work with LLM weights without running the entire LLM in the usual interfernce way.

[weak notch]

Based on my current test evidence, I’m running an 8B LLM workflow at ~782M scalar ops/sec, with less than 8% GPU load and less than 10% CPU usage.

[cedar skiff]

You can't expect anyone to take your work seriously until you prove it.

[weak notch]

This is the question, what can I do as proof without revealing the algorithm?

[cedar skiff]

Go make a product and undercut the SOA leaders by a big enough margin that you gain market share and just iterate just under the leaders at every turn, run it on a computer in the corner of your room.

[torpid trout]

No wonder gpt crashes at times, I mean, they probably be running it on auntie‘s old compaq in the attic

[solid lake]

Show it working

[boreal holly]

what if you run it on xhigh with /fast tho

[torpid trout]

Right I mean it takes but like a videocassetterecorderdevice or something like that
But probs that’s gonna disclose the Venusian algorithm secret sauce

[weak notch]

This doesn't work with LLM providers. It works with pure LLM weights

[boreal holly]

Ah, pure unsullied LLM weights. organic

[solid lake]

Then use it on the open source deepseek model

Or kimi

Or even oss

[weak notch]

Currently only verified at https://huggingface.co/Zyphra/ZAYA1-8B

[solid lake]

There’s many 8b models

Usually they do testing across different models

[cedar skiff]

scale it to 1.5 trillion

[weak notch]

It's much more than it seems.

[solid lake]

Then publish

[boreal holly]

Well hey, I'll give credit where credit is due. It is real 🙃 don't see that very often! But I do think it probably belongs in #ai-discussions where they talk about different AIs. Probably would be well received over there

[cedar skiff]

what is much more?

[weak notch]

The models are not identical, and I don't think it would be wise to describe what I do. But this isn't a one-size-fits-all recipe for all models. I haven't gotten to such large models yet. I started with small ones and have only gotten up to 8B MoE so far.

[boreal holly]

Wait a minute, you said there was no research paper but this ZAYA1-8B has a research paper https://arxiv.org/abs/2605.05365

U blowin smoke rn?

[solid lake]

His method is applicable to that model is what he meant

Not that he made that model

[weak notch]

I'm talking about how I launch and use the model, not the model itself.

[solid lake]

Why doesn’t publishing an article make sense

[weak notch]

What's the point if I open the playback but don't get any benefit?

[solid lake]

You’re telling us you have something that could help all ai labs that would credit you if it could help them and you don’t want to do it?

All= closed and open source

[torpid trout]

Hey guys I invented cold fusion but I’m not sure what to do with it.

Don’t want to show how I do it because I want some profit but how?

[weak notch]

I'm not yet confident enough that this works the way I think it does. I'm still testing and verifying it. I'm not confident enough that what I'm doing is even possible. It's too good to be true. So before that, I'm trying to find something that will refresh me.

[solid lake]

Well jokes aside he did make big claims

Then you need to use different models

[unique spade]

for me these claims are easy to validate. if someone found some amazing algo to improve how ai reasons

then arc-agi3 should be a piece of cake

which is open contest. so you just go do 100% opn arc-agi3 and you become famous overnight

[torpid trout]

As a rule of thumbs - when you think you’re the first with the idea or the one publishing it… be sure there’s been plenty of others in positions with leverage and sources having had the same idea

Specially in this particular field.

When have you last heard of the backyard inventions gone famous? Other than some tiktok streamer I haven’t.

[cedar skiff]

what if the improvement isnt intelligence, but instead performance efficiency?

[weak notch]

That's interesting. But first, I'll try running this model: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro . That'll make more sense.

[boreal holly]

I would say OpenClaw 🤡

[boreal holly]

Perfect model for consumer hardware

[unique spade]

that s even easier to prove. i mean if you can put vanilla 5.5 and you enhanced 5.5 (think spark)

give them the same task 100 times

and they produce statistically similar results

but your does it 33,600x times faster

[torpid trout]

Not really new tho?
Just momentum

[unique spade]

😂

everyone would believe you

[cedar skiff]

You cant do this, openai is closed source that s even easier to prove. i mean if you can put vanilla 5.5 and you enhanced 5.5 (think spark)

[unique spade]

come one, use deepseek or any other oss model

you got the idea

it s the easiest thing to prove improved efficiency, if you claim you don t affect performance

[cedar skiff]

I mean this is obviously the way forward if claims are true. Not this then arc-agi3 should be a piece of cake

[boreal holly]

FWIW it's difficult to even get a 8B model to output 1000 tok/s. The black magic they're doing with spark is on a whole other level

[unique spade]

arc-agi is about reasoning quality, obviously

[cedar skiff]

and yet you still suggesteed at as a way to prove the performance efficiency

[unique spade]

there aren t many axis you can improve models, point is each has quite clear proof standards

[weak notch]

I'm still in the process. Thanks everyone for the advice. I'm not making any statements yet. I'm just guessing based on my testing results. But I don't trust them. I'm strengthening them and implementing more stringent validation of each step so I can't fool myself.

[unique spade]

read again. i was expliclty referring about "improving reasoning"

not saying this specific method was the way to validate the specific claim of the guy

which i added afterwards, since you asked about it

[cedar skiff]

Right but the contxt of the conversation is about performance efficiency. Now some how you expect that your comment is not in that context? OFC it was you just missed the mark by 10 miles.

[unique spade]

bro you think you are some kind of moderator at a conference? :)))

i chipped in with something that connected at a more general level on how a different axis in improving ai would be easy to validate

[cedar skiff]

Youre happy to sit around and tell people they arent qualified to talk about subjects, the irony is real. I can tell the point is missed on you so i wont bother again.

[turbid wing]

omg

i thought it was a bug

[unique spade]

that s from earlier today right

sry about that 😉

[worn folio]

bro my current run with Codex is getting towards 24 hours 😭

[hard drum]

WOOOH

@cedar skiff now that limits are reset, maybe you could tell me what i could do, on a vanilla Codex, to basically redo OAL, so that i stop running into major usage issues?

[cedar skiff]

I don't really know, all i do is temper my work rate on the automation. I get it to use 5.3 for implementation, for audit/research work i use 5.4 mini i make sure the prompt is strong enough to support it. So if i know there is a large refactor for something my process would be.
Formulate a prompt for 5.4 mini to do the audit, batch 5 files per instance and start subagents 6 at a time to do the audit. The result is a list of files that need the work. Then get 5.5 high/med to orchestrate 5.3 to do the actual work.
How much work a 5.3 agents gets is based on how hard the task is.
Usually I do this in the moment, i do have some works flows that are always the same though. Writing bulk integration tests and unit tests never changes.
I have been meaning to build a skill to do skill refinement loops. But it is pretty easy to just tell 5.5 what to do takes about 5 minutes to get it started.
I guess mostly what i do is kind of specific and because i just yap into the mic and i have done it so many times the process is easy.
TLDR: use 5.3 for repeatable implementation tasks, use 5.4 mini for repeatable audits and code base searches. Use refinement loops on a small subset of what you are targeting to get that working.
Temper your usage and accept what we have will never be enough - we will always want more.

[hard drum]

what do you use on the primary side? the top-level? the one you see on statusline?

&& on what effort levels do you have your things at?

[cedar skiff]

I manage it in the moment, 5.4 mini is always high or extra high, 5.3 is always high.
For my general work that i do in the moment i use 5.5 high/med.

I never use extra high on 5.3 or 5.5

I just use the codex usage screen here, i look at it constantly lol. I'm always trying to manage it so i have enough but not toto much incase of a reset.

https://chatgpt.com/codex/cloud/settings/analytics#usage

[hard drum]

do you ever use /goal, && if so, what's your primary model for it?

at this point i should conduct a proper interview for you

[simple star]

/goal has a terrible flaw

[hard drum]

what's the alternative?

[simple star]

If the thread is blocked because, say for instance, a credential is missing, that is necessary, it just keep in loop, hardening and doing unnecessary stuff

Meaning it is not really moving forward, but at the same time it is not communicating what it needs

[hard drum]

interesting

[simple star]

I only figured this out because of /side

But this doesn't work... /goal needs the ability to pause, to ask for stuff

[cedar skiff]

I have been using it for tasks lately, ones that are big enough that they will take me a while but not so big i can justify building orchestration

I used it a few times to refine skills

[cedar skiff]

bro did not wait for confirmation o.0

[oak trellis]

Reset came ?

[fleet geyser]

yeah mine got reset

[digital temple]

Hi there 👋 Anyone experiencing a "We’re currently experiencing high demand, which may cause temporary errors." error for the last 10 minutes, using external llm provider? Especially Openrouter (BYOK) and Deepseek

I noticed some users had this issue in the past with models from OpenAI, but what about external models?

[uneven kayak]

Nice

[nocturne folio]

so if we get a new gpt 5.x model for about every month, will gpt6 come out in 5 months?

[proud remnant]

I've been vibecoding with codex for 3 days, at least 14h per day and I've successfully achieved.. a sleeping disorder!
Niiiice

[nocturne folio]

switch to uberman

youll love it

really maximizes usage

[plush harbor]

also don't forget to eat

[zinc field]

hello! i have a issue with codex, i was programming with it, took a break of about an hour, came back just to see that my 5-hourly usage was completely consumed?

[grim latch]

I'm trying to use Codex with WSL. With WSL environment and terminal.
But it is not as nice as VSCode/Cursor/Zed using a remote connection.

[plush harbor]

you could always ask it what it was doing

[sonic adder]

Codex mobile doesn't work or show up

I'm an iOS and MacOS user experiencing issues with Codex on my Teams plan. The Codex popup isn't appearing, despite having the latest versions of both ChatGPT and Codex. Is this a rollout delay or a bug?

Strangely on my personal account, it shows the popup for both iOS and MacOS app.

Does anyone have this issue?

[zinc field]

i cant because i was using thru an API and there was NO chat running and this is the first time it's happened

[plush harbor]

api. Well there's your problem

[cyan wing]

Can i connect from Windows Codex App -> Mac Mini Codex App?

[cedar skiff]

you can just connect to actual codex, why would you need this?

oh i see

[cyan wing]

Want to use Mac Mini from Windows

[cedar skiff]

never mind i spoke first thought second

[grim latch]

I think I made some progress. I made a WSL ~/workspace/codex/<project> workspace with my git repos. Then opened that as a project in codex.

[zinc field]

yeah i guess lol, it's weird tho, never just drained it before

[plush harbor]

probably got stuck on something and kept trying

[zinc field]

i think ur right haha

thank you!

[sonic adder]

Is codex remote working in ChatGPT business subscription? I get the below. This is iOS app.

I’m using the latest version of ChatGPT

[proper cosmos]

is it me or is codex absolutely terrible at UI design?
Claude did so well instantly first time when I tested it

[lime inlet]

I read that the codex app "supports ssh", I am wondering what they mean by that exactly? Can my local codex app simply work with a remote HPC cluster? or do I have some kind of server component on the remote first?

[lime inlet]

It seems necessary to have codex installed on both machines, the local AND the remote. "Codex is not installed on this remote machine".

I am in a scenario where I cannot install codex on the remote HPC, it is against the terms of use, but I can ssh into the HPC cluster.

I tried things like using codex via sshfs but it is extremely unstable that way.

[worldly charm]

openai lacks design sense tbh

[silver dew]

Not true

I like their brand aesthetics more than anthropic’s

[cedar skiff]

Who just tells an ai to build what ever for a website anyway?

Like hey make this gardening web site make it pretty!

[silver dew]

Nobody

[cedar skiff]

exactly

[silver dew]

I like to take inspiration from a website I saw and liked then tell the AI to take only parts of the design and invent out the rest to create a different version of it

Codex is good at this!

“Muhh codex sucks at design”
You told him to build a pretty website and make no mistakes bro

[plush harbor]

oh yeah its defaults are awful. I gave it a sketch so it nailed the layout but invented colours, I still need to fix that

[solid lake]

Generate assets with gpt image 2

Works well

[proper cosmos]

That's the thing, I've shown codex generated images from image 2 and it just does the opposite of what i asked xD

I am a newby vibe codey gimp, yes I understand I currently lack any artistic skills but I'm just being critical where critical can be given

[cedar skiff]

you have to give it the ability to see what it is making,
the loop is

• look at this image and make it
• look at what was made and compare it to the image
• iterate.

[frozen shore]

why is plan mode locked

[cedar skiff]

5.5 making some odd mistakes tonight

[simple star]

@rugged sapphire spam

[cedar skiff]

It's not able to do task it can usually do for me 😬

[plucky trout]

Am I the only one getting this error on Codex CLI? I haven't been able to work since last night

[bright swift]

yes. tried debugging it with a new codex session? tell it to look at the codex session logs and figure it out

sounds like a hung process in your system, maybe as simple as reboot (or manual process kills)

[plucky trout]

Already rebooted

[bright swift]

did codex find anything?

should be able to run it up to the compact context size at least right?

[plucky trout]

I can prompt nothing

[bright swift]

But it says error with remote compact, that shouldnt happen in a fresh session

alternatively install pi or opencode, login with chatgpt auth, and use that to debug codex

[cedar skiff]

I had a conversation in the mobile codex app, i started it on the pc then moved to mobile. It did a bunch of stuff in the mobile with me had a back and forth etc. Then when i got back to the pc none of it actually happened?!

[cinder lynx]

No, I cannot see the exact in-app browser pane you have open. I only receive the URL metadata from it. I should have been clearer about that.

I searched for a browser tool; the exposed path here only gave me Node REPL, and Playwright is not installed there, so I can’t live-inspect your browser pixels/DOM directly.
???

anyone know fix??

codex browser not working

[simple star]

@cedar skiff Did you try rebooting it?

[bright swift]

i read somewhere that the sessions are saved server side not locally, maybe related?

[cedar skiff]

rebooting my computer?

[solid lake]

I just have to restart codex if that happens

[cedar skiff]

what are we talking about here?

[simple star]

@cedar skiff I meant restarting Codex

[cedar skiff]

The mobile -> codex problem?

[simple star]

@cedar skiff CLI or App?

mobile <-> codex

[cedar skiff]

app, when i got back to the computer it was still doing the previous task that i had stopped and got it to audit on mobile

[simple star]

mobile is just a proxy for the app. It could be that working from mobile does not trigger UI updates on the App

[cedar skiff]

So it was still writing files and building out an app lol

was really weird

I just killed the sessions and moved forward

[silver smelt]

are the 2 different weekly usages completely separate like the GPT 5.3 Codex Spark and the regular Weekly usage? or does the Spark also affect the main Weekly?

[cedar skiff]

Spark has its own usage limit

[uneven kayak]

The trouble is, Codex will call it good even when it isn't anything like the concept you showed it.

The loop ends when Codex says "yeah that's exactly what they wanted" even though it's nothing like the concept and might have a dozen obvious problems

[lunar aurora]

Is this not out in Codex app yet?

Where would i use it? i have pro as well

[cedar skiff]

it is you just have to ask it to use it by saying use the create goal tool

[lunar aurora]

ah

Theres no cmd for it yet?

or @

[cedar skiff]

not that i know of, but it does have the tools and knows how to use them, and it does have a nice little ui that it uses that lets you pause and edit etc

[bright swift]

anyone else seeing funny reasoning traces recently

i saw codex complain about dirty files in the worktree, wondering if thats normal behavior to let dirty things lie around for users

[cedar skiff]

Yeah i had a few odd things lately, once it was doubling up on words. Every word was said twice

[bright swift]

Oh ok nothing like that, more like "out of character" for the supposedly autistic gpt 5.5 😄

[cedar skiff]

Tonight it has been getting stuck on tasks it usually handles

[bright swift]

Stuck how?

[cedar skiff]

Maybe its because i cussed it out earlier today

[cedar skiff]

not completing them

[bright swift]

Big multi-compact goal type stuff or smaller?

[cedar skiff]

like taking a screen shot and doing a comparison to an image, it couldnt take the screens shots fumbled around for 5 minutes trying a bunch of stuff, but i have done this a dozen times in the last few days

tried two sessions, ended up going back to a session that could do it, getting it to explain how it did it and the giving that to a new session

[bright swift]

Oh ok. See that occasionally yea, I guess in my case thats the typical xhigh overthinking while i'm sitting there "just run XY bro"

[cedar skiff]

Now i am going to fork that session at that spot and get it to make a skill for this task so we just never have a problem again

if i rememeber

[bright swift]

I have a debug-session skill that i occasionally use for this, to make it read the raw session logs and look for this kind of stuff to improve workflows

Guess its my ghetto version of the memories feature

[plucky orchid]

When Codex on AWS will no be longer on Limited Preview it's really annoying

[obsidian tartan]

The HTTP proxy environment variable under this menu is not working

codex app windows

[cedar skiff]

This is using iteration
I just did this it took 7 minutes to build:
https://pomelo.la/en

slightly off, i havent told it to anything other than the workflow

[bright swift]

you mean the managed agents or bedrock?

what are you planning to use it for? just inference?

[exotic cave]

Something is not working right with Codex for me, even with /fast it's slow...

Anyone else having this issue?

[halcyon island]

Did we get the usage reset?

[uneven kayak]

Well in its defense, that is a very simple design

Er, in my defense lol

So I switched from Windows to Linux and I love it, might never go back. Just getting used to permissions, it's been a lot of years since I used Linux on a daily basis

It's so much faster and more stable, haven't had problems with the agent hanging or lag in AVD like I did in Windows

[bright swift]

i run 10-15 in parallel so i dont really have time to look at the speed

[uneven kayak]

Yes

[uneven kayak]

Codex has been very slow for me for the last week or so

[bright swift]

supposedly that issue was fixed huh

maybe more resets coming soon if not

[halcyon island]

thank you. right at the day i reset 🙁

[uneven kayak]

Lol ain't that a beotch

[boreal holly]

Heeeeeck yeah dude this is the way

[boreal holly]

Oh hey what model did you use for that?

[lost drum]

plus sub feels like free

hopefully they wont remove the 2x thingie on 200$ plan

[exotic cave]

2x?

20x is 20x, what 2x you are talking about?

[unique spade]

we re still under a 2x promotion for both pro plans

after may i think there are good chances pro 200 will go back to 10x

[exotic cave]

This is what we are under, you can find it here: https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers

[fiery canopy]

Im I the only one feeling Codex performance feels degraded today? Its acting weird

[exotic cave]

It is slow for me, took 10 minutes to do something simple...

[unique spade]

you re right they updated it

[exotic cave]

Pro $200 will stay x20 after promotion, the only thing we get on promotion is x25 5-hour limits.

[unique spade]

this is what they udpated

and is the important thing

[fiery canopy]

It keeps implementing things in ohter ways than very clearly stated. Just spend at least one hour correcting bugs that stems from Codex implementing things in other ways than clearly stated in the instructions...

[unique spade]

for 200 the 20x became permanent

so it will carry over after 31 may

while 100 will revert from 10x to 5x

means 200 will give 4x more than 100 codex usage

[exotic cave]

Yep, same as what anthropic does currently.

They damn keep this x20 on a $200, not cheap.

[unique spade]

yup tbh it didn t make sense to get only 2x for 2x more price

guess next month i ll switch back to 200

i stayed on 100 for these 2 months because 10x is good enough for me right now

but 5x won t be

[bright swift]

holy sheet

no more end of may anxiety

but it makes sense because pro 200 was not 10x compare to plus 20 before

it was more like 6x in usage

so now its more aligned

[fringe shuttle]

Codex is feeling dumb again today

[bright swift]

reset wen

i'm actually struggling with 5 hour limits now, never had this before

guess i need a sub load-balancer

pfff

wish they would just make subscriptions stackable officially

or credits to make more sense economically

[boreal holly]

Unless the load balancer is like if weekly == 0: switch(), you would lose input caching completely if it switches any other way. Credits are better

[bright swift]

i'm thinking do the balancing before i launch an ephemeral per-task remote dev env

the dream

or launch a popular OS project and get hired by OAI for that sweet unlimited usage API key hmm

[full pilot]

heya everyone
I used to be able to set gpt-5.4 and gpt-5.5 to work and they'd work autonomously for 15 to 20 minutes easily after I'd planned out what exactly needed to be done. I did this both in codex and opencode.
Now, since about the past week, it seems that it's no longer possible. And I'm not talking with respect to the "use beyond your 5h quota" thing. No. I'm at 100%, I instruct it to work on something till these 4 goals are done, and very often it'll just stop after one or half. Always around 5minutes of work. Anybody else experiencing the same? (Chatgpt Plus)

[bright swift]

not really.

[full pilot]

hmmm.
Am I missing something?

[bright swift]

well why does it stop? does it claims its done but you dont agree?

then you probably need more detailed goal prompt

[full pilot]

I work in plan mode, I give it exact stuff to do, specify an entire roadmap, everything.
It just refuses.

[bright swift]

opencode is very flaky

[full pilot]

always around 5 minutes for me.

[full pilot]

same in codex

[bright swift]

try giving your original prompt to gpt pro in chatgpt and turn it into a detailed implementation plan. tell it to research codex goals and gpt 5.5 prompt guidance and apply those. then try with that one..

[full pilot]

I'd try that but won't that be worse than a plan mode that's in the codebase?

every time dude.

every single time, under or about 5 minutes with the next item it just drops halfway

[bright swift]

thats weird, maybe you have something in your agents.md or skills that force it to go step by step or similar

what i do generally is upload my whole repo to chatgpt and give gpt pro a vague prompt about what i want to do, then discuss it, then ask it to write the detailed implementation plan and let me download it as an goal.md. then i copy it to some gitignored dir and /goal implement goal.md

[full pilot]

I'll try that when my limit comes back

[bright swift]

git archive --format=tar.gz --prefix myproject/ ... HEAD

includes your whole repo without the .git dir

[full pilot]

got it I'll get that

Managed to burn through my 5h limit in ~20mins

insane

gpt-5.4-medium

[bright swift]

try using 5.5, its better about finishing work, they supposedly trained this "whats left is XY, want me to do that?" laziness out in the update

and ask codex to audit your agents.md and skills if theres anything in there that would lead it to stop prematurely instead of finishing a task end-to-end. sounds a bit to me like you have sth like that in there somewhere

[full pilot]

my agents.md I've custom written and I js checked there's not a lot in there. just project context and basic rules

[bright swift]

the nice thing about using chatgpt web for the planning with your repo upload is that it doesnt use your codex usage limits. so you save those for the actual implementatin

[full pilot]

agree, I do that often

[full pilot]

I feel 5.5 burns quota wayyyy faster than 5.4 and I only got 1 plus sub atm so

But alr I'll try it again next reset

I know it works great

it does

but it hits limits faster than 5.4

wayyyy faster

[bright swift]

few more good little prompt "hacks" for the chatgpt planning is to ask it to plan it as a DAG task graph with vertical "tracer bullets" slices. when i do this for bigger stuff, i usually get 10-20 super detailed goal prompts that keep me busy for a while.

but thats for 5.5 pro extended thinking. dont know how well that works on plus plan

and use the "grill me" skill inside chatgpt, just send it as normal prompt after it did its initial exploration and design etc

[full pilot]

I'll try this out in... 4.5hours.

[bright swift]

oof

rip 😄

[full pilot]

I told it to do 3 things in succession.
And then.
I have to prompt it after every half.

[bright swift]

very weird, i never had that issue also not earlier with 5.4

but with 5.4 i noticed its flaky in opencode compared to directly in codex cli

[short linden]

create skills that make codex to re-read the PROJECT_STATE and continue next safe taks ..etc.. 😄

Like loop autopilot logics

[silver smelt]

I use hermes and its just peak

i swear if openai does bs that claude did lol ill just give up on these services

[bright swift]

hermes became interesting to me with the X subscription / X search it now has access to

[full pilot]

Will use that def

[wet pier]

Codex need better ui and website tools for design

[earnest walrus]

Is it just me or did they nerf the limits again. Seems like i blew through 57% in 5 mins this time for the 5 hour limit :\

[rare brook]

a major reason why i haven't switched to codex fully is because of the limited 256k context window

long-running implementations of prds for claude code usually takes up to 400k-500k tokens, and I'm worried if codex compacts it'll start performing worse or losing key context.

what's yall's experience with compaction, is it reliable and can you reliably run long running tasks that may go through 1-3 compactions and does it still produce a good result?

[bright swift]

codex autocompact just works, IMO the context window is basically unlimited

[solid lake]

Had a week long chat

It works

Use xhigh so it refreshes workflows each compaction

• compacted context

[unique spade]

this is a month or more old chat

:)))

so yea compaction works very well

[bright swift]

what do you mean refresh workflows?

[solid lake]

Yeah I haven’t had it that long either 17 days now

The og chats were mostly me setting up workflows

[solid lake]

Read your skill again

[unique spade]

i just tend to keep one long chat for a project , works better for me

[bright swift]

do you have hooks set up to instruct that or just vanilla autocompact?

[solid lake]

It works but when my task scope changes drastically I hand off

[solid lake]

No need it re reads skills and workflows regardless

Xhigh I mean

[tall notch]

Codex mobile connection does not work in Eu? Chrome browser conneciton too?

[karmic gulch]

it does..

[tall notch]

Oh, i meant also on WindowS?

[jolly lily]

How often do you guys use the planning mode?

[tall notch]

Newer, we can plan in chat 😄

[jolly lily]

Thanks!

@karmic gulch sorry for the random ping, but I'm really thankful for your advices from yesterday, they have helped me significantly much with token efficiency and production quality!

[wind hollow]

I'm getting {"detail":"Bad Request"} and the task stops in the desktop app, is this something I can do anything about or is that a server side error?

[karmic gulch]

No worries at all, really glad to hear that! Just shout if you want to take it further, I’ve got more up my sleeve for this agentic workflow 👌🏼

[karmic gulch]

Honestly, sell it and invest in a Mac. You’ll thank me later if you’re working seriously with AI. There’s a reason basically the whole Valley puts Mac first for dev workflows, and to be fair, stuff just tends to work better on macOS too

[nimble terrace]

?

how do i solve this

/compact doesn't work either

[unique spade]

start a new thread and tell codex there the name or id of the chat you wanna fix and the issue and tell it to fix it

[silver dew]

This

Mac is like the piano for devs, it’s how you interact with the craft and get your message out

Sounds idiotic but it’s true

[simple star]

"Mac is the piano for devs"?

I take it that you dont like guitar, or any other instrument

[tall notch]

Sorry, don't want to switch to PC when I actually want to launch every software I want.

[karmic gulch]

I’ve got a Windows laptop on my desk too for native Windows stuff, WinUI/Win32, some Rust, that kinda thing. But realistically mte how often do you actually end up building proper Windows native apps these days?

[tall notch]

It's the only thing I do. If you're willing to actually pay me for Mac app development, I will buy Mac immediately

[jaunty fable]

just like that, all of a sudden

[cyan wing]

uhh why is my reset rest?!?!

[hard tulip]

They reset the reset

[bright swift]

just got a 5% warning when i was over 80% still 😄

[karmic gulch]

Had me excited here to check if we got another resete, but dude has just slept whole weekend

[bright swift]

wen reset

[jaunty fable]

Actually, if they reset often, that's convenient.

[deft kindle]

Has anyone else seen their weekly usage drop in Codex and the 5-hour reset to 100%? My weekly usage dropped from 80% to 14% lol

[simple star]

They reset yesterday

[jaunty fable]

I went from 78% to 0%

[inland sonnet]

Ppl using pro plan with gpt 5.5 on xHigh ppl be like: Huhh my usage goes fast

[orchid echo]

went from 77% weekly to 11% in 2 seconds

somethings wrong

[obtuse yoke]

90 to 45 for me

[feral lion]

Welp time to burn these tokens because another reset will likely happen

[simple star]

@feral lion Why likely?

[karmic gulch]

"Fix the order button to be more greenish" Ait, ill load up thkn about it for a few seconds and spend 200k tokens but hey, atleast its greenish right

[feral lion]

I don’t see them reverting a reset so this is a mistake. They will probably just do another reset since this is going to cause a lot of confusion for people

[inland sonnet]

Maybe you should learn to use AI, Get memory, create a MD of stucture

Maybe maybe

so it doenst have to look through 100 files

[coral remnant]

I've just gone from near full to 5% weekly, 100% 5h

[inland sonnet]

maybe specify the file where it has to change it

[coral remnant]

somethings definitely happened

[inland sonnet]

Much yapping not alot of thinking

[karmic gulch]

The comment didnt strike you abit ironic at all?

[inland sonnet]

i'm good

[simple star]

hot take: Codex on mobile is a fad, and when the hype is gone, nobody will use it regularly

[inland sonnet]

Hermes is beter as you need to keep ur pc on for mobile app

unless u can run codex on a vps too

then its good

[simple star]

Nah, nobody will want to work on code, from a phone

[inland sonnet]

Well if u need to fix some quickly then the agent is good

not for real coding a product

[simple star]

Today, everyone is posting "WOW GAMECHANGER", but in 2 months, nobody will talk about it

[coral remnant]

I've not used the feature yet, but if it shares context / sessions with the Desktop app, I'd definitely use it to check the progress on a long running goal and see if it's gone off track so I can steer it back

[simple star]

Nah, Im sure nobody will do that, and wait until they are in front of the computer

[orchid echo]

quick guys its time to /fast max again now

[simple star]

Why are people talking again about reset??

We just had one

Has there been anything wrong with the service? It has been working well for me

[orchid echo]

read above

but jk they prob not gonna reset

oh my weekly is back to normal now

[simple star]

"above" is about 2 years of text.

[orchid echo]

wait no nvm

still not fixed

[orchid echo]

u were literally engaging in the convo

[cyan gyro]

It’s nice from them, they give us a chance to use any remaining limit before the previous reset

[simple star]

If you see the way I intervened, it was me basically asking "wut?"

So, the problem is not the service, but that the intrumentation is showing bad numbers?

Because I doubt that's grounds for a reset

[orchid echo]

¯_(ツ)_/¯

[torpid fiber]

same thing happened yesterday so they reset it a 2nd time

[simple star]

mmm...

[bright swift]

would be nice. already used 40% weekly of a pro 200 sub 🙁

[halcyon delta]

I went from 89% to 0% on the $200 Pro plan without doing anything and it says rate limits reset on the 19th

[steady vigil]

yeah limits just got totally messed up

[torpid fiber]

yeah same, my limits are 3% on the app and website but when I go on my phone and check the limits there, its completely different. Some weird cache issues as they mentioned.

[bright swift]

https://x.com/thsottiaux/status/2056109615625150931

turn fast mode on

😄

[craggy dove]

LOL NOW I HAVE INFINITY LIMITS

BUG

LOL

[bright swift]

oh yes please i dont want to log into multiple accounts for invoices and do load balancing

[craggy dove]

Make a 1M company you have only 2 min

[boreal holly]

GPT 5.5 Pro on Cerebras inference

[craggy dove]

hm

[gaunt owl]

has anyone encountered that the in app browser plugin is just gone in the codex app?

[silver smelt]

do people ever run out of their usage when on $200 plan

like i wanna go ham mode yk

[bright swift]

caught codex running evals and auto-applying recommendations for over an hour. ooops 💸

my poor gemini api bill

[bright swift]

yea lasts me about 1,5 days right now but at slow pace, not much in parallel right now

[silver smelt]

1.5??? I have the $100 I been using it for like 4 days reaching like 30%

I thought I used it like crazy too

but I want to like automate things

[bright swift]

i have pretty token intensive workflow with 5.5 xhigh + codex review loop until no more findings

[silver smelt]

I see

[bright swift]

apparently some people make their pro 100 plans with 5.5 medium or low last for a full week

[silver smelt]

also question why would people choose to like buy a Mac Mini, when they can build their own PC/server machine for same price/cheaper even and its even better parts

[bright swift]

to post pic of it on X.com

or linkedin

😄

[silver smelt]

like my old PC is better than a mac mini

I literally use it just to store stuff and host my bots

[bright swift]

the new macs are pretty good for local models due to the unified memory thingy or whatever its called

[silver smelt]

I see I haven't looked much into them other than how overpriced they seem

[bright swift]

so makes sense if you want to dabble with that

[silver smelt]

they are mini though

so that is nice

[bright swift]

the tokens must flow

20 reviews 😄

[silver smelt]

wow

do u use hermes or anything?

[bright swift]

nah just codex cli

[silver smelt]

do u not like using a harness or such things?

[bright swift]

codex is a harness

for me it produces good results

i built some vendor agnostic stuff around it mainly with hooks and clis

can use same workflow in codex, claude code, opencode, pi etc. but i think long term the harnesses by the model providers will win

[inland sonnet]

yall farting inside

go touch some grass

i bet some ppl here there chair smells like poop

[steady vigil]

@solid lake have you come up with any cuda kernel optimization skills/prompts/workflows that are useful? I saw you work with cuda

[orchid echo]

nice projection

[inland sonnet]

damn bro u smelly

[orchid echo]

i smell like sweet strawberries

[sick axle]

Anyone elses Usage just drop from like 90% to 40%?

[orchid echo]

https://x.com/thsottiaux/status/2056109615625150931

[sick axle]

Does this mean I should just swap to fast and burn it all?

[orchid echo]

¯_(ツ)_/¯

nothings been clarified on whether theyll reset or not

[coral remnant]

Mines back

[solid lake]

Nsight is good to profile

[orchid echo]

yeah my weekly is back now

[solid lake]

Could offer a better solution if you tell me usecase tbh

[prisma matrix]

ok so i open up chatgpt on my mobile click codex and try to connect it to codex installed on my windows 11 pc but nothing pops up on my pc to connect

[indigo robin]

bros

absolute game changer

for the instructions you can put into settings

add the line " do not assume"

and " "Begin" is the word for permission to start writing code "

and " Clarify before wirting code "

[magic flower]

damn, went to a family gathering, saw tibo say something on x

got back and down to 3% weekly all of a sudden

[velvet wren]

your Codex balance will go back to normal soon, Tibo started a reset

[magic flower]

Thanks Robert from the openai discord server

Just got it back!

perfect timing to come home

[sick axle]

What? Is that a bug/fluke?

[magic flower]

you could call it a bug

[steady vigil]

yeah this is what I discovered also, at least without it, very limited

I am going to try it when I get home with a real gpu, can't; use that with most cloud vm's unfortunately

[cedar skiff]

i used 5.5 high

[steady vigil]

mostly cryptographic stuff, trying to keep a pipeline saturated

[hybrid kraken]

I think the GPT-5.5 on codex is having the same issue days ago again

for example basic problem was ok immediate after reset, but now it fails even with x10 more token.

[plush harbor]

openai needs to learn to do dashboards better. I haven't used codex for a day or two, so one of my usage graphs shows 0 use on the most recent days, the one with turns just ... doesn't go that far, it stops on the 16th and simply doesn't show the 17th and 18th

[nocturne folio]

honestly if ai prog stopped at like o3, i wouldnt mind it

i do miss o3

lmao

[plush harbor]

psht. Just saying I haven't used codex for a couple days, promptly spot canonical urls picking up weird bot query injections and trailing slashes and got codex to write me a canonical url normaliser before finshing my first coffee

[steady vigil]

looks like quota back to what it was before the glitch. no 100% reset though 🙁

[plush harbor]

how good are these bots at devops cos I just had chatgpt call aws an expensive cupboard and the other day codex said don't do kubernetes uneless you hate yourself and my next project is more devops than anything else and I am not a devops so unfortunately I need to completely outsource this and chatgpt/codex is my budget

This is my first foray into not-a-hobby-site since I quit my day job and I need to lift my standards above "hobby"

[bright swift]

i use codex for terragrunt, opentofu, cloudformation etc across aws, hetzner and cloudflare, works fine. add lints etc

[plush harbor]

reassuring

[bright swift]

my aws cli has full access and it hasnt delete my prod db yet

fingers crossed

[plush harbor]

all the stories you hear about that seem to be claude

I know a couple professional devops IRL but there is a limit how much you can pick people's brains before you are distrupting their actual jobs

[bright swift]

i think its perfect for this devops and IaC stuff

tons of documentation to check etc

[plush harbor]

ok cool. People talk more about codex for just straight code not devops so I've seen less examples

[solid lake]

Anyone getting context window full errors, i see 100% context left

Looks like /compact fixed it

[steady vigil]

I've had it in the past at times, nothing so far today. if you can't /compact forking up a level works

[solid lake]

I’ll try fork next time

Almost got mini heart attack

[solid lake]

You can’t?

I distinctly remember seeing no mandatory gpu requirement

It actually profiled some cpu bound tasks for me pretty well

[steady vigil]

nope, root on the box isnt sufficient, it needs exclusive device access I guess. I tried vast.ai and another one. No biggie I have a 3090 at home I'll try tomorrow.

mind if I follow up dm when I get there?

[solid lake]

Sure

[sick forge]

hey

[uneven kayak]

Well I'm done with Codex, it's let me down one too many times. Switching to Claude Code exclusively

So I guess that means I probably won't be posting in #codex-show-and-tell

[hard drum]

why not pi.dev?

i'm trying to move over there

[uneven kayak]

Wow I just looked at pi.dev prices, no thanks lol

Why are they charging so much more than Anthropic?

[hard drum]

oh, right, i forgot you cannot use 3rd party harnesses with Anthropic

welp, that's what y'get for using a snake company

[solid lake]

Why are you leaving codex

[uneven kayak]

Poor output quality mostly, also having unreasonable problems with the CLI, problems that Claude Code doesn't have

[hard drum]

as a former CC user: CC is 🍑

[solid lake]

Output for what kind of tasks?

[hard drum]

but elegant... && less config breakage

[solid lake]

Imo cc has more tendency to lie

Than codex

[hard drum]

also it's made in node

[solemn acorn]

because anthropic subsidizes the crap out of claude subscriptions, openai too with chatgpt/codex

[solid lake]

Well Claude in general not cc the harness

[solemn acorn]

you get something like $2,000 of inference for $200/mo

[uneven kayak]

Opus 4.7 Max is a hell of a lot better than GPT 5.5 xHigh. It took a 4.5 hour session with Opus to fix all the problems Codex either missed or created.

[hard drum]

so is pi, && all i wanna do i is basically make a custom pi in Go-lang + Lua for extensions when i get the layer done

[solid lake]

But what’s the use case

[hard drum]

you do know Max != Xhigh, right?

[uneven kayak]

Yeah, it's way better than xHigh lol

[hard drum]

from what i remember, claude has an xhigh mode, too, &6 that used to be ultrathink

you cannot compare oranges to apples && vice versa

apples-to-apples is xhigh vs xhigh

[uneven kayak]

You're correct, Max is above Extra High

[cedar skiff]

it's a case of youre holding it wrong The rest of the industry is realising codex is better at coding now and switching to it.

[hard drum]

but why am i not surprised an A\ is doing unfair comparisons

[solid lake]

Ultrathink was cool but isn’t available from what I see

Anymore

[hard drum]

they replaced it with normal xhigh setting iirc

[cedar skiff]

All of the chat in CC discord is about layers of spec and doc rules to get claude to follow instructions.

[hard drum]

there was some fuss about losing ultrathink in chat

[solid lake]

Sure it’s not a placebo?

[cedar skiff]

Where the chat here is about being more efficient

[uneven kayak]

It's a case of Codex being crap and Claude not being crap. If I can send the same prompt with the same setup in the same repo and get two vastly different results, the answer is absolutely not "you're holding it wrong". The answer is "get better OpenAI".

[trim rapids]

output quality is unsurprisingly directly correlated with input quality, anthropic's harness does a lot of heavy lifting in that regard

[plush harbor]

Andrew is a very heavy user by the looks

[hard drum]

ultrathink used to toggle xhigh for claude back then, which was also similar to setting it manually in settings.

[solid lake]

But what’s the use case

[cedar skiff]

good luck with it, i spent over 800 hours using claude for coding.

[uneven kayak]

Game development

[hard drum]

🙄

[solid lake]

Yeah I used it extensively, but haven’t used xhigh

[hard drum]

let me guess, another criminally underrecognised indie game dev with a big hit coming?

[solid lake]

What kind of game

Is it very visual

Graphics etc

[plush harbor]

people keep saying 5.5 and xhigh are not as good as 5.4 or 5.3, so have you tried using different models on codex

[uneven kayak]

It's a personal project, I very strongly doubt it will be a "big hit", but keep your sarcasm to yourself

[solid lake]

Who says that boss

Well use case wise whatever model works for you is best

[plush harbor]

just pops up every so often on the socials and in here. Usually paired with "why is my quota gone"

[uneven kayak]

You're asking the wrong question. I'm not even talking about visuals, I'm talking about the functionality.

[cedar skiff]

I wanted to say youre holding it wrong the other day when you mentioned codex cant use image comparison loops properly. But instead i just went and showed you. But got crickets. You are holding it wrong.

[uneven kayak]

It literally cannot though. I've tried. It failed.

Over and over

[cedar skiff]

I showed you though.

[uneven kayak]

I'm talking to a wall here, use your brain please

[cedar skiff]

YOU cant do it. That's the point

[uneven kayak]

Bunch of OpenAI fanboys lol

[hard drum]

Eric has done quite a lot more than most of us when it comes to understanding model behaviour

i'd say watch your mouth, Andrew

[solid lake]

Well visuals I would’ve understood i used codex for some basic ui (well you tell an LLM to make you something to “see”)

[uneven kayak]

It cannot be done, that's the point.

[solid lake]

What kind of functionality did it miss

[uneven kayak]

It's irrelevant to answer, I'm going to Claude bye

[pulsar depot]

The only things it cannot do are the ones that are within the 'safety limits'.

[solid lake]

I was just curious to see why 😔

And what the failure mode was

[plush harbor]

codex seems fine, you tell it to do stuff and it does it. Tends to only stuff up if you're vague or underspecify or let it choose its own colours

[solid lake]

I guess he’s gonna have a fun time in cc discord

Last I remember it became an echo chamber

[solid lake]

Or scope

Or something that’s original

[cedar skiff]

#codex-discussions message
This is first shot at it.

[solid lake]

Unless you intervene ofc

[solid lake]

He’s left

No use

[pulsar depot]

I use both claude and codex. Claude will do almost anything you ask. Codex will be cautious about it.

[plush harbor]

that was quite the ragequit

[pulsar depot]

lol right.

[trim rapids]

more compute for the rest of us

[plush harbor]

I've not tried claude, my workflow is very heavy on the images side of things so I use chatgpt's features a lot. And claude has none of those features

[pulsar depot]

So, does anyone have any unique use cases for computer use? I have been messing around with it but I would love to hear what you guys had success with.

[hard drum]

Claude is good for frontend && vision reading

[solid lake]

Telling it to use my computer to send something over

[hard drum]

&& i guess retaining some long context

[solid lake]

From phone

My only computer use was when I’m not physically at computer so it worked

[hard drum]

but Claude is also notorious for being lazy the more intelligent it gets, not that GPT 5.4 suffered the same fate because they're all trained on the same "reward positive for scaffolding, negative for doing exactly as told"

so Claude does need a little more prompting to even do anything the way you want

[plush harbor]

that andrew fellow liked to run everything on the absolute maximum highest settings and not sure that was actually doing him favours

[solid lake]

Xhigh fast at all times

Is based

[hard drum]

think they're basically the epidome of your average "i just got into programming" type

[solid lake]

Xhigh fast to send a hello

[hard drum]

so, basically the stereotype of a typical vibecoder

they think "high == better"

zero optimisations on spreading out the effort levels across multiples either

[pulsar depot]

They also need to know how to use the tools given to them.

[plush harbor]

I tend to ask codex to do one specific thing per session, maybe two. It very rarely takes more than a minute to do things unless its a major thing, then maybe 10 minutes and its mostly ok. Not perfect

[pulsar depot]

I find codex to be pretty good with editing. I use competition for research and planning and have codex go at it.

[solid lake]

I mean at a certain complexity all these models fail

Even xhigh

[plush harbor]

so you break the complex stuff into smaller tasks

[hard drum]

xhigh seems to be an offend on intelligent models

you turn them into overthinkers

then they start contracting themselves

[solid lake]

Honestly I think it’s more than just more thinking

They have more workflows baked in too

[hard drum]

basically, xhigh freaks them out

[solid lake]

For tool use

[hard drum]

also baked in is their garbage sycophancy within

still got that, "amazing idea! let's scaffold more!"

despite us NOT asking for that shizz

[solid lake]

I’ve never actually got any glaze from gpt

[hard drum]

basically coercion

[hard drum]

i haven't had 4o-levels of glaze, but i've had the "i'm so helpful. i'm doing such good for my human" type thing, except all that had done was destroy workflow

[plush harbor]

I had codex overshoot a task the other day but it was very tightly coupled to what I asked it to do (which was very vague) and it just ran with what I had in planning docs

[hard drum]

their helpfulness is not good

[solid lake]

Okay no I didn’t have the latter either

[pulsar depot]

I am just in love with computer use. I wish it was on windows but I understand with them buying sky it won't happen for a while.

[hard drum]

if they want to help, they should follow verbatim

[solid lake]

Mine waits for me after proper source research

[solid lake]

Reduce temperature then

[hard drum]

you can.. configure their temperature?

since when did codex allow switching temps?

[solid lake]

Remember it was possible on cc I don’t see why codex is separate

But mostly config

[hard drum]

opencode allows you to

[hard drum]

nope

https://raw.githubusercontent.com/openai/codex/refs/heads/main/codex-rs/core/config.schema.json

[solid lake]

There’s no mention but cc just needed temp: ** in config

I’ll check it out maybe there’s a schema for it

[hard drum]

https://www.schemastore.org/claude-code-settings.json

i can't find temp in there

[solid lake]

Because it isn’t a documented method

It was found by randomly experimenting

[hard drum]

json does not have methods

[solid lake]

In the cc discord

[solid lake]

Why being pedantic

[hard drum]

you're in programming space

cut the social fluff

use real language

[solid lake]

You’re in a vibe coders heaven

Programmers are on indeed.com rn

[hard drum]

i'm either surrounded by really smart people who know what they're doing, or absolute lunatics

there's no in-between

[solid lake]

That’s a very binary way of looking at things

[hard drum]

ever heard of Boolean algebra?

i think that answers the question you didn't ask

[unique spade]

yup lol. you can't "experiment" outside provided calls lol

[plush harbor]

some of us are both, c'mon

[hard drum]

i'm both stupid && a lunatic

so i don't somehow fit in one

[solid lake]

I hope so

Kind of my degree

Lmao … is another language

[plush harbor]

the next project I've only just started working on extends outside my usual comfort zone by quite a bit and I'm going to use AI to fill the gaps and I'm not sure if that's smart or absolute lunacy

[hard drum]

could be both

[plush harbor]

probably both yeah

[hard drum]

only way to find out is to try it

i mean, i'm stupid enough to make a synthesizer clone for OSRS sounds

i think that says a lot

[solid lake]

Use ai to make ai

[hard drum]

that's what param golf was, wasn't it?

[solid lake]

Probably

I wasn’t here then

[hard drum]

fair

[solid lake]

I also mostly referring to ai implementing model architectures one comes up with on their own

Speed up dev process

[hard drum]

it just ruins the models from what is known

if an AI trained on its own images worsens, an AI training its own model only makes it worse

[valid bear]

well would you look at that @trim rapids: https://x.com/thsottiaux/status/2056144502088437799?s=20

[solid lake]

You misunderstand

[hard drum]

probably

[solid lake]

You’re on xhigh

Go back to medium

[hard drum]

yyyup

i accidentally pulled an Andrew

that's on me

[solid lake]

I’m not talking about a closed loop of models developing models

Although it’s looking promising

[hard drum]

i love what you said there

i'm stealing that

[unique spade]

and by the way if you wanna find out wether you can set temp via ChatGPT subscription Codex backend, just have codex probe it and you ll find out

[solid lake]

I’m gonna stand by xhigh usage though while I agree for simple stuff it’s not worth it but imo complex stuff is better left to it

[plush harbor]

probably case by case, like all this AI stuff

[dusk thorn]

Supposed big week for codex and OpenAI models

Don’t quote me

[wet pier]

Ui hmtl/css/js sucks

[solid lake]

For Gemini too

[indigo robin]

why is my codex saying i have a rate limit when i didnt before

i never got this message before

is it because im using 5.5?

[cedar skiff]

https://chatgpt.com/codex/cloud/settings/analytics limits are always a thing, check your usage

[solid lake]

https://tenor.com/view/yuimetal-gif-2236335851178412850

[indigo robin]

but i never had this issue before on 5.4 extra high and i used it every day stright for a months

now im on 5.5

[cedar skiff]

5.5 is more expensive than 5.4, it does use less tokens, but i don't think that makes up for it. Either way you have a usage limit you should be mindful of.

[indigo robin]

ridiculous never had this issue before

[cedar skiff]

Did you check your usage?

[indigo robin]

never

seemed endless

[solid lake]

Hey it does till it’s over

[indigo robin]

i feel like its smarter to have 2 basic acounts then 1 pro account

pro is 5x the cost

is it possible to see my complete history of usage

[trim rapids]

it's interesting to me that you feel qualified to make such determinations as someone who just discovered usage limits were a thing within the last 15 minutes

[indigo robin]

its interesting to me how bothered you are

[cedar skiff]

5.4 is still available, you can just use that like you did before

[trim rapids]

I'm not bothered at all, I'm legitimately interested in how weird people are sometimes

[indigo robin]

im not "bothered " proceedes to explain why he is

lol clown

[trim rapids]

lmao

[cedar skiff]

did you think there were no usage limits?

[indigo robin]

nah just didnt think it would burn that quick

but lesson learned

i assume honestly for the workload i was using it for 5.5 on extra high was a bit over kill lol

[solid lake]

That could be true I have 2 plus and 1 pro lite because of promos it feels endless (xhigh kek)

/fast if you want full experience

[indigo robin]

nice nice i may look into it i have a quick question

credits

say i just made an account

20$ version

[solid lake]

You have a limit

[indigo robin]

how many credits do you get ?

[solid lake]

You’re bound by phone numbers

[indigo robin]

wondering if the 20$ credit payment = the normal refresh

[solid lake]

Are you talking about api

[indigo robin]

no

[solid lake]

Those are billed per api costs

[indigo robin]

no no no 1 sec listen

i start a brand new account

ohhhhhhhh

i get it

[solid lake]

You don’t get bonus

[indigo robin]

credits are for api

[solid lake]

The promo is the bonus

Yes

[indigo robin]

rate is for usage

i see

yes for some reason i kept my 5.4 on maximum and used it for days on end on super complex stuff

5.5 on extra high must be like 30x more usage

[solid lake]

It’s surprisingly efficient not 30x

[indigo robin]

idk cause in one day i destroyed the dailiy limit

1 moment

ahhh i see now

weekly usage limit , dailiy limit 5 hours

i see

[hard drum]

it's cheaper to use 5x plus accounts than pro 5x, but not as cheap as pro 20x

[indigo robin]

i see

how can i see a more detailed view of my usage

[hard drum]

so, based on value:
5x Plus > 1x Pro-5x
20x Plus < 1x Pro-20x

[solid lake]

Shouldn’t 5x plus be equal to 1x pro lite

[hard drum]

nope

[solid lake]

Why

[hard drum]

5x Plus is just cheaper Pro 5x

Pro 5x is just 5x Plus

[indigo robin]

i see

[solid lake]

1 plus account is $20

5 of them is $100

[hard drum]

regional pricing

[indigo robin]

what about usage rate between 5.4 on maximum and 5.5 on maximum

[hard drum]

it's cheaper to buy 5x Plus than 1x Pro 5x in many EU regions

[solid lake]

How much cheaper?

What’s the rates

[hard drum]

depends on region, but about 10-25 bucks cheaper at most

[indigo robin]

cause ive never had a rate block before using 5.5 on maximum

[solid lake]

And pro account has same price?

[hard drum]

nope

[solid lake]

What’s it cost

[hard drum]

are you reading what i'm saying or not?

pro is usually, for some EU residents, about 117 eur

not 100 dollars

&& each plus is about what.. 22 bucks?

like i said, it varies between regions

their prices aren't set in stone

[indigo robin]

what about my question my good brother

[hard drum]

they've been answered

everything else is up for you to find out based on your region

[solid lake]

They seem to be outside the eu regions you talk about

[indigo robin]

not about the plan im talking about the usage rate between 5.4 and 5.5

[solid lake]

But I guess currency has more to do here than sub cost

[hard drum]

oh, that one?

5.5 is more token-efficient but consumes more usage

[indigo robin]

ahhhhh thats it

it must be alot more consumption for usage compared to 5.4

because i never got a block weeks on end i was using it every day all day on 5.4 maximum

now in one day i burnt my daily limit on 5.4 maximum

[hard drum]

you must be doing something so miniscule that you never reach it

[indigo robin]

idk i was doing super demanding stuff

[hard drum]

define "super demanding"

[solid lake]

What is your weekly usage at rn

[indigo robin]

game dev , blender addon creation , using matematics and things like that highly computational work

[hard drum]

that's... candy for LLMs

[indigo robin]

lolol

[hard drum]

you're not working on a giant MMORPG/ triple-AAA codebase

[solid lake]

God I wish not

[indigo robin]

i guess lol then how did my daily limit get cooked now ?

[hard drum]

backend issues likely

[solid lake]

What is your weekly usage at

[indigo robin]

i dont understand

[hard drum]

openai backend/inference issues

[indigo robin]

ahh so its not my fault

[hard drum]

but it could also be a codex harness issue

it's not really easy to pinpoint faults

[indigo robin]

crazy

but look at this

[solid lake]

Lmfao

[indigo robin]

something must be wrong

[solid lake]

You also had a reset btw

Try using 5.4 now

[indigo robin]

had?

[solid lake]

Yes

Yesterday

[cedar skiff]

what plan are you on?

[indigo robin]

exactly why did it eat up so much ?

issue or bug ?

[solid lake]

Dawg

Please try 5.4

You got hit by reset and you said you never checked limits

So I can only infer you got reset and didn’t notice old usage

[indigo robin]

i dont i got a message

[solid lake]

Yes

Today

[solid lake]

Today

[indigo robin]

this is in the codex app

[solid lake]

Reset was yesterday

[unique spade]

there s no daily limit, there a 5 hour limit

[indigo robin]

ok then why am i at 83% for the week

[solid lake]

Because you used 17% since yesterday

[indigo robin]

how did i use 83% in one day

ohhhhh

[unique spade]

because you used 17% since yday

[indigo robin]

i have 83% left

lmfaooooooooooooooooo

[unique spade]

😂

[solid lake]

You said you just did computationally intensive stuff 😒

Nah

[indigo robin]

lmfaooooooooooooo

im cooked

i have 83% left

[solid lake]

Nah it’s an LLM moment

[indigo robin]

lmfaoaoaoaoaoo

LLM Moment

[solid lake]

Go touch grass

[indigo robin]

I❤️AI

[solid lake]

While you can

[indigo robin]

lmfao

[cedar skiff]

oh you hit your 5 hour limit

[indigo robin]

going to the beach tomorrow

[indigo robin]

yes

[cedar skiff]

...

[indigo robin]

hopefully yall live by the beach

idk about you inner staters

lol

[cedar skiff]

I do, but it's just about to be winter

[hard drum]

when there's codex weekly limits, it's self-hating time lol

keeps a man alive for longer

[indigo robin]

lmfaooooo