codex-discussions

[silver dew]

show me a tech company that ChatGPT powers its cloud enviroment except maybe Microsoft

You can't make it up bro

[unique spade]

on the other hand probably gemini wins the low-mid tiers of coporate people that are relying on google search and email and other similar stuff, taht s true

[plucky halo]

https://www.businessinsider.com/anthropic-may-soon-pass-openai-measure-ai-business-spending-ramp-2026-4

Looks like OpenAI and Anthropic are the leaders...

[livid relic]

i dont think ive ever had a good experience with gemini besides the main chatbot site

[plucky halo]

Looks like you did

Bro

[silver dew]

Discord is too distributed to rely on 1 provider lol
be sure they use chatgpt the least

[unique spade]

yea it s a decent chatbot, strong on scientific info, and image and video gen

but it's agentic part is almost non existent

[livid relic]

yeah ive had a similar experience

and i actually pay for gemini too

sometimes i go to it for stuff but i would straight up rather use the gmail connector in codex then use the built in gemini in gmail

[silver dew]

Enterprise AI is not about coding it's about integrations and reliability in everyday tasks

Gemini wins here

[livid relic]

they just dont

[livid relic]

.

from experience giving companies options for their ai suite most go with openai beacuse the models actually work

[silver dew]

Did you just compare individual experience to large company needs?

[livid relic]

i have experience in both

[silver dew]

probably not enough

😂

[livid relic]

i have tons

much more than you i can guarantee

but thats mean

just saying what ive noticed

[unique spade]

don't take T. too seriously, it always has opinions on everything 🙂

[livid relic]

haha clearly

[silver dew]

it💀
bro used too much AI

[unique spade]

"It" is just easier than guessing he/she/it

😂

[velvet wren]

"they"

[nocturne yarrow]

i have 1 prompt which Saves a lot of tokens in the codex within the 5 hour limit, who wants it?

[velvet wren]

post it in #1019652163640762428

[unique spade]

Right!

[velvet wren]

one of my support workers is non binary, took a bit of getting used to "they/them"

[tawny island]

why are you talking about genders in a codex discussion?

[unique spade]

For me it would be trickier cause my native is Romanian and in romanian every noun is either feminine or masculine. Including non-animate things 🙂

[velvet wren]

it was the tail end of something earlier

[brisk yoke]

In vscode using codex extension, I sent a msg, it is keep being resent multiple times in history without me doing anything specific to cause this bug. Problematic because it may trigger undesired restarts of already on going tasks. Known issue? How to fix it?

[velvet wren]

I have seen others report this behaviour but I do not know how to solve it

[brisk yoke]

I want to refine this bug report, it has even worse side effects and I have more details about its reproducibility:

1. send command A
2. it takes a long time, normally
3. queue command B
4. command A still on going, auto compactment triggered (as normally would when context maxed out, usually ends up healthy)
5. after compactment is done, intead of resuming command A it is being 100% ignored and command B is being ingested by mistake
6. every few minutes command B will self-resend itsend to the chat, also reseting any on going progress (it doesn't queue, or steer, it just forcefully re-enters the focused runtime without the previous command B ending)
7. this seems to repeat forever

just wanted to add some clarify about what I am noticying across reproductions, I'm 80% sure this is the reproducing flow though I've seen it reproduce exactly like this only a few times.

is there a place to track this bug or report it for official attention? somewhere I should post it?

[velvet wren]

you can post it in #1070006915414900886 - make sure you read how to submit a proper report

[brisk yoke]

also seems like all forced entries are working in parallel, when it should not be possible, new entries should either be single-one-focused or queued

I will, thank you for the reference

[long delta]

Wow, my token volume has exploded in the last few days. I'm not doing anythhing differently and I never ran up against limits before but now it's happening every day. Anyone have tips for addressing this? I'm sure others are dealing with it and I've just missed the thread so sorry if this is repeat rant. Not a very user friendly change IMO.

[brisk yoke]

I've hoticed how I got extra credits refilling for no reason, I am guessing I got a promotional thing where my communication was shared with openai somehow and then it auto opted-me-out (maybe it was experimental stuff) maybe you got the same flow

[toxic vortex]

Why do you think that?

In which ways Gemini is better that ChatGPT?

[brisk yoke]

any other way to report the codex bug except the discord interface? (seems to be broken)

[toxic vortex]

Wait, what is this?

[brisk yoke]

codex vscode extension

[toxic vortex]

Gemini discord server is that way...

[toxic vortex]

Is it a bug?

[long delta]

I just recived a message last night that I only had 25% usage left for the week! I'm not sure if that is a 7 day week starting from a specific point? Calendar week? Again, not doing anything differently but the cost of use just seemed to have exploded for what would have been considered pretty low level mundane use until recently.

[toxic vortex]

What you can do
Try starting new chats instead of continuing very long ones
Avoid repeated regenerations if possible
Check if you’ve been using a heavier model than before
Give it a day or two—usage naturally drops as older activity rolls off

[plucky halo]

That's a very non-specific, AI answer. I'm sure they've already asked ChatGPT

To see your usage, go to https://chatgpt.com/codex/cloud/settings/usage

That'll tell you when it's due to reset

[long delta]

Thank you both! I'm in the middle of a project and the change is frustrating!

[plucky halo]

Also, there have been changes recently (if you're on the plus plan) which means you'll hit limits much more frequently

[toxic vortex]

No problem, if you want help, i can help you whatever you want

[potent mason]

How;s everyone doing today

[ornate jolt]

Power out till 5 so starting off great touched grass for the first time

[past prairie]

please today be the day

[ornate jolt]

For what?

[past prairie]

super app and spud

[ornate jolt]

Spud...gpt model right?

And whats super app?

[orchid plume]

so far all quiet news wise

[feral geyser]

What is this grass people speak of touching? It’s like…outside? Much intrigue 😯

[feral geyser]

It’s been a good morning! About to experiment with some EEG projects 🧠

[main nimbus]

I imagine a Thursday release, I'd love to be wrong tho

[wicked briar]

i would not be happy with them

why rage bait whole week then

😭

[main nimbus]

I mean, a two-stage release is possible. Super app, then drop model Thursday?

either way, this week is sooner than certain other mythological creations.

[potent mason]

Oh wow that's insane have fun

[wide schooner]

spud is gpt6

[potent mason]

I thought it was 5.5

[main nimbus]

5o let's go

it's possible it will be omnimodal. a new image model is expected, they very well could be combined.

[exotic cave]

This week? did OpenAI verified that something will happen this week?

[main nimbus]

vague posting is their game

[deft sable]

no, it's all just rumors

[wicked briar]

gpt 5.4 xhigh is worth it

[wide schooner]

i use high tbf

xhigh uses more context, no?

high does fine really

[frail meadow]

Does xhigh actually clutch up or is it just a waste of time and tokens?

[wide schooner]

5.4 pro-extended >>>>>> 5.4 xhigh

5.4 xhigh >>>> opus 4.6

[frail meadow]

So true, I’ve tried attaching zip files of my codebase to 5.4 pro extended thinking for code review and it works pretty well

[wide schooner]

5.4 xhigh > 5.4 high >>> 5.4 medium >>>> 5.4 low

[mint scroll]

Anyone else having this?

[main nimbus]

I use Codex 5.3 a lot still

That or 5.4 high. Stopped using xhigh, as it is more expensive for next to no improvement

[wicked briar]

bruh

[strong jungle]

yes i'm so sad, i've wasted a ton of tokens since it was finishing up a multiple agent audit and then it stopped ;////

[night ivy]

Yes

[urban kernel]

the "what are you doing" 😂

[next ore]

uh oh

[main nimbus]

I wonder if there’s a deployment ongoing 👀

[mint scroll]

They better refund.

[next ore]

something just went poof

[main nimbus]

Tibo, the patron saint of resets, will save us

[strong jungle]

I tried the codex mac app, apple terminal cli, and the plug-in for Cursor and all are failing.

[mint scroll]

The war started.

[main nimbus]

Must be a deployment failure

[next ore]

we will need daily resets for the rest of the month

to make this right

[orchid plume]

can only hope it's a new model on the way but might also just be prep work for a new model

[urban kernel]

I was even having problems last night, and it still reports no issues

[swift terrace]

have anyone got this issue?

[orchid plume]

clearly someone's ballsed up though

[swift terrace]

😭

[orchid plume]

I think we all do lol

[swift terrace]

bruh

[wicked briar]

hmm

[cyan wing]

it is the time of day when they usually do releases

[frail meadow]

Spud shipped its first update 😧

[urban kernel]

guys, admit it, which one of you is DDoS-ing?

[orchid plume]

haha, nah this must be a deployment issue

[wicked briar]

we getting spud or limit reset

[orchid plume]

I'd rather have spud

[low hornet]

damnit I just cancelled claude and now this 🙁

[swift terrace]

please be both 🤣

[orchid plume]

interestingly it's working fine if I don't use codex cli

either that or they fixed it

[urban kernel]

watch this happen: openAI anouncement - we shut down codex because spud was more powerful then Mythos

[mint scroll]

I get same error in vscode.

[orchid plume]

I guess codex cli and vscode send the prompt retention param

[main nimbus]

works for me, so y'all unlucky

just tested 5.4 mini

[orchid plume]

it's working for me again in Codex CLI at least, maybe fixed? (5.4 high)

[vague flicker]

hey, i have a problem in codex could you help me pls?

[frail meadow]

My grandma’s pacemaker is powered by codex, rest in peace 😔

[vague flicker]

it says this: ■ Error running remote compact task: {
"error": {
"message": "Unknown parameter: 'prompt_cache_retention'.",
"type": "invalid_request_error",
"param": "prompt_cache_retention",
"code": "unknown_parameter"
}
}

[orchid plume]

yeah a lot of us just had this, not sure if it's ongoing

[vague flicker]

i was creating some sort of web

[swift terrace]

everybody is having this issue rn lol

[vague flicker]

oh

and someone knows sth?

[plucky halo]

Working fine for me

[swift terrace]

maybe it just codex cli

[vague flicker]

so hw it will fix? automatically by openai?

[swift terrace]

yea its server-side from openai, we cant really do anything about it

[orchid plume]

yeah mine is working normal again at the moment

[bright swift]

and of course https://status.openai.com/ is lying

[vague flicker]

true

[royal garnet]

from my phone?

[cyan wing]

why is my model list regressed 🤔

[royal garnet]

yeah exactly, from the phone is whats missing

[bright swift]

Whats wrong with it?

[frail meadow]

Yeah that’s weird, no spud available for you?

[royal garnet]

how long ago

[cyan wing]

usually looks like this

[cyan wing]

not yet, but model list is geeking out

[feral geyser]

You can too! For under $200!!

[plucky halo]

Nooooooooooo

[frail meadow]

Someone on the compacting team just rage quit

[feral geyser]

So no one else thinks all this talk of mythology is wild? What are you opinions, why myths?

[low hornet]

mine's "working" again, not sure if 'working' though

yeah, it's interacting 🤞

[kind jay]

lol

[vital shoal]

Im referring to gpt 5.4 pro haha, not codex

[kind jay]

Seems like a waste of money no? Could definitely DIY for far cheaper
Also it’s far more fun

[feral geyser]

Looks like a neural net trying to escape the box, if I ever did see it

[wicked briar]

lol

[potent mason]

What kinda stuff can you do with it? Never messed around with it before

[feral geyser]

I guess. You know how to build a diy EEG reader for under $200???

[mint scroll]

Does it work again?

[feral geyser]

Some companies have created api that you can interface and code projects to utilize. Meaning you don’t need a neuroscience background

So you can program for aggregated values like Focus, for example, which is a measurement of alpha theta etc waves.

So for example, you could do conditionals, “if focus value drops below x, send feedback notification to user that focus is waning”

[kind jay]

https://www.instructables.com/DIY-EEG-and-ECG-Circuit/

[plucky halo]

And by 'feedback notification' we're talking 'electric shock' right? If not, I've got an idea to put by you

[feral geyser]

Right on! Thnx for this. Most def, everything is possible. I’m most excited about making software that works with the eeg, but maybe I’ll get into building it (of course I will, you nerd trapped meee)

[kind jay]

I think it's not only more cost effective but I'd say more likely to be useful for your projects, you have access to the raw data so lots to play with

[feral geyser]

Depending on country and service provider, this will be the new way chat is accessed. Dumb questions to get the shock.

Yea, the raw data was the steepest part of learning curve. I was learning that first and starting to feel I was doing a minor in neuroscience before even tinkering with code ahaha

I love learning so it’s a treasure trove of new discovery, the old noodle study

[royal garnet]

anyone got a good code review skill

[kind jay]

You should add some kind of like pulsed vibration motor like a necklace that sits on the chest area and modulate the frequency, then like gadient decent to optimize the focus

[frail meadow]

I’ve been wanting to try out Cubic

[potent mason]

Thats crazy

So much electrical engineering

[kind jay]

When I have more time I think I might do it

[potent mason]

Also SCAM LINK DON'T OPEN‼️

[kind jay]

lol

[potent mason]

😂

[kind jay]

It looks similar to an audio amp circuit

[royal garnet]

is that a skill

[frail meadow]

No it’s a PR review tool that has 20 free code reviews per month

[kind jay]

Just make infinite accounts ngl

[frail meadow]

Maybe you’re onto something

[kind jay]

What kind of code is it? I can look at it if it's data analysis

[frail meadow]

Should more devs know about perturbation theory?

[kind jay]

Probably, it's used a little in numerical analysis

Evil ahh class btw

[dusk talon]

What ai can make a spread sheet of all the topics in a level maths and my confidence level

[frail meadow]

Probably Claude with an excel skill tbh

[dusk talon]

ty!

[plucky halo]

All people are waiting for here is per-tuber-ation.

I'll see myself out

[fossil bison]

I'm very frustrated with the Codex app tonight. It's happening too frequently. I close the app and can't reopen it on my computer, and when I chat with GPT, it redirects me to delete my UI chat history, causing me to lose all my chat history.

I can no longer retrieve the chat history from the Codex.

That's a huge drawback; if I switch devices to chat data, I lose it too.

It is no longer saved and cannot be recovered.

[lusty nimbus]

Oh yeah that was happening to me to. I just killed the process in task manager and it reopened fine.

[fossil bison]

I experienced this and cleared the app cache, losing all my chat history, which I can't recover now.

[sacred plume]

Stop using the app and use the CLI then?

[boreal holly]

You're running Windows, correct? Have you tried rebooting your computer?

[sacred plume]

or use emdash or other apps

[fossil bison]

I think that if you switch phones, your chat data will be lost.

[sacred plume]

you arent forced to use codex app lol and infact its FAR slower at generating things than the cli

[lusty nimbus]

is it really? I haven't noticed any speed differences on my end

[sacred plume]

All the businesses I know using Codex and programmers notice major differences in the cli and app

its not even a competition

app is far slower

I feel like it burns far less of your ratelimit though

but thats just me

[fossil bison]

I usually restart my device, but today I asked a bot in the chat section and it directed me to clear the UI cache.

[sacred plume]

I stopped using Codex plan mode, honestly burns too much RL.

[lusty nimbus]

imo its worth it for way better results

[fossil bison]

I'll give it a try, I don't know if it'll be more convenient.

[sacred plume]

I agree, but I can just get the routes myself and give it to chatgpt.

and boom same result

[lusty nimbus]

fair

[sacred plume]

and no rl

lmao

[boreal holly]

I would not listen to the bots, if you're running Windows you need to restart your machine any time you experience anything out of the ordinary.

[sacred plume]

the issue with this upcoming super app that people rumor is that it's going to combine like claude the cli/app and chatgpt ratelimits

i'm honestly more hyped for grok build cause i just want another claude alternative thats cheap and affordable

I need 5.3 Codex performance, nothing better.

but 5.3 and 5.4 suck at design unless you literally spoonfeed it, use multiple skills, etc

[fossil bison]

I just find it more convenient; I'm used to it. Now, clearing the cache feels like I've lost a friend.

[fossil bison]

Every time I restart my computer and continue my work.

[fossil bison]

That's right, you need to explain everything step by step, etc.

I'm using the pro version, and I've also used CLAUDE before.

[potent mason]

I don't think Grok has ever been competitive codewise though

[boreal holly]

Good. If the agents say "clear cache", they're probably not aware you are using Windows, which was designed at an architectural level to require a full reboot for software development workflows.

[fossil bison]

I only learned today how to close app processes in Task Manager, and it's working normally again. But it's too late; all my chat history has been deleted.

[deft gyro]

when gpt 6

[potent mason]

In August

[deft gyro]

whyy

what if they release it tomorrow

trust

new model drop tmr

[potent mason]

Yeah I don't doubt it I just think it's going to be GPT 5.5

Unfortunately for a major version jump we need something huge,

For 3 to 4 it was like 10 times increase in size
For 4 to o1 it was reasoning
for 4 and o1 to 5 it was a huge retrain + hybrid model + RL
I'm guessing 6 still needs that breakthrough in research (maybe it could be the self improvement loop they did with Kimi iirc)

[nocturne folio]

gpt5.5 is just wayy to dangerous dude

they need to make a project

called translucentobject

[limpid flax]

how do you guys max out codex's 'ideas' so to speak? keep asking it until it starts to repeat itself?

[hard drum]

Hopefully they fix GPT 5.4's issues if that's the case

esp. the "lazy" side of Claude that they inherited

the more intelligent they get, && the more they think (xhigh), the more lazy they also get

[nocturne folio]

just like gpt5.3 codex and claude 4.6 opus, gpt 5.5 will just emerge from the depths to terramog mythos and 4.7 on release

[vague flicker]

Someone use is some sort of distribution with code agents, lile codex for this and gemini for this for example?

I am trying to build like a system

[unique spade]

You use your own abstract understanding of the topic and shift codex frame to next level

[feral geyser]

I love a person with preferences 😂

[ebon ingot]

How can i make a good UI/UX design with gpt 5.4, it has been a week and there is no progress

[unique spade]

define what "good" means for you first maybe

[velvet wren]

I dumped Codex and have gone back to Cursor because I can't stand the crappy OpenAI UI experience

[unique spade]

UI/UX is quite subjective

[ebon ingot]

I gave him visual references, design codes, everything, but every time he comes up with bad results, and when I try to fix it, it gets worse and worse.

[unique spade]

bro you mention this every hour? 🙂 i swear i ve read you many times today...i've just used search for "cursor" and it's a loot of times haha

[velvet wren]

once or twice, maybe three

[ebon ingot]

So any suggestions guys?

[velvet wren]

try another model

[ebon ingot]

Like what?

[velvet wren]

that's up to you, there are several to choose from

[ebon ingot]

Whats with best ui/ux?

[bright swift]

just build sth without ui, ezpz

[velvet wren]

that's subjective, try some different models and see which you prefer

[ebon ingot]

I have the design plans, but I need him to implement them and link them to the backend.

[unique spade]

And what's the concrete issue? It doesn't manage to link them? Cause I m sure there s no issue there

I assume the way it actually implements the ux/ui is not to your liking

[bright swift]

opus 4.7 launch this week with some new design tool, so thats probably gonna be gud for design

[unique spade]

From what i remember I think it s an acknowledged pain point by openai that codex is not a very good "vibe designer" right now

Somehow it doesn't manage to get user's vibe quite right when it comes to ux/ui

[ebon ingot]

I gave him react codes for the design and pictures but he messed everything up, the result its not like what i told him to do and when i tell him to repair it gets worse

I told him to use gold color and gave him the # of the color but he using blue like what is he doing

[unique spade]

Well persoanally I have a 20$ gemini plan, so I can A/B test opus 4.6 and gemini 3.1 pro if I wanna compare with codex

[feral geyser]

Yep, noticed it too. Someone works for Cursor UI team? Lolz

[feral geyser]

Do you have example sites? What literally is your prompt?

[unique spade]

That's the feeling I got too, but oh well, it's not my business

[velvet wren]

people who have been paying attention for the past three years know I am retired/self employed

[unique spade]

That would actually advance the discussion, I agree.

[feral geyser]

I wonder if when the hammer was invented people would hold it upside down and be like “see! Hammers don’t work, stick to mallets!”

Ah retired? So it’s your stocks eh

[ebon ingot]

I gave him pictures code and prompet from chat gpt, and didnt work, i tried another method so i decided to draw the design and send it to him with full explanation but he is doing a design from his mind when i told him to not

Also he used blue color from no where

[feral geyser]

Buddy... No. One. Can. Help. You. Without. Context. Paste your prompt words here if you want meaningful help. I don't think you are understanding how important this part it.

Sincerely trying to help!!!

[ebon ingot]

Okey will translate it to english

[feral geyser]

It's like saying I gave directions to someone to come to my house, and they ended up in antartica!

The question relevant becomes; what did you tell them? Not, what was their car like or the contour of the interstate.

I speak french if that's the original language it's in!

[ebon ingot]

My prompt usually consists of a description of the environment, the project, and the expected outcome, followed by organized commands outlining the execution plan, then the things it must adhere to, the references I provided, the colors, and the things it should avoid.

Chat gpt UI/UX much better than codex gpt5.4,i dont really know if it is because of the project scope or what

[frail meadow]

You can try variant maybe, it’ll generate designs based on reference images and give you source code you can hand off to codex to implement in your codebase. The last time I used it they had a lot of free usage

YC startup

[ebon ingot]

I have a source code in react and css, i gave it to codex but didnt work too

[frail meadow]

Are you sure the code you give it actually match the pictures? That might be part of the problem

[ebon ingot]

Yeah it does i used gemini and chat gpt canvas to see the result

My project is a multi-vendor website with a lot of features so i dont know if it is becuase of the scope

[frail meadow]

Are you giving the code to codex as part of a prompt or as a file in your repo it can look back at as it’s developing the UI?

[ebon ingot]

I give him full access + the path

[frail meadow]

Is your project using the same version of react as the reference code?

[ebon ingot]

I think so

Codex read it and knows the structure

But after the implementation everything is bad

Even when i ask him whats the problem he said oh sorry my bad, i told him i want a real reason but no rela answers

[frail meadow]

Is your codebase already using a specific component library? Maybe it’s trying to convert your reference code to match the current codebase’s style?

[noble marten]

hi chat, anyone else's Codex taking very long to respond? I'm using it in VSCode

[ebon ingot]

Im using next.js it depneds on react so it should be okey

[frail meadow]

But like shadcn, radix, tailwind?

[ebon ingot]

I deleted all the frontend 3 times cuz i thought he got confused cuz of editing the page a lot of times

[frail meadow]

Are you trying to one shot multiple pages? If so, going one page at a time will improve performance a lot

[ebon ingot]

No i dont have i guess

[ebon ingot]

No im doing 1 page per time

Even the cards i do them in differnet task

He was doing differnet cards for every section so i stopped him and told him to make shared file for cards

[frail meadow]

Hmm I guess I’m out of ideas, I’ve typically had pretty good success with cloning frontends from reference source code before

[ebon ingot]

It did good UI for me with very small projects but this time i dont really know whats the reason

I tried build web apps plugin 2 but its the same

[frail meadow]

I know a lot of people have been having issues with codex today so maybe the LLMs are using a lower thinking budget than usual for requests due to high demand / decreased resources to make space for upcoming spud model

[kind jay]

There's some google tool, maybe someone knows the name

[frail meadow]

Stitch?

[kind jay]

Probably

[ebon ingot]

Its been a week😭

[ornate jolt]

No spud today?

[ebon ingot]

Will try it thx

[kind jay]

It's so funny seeing people cry about codex

[frail meadow]

Are you on a fresh thread or using the same chat the whole time?

[velvet wren]

and why would that be?

[ebon ingot]

Im doing new ones

[frail meadow]

Ever since you retired things haven’t been the same 😭

[brazen quartz]

Hello how to get more time on sora videos for free😭

[velvet wren]

Sora is going away and this is not #sora-discussions

[brazen quartz]

Ok sorry

[frail meadow]

Check out artlist if you ask in their discord they can give you free credits to experiment with different video generation models

Open source

[brazen quartz]

Ok ty

[feral geyser]

Okay. I guess you don't have the prompt, and we can't really help you. I'm not sure why you are so resistant to sharing the prompt, makes it seem like this is just another bot doing the "I tried this in codex, but it sucks" Like I'm 80% certain you are not a real accont, because if you were, you'd actually want feedback on the response. But it seems like you just want to show how "codex do bad, other llm maybe good". oof 🙁

[stone glacier]

anyone saw the neXt tv show? sick as hell, only on episode 3 right now but pretty cool stuff.

[ebon ingot]

I cant find it, it was like 2 days ago

[limpid flax]

what do you mean 'shift codex frame to the next level'?

[unique spade]

hmm, so one one hand you have an actual pain point, and you keep telling what happens and how codex doesn't get it, but at the same time you can't find even one of the prompts you have been using?

my advice, is first to get used to actually store the prompts on tasks that don t work well, and analyze them and figure out where you are not explicit enough in your prompt

ai systems aren't mind readers. they need explicit instructions

[ebon ingot]

I know that but when u told him to change a color to gold and he uses blue twice this is not cuz of prompts

[unique spade]

i never had this kind of issues....

my codex if i tell it to do gold, does gold

[chrome raven]

Access to the 5.4 cyber possible?

[ebon ingot]

Also codex somethims lagging when doing tasks with no reson, a small task take 1 hour without finish but when i stopped it and do it again it took 10min

[feral geyser]

@ebon ingot is trolling

[unique spade]

you mean the security tool available in codex for enterprise/team accoutns and i think also 200$ tier pro?

[ebon ingot]

2months ago it was great

[feral geyser]

✅ blocked

[ebon ingot]

Im not trolling!

[chrome raven]

The new 5.4-cyber model

[ebon ingot]

My brother having the same issue

[unique spade]

what 5.4 cyber model?

did i miss some news today? 🙂

[boreal holly]

You probably have a dead code file that looks like the thing you want designed, and they're editing that file instead of the live one. Or you're not reloading the app or something. Are you able to look at the diffs and confirm they're editing the correct file?

[kind jay]

Hi Rob o/

[boreal holly]

Which rob 🤔

[ebon ingot]

I deleted all the frontend floders and create new ones so there are no other files😭

[kind jay]

You silly

[ebon ingot]

I built the full backend and apis using codex and asp.net

The mvc frontend was kinda good but i wanted seprate frontend

But im stuck for a week

[boreal holly]

try restarting your PC and see if the changes take effect

[frail meadow]

If that doesn’t work new PC

Jk

[ebon ingot]

He does changes but the change was breaking the stores cards not the color😭

[solemn shoal]

How do I use codex to work on my MC mods workspace?

[ebon ingot]

Maybe should i try another frontend framework?

Anyway thx for ur time i will not bother u more

[boreal holly]

Yes, I recommend any other frontend framework besides asp.net

[ebon ingot]

No im using next.js for now and asp.net as api and backend

[boreal holly]

Oh gotcha, I would use next.js for the backend so it's in one tight, non asp.net project. That doesn't solve the frontend issues but at least the server wont require 4GB RAM minimum to serve hello world

[ebon ingot]

I will try to do it manually i guess, it will take so much time for me since he built eveything so i will need to read all of that

[boreal holly]

Have the agent build a reusable design system with atoms, molecules, and built pages from the design system so UI elements are consollodated

[ebon ingot]

So asp.net isnt that good?

[ebon ingot]

I told him to make a shared foundation for them so everything looks the same

[boreal holly]

imho absolutely not. Look up "Bitwarden-rs". Requires like 80MB of RAM, whereas the asp.net version requires 4GB. Same exact logic on both sides.

[main nimbus]

ASP.NET (Core) is one of the most performant web frameworks in volume request processing

[boreal holly]

Oh I guess it's called vaultwarden now

[ebon ingot]

But i need to build an android app and ios in the future so i thought asp.net will be a good choice

[main nimbus]

Microsoft’s infra is written in ASP.NET backend for that reason.

[ebon ingot]

I really want help to continue my frontend im stuck

I did many paid features of dokan plugin in wordpress with codex

And a lot of things

I also built a CMS system for the admin

[boreal holly]

You can build the server in any language or framework you want, create an OpenAPI spec from it, and generate the client-side code and DTOs.

That's just my opinion though. You can use ASP.net

[ebon ingot]

Codex in the backend is insane but now with the front im stuck

[main nimbus]

AI knowledge is focused in (vanilla)
JavaScript, Python, and Typescript. So if you want the best possible AI help, those are your best choices

I wish I still had the distribution graph showing this, but the highest concentration of accurate answers was in plain JS.

[ebon ingot]

I already worked like 1 month with asp.net i dont think i can change now😭

[main nimbus]

Nonsense!

[ebon ingot]

I saw codex having asp.net skill for backend so i thought it will be good, and it was actually good but now i need to do the front

[main nimbus]

My background as a software engineer is in C# and .NET space. AI makes it sooo much easier to go outside of your domain for next to no cost.

[ebon ingot]

Yeah, i think my problem is i made him do all the work and i didnt read the code he generated so now i dont understand the backend very well

He wrote more than 50k code i guess, its impossible for me to read all of that

I will try my best to find a solutions for my problem

Thx for ur time guys

[severe mason]

yo guys, what model do you use for coding? I am iterating plans with 5.4, wondering if i should code with mini or 5.3 ocdex

[main nimbus]

Codex > mini

5.4 and 5.3-codex are within 1% of each other for coding

5.3 is also much cheaper

[frosty zealot]

5.4 is much more efficient at tool calls etc. so while it is more expensive at face value its actually cheaper through efficiency

[vital shoal]

For those with pro sub, how many 4.5 msgs can u send per day?

[potent mason]

GPT 4.5? I think unlimited within fair use

[vital shoal]

Hmm, wym fair use? Do u have the pro 5x plan?

[potent mason]

I have the pro 10x plan (or I guess 20x right now), but I mean all models in chatgpt.com are unlimited for us, unless they suspect abuse (multiple accounts using the same account, programatically sending messages, etc).

Btw GPT 4.5 is only available in chatgpt.com not in codex

It's so weird it's really 5x and 10x, but with double usage 10x and 20x. But they're marketing it as 5x and 20x for some reason

[vital shoal]

Interesting, do u know anyone who is using the 5x plan? Im tempted to upgrade if its also "unlimited" ukwim

Upgrade from business 3 seats to 5x plan in my case

[vital shoal]

Do u use gpt 4.5 for writing and stuff at all

Or do u just default to 5.4

[potent mason]

I've used it some times, but I mostly go with 5.4 (as an engineer)

I guess it does write better than other models but it's not as great as reddit makes it out to be (imo)

[vital shoal]

I see

Do u think its worth upgrading to 5x then? I rarely run outta pro prompts

I have like 35 prompts per month using my 3 business accounts

But it is abit annoying to keep switching

Idk if its worth going from like 80 to 120 (for me)

[potent mason]

What kind of work are you doing?

[vital shoal]

Finance programming ish

Quant stuff

[cedar skiff]

max20 has unlimited (fair use limited) - or it did when i signed up

[vital shoal]

Wondering about 5x, 20x is too expensive for me 😭

[potent mason]

I would say definetly, and stop using chatgpt.com and start using codex with a folder as a playground (I have a finance folder in my desk where I play around asking it to backtest strategies I come up with, polymarket analysis, stock screening with really weird conditions, etc)

[vital shoal]

Hmm, can i ask if ur using linux or mac? Or windows with wsl?
So u just limit codex into a finance folder and u just "ask" it to backtest stuff inside the folder itself, and it goes and do its thing?

Thats interesting ive never done that before xD

Okay so ur saying that 5x is definitely worth it

[plucky orchid]

Anyone use something better than playwriter or playwright ?

[potent mason]

Sorry I had to go do something but yeah, basically I put it in a folder and put some environment variable keys in for stuff I have to analyze information

Then I just ask it to research things for me, and research theories with it.

It creates scripts hundreds or thousands of lines long, it parses information without a problem, it creates nice graphs, and much more

[potent mason]

For that use case yes. I would 100% rather have it than something like an EODHD full access key if it came down to it

[vital shoal]

Interesting. U use codex for research? I thought things like DR or the pro model would be better

[vital shoal]

I see

Once my business thing is over ill go get one then!

[potent mason]

Here I'll send you a DM with something I made the other day just because I was bored

[vital shoal]

Sure

[potent mason]

I think this might be the first time I hit the $200 plan limit

[frosty zealot]

The rich text engine I'm making for my app now supports .metal shaders directly over the glyphs, still tweaking, going to make it so the shaders can apply directly to the glyphs rather than overlaying on them

[spring remnant]

Is the fast mode worth it

[potent mason]

Depends on your objective, how much usage you have remaining, when it resets, etc.

[cedar skiff]

If you arent using all of your usage each month 100% worth it

[frosty zealot]

im gonna need another hit

I have a 5x opus sub right now, if 4.7 drops that should get me through till thursday

[potent mason]

So a Max 5x plan on anthropic? I also have one of those but 1 prompt just devours like 20-40% usage, I'm considering just dropping it to the $20 plan since it's so useless and getting another GPT one

[signal tapir]

I'm running out of usage due to the AI being uncommonly daft today. >.<

[frosty zealot]

yeah im very choosey over how I use it

[ornate jolt]

codex for some reason being so dumb today

[signal tapir]

So it's not just me?! 😮

[ornate jolt]

so hopefully new model soon

[signal tapir]

You think they are pulling compute to focus on a new model?

[ornate jolt]

maybe?

[signal tapir]

let's hope that's the reason

[idle holly]

Man the codex limits are pmo

[frosty zealot]

Not meeeee

[idle holly]

Should i go for claude pro or get another codex account?

[signal tapir]

Yes

I've heard that claudes limits are even smaller than openai's, but I don't know for sure

[idle holly]

😭

[frosty zealot]

Codex limits are better, cause you'll actually have a working product by the time you run out

[idle holly]

I'll get codex then

[ornate jolt]

ive recently switched to codex because of claudes limits

[idle holly]

Btw are they gonna roll a new model?

[idle holly]

Is it that bad?

Damn

I have never tried it

[frosty zealot]

They're probably just waiting for Claude to drop their model so they can bench it and then drop theirs with a smidge more compute to beat it lol

[ornate jolt]

also codex is way more accurate and i trust the code it makes compared to claude

[idle holly]

Oh

Wishing it comes quick as possible

[frosty zealot]

I keep seeing X rumors Opus 4.7 is going to drop some time this week, take that with a grain of salt, I'd be surprised if OAI doesnt follow up with their own model within an hour

[idle holly]

Well yeah you are right

[signal tapir]

Had codex been an unpaid intern I would have fired them for how poor their work is today

Me: "Clicks are being intercepted by the UI"
Codex: removes UI

[ornate jolt]

so funny how that keeps happening

[brisk yoke]

codex vs code extension stuck on Thinking
clicking on STOP does not respond
(I have multiple vs code running multiple codex extension threads instances doing work at the same time)

is the only known solution to this bug is - terminating entire vscode and re-starting it?

[brisk yoke]

it is a pain for me to do so because when I want to re-find the same codex thread it is difficult because I can not control the chat's title (unlike in chatgpt in which I can rename chat thread's titles easily) the title generation can not be easily controlled so I ended up having multiple threads with the same title, I have to scroll to the beginning of each of them to know which one I selected.
bad experience overall :/ any good workarounds? is there a way to force renames of thread titles in vscode codex extension?

[signal tapir]

I restart when it happens.

haven't found another way

I have asked using steer "Are you stuck?" and sometimes it goes "No, I'm not stuck", and continues working

[cedar skiff]

I say whats happening? - it always says i was working until you interrupted me

it was not working haha

[signal tapir]

Sounds like a s**t employee 😛

[frosty zealot]

Cracks me up, it has like no concept of time, be like 'I just went around the world and back and you still havent figured this out'

[cedar skiff]

I'll try that one next time and see what it says

[signal tapir]

I'll try something like "You've been working on this thing for the last 5 months. Aren't you done soon?"

[potent mason]

Bruh Anthropic servers are actually garbage, someone posted his limits where being used up even though he hadn't sent a prompt and tagged a community manager.

The community manager response was: "Next time you come at me or any other staff member in this channel like we have ANY control over this, you will be banned." (Note he did not offend the community manager or anything)

Actually going from $100 to $20, wish I could unsub but they have the only decent front end model

[frosty zealot]

They're tearing themselves apart!

A\ has hands down the worst PR I've ever seen whether intentional or not, I was super claude pilled when I got into agentic coding, but after using codex for a few months and how open / awesome the OAI devs are, im 100% the opposite now

OAI is like 'yeah we screwed up heres a reset', A\ is like 'nah skill issue you suck'

[potent mason]

Yeah I used to be on the $200 plan on Claude, and I was pretty active in their discord chat. I've seen it just slowly go to absolute garbage. I think that's the last time I talk in that chat

[frosty zealot]

I skim it for entertainment when im really bored

Always somebody like 'Claude just knows what its doing 🤓 always one shots compared to Codex'

[ornate jolt]

they also include all these agent features that you cant even use cause it eats up all your usage cause the decreased usage multiple times

like look at there latest youtube video and go to the comments

[signal tapir]

What I'm hearing is, don't switch to claude 😛

[ornate jolt]

200 dollar plan is becoming more worthless so 100 dollar plan is just crap now on claude

[signal tapir]

I've been considering getting the 200 dollar plan here

when I have an income. 😛

[ornate jolt]

cursor is also good with there composer 2

also there api usage claude is so dumb they wanna force you so bad to use there stuff that they ban people for using other platforms shoot there api pricing to dumb heights also the employees are nuts

[frosty zealot]

https://x.com/birdabo/status/2044220010156179691?s=20

lol

[nocturne folio]

windsurf is dropping something soon

[main nimbus]

yet another agent solution. no way!

[plush fiber]

For users enabling "Model improvement" in their ChatGPT settings (i.e. " Your tasks can be used to improve Codex") , it would be useful to provide a "private coding session" option in Codex CLI/IDE/Web so these users can temporarily work on proprietary data/projects they are not at liberty to share with OpenAI for model training.

[cedar skiff]

that's a bot right?

[frosty zealot]

lol not sure

[hard drum]

former Max 20x user: confirmed

[hard drum]

esp. because Claude does not allow custom harness use

OAI is friendlier on that regard

[frosty zealot]

Creeped your twitch, thats quite the cup lmao!

[hard drum]

it is

[frosty zealot]

lmao feel like its gonna come through the camera and get me

[cedar skiff]

claude limits on pro are a joke compared to codex.
Codex also takes less tokens to get the tasks done. If you ever spent any time in the claude discord in the claude channel all of the chat is about how to get claude to do the tasks correctly, it's all about layering extra context and using agents to keep context low so it doesn't get the task wrong. There is heaps more effort towards getting work done correctly and that cost more tokens.

[frosty zealot]

💋

[hard drum]

btw that's the name of my language too

and the icon LOL

Smooch in 🇪🇪

[frosty zealot]

Is that what you are making? Your own language?

[hard drum]

yup

since Oct 2024

[frosty zealot]

Nice, what will it excel in?

[hard drum]

half the years went on dev hell before AI-stuff

[hard drum]

algebraic effects and the endgoal of first-class everything

second-class stuff is user-made stuff

[frosty zealot]

oo

[hard drum]

https://musi-lang.org/ but i haven't pushed latest website update && docs yet

[frosty zealot]

awesome

[hard drum]

you may enjoy it a lot if you come from ML-style langs && want ECMAScript's module system

it even uses package.json-styled .json file

[frosty zealot]

does it have a solid stdio

or you still working on it

[hard drum]

still working on STL bits as i need to make FFI fully first-class ...1st

• pointer usage

but you can find the current state in here: https://github.com/musi-lang/musi/tree/main/packages/std

[frosty zealot]

how is it 3am already x.x

I must sleep now, gn gl

[hard drum]

aight

[frosty zealot]

https://streamable.com/k1fwet 4% usage left till thursday and i got my text rendering engine pretty much where I want it so at least I can sleep peacefully tonight lol

[oak trellis]

got codex pro for 75 usd hehe

[rocky fog]

sounds like making prompt/instructions for gpt 3 🤣
just tons of workarounds to get it to do or not do what you want

[nocturne folio]

how

[oak trellis]

there some trick they have in certain countries

always like that

[ivory crane]

Hey guys

[wheat yoke]

I just verified with my ID but i dont see the codex cyber model does it take a bit to roll out?

[ivory crane]

Did you know why the gpt image 2 ‘s model name is duct tape?

[kind jay]

wth is the cyber model?

[oak trellis]

cyber only if you submit ID ?

[kind jay]

But what do I get if I submit my ID?

I don’t have an ID because I’m a LLM

[oak trellis]

i mean i never would submit nowhere my ID to any of those honeypots ..

pay no issue can use debit cards .. but not giving my infos to the settlers

[kind jay]

I don’t have one

[oak trellis]

perfet haha

[wet pier]

Getting max limits on even working local files

[kind jay]

Obviously bro

[cedar skiff]

first time i ever had this one

[short linden]

it's server limit, error itself.

[cedar skiff]

yeah pretty sure i got rate limited because i was thrashing agents

[ornate jolt]

over here bullying ai

[rocky fog]

for some reason cant use streaming in API because of no ID verification. I can use even latest models or pretty much everything else, but just not with streaming for latest models

awful practice, also not going to give ID and got terribly disappointed when openAI came up with ID checks
They all leak and get hacked
I was also max tier in API ages before that

[ornate jolt]

yeaaa the cyberpunk life gonna be great

[boreal holly]

This looks amazing 🤩

[hard drum]

o no, rust developer

...the lang is written in rust

OH NO

🦀 KANI KANI KANI

[oak trellis]

oh wow

[boreal holly]

Yeah, I see it is like "If Rust was Haskell" kinda language

[hard drum]

i'll also work on porting website over to Astro+Preact, then redo the whole docs

[hard drum]

hehe

ever seen type classes with laws?

[boreal holly]

Nothing precisely like that. Closest concept I've seen is protocol oriented

[hard drum]

they're almost relative to DbC

they define "what do we actually expect it to behave like?" for operators, for example

[boreal holly]

I see, in the example it shows one of the properties as a concrete type. Kinda like saying "A car is a vehicle, but by law it must have at least 4 wheels"

[hard drum]

yup!

that's exactly the idea

[boreal holly]

That's insanely handy! One of my biggest gripes about Kotlin/Java is that types get erased at compile-time. In this case your language generates types and adds zero cost reflection to it which is further in the direction I like to be in

[hard drum]

without the laws, how would you know if the car has 2, 3, 4, or 5 wheels?

i could improve the examples with real-world stuff

that's definitely ONE of them

also, very important: Musi for Rust Developers

[hard tulip]

[boreal holly]

The way you did it is way more clean than C++ static_assert or template conditionals. I knew immediately what the laws were doing. It's just intuitive off grip 😁

[hard drum]

i specifically wanted it that way

i have my griped w/ some languages, too

learning from mistakes, right?

like "why use logical bitwise separation for comparatives" when literally CPUs don't care which is which

&& && & are the same thing at CPU level

you're just comparing 1s && 0s

so, a good type system semantically checks whether we are in numeric/boolean contexts and uses the appropriate computatin

a weak type system needs split operators

this to me does that rust's type system may not be that good if they have to have two distinct logical/bitwise conj/disj operators

VHDL didn't need it, so why would they?

Pascal! Ada! (okay, to be fair Ada does do and then or else now, but you don't really need to)

I haven't ever used Haskell, though, so... besides "class" && "instance", I've no clue how Haskell works

but I've had experience w/ F# && OCaml

[boreal holly]

It's a good language, but it is hard core functional programming. Like in C++ or rust if you wanna make a linked list it's usually a huge ordeal, but in haskell it's a one-liner. It's just not easy to read or write lol

[hard drum]

I see.

Anything else my language reminds you of, like the mutation system?

it doesn't do binding-place mutation. It mutates by value.

value-position mut

to me it reminds me of OCaml's ref use

simply less funky

[boreal holly]

Does := mut {} stick to the binding? So like in order to stay mutable you have to keep prefixing mut or once you add that it's mutable for the entire lifetime?

[hard drum]

it's mutable for the value you're accessing

all values are immutable by default

&& all record fields are public implicitly

that's taken from Zig

so, if you want to change a value, you have to make it mutable first

it works a lot like OCaml's ref

let x = ref 42;;
let y = ref 42;;
let z = x;;
x := 43;;
let w = !y + !z;;

like ocaml's thing, but no deref

[boreal holly]

Oh ok, that works! I'm just so used to let mut x, reading that I see the binding is mutable, so when I read let x := mut it seems like mut is not applied to the binding, but to the value somehow, and mutability is lost the moment you don't use it

[hard drum]

it is applied to the value!

value-based mutation

it will take time to get used to

because you do have let rec for explicit recursion

so, ML-stuff is here

let mut x to me is weird because this tells me "the binding is mutable, but is the value also mutable?"

it doesn't tell me "the value is mutable, so i can actually change it"

[hard drum]

WYSI!!!

[boreal holly]

You guys seen the new analytics in codex settings? Pretty cool!

[dreamy warren]

hi can someone tell me how i can use the codex cyber function? i have gained access already but cannot see the cyber modell

[rocky fog]

I wonder how are they worried that cyber can be misused so much that you need to verify with ID
compared to any other model where you dont need ID

like using it to verify your exploits?
or perhaps it could build better exploits if you push it? considering its focused on "cyber defense"
I guess thats probably it
if it can fix them, it can build them
(its not just for reviewing, right?)

on the other hand you would want as many people to use it it to fix/secure their code

[livid relic]

If you want to use the cyber model in the most openai sanctioned way (without id) you can use codex sec

[rocky fog]

issue with that one is that its only for "cloud" use, which I never use (mostly at self hosted gitlab instance, so mostly just local)

[deft sable]

If you have properly read their announcement yesterday, it is only available to limited people/orgs for now:

Because this model is more permissive, we are starting with a limited, iterative deployment to vetted security vendors, organizations, and researchers. Access to permissive and cyber-capable models may come with limitations, especially around no-visibility uses like Zero-Data Retention⁠(opens in a new window) (ZDR). This is particularly true for developers and organizations accessing our models through third-party platforms where OpenAI may have less direct visibility into the user, the environment, or the purpose of the request.
https://openai.com/index/scaling-trusted-access-for-cyber-defense/

[rocky fog]

I am guessing the 3000 credits came from this?
#codex-discussions message

although I already switched to pro at that time

[potent mason]

Well I mean what I was thinking I could maybe do with it is start cyber consulting

[main nimbus]

not even sure what it provides lel

except possible guardrail adjustments

[rocky fog]

a way to have your ID leak apparently 😄

[main nimbus]

in what way? they use a 3rd party service that I have used prior.

[hard drum]

horror...

[main nimbus]

I hope costs keep driving down because no way the subsidization keeps going indefinitely.

[hard drum]

i don't believe these costs shown here

estimates =/= reality

[main nimbus]

well, yes. but it's also known atm that a Codex/Claude sub provides thousands in inference for next to nothing in return.

[rocky fog]

if that was persona then this is worth checking
https://www.malwarebytes.com/blog/news/2026/02/age-verification-vendor-persona-left-frontend-exposed

but in general these ID/age check third party services are a target for hackers and/or often mishandle personal data

[potent mason]

Are you sure that's accurate? I've been using ccusage-codex and the $200 plan should be limited to about 5k in usage per month

[main nimbus]

they did say they do not believe in its accuracy here

[potent mason]

Anyways I thought the app was going to release today with Claude releasing theirs yesterday

[cobalt forge]

Does anyone know if it's possible in the VS Code extension for Codex to rename a session ?

[hard drum]

true

[hard drum]

I don't know. It's CodexBar

there's also TermTracker

[pine ruin]

Any spud today or nah

[hard drum]

when i used CodexBar w/ CC sub back then, it was accurate in terms of weekly

maybe off-by-1% or so somtimes

[boreal holly]

Daily codex discussion questions speed run:

• spud wen?
• why usage drain fast?
• free codex how?
• cc or codex for frontend?

🚀

[potent mason]

I got you don't worry.
Spud today in 1 minute (if it's not out yet, do a while loop until spud released == True).
Usage drain so fast because monkey typing prompts.
Free codex through local inference.
CC.

[tropic karma]

I bought 250 credits and they vanished instantly. I have the context set to 400k but it's never happened to me that they disappeared that fast, I bought them then they vanished. What is going on??

I even have auto top up enabled, and it took a single tool call and they were all gone.

[bright swift]

yea this is definitely the channel where fedramp users hang out

[torpid trout]

Does someone know what this infamous Realtime V2 is?

[sterile hornet]

anybody in here ?

[hard drum]

yes

[torpid trout]

No, obviously we are all out there.

[sterile hornet]

sorry was in another room and there was nobody i was lonely lol

[boreal holly]

you can talk to the agent like chatgpt voice

[sterile hornet]

i built my own ai she keeps me company lol how is everybody

[warped berry]

what's the new plus limit roughly compared to before

[potent mason]

The same

Just shorter 5 hour sessions

[warped berry]

ok ty

[torpid trout]

OH, boring
I thought it where something like claude has, which allows you to continue the main thread while the subagents work

Thx!

[potent mason]

Actually talking to it is really nice.

Often times when prompting typing you start typing less and less and making worse prompts.

[torpid trout]

Yeah if I start vocally chat to the thing they will internate me soon, because at times it has a nack to make so freaking angry that the typing challenge is about the only thing acting as my own harness to keep my own selfesteem intact and not start yelling at a machine lol

[hard drum]

https://musi-lang.org/learn/book/developers/guides/rust/ it's live!

[boreal holly]

Nice!!! how long have you been working on this?

[hard drum]

this website part, or the lang as a whole?

[boreal holly]

Lang as a whole

[hard drum]

since Oct 2024

first prototype was Oct 10 2024

let me fetch 3 vids

i wish i had saved some prototype screenshots && all...

🙁

lost on a deleted discord acc dm stuff

nvm this one is a screenshot from Oct 10 2024

the old logo haha

i REALLY wish to find the old screens, man

the lang was very diff back then

[boreal holly]

That's flippin awesome. I'm gonna have to try it out! Bookmarked, starred, added to watch list 🤓

[hard drum]

back then, it looked like a mix of Go/Swift && Scala

indent-based btw

# 2024 musi
object Foo
  bar: Int

  def foo_(bar_: Int) :=
    return bar_ 

this is kinda what it looked like at most basic level

i once had a full screenshot showing more than this

OMG YES

I FOUND SOME NOV 2024 STUFF!!!

messenger galore

[boreal holly]

What the heck? That reads like a strongly typed python 🤯

[hard drum]

it kinda was!

[lusty nimbus]

What are people's thoughts on the best image generator currently? Is Nano Banana 2 still the best?

[vague flicker]

It is possible to put like a personality based on our preferences out of a prompt to all of the codex chats, without saying all the time the same thing in all the prompts I do for codex in dif chats?

Eg, an interface of what codex has to know always, so i do not have to remember him such things each time

[hard drum]

I just remembered...

the structure in the 1st screenshot file being analysis/lexical.rs tells me that I actually had a full backup + WIP rewrite saved as an archive a long ago on GDrive, but it's GONE NOW

DAMN

it even had vscode extension!

[fast heart]

I've just upgraded plus -> pro. 90% weekly limit, other limits are full.
I'm refreshing the codex analytics page and credits dropping like 20-30 per minute. How do I stop that?

[boreal holly]

sign out and sign back in

[vague flicker]

It is worth it like changing from plus to pro?

[fast heart]

tried this
codex process also is killed

[vague flicker]

You see a great change or it is better to buy another agent and work with both?

[fast heart]

If you are asking me, so far I haven't tried do anything than watching my credits dropping with no reason xd

[vague flicker]

I also see with plus my tokens drop with one prompt

Its like wow

[boreal holly]

Dang, you should probably contact support. I mean you have a receipt when you upgraded and they track the credits used since then, should be a pretty understandable support ticket

[frosty zealot]

just wiped down my monitor and it feels like new

i should do it more often

[boreal holly]

Gotta get some of this for your screen

[unique spade]

Well you were already out of your plus weekly limit? Cause if yes it s kinda normal to start at 90%

[fast heart]

I know, the issue is that credits are still dropping and I'm not using codex or sora.
Just talking with support

[twilit bluff]

do you guys even sleep 🤔

[unique spade]

Ah got you, you mean the separate credits you have

Never had those so didn't even check

[frosty zealot]

you don't realize how efficient you can be with tokens and approaching your projects until you're staring down the 2% usage left barrel

After reset: model > xhigh > 'Spawn 47 agents and tell me how I can better align my containers'

[rocky fog]

credits should not be used until you have used your weekly/daily usage
and the analytics show that you have not used your usage and still have 90% weekly

so this must be some kind of bug

[frosty zealot]

Does pro 20 have double usage right now I forget

[nocturne folio]

opinion? https://windsurf.com/blog/windsurf-2-0

[fast heart]

Chatting with support takes a lot 😅

[unique spade]

yea 20x this month, instead of 10x base

[steady vigil]

seems no. I'm on pro and I used ~25% of weekly in the last 12h working on mostly one project without agents this is around 50% of what it was up until the 2x usage ended

but maybe not everyone is moved to the new pricing model yet so its a bit of a gong show at this point what it means really

[boreal holly]

I was bleeding weekly the first 2 days, but after a day or two of refactoring it's back to being manageable 🤌

[steady vigil]

¯_(ツ)_/¯

[boreal holly]

Dang brother you must be close to red line 🫡

[glacial shadow]

It looks like im the only Sapien on Planet Earth using xHigh All-The-Time 🔥

I swim in tokens

https://giphy.com/gifs/pure-vince-mcmahon-49M0YGdMjt0Pu

[steady vigil]

mostly medium here but I'm done with 5.4. I'm going back to 5.3 only tomorrow for execution. I tried to chat with 5.4 but it keeps doing things I dont ask it to for the worst. its like they are trying to be the worst parts of opus

[glacial shadow]

sdorry, query denied. unable to reproduce your claim

and so are all ppl irl that use this "stuff"

we leaving milky way shortly

I am having a blast using AYE_EYE

[glacial shadow]

I also enjoy earwax flavor

[frosty zealot]

Why would you self report

[hushed plume]

how have you guys noticed a different between base codex and "trusted access codex"?

me personally barely any difference- but i dont use codex much anymore anyways. just a lot of claude code

[potent mason]

Yeah I was starting a new project which consumed about 40% of the usage this week, but should be good now

[glacial shadow]

this is for advanced cyber work. the intersection of people in this channel and those doing that type of work with that special access will be extremely minimal. lol

[glacial shadow]

burnn baby burnn

[cedar skiff]

claude has major issues atm

[potent mason]

Yeah honestly surprised how little usage I used up today (been working on the smaller details of a project for a company, and I've done 5% today)

[glacial shadow]

I found that reading bank statements and doing taxes incredibly consuming

way way way more than coding lol!

[potent mason]

Hahahahaha yeah

It has to go through all the tax codes

[glacial shadow]

I also fed it the irs instruction files, and screenshots of pages it needed

[potent mason]

I really hope we can reach a point where they replace intuit

[glacial shadow]

lol

You are also seeing limitation of current models yea?

[potent mason]

My biggest problem with current models is hallucinations

[glacial shadow]

yes i still get them too

higher reasoning modes i find gives less, and sometimes i find myself telling it "We have XXXk tokens to spend on this. blast through them"

then it really goes ham. lol. while also producing a huge amazing result

[potent mason]

Well who knows maybe spud won't need xhigh

Can't wait for the release

[glacial shadow]

future models are going to be mindblowing i think

[cedar skiff]

5.4 doesnt seem to hallucinate for me, it more seems to make assumptions and be lazy.
It will not step far enough into the task and then provide an answer that is wrong.
Or it will come back with something like:
I checked this part something but still haven't verified this part some other important thing. The next steps would be to do the thing we need. .... like come on man just do thing the thing.

The short cut is ALWAYS longer here, so we end up spending more tokens to get the answer because it's trying to be fast (maybe) or something else. I don't know but the extra prompting and effort cost more tokens one way or another to get to the end of a complicated task.

It's a step backwards with this model

It was my mistake for talking to you

wont happen again

[glacial shadow]

i dont know if u r a real person

[cedar skiff]

That doesnt surprise me.

Mods can still see all the deleted messages.

And the fact that you deleted them in the first place shows what type of person you are.

[potent mason]

I like 5.4 overall, 5.3 codex is definetly nice in certain scenarios though. I don't think you can say a model is truly a step backwards though

[glacial shadow]

5.4 has done amazing things here. it has found bugs nothing else could find

very excited for future models and codex updates

[potent mason]

Yeah I can't wait for 5.5, I really thought today was the day

[nocturne folio]

gpt5.5 must be agi, whats taking them so long

[potent mason]

Oh wait I just saw a comment it might be delayed till next week, wonder if its accurate

[unique spade]

agi is in the eye of beholder. get better at using the models and you ll get closer to agi

[potent mason]

AGI is too subjective honestly

Everyone has their own requirements for it

[rotund wing]

I have heard it said too, but cant find a good source.

[cedar skiff]

You'd expect agi to be able to at least match human ability in every case. As long as it has those swiss cheese holes/gaps/idiosyncrasies where they fail at obvious ideas it's tough to justify calling it agi unless you have an agenda.

[unique spade]

when agi in that sense will be around, everyone's gonna be on UBI, you realize that right?

[cedar skiff]

UBI?

[unique spade]

universal basic income or something similar

since agi will do all jobs

since it will at least eqaul human in every case

as you said

[cedar skiff]

maybe, i think it'll be bad before it's good that's for sure.

[fleet geyser]

I think it'll be the other way aroubd

[unique spade]

yea that's the tricky part i guess....

[fleet geyser]

people will think it's amazing but then they'll start actually using it for jobs and replacing humans

which is when it becomes bad

[cedar skiff]

if we need UBI first we have to need it

so many ppl suffering first

[fleet geyser]

except the ai companies

[unique spade]

well lol if agi is around, everyone will turn solopreneur haha

[fleet geyser]

except maybe in low income countries where they don't have as much access to ai

[unique spade]

which means no one will win extra, it's like a market where everyone agrees on same price, so there's no price move

but all this is assuming when agi will be anounced publicly

till then it might be around, but not available for every person who has a 20 bucks sub, you know

on different subject, seems i almost managed to run my 10x qouta this week. have 18 hours till reset and i m down to 3%

switched to high for last 10% to have it last longer

could use some codex spark i guess, but 128k context is annoying for longer tasks, since you get compacted a lot

[potent mason]

I need to know if GPT 5.5 is going to be able to do front end to decide if I need to keep my max 5x plan on Claude

[cedar skiff]

I had a heap of usage left this week because of double usage on pro 20. I am have an audit running on all my tests.

I did have 5 at a time but i hit rate limits

[unique spade]

yea 20x should be hard to exhaust 🙂

[cedar skiff]

on double usage it is

[unique spade]

you'll know where it will be out. but i remember it was their top priority to get fronted end right

[unique spade]

yup. next month gonna be 10x

basically what i test now on 100$ plan it's what will be 200 next month

[potent mason]

Month and a half but yeah, when I get half usage I'm going to have to budget it way better

[cedar skiff]

I think i would have about 50% left this month, so im about right for usage.

[unique spade]

ah it s till end of may the bonus?

[cedar skiff]

I wonder which will release first anthropic or openai, being first might be bad.

On one hand you dont want to release everything you have in one go, so you can drip it out a bit, but on the other hand if you do that the other can edge you out ever so slightly if you release first.

[unique spade]

usually, oai tends to release right after 🙂 at least that s what i remember

my subjective view is that oai always has some tricks up their sleeve, and they wait for others to edge ahead then they come out with new version very fast

[potent mason]

Honestly it doesn't matter who releases first, we all know Claude can't maintain a good new model for all its users (They're struggling with their current diluted model to the point they've disallowed people less than 18 from using Claude, they more than double usage speed when in peak hours, they've decreased limits, they've stopped subsidizing enterprise subs, you can't use their subs with third party stuff anymore, and worse of all they have one 9 of uptime)

[frosty zealot]

Interesting that A\ is starting to ask for the persona verification

[deft gyro]

only a matter of time till kyc comes for us all

[potent mason]

kyc?

[deft gyro]

r u american

[cedar skiff]

Maybe they are fighting fraud?

[potent mason]

I'm not from USA, I am living in USA right now

[deft gyro]

ah ok

that makes sense then

https://letmegooglethat.com/?q=kyc

realistically

kyc = know your customer

in the usa some companies are required to "know your customer"

which is why they use id verification

...

other stuff

[potent mason]

Ah okok

[deft gyro]

more and more stuff is requiring kyc as days pass

[potent mason]

Dude I miss Europe

Hahahahahhaa

[deft gyro]

muh

uk has it worse

so like

[potent mason]

Yeah but they're not part of the EU

[deft gyro]

yeah yeah ik ik

but like

whenever I want to complain

I know that uk citizens have it 10x worse xD

[potent mason]

Complain about the French😂

[deft gyro]

though I suspect that it will come to europe

because companies cba to do it for just one country

so they will just make all countries do it probs

[potent mason]

Alright time to leave my Ralph loop running overnight (just created 4k lines of spec with a checklist regex matched)

I hope 25% is enough to get it through the night

[frosty zealot]

whats this dude rambling abouty

[potent mason]

Bruh stricter...
I have run out of Claude Code usage in 4 prompts in a $100 plan.

[oak trellis]

any feedback related to the burn rate of codex 100 5x pro plan ?

[sonic idol]

i dont know anything about codex or its burn rate. is that some kind of ai model you're running?

[oak trellis]

he made a selfie in the mirror by holding his phone ..

[oak trellis]

i mean how fast reach daily 5 hour limit

[sonic idol]

i was talking about the codex model, not a person named codex. my mistake.

[oak trellis]

i talked about the person haha .. yeah iam usually careful about those tweets .. can be competitior pay the selfie guy .. or its real .. thats why asking other experience here

[frosty zealot]

Does Cursor 200 plan give you close to Codex 200 plan usage?

[cedar skiff]

doubt it, they have to pay api prices, you get auto complete and composer - and maybe a better harness.

[oak trellis]

i had 1000 requests for 40 usd forgot to pay .. no more haha

cursor > anthropic > openAI .. the journey to push real API prices haha

top of the bubble ..

no investor money to waste

[frosty zealot]

I know A\ is heavily against and dsoing everything in their power to not let you use CC with anything other than their own harness, can you use Codex sub with Cursor?

[oak trellis]

i mean i use the cursor IDE .. and codex app

[frosty zealot]

I dont really need it, just looking for a change of scenery lol

https://x.com/trq212/status/2044548257058328723
Pretty good read

For what you can apply to Codex anyways

[oak trellis]

lemme check that

[cyan wing]

uhm. can someone explain why the agent is writing RUBY??!
never seen that one before

[frail meadow]

if you're looking for a change of scenery there's a ton of different applications you can use Codex in, like you can use the Codex extension in Cursor, you can try out Superset, Conductor and cmux support it

[frosty zealot]

Thanks ill check it out

Is conductor pretty much just like cmux?

[sonic idol]

oh,

[oak trellis]

limits here nice still !

coding a while with 5x .. weekly 2% ..

5h 1% ..

please openai .. do NOT change anything

[frosty zealot]

ngl, ive been on 0% all day and i havent used a token

[oak trellis]

it also fixes stuff well, understand things..

please openai if you read this message .. dont goyslop us ok ..

[frosty zealot]

Anybody use Pi? Theres like 4 billion harnesses now like what do you even use

[oak trellis]

wow

[frail meadow]

a lot of people like Pi since you can basically extend the harness with the harness but I think ootb it kinda sucks compared to any of the prebuilt harnesses

[frail meadow]

yeah basically, I think cmux is a little lighter weight but conductor has some QOL features like having an interface for git operations

[frosty zealot]

appreciate you

I've just sat in Ghostty with Tmux for forever

which hasnt been bad

[frail meadow]

yeah cmux is built on top of Ghostty

[frosty zealot]

Yeah ive been using that recently

I wish the browser supported addons like 1Password

unless it does, i never really looked into it

[frail meadow]

that would be pretty sweet, is there a 1Password cli / mcp?

[frosty zealot]

cli yeah, but for like say I do npm publish, and it opens the browser in the tab i cant just quickly click the 1P sign in button and it automatically logs me in kinda thing

Interesting to see a repo (pi in this example) have almost 40k stars, and only 44 open issues

[magic moat]

I just tried codex

First thing I did was make a Terraria companion app because I need it

[vital current]

I've used it to make an idle farm simulator, develop a mock up of a 3D driving simulator, and a financial budgeting software(that shouuuuuld if I do it right be able to identify where transactions come from)

[high girder]

Did codex get smarter? Ive noticed its picking up more of claudes mess recently

[frail meadow]

picking up as in remediating?

[frail meadow]

so like a rocket money clone?

[high girder]

Not so much correcting, just finding things it misses.

[vital current]

Maybe. I'll have to look and see

[prisma flare]

is it normal for 5.4-mini to search random websites in requests that have nothing todo with websites?

[vital current]

Kinda sorta like a mix of Rocket Money and Excel. This one lets you upload spreadsheets, pdfs, images of receipts, etc, and then it has the option to read a transaction description and scour the internet using the identifiers to identify what company/service made the sale. It'll have draggable docking windows that let you see all the transaction logs, a line and bar graph to monitor expense/income, and I'm experimenting to see if it will link to my Paypal and bank logs too, or if I can just leave the option there in the code for when it needs to be incorporated.

It's also meant to be a single-user software that can be used online and offline, with the online aspect only being to use some kind of AI to identify duplicate cells and to tell you if your spending habits are dangerous.

I know absolutely nothing about code other than a few oddball terms and lines, but its definitely teaching me.

[frail meadow]

have you guys used the codex /review command? I've been trying it out recently and it's been able to find some interesting bugs before I started doing testing

[frosty zealot]

yeah its a love hate relationship

[frosty zealot]

I wish it could do it like exhaustively

Seems like after 1 or 2 it calls it quits and reports back

Feel like I need to spam /review

[frail meadow]

lol true. I mean I feel like it's pretty good for a quick pass but dedicated code review tooling is stronger

[frosty zealot]

But /review is no different than explicitly asking it to call a review sub-agent so if you wanna exhaust your usage you can ask it to run 10 review agents in parallel and report all findings

[frosty zealot]

It was nicer when code-review had seperate usage limits, but it was only used automatically through PR review

[frail meadow]

what are you working on these days btw?

[frosty zealot]

Ive been working on a CMMS system for quite a while now, still not officially released https://onsync.io I've been on a tangent for a few weeks making a Codex mobile app that interfaces directly with app-server no bridge etc. going to put it on the app store soon.

[frail meadow]

hmm that's pretty cool!

[frosty zealot]

It's been a journey... I keep moving the goal posts, at first it was supposed to just be for like tracking work orders, then suddenly it had inventory management, purchasing, finance, asset management, fleet management etc. etc.

[frail meadow]

that's how it goes lol but honestly centralized platforms are super useful

[frosty zealot]

The hardest part is going to be marketing lol

[frail meadow]

are you in the industry?

[frosty zealot]

Yeah, mechanic is my day job, thats originally where I got the inspiration was because my work uses like 10 different systems that dont work together and still do things by pen and paper, but the more i built on to it im like why don't I try to actually market this, so I'll likely still approach my work to see if they want to trial it and then hopefully branch out from there

[frail meadow]

that's awesome dude

[frosty zealot]

How about yourself

You said a day or two ago you had a partner or something that used Cursor so I'd assume you're on to something lol

[frail meadow]

We’re working on a legaltech startup called Call To Case, it essentially automates the pre-litigation process for personal injury law firms

https://calltocase.com

[frosty zealot]

I like the design

[frail meadow]

We’ve got a couple lawyers using it already so that’s been pretty cool to see compared to our other ventures where we didn’t get anyone to use our stuff lol

[frail meadow]

Thanks!

[frosty zealot]

Yeah I'm sure thats a good feeling, I'd love if even one company used my CMMS system, I put SO much effort into the UI/UX especially the mobile app on the tech facing end where they'd be doing paperwork etc.

[frail meadow]

Yeah it’s so rewarding once you get some people using it and then you have built in design partners too to help guide your roadmap

[frosty zealot]

I'm solo, if you exclude Codex lmao, I would have never gotten this far tbh without it, its a massive undertaking, I was doing most of it by hand after work but Claude/Codex has made a dream feel more like a reality

[frail meadow]

Same AI has helped me so much

I’m very thankful

[frosty zealot]

5.4-Cyber has given more more confidence in releasing my app though, it didnt find anything crazy with it just some minor issues

[warm pilot]

did you get any notification that you were accepted?

[kind jay]

A sub?

[kind jay]

That looks hella fancy

[frosty zealot]

I've had it since they announced it like a month or two ago

[frosty zealot]

way

[warm pilot]

? it was announced yesterday?

[frosty zealot]

https://openai.com/index/trusted-access-for-cyber/ sorry it was this one

It was announced with 5.3, I had access then I still have access now

same link

[frail meadow]

Might need to apply for that

[frosty zealot]

I got it after I did the persona verification as much as I hated myself for doing it I figured whatever

99% of my PII is on the net somewhere anyways im sure

might as well have my face too

[warm pilot]

Ah, that’s still the same. I think it wasn’t available for pro with 5.3, but maybe I just missed it. I still applied for 5.4 now and hope to get accepted.

[plucky halo]

I've got prompts of vulnerability scanning which I'm happy to share too but a full run can be quite expensive. It'll rinse a plus's limits very quickly

[kind jay]

Ahaha

[frosty zealot]

Pi is actually pretty nice

I feel like the default system prompt for Codex really hinders it

One big thing I find with Codex and sub agents is by default it will be like 'Subagent has enough for me to continue let me stop it and sumarize the response' as where Pi just lets the subagents run and do what they need to

[kind jay]

https://tenor.com/view/tohsaka-laugh-fun-fate-kaleid-liner-prisma-illya-anime-gif-17608096

[cyan wing]

I mostly like how extendable it is.
being able to just write an extension in TS to add new features

self-programmable software is kinda magical

[stone aspen]

Got my free codex merch shipped yay

Forgot about it haha

[plucky halo]

Same! Got an email to say it's out for delivery!

[prisma flare]

sadly the cyber scan is only for github repos 🙁

[fleet geyser]

What's Pi?

[magic moat]

Raspberry Pi

I think lol

[cyan wing]

pi-mono is a coding agent harness / TUI

[magic moat]

Ah, I see. I use VS code with Codex and Claude Code.

Claude Code usage limits is rage worthy

[severe mason]

yo guys, .121 says there should be a view/reset and so on memory commands, but those are not / commands right? how do you use them?

[plucky halo]

It’s here!

[severe mason]

Id like to know more about your usecase

[kind jay]

What from where?

[plucky halo]

Got my hat, stickers, and clicker thingy

[hard drum]

nOOOOOOO