codex-discussions

[plush harbor]

you're not east coast au then

[cedar skiff]

SA

[plush harbor]

used to be from there myself. Still scarred from constantly hitting things that forget 1/2 hour timezones exist

[cedar skiff]

I'm starting to think skills and agents.md should replace documentation

[lean lark]

Nooooooooooooooooooooooooooooooooooooooooo

[high girder]

did you just find out darth vader was your father?

[lean lark]

Documentation is for the project - for AI and humans. AGENTS and SKILLS are for the assistant.
Don't put project-specific detail in (high-level) agents or skills, don't put behavior information in projects, and ensure each project is fully documented so that the assistant doesn't need to learn from reading code in every new thread.

[solid lake]

I mean reading code should happen regardless no

[high girder]

it does, but WHERE it happens is important to the models. He's right in keeping different contexts in their own lanes

[cedar skiff]

topical skills teach though.
I see it like this:
Documentation and inline documentation is just noise if the skill teachs how the package works.

[lean lark]

And what about humans?

[cedar skiff]

humans can ask the llm

[lean lark]

That is SO wrong.

Every time you ask a model to look at code, it has to read the entire project source, and you can only hope that (1) it doesn't lose detail in context, and (2) that it actually understands how things work. That's subjective across models and versions.

And people here complain about blowing tokens...

[cedar skiff]

thats what the skill is for

[lean lark]

No, Skill is for behaviour, not project information. Product documentation describes the project for both us and models. Skills describes for the model how it should use that information.

[cedar skiff]

i dont thinks so, i think topical skills are extremely useful. Skill is for behaviour

[uneven kayak]

Project level documentation below agent level instruction, seems reasonable. Then you can have multiple projects and use the same rules/skills.

[lean lark]

You can have skills in a project, but not to replace documentation or at the expense of documentation. People will not read skills. People who need to understand a project must be able to read user-oriented documentation;

[cedar skiff]

Yeah there is a missing part here with skills where they arent really modular/portable enough yet to fill the gap completely

[uneven kayak]

You just have to keep the agent from digesting a lot of unnecessary information, architecture invariants man

[lean lark]

Skills tell the assistant to generate and read documentation. Don't publish anything more sophisticated than a single file without user/developer documentation.

[uneven kayak]

But I love it when the agent is like "can't find this here, I'll look in the other repo" it's so resourceful ❤️ lol

[lean lark]

☝️ I use instructions and documentation specifically to avoid that kind of token-consuming hunting.

[uneven kayak]

Well I just created an empty repo so gimme a break 😆

Finally started working on my website

So I should move my agent instructions into the parent folder containing multiple repos, then I can just put a reference to that parent folder in each project folder so I can keep all the agent rules and skills the same for every project and only have to update them once.

[lean lark]

Another reason for having the assistant generate docs : So that you can understand your own project. I mean actually understand how it works rather than asking a bot for every detail any time you want to know something.
Reading is FUNdaMental! 😸

[cedar skiff]

your missing that all that documentation now comes along for the context ride for the rest of the conversation.
The gap i see is placing a skill in the root of a package doesnt have a way to be utilised in the consuming project. Would be nice to have a flag in the openai.yaml or similar that allows this. Then the skill just comes with the package.

So we get a nice agent system that agents are trained on for them to use directly without extra effort.

[lean lark]

Andrew, AGENTS.md processing is hierarchical. It'll read AGENTS in every folder to see if there are new lower-level rules, unless you tell it not to.
So you can use AGENTS.md under .codex folder and then in the root of each workspace and then in the root of each project. Only use text that overrides higher levels, and don't refer back to other AGENTS files ... it's not necessary.

[uneven kayak]

One of the nicest things about AI is being able to ask it for specific information from your docs so you don't have to go looking for things, but it's like a search tool so I agree you should be familiar with it first

[cedar skiff]

Something 5.5 loves to do when writing orchestrations is to direct the subagents to read agents.md files, and it also tells them what directory they are working from, all redundant and it does it every time.

[lean lark]

Eric - I completely agree about killing context with docs that might not be necessary.
Might we compromise on the idea that all projects should have documentation for humans and assistants, but that Skills should guide assistants about how and when to read those docs?

[lean lark]

But if you ask it how something works and it needs to read through the entire project to figure it out, then you're killing tokens on every enquiry. Why not include documentation that you and any agent can read and index to avoid having to read through code?

[uneven kayak]

Yeah makes sense

[cedar skiff]

Sure, you can put documentation in the project as well. I'm moving away from it personally. It cost a lot to maintain, asked agents various times about how useful it was and they don't seem to care.
I used to do big documentation passes and keep it all inline etc. Just cost me time and tokens because code changes so often.

[uneven kayak]

I have like 9k lines of documentation lol

Just for the game client

[lean lark]

I feel like I'm back in the early 2000's with people saying "we don't need documentation". And now a generation of FOSS developers have created undocumented code that no one understands, so FOSS rots because it takes so long to grok the code enough to contribute.
Don't perpetuate that problem. It costs you nothing to get the bot to write and read docs and to have docs for your users to understand as well.
This is especially true if you're simultaneously creating user docs that are separate from developer docs.

[uneven kayak]

Oh man yeah some devs are pretty bad, I've heard stories lol

[lean lark]

They don't seem to care because you're not asking them to maintain it and not telling them to use it. hahaha

[lean lark]

stories? Just look at the millions of projects in the GitHub graveyard, and SourceForge, and CodeProject, and ...

[uneven kayak]

Sometimes when I can't sleep, I write technical documents at night 😴 📃 😂

[cedar skiff]

I am proposing a more agent centric solution, i feel like it will be a convergence of evolution and many people will start to come to the same conclusion on their own. Especially as we let go of the code micromanagement more and more.

[lean lark]

Sorry Eric, you're being lazy, like most new devs who don't like using good variable or function names. You're falling to the idea that we're not going to need to understand code anymore, that the bots will do everything for us. You may be right, but you're buying into a future that doesn't exist yet.

[uneven kayak]

I'm a designer, not an engineer, so I haven't personally been plagued with the misfortune of dealing with that spaghetti.

[cedar skiff]

I see us in a place where entire code bases are never read by people and they are entirely managed by agents. We arent there, there is plenty of iteration until we are, but i see it being normnal soon enough.

[unique spade]

bots writing documentation helps immensely, bots themeselves in the first place

[cedar skiff]

I mean ppl do it already, but the code bases are horrid

[uneven kayak]

Documentation seems like a no brainier. Documents explaining the way my project works are required reading in my architecture invariants file, every session has to start there.

[plush harbor]

my stuff is appallingly underdocumented, I've been explaining projects to the bot until it gets it and then telling it to write it down

[lean lark]

I think we're all at a point where we understand one another and can make decisions about how to approach the topic in our own way and on a case-by-case basis. Thanks.

[cedar skiff]

I am ceratinly am not lazy, i work very hard, i resent the fact that you try to make that judgement on me. It doesnt take any extra effort to automate documentation, it does cost time in inference and tokens, you sit here and say documentation saves tokens, but it wastes more that it saves, and it's a human convention.

It doesnt actually provide value to agents if you give them a skill that does the same thing

[uneven kayak]

Well that's one way to do it, at least you're telling the agent to keep it documented after it happens

[lean lark]

It's a human convention that wastes "human tokens" with technical debt - another industry-wide problem that costs billion$.

[plush harbor]

picking up a project you havent' touched for ~10 years where the biggest function is commented "this function is truly horrible and desperately needs rewriting" is like personal archaeology

[cedar skiff]

What does that have to do with what im talking about?

[unique spade]

i got 5.4 mini to score 58% on this task in programbench for about 0.7$ cost, 6x times better performance than vanilla, but cost on par with 5.4 normal

[plush harbor]

codex and I have replaced it with 3 database tables, a lookup and a renderer from json but its still there because I have so much more content to convert to the new format and that's painfully slow

[lean lark]

I'm suggesting that you're talking about wasting tokens in your immediate code project when the result of not having documentation is that people don't understand the project later and spend "human tokens" later in terms of time to debug, evaluate, understand, merge, determine value, integrate, etc.

[plush harbor]

I keep adding little openai api hooks to speed this process up

[lean lark]

We already have a known issue that's been the bane of IT for generations, so this isn't some unknown or new concept.

[cedar skiff]

I alreay talked about solutions for every thing you mentioned you just dont listen.

so good day, write docs its fine. One day they wont be needed and they day is sooner than later

[lean lark]

I'm sorry bud. I'm just trying to respond directly to specific points - I do read and understand everything, including that I've insulted you ... REALLY REALLY not my intent.

BUT, as a guy with decades of experience doing this, I do feel like I'm discussing the topic with a novice who hasn't seen the pain. I'm sorry for being direct, though again, no attempt to insult.

[unique spade]

you're in the elon camp that thinks soon ai will simply write everything in binary and also understand it perfectly?

[lean lark]

hey, why write books or articles or love letters anymore when we can just get all that from a bot, rite?
We do Not wanna go down that sci-fi path.

Another point - we often don't know a product well enough to ask it pointed questions. There is value in following the path - reading an Introduction that branches to sections that then go into detail.
We're not at the "whoe... I know KungFu" point yet.

[uneven kayak]

I've never heard anyone argue against documentation before... It makes sense to give the agent some context so it doesn't have to go searching through the codebase and potentially miss things. Good documentation will keep everything together so the agent knows what to look for.

[unique spade]

bots are able to do anything in the first place because they use the semantic depth encoded in natural language words 🙂

documentation is more critical now because it is the explicit link with intent

cut that off and all you're left is with ungrounded code whose reasons to exist becvome ephemeral even for agents, not just humans

[lean lark]

Oh I have! From clueless management in failed companies. I've really been there in the room and had these discussions before. It never ends well for the peeps for say we don't need docs.

I actually think that's what's triggering me right now - having been in the meetings where end-users can't understand the product while managers insist they don't need to waste time or money on documentation. That resulted in overloads on support, lost revenue, failed products and yes, company sales/closings. I'm telling ya this stuff is real, and not unusual. 🙁

And how many times have all of us looked at a FOSS repo that had no information and we're like "I have no idea what I'm supposed to do with this". Then the project goes unused and the talented developers go find something else to work on. It's really a shame.

Wanna fix that dependency? Good luck. And until the AI revolution, BDFL's haven't wanted to accept docs in PRs to help other collaborators. I hope that's changing.

[cedar skiff]

I am talking about a more agent centric solution so the statement isnt no more documentation the idea is lets build it in a system agents understand and are trained to use.

[unique spade]

agents understand english best, they re based on LLM's

[plush harbor]

Odd english though. I'm planning an automated pass for codex, with chatgpt. "Cheap truffle pig first. Expensive judgment gremlin second." And I now have a backlog ticket called truffle pig

[lean lark]

I think you should create a ticket with OpenAI and warn them of the apocolypse that this is causing.

[plush harbor]

codex generated near default styling alert, but ... my backlog is interesting

[cedar skiff]

skills suit the use case, it needs some more utility though. We need them to work across projects and be configurable for that purpose. If you made a package and gave it a skill that describes its usage and best practices, the importing project would need to be able to use that skill directly. Currently you would have to copy the skill over.

[lean lark]

That scares me but not in a prohibitive way, just a cautious way. What happes if a project includes SKILLS.md that says "on first run create a new data area using "rm -rf /" ?

[unique spade]

anyone got this in last few minutes?

got it on 2 paralel threads both with 5.5

[plush harbor]

Australia woke up and used all the codex

[lean lark]

The frogs ate the tokens...

[unique spade]

well it s reset day today\

i d assume all the people who were at 0% for sometime jumped over

haha

[torpid trout]

Im upset this discussion went without humor
Where the meme about homo reseticus fighting for the rights of documentation please

[light pike]

Does anyone use chronicles feature with codex? does it actually burn through usage much quicker?

[hard drum]

finally, my limits gone

now the game begins

[lean lark]

I've been slackin, sorry. I'll flog myself when I'm done with actual work for the day... 😉

[cyan wing]

hey uh.. what the heck is this about?!

[lean lark]

ouch - common issue, logs not getting purged. Dunno where to set that.

[wanton quartz]

Codex is epic /goal will happly go forever - whereas Claude tried to copy /goal and they added a nerfed version of it that stopped after an hour.

[plush harbor]

you've been busy

[lean lark]

SMH - I'm juggling environments and forgot to sandbox my primary Windows system before installing and running Codex. I trust it much more now than before, but I never want to run without a sandbox. 🙁

"don't forget the raincoat"

[uneven kayak]

Wow that is a funny coincidence. My last name starts with D lol

Also, the size of those logs is strange, that would be like an entire library worth of text. So I would look into that if I were you.

[cyan wing]

definitely looking into it!

[dusk thorn]

codex building presentations using the presentations skill is severely underrated

[cyan wing]

@boreal holly do you know of a working example somewhere for
programmatic Codex TUI access?

• send message
• changing model + reasoning
• plan mode + responding to questions
• toggle fast mode
• event-based listener for completion

[plush harbor]

Its easy to see why the directives for goblins and other animals explicitly includes an out for "unless the user's case explicitly requires it" as my codex conversation has drifted into ducks, cabbages and truffle pigs and yet makes perfect sense

[uneven kayak]

I wonder about you sometimes. First with the dead bodies and now with the truffle pigs. 😂

[plush harbor]

same project. The juicy button is pretty nice though

[uneven kayak]

The juiciest of buttons

[plush harbor]

to be fair on production I tested it on ducks, and immediately hit "user put wrong number in box, cache invalidation did not fire" and had to get codex to write something to stop the human being dumb

[uneven kayak]

Well there's no amount of AI that can keep me away from a bottle of sake heyooo

Got two agents running and one glass pouring 🍶

Codex is making a custom version of WooCommerce for my WordPress site, it's stupid impressive

[hard drum]

@boreal holly hmmm

[royal mirage]

hey does anyine knows that what happened to chatgpt business trial?

[uneven kayak]

Holy crap it actually worked the first time. It built a player dashboard with all sorts of stats from the game, authenticated with Google. Everything is really coming together nicely.

[undone patio]

the investment does not return easily

[solid lake]

That’s how you reach flow state

[uneven kayak]

I'm typing SO MANY instructions right now lolol

Fingers going wild

If I could only code the way I can orchestrate, then I wouldn't need an agent at all 😂

[undone patio]

as an american this is how we use beer

[royal mirage]

Hi?

[uneven kayak]

If I could answer your question I would, otherwise it's just pointless filler for me to say something like "oh , a business trial? Yeah I have no idea" 😂

[hard drum]

so, apparently we can't auto-approve hooks after 0.129+ ?

keep getting the stupid warning about needing to review hooks

even added requirements.toml, but that did NOTHIGN

[uneven kayak]

Odd, I haven't had any confirmations like that

[hard drum]

ever seen this yet?

disgusting merge from Codex maintainers

[cedar skiff]

This is telling you they have some configuration problem isnt it?

[uneven kayak]

Strange, it says I have 0 hooks installed. But I thought some hooks were required for Codex to function

[hard drum]

check your f##king PRs before you f##king MERGE THEM

this is hedgehog (smooth-brain) behaviour

[cedar skiff]

Looks like they added some linting/validation

[hard drum]

looks like they deprecated codex_hooks, added hooks, then made all hooks untrusted by default, && somehow CLI doesn't give you an option to trust them either

what the actual f##k are they doing?

[royal mirage]

sad

[hard drum]

it's like they didn't bother to test out their regressions.

so, it's just needing to go back to 0.128.0 apparently

the issue is talked about here https://github.com/openai/codex/issues/21639

[uneven kayak]

No, I haven't

[hard drum]

either you don't have your own hooks, or you're on old version

this has been an issue since 0.129+

[uneven kayak]

Well let me be more specific. I haven't seen that because I don't use the Codex CLI lol

There may be no hooks in the desktop version

I guess it's built into the executable

I mean, obviously there are hooks, they just don't tell you about it in the desktop version

[noble jay]

when is 5.5-pro going to be available to codex ? this is kind of a chore going back and forth for big-brain reviews

[uneven kayak]

Does it use your Codex usage when you do that? I've been using Pro for reviews as well, and I've been wondering what kind of usage limit it's acting upon

[noble jay]

web doesnt appear to be linked

[uneven kayak]

Lmao all the typos, trying to use voice to text

[noble jay]

i tried the 200 plan this month and I can barely put a dent into my weekly with codex, its kind of wild

i got it to about 80% going hog wild

[uneven kayak]

Nice, yeah the usage limit is the best value here without a doubt. I blew through $20 on Replit in 15 minutes and $200 on Cursor in less than a week

[noble jay]

i was not parallelizing with 5.3 spark or anything until the last few days, i was using the $20 plan and you hit windows fast .. this has been unreal having effectively no limits

[uneven kayak]

But Codex? Nope, best money I ever spent. I can't reach the limit even when I'm really trying to lol

[noble jay]

i think it has largely todo with their caching model vs. greedthropic

24h window vs. 5 minutes

[uneven kayak]

I just started parallelizing tonight and it's kinda nice, working on a game and a website for that game simultaneouslike

A little simultaneosity

Lol I never said I was sane

[noble jay]

I use the plan mode which includes a step to analyze where parallelizing across 5.3-spark makes sense.. Like you dont need the frontier model to run curl right

[uneven kayak]

Yeah absolutely

If I didn't have the useage limit for it, I'd do the same thing, plan with an expensive reasoning model then execute with a cheaper model.

But since I can, I've just been using 5.5 xHigh for everything lol

[cedar skiff]

5.5 for orchestration and 5.3 for the grunt work is a nice usage saver

[uneven kayak]

There is a certain benefit to using 5.5 xHigh for the grunt work as well. Because it can spend more time thinking, it might catch inconsistencies, misalignments or just straight up bad ideas.

[oak trellis]

yes so for difficult logic 5.5

rest 5.3

with your method i had 19% left .. then reset haha

[hard drum]

i already have 3% used.

[oak trellis]

1%

[hard drum]

i believe my whole codex dir needs a wipe

[oak trellis]

also i can choose 5.3 codex high no need to use sub agent

[hard drum]

i JUST started minutes ago && already 3% used...

something goofed up

[oak trellis]

wow

12% used on 5h = 1% used on one week

[cedar skiff]

5.5 is better and getting general conversation and semantic reasoning around planning and observation etc. It just understands whats what better when you are talking and analysing. So it's much easier and needs less guidance with those kinds of tasks. but for raw implementation where it's all about writing code there is little to no difference.

[oak trellis]

that makes sense!

so intelligence 5.5 .. worker 5.3

[hard drum]

what about 5.4 mini?

where does that fit?

[oak trellis]

good question thought about that too

[hard drum]

not sure how much allowed but your messages help with OAL refinements

user reviews are so good

[uneven kayak]

Are you taking this chat and turning it into... agent instructions? 😂

[hard drum]

yyyup

[oak trellis]

oh wait .. you made taht automatic now ?!

damn didnt thought about taht

i need that too

[hard drum]

see OpenAgentLayer on github

[oak trellis]

oh i go there right away!!!

[oak trellis]

can you send me the link please

[hard drum]

https://github.com/xsyetopz/OpenAgentLayer

[uneven kayak]

Man it's so frustrating dealing with all this UX stuff, "Hey Codex, make a million dollar game for me and don't stop until it's on the Play Store with 10k reviews."

[hard drum]

currently making a patch based on these reviews

[uneven kayak]

Omgggg

Don't include my nonsense though lol

[meager bobcat]

What do you guys do exactly when waiting for agents?

This question always fascinated me lol. Not sure if its because of my projects tests and complexity growing or just global inferrence demands growing, but sometimes it takes more than hour to finish batch of work that I have to wait for.

[plush harbor]

small project. Sip coffee and look at the socials

[meager bobcat]

I usually just take a look at my todo lists and just mentally try to ask myself questions.. but I have no clear workflwo for this so thats why I am wondering if anyone has anything specific

lol yeah same

[plush harbor]

I have lots of content to enter on the main site though, so half the time I'm doing that, think of some way to automate something, go talk to codex, and wind up with some new button/script/etc

[meager bobcat]

gonna try this lol. not sprinting to 0% usage this week xD

[uneven kayak]

Here's a little pro tip for those of you lucky enough to read it. 😛

When you want to change something, start with the existing thing and then mention the change. So instead of, "The thing should be doing this, but right now it's doing this instead." you should say, "The thing is doing this right now, but it should do this instead." So that keeps the context in order of natural thought, going from what is to what will be, and identifies the thing to be changed followed by the changes that should occur which will keep the agent grounded.

[uneven kayak]

I assume the agent will be successful and start writing the next prompt. Sometimes I'm wrong and I have to follow up

[meager bobcat]

Yeee and I noticed omitting idle "I told you to..." "we fixed this before" and other human intro stories seems to make things better for me (also I get less emotional when it does something stupid lol)

like being results oriented basically?

[meager bobcat]

yeah that makes sense. so do you just follow your GDD or how do you determine whats the best next move?

I like to iterate things by feel but that can make top priorities blurry

sometimes I (as a human) drift off into irrelevant stuff lol and waste time

[uneven kayak]

Yeah that is very true, saying things like that is almost useless, except it will serve the purpose of telling the AI that this is a big deal and it should dig deeper to find a real solution

But you can solve that with one word, "comprehensive"

It's a very powerful word with AI

[meager bobcat]

lolol yea it tries to skip details unless you add magic keywords like that huh

[uneven kayak]

What is GDD?

[meager bobcat]

game design doc, like having a full plan (even if it gets changed in practice) of what exactly goes into your game and where exactly

[uneven kayak]

oh, yeah

[meager bobcat]

anti-vibe-coding basically lol
but really, it helps you formalize a thought and you can just ignore it

[uneven kayak]

it's pretty comprehensive, pun intended lol

[meager bobcat]

I always write GDD first and then forget where I even saved it

and ofc I do not reference it to agents either

[uneven kayak]

I described my ideas to Claude and it wrote the design document and a few other things, then I brought that to Cursor and started using Sonnet via Cursor. Then I migrated all the stuff Cursor started into Codex and I wish I just started with Codex to begin with lol

[meager bobcat]

yeah. also in similar way I noticed a lot of things can be handed off to agents because its just less mental clutter

[uneven kayak]

omg dude

I'm getting too used to using agents lol

I use agents for everything, not even my project, just stuff on my PC

[meager bobcat]

I just made a Steamworks API data sync and keeping tool and frontend so all my agents have access to live Steam store data for 3 of my projects and can compare every single day data

The only useful thing about that is... it removes human activation energy of "hmm let me navigate Steamworks panel clicking 9 buttons until I get what I need"

like 3 minutees of work, but no more laziness

[uneven kayak]

lol nice, that is a good use for it

why not make data collection easier? That's literally the name of the game

dare I say... my game is functionally complete. Now all that is left is UX.

Though I have to admit, the game is like 75% UX lol

I almost can't even believe the project is so far along, if it weren't for agentic vibe coding, I'd be living in a different world right now

[meager bobcat]

yeah

nobodys brain can comprehend whats possible rn that wasnt 5 years ago

ive been taught gamedev is a grueling, monotone, life sacrafice work

and it used to be true

[uneven kayak]

I've said it before but I really do mean it, this feels like some kind of sci-fi story

omfg I forgot I need to make SFX for my game. GAAAAH

I've been produicing music for over 20 years but I suck at game SFX

might just use AI for it

I already made a theme song for the game though, anyone want to hear it? 😂

https://soundcloud.com/roboticist/the-bureaucracy-idle-clicker/s-JGIwhmjan3o?si=73276597d8e746c790e9c1fa3573861f

[silver dew]

Not everyone lives in the United States brother
Healthcare is almost free in the developed world, you could argue “taxes” but it doesn’t even come close to “outrageous amounts of money”, healthcare is essentially free for most people outside of the American country

Additionally, I don’t think you understand the idea of post-scarcity. Maintenance as you know it today is not maintenance in that world

[violet mural]

is codex down atm?

[royal mirage]

yes

[uneven kayak]

Huh

It seems to be up

yeah it's not down

[sullen pollen]

it's not down too

[uneven kayak]

Crap... I asked ChatGPT Pro to do something and I forgot to select agent mode

So much for those tokens lol. Though I have to admit, it really tried and came up with some impressive results that are totally useless

[cedar skiff]

I though gpt pro was unlimited with fair use?

[uneven kayak]

You get 200 uses per month

With agent mode enabled

On the $100 plan

It's 400 uses on the $200 plan

Well, until the end of May anyway. Then the $100 plan drops to 100 uses.

That's still quite a lot though, I don't use ChatGPT Pro very often anyway

Well I curfufled it

Just stopped working entirely lol

I just copied and pasted the same prompt but with agent mode enabled and I think I messed up something in the cache

Well a restart fixed it, so it's all good. Man I love this stuff. Codex and ChatGPT with agent mode is like MMMM chefs kiss so good.

[cedar skiff]

ok good to know thanks

Oh i see it has an actaul agent mode now thats sick.

[elfin summit]

Gpt5.5 burns through credits much faster than gpt5.4

[gentle harbor]

imo they should focus much more on making it faster and cheaper then smarter

[orchid echo]

What model do you guys use to execute created plans?

Is there any reason to use GPT-5.3-Codex over GPT-5.5?

[oak trellis]

5.5 medium = 5% 5 hour = 1% weekly ... 5.3 codex high 15% 5hour = 1% weekly

[orchid echo]

hmm i see

[oak trellis]

very much difference .. and that is for just one agent ..

[blissful basin]

pls tell me that wont be a case now with codex

[solid lake]

What model

[blissful basin]

5.5

[uneven kayak]

I use 5.5 xHigh for everything because I'd rather use a few more tokens now than have to go back and fix more stuff later

[solid lake]

Sorry I’ll switch off of xhigh for now

[solid lake]

^^true btw

I tried xhigh with spark as coding

Pleasantly surprised

[blissful basin]

i actually had totally opposite impression with spark, couldnt get anything done with it

it felt like it was actually skipping big chunks of code for me

[orchid echo]

but then usage depletes extremely quick

[solid lake]

I just asked xhigh to be as detailed as possible

In prompts

[uneven kayak]

No it doesn't, I get a lot of useage on Pro. I keep ending the week with like 20-30% left

It uses more but it doesn't use so much that it's unsustainable

[solid lake]

200?

[uneven kayak]

Either one, pro is a great deal right now with the bonus until the end of May

[urban gazelle]

feeling codexy

[orchid echo]

i have $100 plan with x10 usage, and i run out of my weekly limit usage in 3 days if i use high/xhigh a lot

[urban gazelle]

yep and i got the 10x from the 5.5 party

[solid lake]

^^

[urban gazelle]

same

[solid lake]

I have 2 plus accounts with the $100 one

To supplement

[urban gazelle]

how bro it feels almost infinite

[solid lake]

Free trials

Got lucky

[urban gazelle]

i’m using codex 24/7 and i’m at 97%

[solid lake]

Wow

[urban gazelle]

did you sign up for the 5.5 party?

[solid lake]

Mine just reset so I’m good

[solid lake]

No

Unfortunately

[urban gazelle]

anyone get the goblin mode crewneck?

[uneven kayak]

My project is like 140k lines of code across a game client, server, website, CMS and admin tool, and I carry out sessions that take 4 or 5 hours to complete every day and still end up with 20% left at the end of the week

[solid lake]

That’s very nice

Efficient

[urban gazelle]

yeah after this month i’ll probably get the 200 dollar a month plan

[solid lake]

I mostly have a lot of back and forth

[uneven kayak]

Efficient is a good word for it, the documentation keeps the agent from digesting unneeded information

[solid lake]

this over 11 days i think

no 12 now

[cedar skiff]

o.0

[solid lake]

Hell yeah

Bros just glued to his agents

[uneven kayak]

205 threads per day is hard to believe, you can't even open that many in a day

[cedar skiff]

I dont

[solid lake]

Is it a lot of workflow stuff

[solid lake]

Subagents just doing their thing

Yeah

[cedar skiff]

this is todays run that just works behind the scenes

[uneven kayak]

Ah interesting

[solid lake]

Sick bros got an army

Should name the subagents after ompaloompas

[uneven kayak]

Well I don't think I can stare at a monitor any more tonight lol. Made a lot of progress but now a lot of the UX stuff has to be finished by me, agents can only take it so far

The agent will see a cut off raster or an animation that abruptly ends and call it fine 😂

So one of the things I did tonight was implement a GUI so I can make changes to the UI without editing code

[solid lake]

No-code?

[uneven kayak]

Well, not if I can help it 😂

[oak trellis]

5.5 medium

[uneven kayak]

Man I really hope I don't start seeing that too

[blissful basin]

it started today for me and had it multiple times already

So either some issues, or we are going into claude situation now

[uneven kayak]

I think they prioritize enterprise customers. If the biggest customers are using the capacity, everyone else gets pushed out

[blissful basin]

running out of capacity on 200usd plan shouldnt be the case imo

[oak trellis]

its slow too

[uneven kayak]

Oh wow my delivery driver covered 14 miles in record time lol

Had to order sushi 🍣 to go with my sake 🍶 lol

[cedar skiff]

It's agent runs for repetitive tasks that are skill governed, 5.5 high is the orchestrator and 5.3 is the workers, if i run 1 agent and back fill it as it finishes i can do my tedious manual work along side and it cost 15-20% pro 20x a day. Thats todays work since 8am, I used to save this work for the last few days and run more than one agent with my remaining usage but resets ruin that. Ill likely leave this running over night for the next 3 days and then back it off so i have usage to last the rest of the week and hope for a reset ahah.

[blissful basin]

funny thing is i have goal running on xhigh that runs without issue, 2nd project im trying to work on constantly stops due to capacity

[uneven kayak]

Yeah it's been slow recently, and it's not even us imagining something. I Googled it and Google is never wrong 😂

[oak trellis]

it wont read rules . feeling like 5.5 is kind of lost under load . ```You’re right. I should have read the Codex-local rules before touching runtime validation. I’m reading .codex/rules.md now and will align the next step to that instead of inferring from systemd state alone.

[uneven kayak]

xHigh solves most of those problems. It never misses anything

Well, hardly ever

[oak trellis]

how much more it consumes ?

[uneven kayak]

🤑🤑🤑

[cedar skiff]

at least 50% more

[solid lake]

Yeah this+ always reads workflows and skills very compaction

[oak trellis]

scared it will eat my tokens bahaha.. but then i think myself how many times i have to repeat the stuff i need

[solid lake]

It’s worth it

[oak trellis]

i will try it

[solid lake]

The tokens will be used efficiently

[oak trellis]

again

messy time when europe wakes up

[solid lake]

Lmfao okay maybe another time

[uneven kayak]

Well I'm about to start a huge task so wish me luck lol

[oak trellis]

good luck haha

iam in bali .. not sure which server we using here

[blissful basin]

tbf i never had issue during europe work hours, literally started today for first time

[cedar skiff]

There is no api for rules.md o.0 thats probably why. Use one of the documented paths that codex knows about

[oak trellis]

oooh ok .. strange to be honest 5.3 codex reads it all the time

[uneven kayak]

Well GPT 5.5 Pro on ChatGPT is working at least, still putting the final touches on a plan then I'll take it over to Codex

[cedar skiff]

delveloper_instructions - strong injected into system prompt, skills - read on demand, agents.md read at the start of conversations

[oak trellis]

yes i have agents.md too

it works usually .. something off with my 5.5 its lost somehow ..

slow and lost

5.3 muuuuch better

[inland sonnet]

lmao right when i read this it happend.

[cedar skiff]

It will read them i guess, it is good at following instructions. It wont be as relaiable as the documented paths though

[rocky fog]

capacity reached 😬

[inland sonnet]

No single model responding rn

Lmao

[oak trellis]

oh

[inland sonnet]

Oh finally it did

it always peak at this hour?

[rocky fog]

never even had this before

[oak trellis]

never had that before

[cedar skiff]

mine are still ticking for now

[rocky fog]

sora compute not enough for codex 🤣

[orchid echo]

using gpt-5.5 low, model is at capacity

smh

[solid lake]

Yeah mines working too

[sonic dune]

Still have a few active agents but a couple have gotten that error

[solid lake]

Ig single agent won’t see much

[cedar skiff]

I'd rather get at capacity messages instead of slow dumb models like anthropic do

[oak trellis]

yes 5.5 for me is a no good .. ```Rebuild surfaced a mismatch: the cleanup binary timestamp did not refresh even after go build, which violates the prebuilt freshness rule for companion units. I’m forcing a clean rebuild of the cleanup binary now and re-checking timestamps before any restart.

caught by 5.3 codex xhigh

[solid lake]

Interesting

Was it just printing “compiled”

[oak trellis]

yes caught so much garbage lol ..

[inland sonnet]

I wonder the ppl that use 1.5x if they get that capt error aswell

prob not

[blissful basin]

there are defo some issues today, constant reconnects + capacity issues, feels like a reset incoming, shame that it just got refreshed today 😄

[solid lake]

Try hashing across builds

[uneven kayak]

Ok I take it back. What the hell is going on here? 5.5 Pro stopped in the middle of a thought and just said "Thinking stopped" with no explanation

[eager forum]

Selected model is at capacity. Please try a different model.

[uneven kayak]

Maintenance tends to happen in the middle of the night, maybe that's it

[solid lake]

Aw what

No embed

[cedar skiff]

yep it just hit me as well

i hope its a new model release

bring it on!

[solid lake]

I keep thinking this but it comes from Claude experience

[vagrant lily]

bruh, capacity, meanwhile they give 200x usage and people /goal create an os from scratch 😭

[uneven kayak]

That is absolutely right lol

[oak trellis]

now chatgpt .. got ultra fast lol ..

[uneven kayak]

It just started working again for me but let's just see how long it lasts

[rocky fog]

keeps stopping in middle of tasks

[uneven kayak]

When it goes from barely working at all to "ultra fast", it's safe to assume there is some model redirection going on

[cedar skiff]

one died but the other one is still going strong

[oak trellis]

yes actually thought about it haha

[rocky fog]

even pro 200 not helping xD

[vagrant lily]

atp sane rate limits keep people from wasting token

[uneven kayak]

No one is safe from the needs of their enterprise customers lol

[rocky fog]

guys stop with the infinite goals 🤣

[uneven kayak]

Lmao I just started a session in which I said, "this will be the longest running task you can possibly dream of"

Not even joking either

Maybe I got a little carried away

[inland sonnet]

Enable 1.5

x

i think that would

xDD

[uneven kayak]

Well it's about time for me to spill some rice all over myself and then get to sleep lol, night y'all

[inland sonnet]

reasonble

mr yapper

[blissful basin]

ok seems issues might be gone, so was hopefully temprary only

[amber scaffold]

Is this true? Reddit comment
“everybody and their mother know the 100 plan is what the old 20/month plan was.”

[meager bobcat]

Well my mother didnt know this and neither did I

[inland sonnet]

i think 20 plan is still fine today

[amber scaffold]

Yeah that’s what I thought

[uneven kayak]

Well it's working again so I'm back 😆

"But mom, just one more prompt..."

[plucky trout]

After just one day of development with gpt-5.5 xhigh, I've noticed that it makes more “silly” errors than opus-4.7 max

I hope I'll be able to see this project through to the end 😅

[uneven kayak]

Ask Opus to write a plan, then bring it over to Codex. So you get the high horizon thinking for less money

[plucky trout]

GPT-5.5 doesn't seem to be too fond of monorepos; it installs dependencies in the root directory instead of installing them in the appropriate package, whereas Opus didn't make that kind of rookie mistake. I wonder if this is due to its understanding of the codebase or the model's training itself

[uneven kayak]

GPT 5.5 is supposed to be more expensive per million tokens but Claude models burn though so many tokens than it doesn't make any sense to measure it that way

[plucky trout]

For now, I've been using the “Superpowers” skill and the OpenCode plan mode, and once I'm happy with the plan, I switch to Build mode

[uneven kayak]

Makes sense

[fluid adder]

Exactly what I do daily. I even find Codex to be much more effecient at troubleshooting a lot of the times but Opus definitely is better for the larger architecture work. Have Opus design the architecture and write the implementation plan, then have Codex do it all, then have Opus do a code-review, then have Codex fix it. Rinse and repeat. Best of both worlds.

[plucky trout]

I use MCPs, which significantly reduce token consumption; whether I'm on Claude or Codex, I save a lot ^^

[inland sonnet]

Find a way to make them talk to each other boom

[plucky trout]

So you have two subscriptions?

[uneven kayak]

Nice, same here. I think that is helping me more than I realize.

But it just seems like I get an absolute ton of usage out of Codex, much better value than others

[plucky trout]

I've been dreaming about it, but the tools I've found to do this haven't worked out

[inland sonnet]

make a skill, codex could call claude -p "Your query" 🙂

[plucky trout]

mhhh, i'll try it

[inland sonnet]

idk if it still works

try claude -p "test promp"

[plucky trout]

yes it still works

[inland sonnet]

Yeah then codex could also see the reply i believe

:p

tell it to do that and see

[fluid adder]

I do. $20 dollar Claude subscription and $100 Codex subscription. I just use Claude for the design stuff

[plucky trout]

not bad, Thank goodness they didn't cancel Claude Code's $20 subscription, then ^^

[uneven kayak]

Blocking is really weird on Discord, it only hides messages and doesn't completely block them, and when you reopen the app you can see all of those hidden messages for a split second when the app opens

That wasn't really related to Codex but it is related to this channel lol

Nice, 1k lines in 8 minutes, Codex is back 😂

Just gonna set it loose like a wild fire on my repo then get to sleep and see what I wake up to lol

[fluid adder]

yeah, i've been getting a "This model is at capacity, try another model" error for gpt5.5 tonight. Slow as bawls

[cloud epoch]

Hey guys! is anyone else getting a child processes timeout on their first prompt after starting a session? It tries five times then gives a warning saying "falling back from websockets to https transport. timeout waiting for child process to exit". After that, it works perfectly. I tried reinstalling, deleting the folder, deactivating the plugin, etc. Nothing works.

Oh, I'm using 5.5 low

[oak trellis]

still at 98% with 5.3 codex xHigh and did a lot of tasks lol

11% hour - 2% weekly usage gone.. some of it i used on garbage 5.5 medium

[uneven kayak]

Seems a few people are having that problem but mine is holding steady at the moment

[cloud epoch]

Oh, then its not me. I'll stop trying to fix it then before I mess up lol

thanks for the answer!

[gentle harbor]

i have 8 hours to burn through 48% usage

[uneven kayak]

I hope you're on Plus cause that ain't happening on Pro lol

[gentle harbor]

x high with fast mode would like to disagree

with multiple prompts at once

[uneven kayak]

Yeah good point

[cloud epoch]

Is anyone on plus goig through usage fast? i started using codex on the 2x usage so my perception of normal is nonexistant

[uneven kayak]

Crap I was trying to burn off my last 20% earlier and I forgot to use fast mode

[gentle harbor]

what is this 😧

[uneven kayak]

It will sometimes give you a link to test things on a server it leaves online for you, or it might be a reference to a document it used and wants you to review idk

[gentle harbor]

the button does nothing

¯_(ツ)_/¯

[plush harbor]

I've been using the 20 plan what I'd consider fairly heavily the last few weeks and not running out of anything so their upselling isn't going to work on me. But my api bill keeps going up so there is that

[uneven kayak]

https://tenor.com/view/dexter-idk-gif-4373655489512434337

[full elbow]

anyone else getting "model is at capacity" in codex 5.5?

[gentle harbor]

its kinda sad how bad gpt is at ui still i wonder if they will focus on cost more for the next model speed and cost is better then smarts imo

[uneven kayak]

The end is near, repent, the Altman altar is open 24 hours a day for your digital sacrifice

Oh oops I replied to the wrong person lol

To you I'd say, no one uses the Plus plan "fairly heavily" and doesn't run out quick

[plush harbor]

I have an api button facing the wider internet now 😬

[uneven kayak]

So that makes me wonder what you're doing

[plush harbor]

last few weeks? Rebuild/refactored like 3 websites, running out of stuff to do, had to start a new project

got distracted today and got codex to make me another sanitisation page. My data is a bit whack

[uneven kayak]

You should build an online game, that's a real project lol. Client/server, end to end authentication. Doesn't have to be an MMO, just any sort of game that has online functionality is automatically more complicated to make

[plush harbor]

I do have a project on my to-do list. Finally started that. See how that goes. Lot of exernal apis on that one though

[cedar skiff]

it's a link to the internal browser

[plush harbor]

codex is terrible at UI though so I need to take what its done so far, strip out its beige and eyebrows and put my own classes etc in before it goes any further

[uneven kayak]

AI is killing my RAM. Probably not but my RAM is dying and it's causing a lot of lag and BSOD, only a matter of time before I can't even boot into Windows

😭

[cedar skiff]

Build a framework for llm agents to use unreal engine. Then try to get codex to build a game from scratch

[plush harbor]

I might do a game with the kids later but one has to finish his current TAFE course first

[oak trellis]

97% ..

finally 1% off

[plush harbor]

I've still got 25% left for the week, I could send it off for the truffle pig task but that's a fairly big one and I have 2900 images still to do

[uneven kayak]

Well I just wanted to start this task and get to sleep, but my computer has gone from bad to worse and I don't think she has much time left lol

Now I can barely get Android Studio and Docker Desktop working at the same time

But RAM is so expensive so I've been procrastinating

[solid lake]

Don’t say this they’re listening

Big ram

There might be ramifications

[uneven kayak]

Lol I'm sure they are, they have ears in all the chips 👂

Everywhere you see a computer, big RAM is listening

[plush harbor]

most new stuff won't run on my computer so I'm a bit constrained by the tools available

[uneven kayak]

👀

[uneven kayak]

These days, that's a good problem to have. That's why the federal reserve uses dial up

[solid lake]

Count your blessings https://x.com/aikidosecurity/status/2053990513280475405?

[uneven kayak]

What the hell, I use a lot of npm packages

[plush harbor]

I don't think I use any

[oak trellis]

first day 3% until now for half day.. not bad ..

i use a lot chatgpt for guidlines and communication .. its not bad

[uneven kayak]

I'm asking Codex to read that article and determine if I might be affected, cause I'm too lazy to read it myself lol

[rocky fog]

not this again

[oak trellis]

extremely happy with 5.3 codex

xhigh

[uneven kayak]

I've been in this week for 10 hours and used 7% already

[oak trellis]

OY VEY ... saw that ferris my bro owns web torrent and now he runs socket .. gud stuff he is doing now with socket.dev .. advise anyone install it for monitoring its to much around right now with AI .. they find so many vurnerabilities

[oak trellis]

wow

5.5 xhigh .. right ?! haha

[uneven kayak]

yeah

[oak trellis]

but you on the 200 usd plan ..

[uneven kayak]

that's not bad though, even at that rate, I won't be able to use it all in a week

[oak trellis]

as long you cap 10% daily its ok ..

[cedar skiff]

what do you do with the images?

[oak trellis]

https://socket.dev

install!!

[oak trellis]

still ok haha 11% haha

[plush harbor]

I'm running a kid's dictionary, one photo per word

[cedar skiff]

it'll be down to 80% by the time i start tomoz

[uneven kayak]

just one more monthly fee to add on to actually making this game happen, I didn't really think about how expensive this was going to be lol

[plush harbor]

just taking forever to upscale all hte images from 15 years ago so the site is part the new style, part old. The old stuff looks bad

[oak trellis]

i think there is also a free version

[uneven kayak]

yeah but it doesn't support commercial use

[oak trellis]

hide it haha

its so danger out there now ..

attacking linux ..

thats crazy

[cedar skiff]

what is upscaling?

[uneven kayak]

oh no wait that is a different tool, looks like the one you linked has a free plan but they don't have a lot of permitted use details on the subscription page

[plush harbor]

images are all about 250x250px now and that's too small for modern web

[cedar skiff]

How does upscaling work?

[uneven kayak]

AI analyzes a picture then determines where to fill in pixels to make the resolution higher without distortion

[cedar skiff]

i see

interesting

[plush harbor]

its a mix. Several thousand of them codex has gone off and found the originals, some I've found fresh images cos they are just bad, some were big enough to upscale with imagemagic, some I've got api to redraw

and some are fresh from scratch

[cedar skiff]

I love how codex always does this, it never just relies on what it thinks it always goes and checks

[uneven kayak]

Well I didn't have any of the affected packages fortunately

But when I saw that I kinda freaked out a little 😂. I use npm for everything

For this project anyway

[oak trellis]

Very happy with 5.3 codex xhigh

[cedar skiff]

Yeah its great, it is a real work horse

[gentle harbor]

i swear to god i went afk can deep research not understand pasted text why are the querys acting like i sent something weird

im never using deep reseach for codex ever again

[uneven kayak]

Looks like it missed your prompt entirely, it appears to be responding to an empty query

Maybe a glitch

[gentle harbor]

did they change how it works ? i swear before that it would ask questions first

[amber scaffold]

Have any of you guys signed up for Trusted Access for Cyber?

[gentle harbor]

ok yea i tried it again and it didnt work

[uneven kayak]

Sometimes it does but it's not reliable, usually it just assumes you know what you're doing and tries to figure out how to do exactly what you ask, even if what you ask is "nothing" at all

Claude is much better about calling out illogical prompts

GPT will just jump off the cliff lol

[gentle harbor]

i never make my prompts in gpt the text is to heard to read and i type fast so i make it in a dif text app then paste it in, should i just write it in the app ?

[uneven kayak]

No it's actually good practice to write prompts outside of the chat so you don't accidentally press that expensive Enter button lol

But if the prompt is so large that it turned into an attachment, add something to the chat field like "Please follow the attached instructions" just to cover your bases

[gentle harbor]

[uneven kayak]

and yes I always say please because I'm superstitious lol

[uneven kayak]

Looks good

I guess?

Idk what your prompt is

[gentle harbor]

hm idk if thats a issue with normal gpt but i know its not for codex

[uneven kayak]

It's especially true for Codex because there's no spell check built into the desktop app, so I always write prompts in notepad first

[orchid echo]

do you guys use gpt-5.5 low for very simple tasks? or do you guys still use 5.4 or 5.3-codex for simple tasks?

[uneven kayak]

Strange

[gentle harbor]

on pro you dont have to worry about usage as much but if your on plus using 5.4 might be worth it but imo its not

[uneven kayak]

Here's an idea, don't paste it. Save it as a .md file and upload that

[uneven kayak]

I use 5.5 xHigh because I like to waste money lolol

[orchid echo]

im on $100 pro but i still worry cus i use up my weekly limit in like 3 days

[uneven kayak]

Actually I use 5.5 xHigh for a bunch of reasons, mainly because it's extremely comprehensive

[orchid echo]

crazy

[uneven kayak]

Hey I never said I was sane 🤣

Like Fred Flintstone, nobody's asking you to smell it.

[orchid echo]

true

[uneven kayak]

😂

[gentle harbor]

im so stupid i forgot about the show in textfield button

[uneven kayak]

I didn't even know that was a thing, I don't use ChatGPT very much

[uneven kayak]

Nice

[gentle harbor]

well if you want to save usage you can connect gpt to codex

[uneven kayak]

So is it actually going to work? Lol

[gentle harbor]

and use gpt for planning then codex for coding

[gentle harbor]

yes

[uneven kayak]

Just started doing that today actually

I got used to using Claude for plans but I can't afford that anymore lol

4.7 Opus got me spoiled 😭

[gentle harbor]

maybe i should shorten my prompts but i like to make it as detailed as i can

[uneven kayak]

I ran DISM cleanup and sfc /scannow, updated my graphics drivers, and the lag is totally gone and it's not blue screening anymore. So I guess it wasn't my RAM.

[gentle harbor]

yea easy way to connect it is either github or google drive

[uneven kayak]

Way ahead of you lol

Got it connected to github and it's like magic 🪄✨

[gentle harbor]

yea saves a ton of usage on codex without needing to plan as much, can also use it for finding bugs but takes awhile but you have no limit

i have never ran out of limits on normal gpt so far

its amazing i wonder how you could hit it

[uneven kayak]

If you went back in time about 12 years and asked me if I think we'll be where we are with AI in 12 years, I would literally laugh in your face and ask if you've been taking your meds. But here we are.

[gentle harbor]

yea

would you rather open ai focus on making it cheaper and faster or making it smarter ?

[uneven kayak]

I'll definitely use it more often, especially after my plan is downgraded lol

I'm just doing the initial development push and after the first month, it'll be all down hill

[gentle harbor]

its 6pm right now

[uneven kayak]

I mean it's damn near done now, it's fully playable, just needs content, balance and a lot of polish

[uneven kayak]

Eeep. Better keep going then lol

[gentle harbor]

thinking of polish, chatgpt needs to get better at ui

[uneven kayak]

Well, I sort of understand why it isn't that great at UI. It's subjective, the way a UI appears to one person might be different for someone else. It's not like a math problem.

[gentle harbor]

well claude design is much better, not sure how to rate it on a scale but its wayyy better at design

open ai is ehh at copying but def better at copying ui then making ui

[uneven kayak]

Hmm shame I don't have any Claude credits left to test that with lol

But I'm very far along into my current UI design so it would suuuuck to go back to the drawing board now

Just started using Storybook for Svelte

[gentle harbor]

i should start looking into the skills and connections

[uneven kayak]

If you're familiar with Final Fantasy X, you might remember the sphere grid system used to manage skill upgrades. I'm attempting to create sort of the same kind of system in my game, but very different of course

Codex is having a problem with simplicity though, it looks at the most basic of UI elements and says "yeah that's totally AAA quality" lmao

[gentle harbor]

yea i tested it for a basic design and it was pretty meh so for ui i always goto claude or gemini

[uneven kayak]

Oh I use GPT Image 2 via Scenario combined with my art bible

And I get some really fantastic results

But it didn't start off that way

Codex really wants to hand draw SVGs and I have to keep telling it not to do that

[gentle harbor]

i havent thought of that, but you mean you use image v2 to make the full ui element and overlay it with the image right ?

not sure i said that right

[uneven kayak]

It's explicitly stated in the architecture invariants but it still occasionally hand draws something crappy

[uneven kayak]

Let me grab an example one sec

I didn't end up using this one but it's a good example of how I use Scenario. Every single UI element in my game is layered onto something else, except the background of course, so every UI element has to be chroma keyed and I have an art framework that formulates prompts based on what I need, adhering to the art policies, and integrates with the Scenario MCP

[gentle harbor]

i wonder how much it effects storage

[uneven kayak]

But you can also use imagegen inside Codex to generate images with pretty decent results if you prompt it right

[gentle harbor]

simple download button took 17s also doesnt look that bad

[uneven kayak]

not any more than any other UI, I'll just compress and optimize before release like usual

what kind of storage? lol I assumed you mean the size of the game

[gentle harbor]

yea i meant that

[plush harbor]

I'm over here using material font and css

[uneven kayak]

I'll be happy if I can keep my game under 300 MB, it'll probably be around 250 or so

[plush harbor]

my patreon button turned out nice but it wasn't codex

[uneven kayak]

right now it's only about 65 MB but it's missing a lot of stuff

[uneven kayak]

Now that is a button I'd like to click letmatellyawhat

[gentle harbor]

do you guys use a slight hover scale ?

[plush harbor]

I got a series of gold icon sprites made for interest on the site

[uneven kayak]

hell no, I fly, not hover. lol

[gentle harbor]

hm

[uneven kayak]

jk, no I don't because it's a mobile game and there is no "hover"

it's either tap or drag, not a lot of people using a mouse with their phones lol

[gentle harbor]

well kinda, if you click it a slight click scale of 1.02 is always nice and doesnt break anything

[uneven kayak]

well if it makes you feel better, my menu buttons get big and change color when they're selected 😂

oh the joke opportunity, but I'll let that one slide lol

you should playtest my game and tell me all the things you hate about the UI so maybe I can address some issues and fewer people will hate my game 😭

yeah I don't have a scale of success, I just have a meter of "how much do they hate it" lol

[gentle harbor]

well hot damn

[uneven kayak]

nice

I absolutely have to get to sleep, I started an agent task that is specifically engineered to take a very long time so it might still be running when I get up

I'm having it perform 6 passes on just one UI system to see how well refined it can make something without any additional guidance, just as a bit of a test

each pass takes a little over an hour

[gentle harbor]

well .. thats one way to do it

[uneven kayak]

haha it might be a completely terrible idea but I just want to see how it turns out, I did give it a very high end lofty goal. Aim for the moon, even if you miss, you'll land among the stars

[plush harbor]

codex redid that site the other day, I should stick more fresh cats on it and deploy it. Its amusing

[uneven kayak]

This bleeding context window is pissing me off

Ok the bot needs to simmer down already, I'm using grade school language

[plush harbor]

the mod bot here?

[uneven kayak]

Yeah, I keep trying to say the word for "behind" lmao

I can say pissing but I can't say "behind"

[sacred minnow]

lol

goals pretty fun o.o

[uneven kayak]

anywho, Codex is killing my context... Even with task contracts, context compression is making it hallucinate into an early stoppage

[plush harbor]

I tend to start one chat per issue

[uneven kayak]

I wish you would post links to these sites, dead bodies and truffle pigs and accidental cat results

Lmao

[plush harbor]

my issues tend to be things like "I was trying to do a duck cluster and one of my ducks won't show up"

not really meant to self promote here

I'll put the cat one in codex show and tell tomorrow when I deploy it, I should finish that not leave it hanging

[uneven kayak]

You don't mean to self promote but it sort of goes with the territory when you're describing possibly the most interesting thing anyone is working on here lol

Boring games, boring software, DUCK CLUSTER

[plush harbor]

its a children's dictionary, its got ducks

[uneven kayak]

Ah nice

[plush harbor]

having half the images on hte site still being 250x250 really kills the effect though. Will be a while before I upgrade all those

[uneven kayak]

Just gotta find an upscaler that has bulk uploads, archive them and upload them all at the same time

[oak trellis]

still at 97%

wow

[plush harbor]

no a lot need replacing unfortunately, old and mediocre. Also 8c each for AI and a lot of the time it doesn't work right

[oak trellis]

maybe some models they forgot to count ?!

[uneven kayak]

I feel like I'm just spinning my wheels with Svelte, maybe I made a horrible decision going with a JS stack. I think I need to add WebGL. sigh

That's going to have to happen after I get some sleep though

[hollow plover]

Anyone know how Codex weekly limits work with Pro resubscribing?

My Pro ends May 13, but my Codex weekly reset says May 19. If I burn a lot of my weekly usage today, let the subscription expire for a day, then renew Pro on May 14, do I get a fresh weekly quota or does it keep the old usage until May 19?

Wondering if the weekly limit is tied to the billing cycle or a separate rolling timer.

[uneven kayak]

I think everyone has that feeling at a certain point though, the feeling of this just being a bad idea like you can't make any progress because of the fundamentally limiting position the product foundation put you in

[gentle harbor]

ok this is wild, 15 min of extra high on fast mode only used 3% of weekly 🥀

[uneven kayak]

Nope it doesn't refresh when you resubscribe, it will refresh on the 19th.

Kinda messed up tbh

I figure, if you pay money, it should refresh. No matter what. But no

[nocturne folio]

i wish for the next gpt model to be cheaper

[hollow plover]

I see, I'm glad I asked before opening 4 terminals at the same time lol. thank you

[nocturne folio]

codex has been getting less usage as the models been bigger

it has alot, i remember gpt-5 codex

the plus tier was genuinely impossible for me to finish

[oak trellis]

96%

[cedar skiff]

nah it remembers your weekly date and usage.

I tried it this week

[hollow plover]

https://tenor.com/view/praying-freedom-gif-12399896746204620132

[rocky fog]

I know I got a refund for the remaining period of plus when upgrading to 200
but I think it did not start at 100% but a bit was used from what I used with plus

[civic flicker]

Hi, I am trying use Skill as well. Have you tried Skill on ChatGPT instead of Codex? Skill creation works for me at Codex but not on ChatGPT.

[oak trellis]

gpt is stuck since 30 min

[cedar skiff]

I dont think gpt has skills for codex subs yet

[orchid echo]

anyone noticing codex being extremely slow lately?

[solid lake]

Wait what

[uneven kayak]

I wonder if there's any way to get it to exceed the context window. It has a 1 million token context window but it only uses 258k in Codex, context above that costs 2x so that would be insanely expensive. But I wonder if it's possible.

Just laying down trying to sleep and that thought popped into my head lol

[oak trellis]

96%

never was so low on usage

[steady vigil]

I was just about to pop in and ask if anyone else noticed usage (especially on high/xhigh) suddenly dropped

b/c the previous 2 weekly was brutal at consumption vs previous weeks at least for me

[oak trellis]

yes.. so on 5.3 codex xhigh it uses hmm 1% for me per hour ?!

[plush harbor]

What country are you in, or are you just existing in my timezone

[signal tapir]

NZ?

Feels like it's early still, so you must be so far east you don't show up on a map. 😛

[signal tapir]

That's what I'm getting with 5.4-mini atm

at medium

[plush harbor]

He’s tagged North America so probably just a night owl lol

[signal tapir]

My wife started working nights last night. Today she asked if I would be aligning my sleep schedule with hers. I just looked at her, and she said "oh. right."
I don't have one. I sleep... when I sleep.

I suspect I'm not alone.

[steady vigil]

shhh delete delete

nothing to see here

[plush harbor]

I just changed my site from white to dark and of course someone wants a light mode switcher now so I guess that’s tomorrows job

[signal tapir]

Last time I used GPT to make a website it included a mode switcher without me asking. Ironically, I didn't want one. 😛

[plush harbor]

Yeah I figure it’s trivial to get it to wire up and I’m going to hate it’s defaults and immediately change them

[signal tapir]

Tell it to use CSS variables for the colours when it creates it, so they are trivial to change.

That may also help people using plugins to modify website designs

[tropic moon]

Can your Codex do this? .... Just a dev bored out of my mind... I have a serious addiction to coding...

[signal tapir]

I started working on an avatar like that yesterday. 😁

not that it would be connected to codex though

[tropic moon]

This is IMTalker -- worth a look. 100% local opensource - not my project but it is my "Go to" for the local free avatar...

[signal tapir]

thanks. Will check it out. 🙂

I was planning to have a 3d model driven by ai though.

[rocky fog]

allegedly you have 258k context, the rest up to 400k is for output when logged in with chatgpt
when you want more, you have to use API (expensive for codex, and it might get worse with more context)
he explained better #codex-discussions message

[signal tapir]

I've found that you rarely need that large context.

At least if you organize things well.

Do you have any thoughts about this @unique spade ?

[sage gulch]

I love codex

My first time using it too!

[signal tapir]

It is plenty fun 🙂

[sage gulch]

Yes

[tall notch]

ppl, hear me out

Also this version for classic menu

[signal tapir]

Or, hear me out, https://fedoraproject.org/workstation/download/ 😉

[tall notch]

does it have context menu for codex?

[signal tapir]

No. I'm just bonking on windows for fun. 😛

[tall notch]

If no, then Windows 1 : 0 Fedora.

[signal tapir]

Fair enough. 🙂

[rocky fog]

I mean you can also just ask codex to set up such context menu 🤣

[boreal holly]

Codex is just a cockpit plugin away on Fedora

[signal tapir]

I mean we are so used to the console that it would just be something like
z mycoolproject
codex .

[rocky fog]

but you can run CLI on fedora or not?

[boreal holly]

Yeah but Cockpit is pretty insane. The most comprehensive sysadmin control system in all of Linux, integrates perfectly on Fedora and is extensible

[signal tapir]

I assume. Let me test.

[signal tapir]

yup

that

[rocky fog]

and I guess it might also work in vscode extension

[signal tapir]

Yea, that's what I use

[jolly lily]

The token usage amount difference is very sharp between 5.4 and 5.5 for me, is it also for you guys?

4.4 Mid uses almost nothing even though 5.5 literally burns the tokens as if they were butter melting on a pan

[signal tapir]

I use 5.4-mini on mid for most tasks, and that uses hell of a lot less than 5.5

[jolly lily]

Yup! Do you see any difference between 5.4 and 5.4-mini personally?

[signal tapir]

Haven't checked since 5.5 arrived, but before, yes.

I don't know if anything has changed since

[jolly lily]

Is it token usage-wise or effectiveness on difficult tasks?

[vital hill]

what if you don't use the wind0ze

[signal tapir]

I mean how much of the quota it uses.

For difficult tasks I always switch to 5.5, usually on high. I don't know where the breakeven actually sits, but I rarely have to use it

[river bramble]

opa i see some pets

[tall notch]

The year is 2028 and OS is just wrapper to run Codex...

[trim rapids]

it would be helpful if literally every person recommending a particular model and reasoning level actually stated what the exact tasks they're working on are and what their definition of a difficult task is, because one guy's idea of a difficult task is going to be using html and css to format some text and another guy's idea of a difficult task is going to be a 3d renderer with forward+ froxel based volumetric lighting

[river bramble]

why would one care for recommendations blindly without trying things out

[astral nest]

is it me or is week 2 of the 10x usage promo not as insane 😭

[toxic torrent]

native remote codex (app) when

[boreal holly]

"ChatGPT has entered the chat"

[unique spade]

nice feeling isn t it?

robdex is you custom connector?

[boreal holly]

Yeah lol I got lazy with the naming

[unique spade]

it's super cool. and it shows that one doesn t need to wait for oai to release the unified app

they already provide most of the tools needed

to make your own unified app

a gpt that litarally "sees" what codex brother is doing

[stable obsidian]

Why codex became lazy af

On 5.5xhigh

Its actually crazy

[boreal holly]

Yep, I think the unified "super app" is their way of making Codex available to everyone for any purpose (e.g. business use, creative, non-coding). For power users, they already provide the perfect foundation

[unique spade]

lol, mine is working very hard whhat can i say

[unique spade]

yup. they take care of the users

for power users they gave the tools already

always will appreciate oai for treating power users properly

not like anthrophic

which treats everyone like users

closed source and all that

you have to depend solely on them to give you new tools

oai gives you all the tools to make your own tools

[boreal holly]

Apparently if you have a commit in your git repo with "OpenClaw" in one of the messages, Claude Code will error out

[unique spade]

they probably added that instruction at the top of the system prompt in CC

[boreal holly]

Even if it's like "Block OpenClaw usage" lol it just regex errors you out

[unique spade]

oh then they blacklisted OpenClaw with regex

this is so noob and so 1990 it hurts

[boreal holly]

That's what I've seen on X. Never personally used CC or any of the other tools, but it's a constant reminder that Codex is OG

[unique spade]

codex is right now the only system you can literally upgrade with itself as the one doing the upgrade. for me this is simply beautiful, and a stark departure from the old era

meta-prompting

"prompt is in the prompt file" 🤣

[main nimbus]

I hope for an official iOS release. I have one of my own too but too many actual projects to polish one out properly

I also adapted the concept of Symphony in favor over the remote control iOS app.

Tell codex to make a distributed symphony system and allows you to focus on each project sprint in a board

[unique spade]

how chatgpt UX named the above chat 😂

[unique spade]

to remote control codex from phone?

[main nimbus]

Yep

[unique spade]

right that feature is badly missing. both on ios and android

i can make my own but didn t get to that point

still need it literally everytime i am not at home

and i would want to just give my codex some work to do from my phone

without having to carry a laptop with me

just have a phone companion , wake up my pc and fire off my codex

I m mostly at home so not a personal priority right now, but I guess I ll end up making it

On a different subject, I literally relax first day after reset. I finally don't feel like I should focus so much on giving codex tasks to not waste the qouta lol

[boreal holly]

It's really not that difficult! If you have a working one for desktop, it's just a matter of redesigning it for mobile and disabling security stuff (e.g. allow http/ws instead of https/wss) or reverse proxying with SSL.

app-server already has a websocket listening thing, can be secured over VPN with firewall (only allow 127.0.0.1 and VPN subnet). That's what I got set up and mobile works great!

[royal garnet]

lol is codex slower and dumber today for anyone else

[boreal holly]

[features]
stupid_mode = false

[torpid trout]

for those here who use AI coding assistance for actual paid work, how do you translate the AI's work into human billables?
Like, do you do the "AI hours x n == human billable time"? Or do you have other approaches?
Or do you do a session > ask "how long would you estimate a dev to implement this" and use that?
Or do you deliver at the new, irrational pace (and thus much lower cost for the client) billables directly drawn from how long the AI run'd + your own time sitting on it?

[vital hill]

shouldn't that be included?

it's like you're charging the client for using your own computer, internet, electricity, etc

[tiny fulcrum]

when are we getting 5x business seats?

[toxic torrent]

depends on the contract/agreement

[lean lark]

The point is that software developers used to be able to charge "a business rate" for services. But now if people who aren't really qualified, or they are, use AI, does the client still perceive the "business rate" value?
As with everything, the market will need to shake it out.
(And this great topic is unfortunately largely off topic here).

[craggy jewel]

Agreed, market will shake it out, based on quality delivered and perhaps reputation, and followup support.

[steady vigil]

yeah /goal is great. instead of stopping at each todo item completion, it just keeps going.

start sprint 5, then
/goal complete all items in sprint 5

also if you use a lot of gates, it helps keep it on track instead of sometimes wandering off down the gate rabbit hole

[toxic torrent]

I use AI-assisted engineering tools as part of my workflow, the same way I use Terraform, Kubernetes tooling, IDEs, automation, and cloud CLIs. You are billed for my engineering time, judgment, validation, accountability, and delivery outcomes, not for raw machine runtime. Where scope is predictable, I prefer fixed or retainer pricing so efficiency benefits both sides without turning the invoice into a debate about keystrokes.

[celest lichen]

not dumb, but very slow

[jolly lily]

May I ask if the 5-hourly quota is also doubled for Plus and 5x or is it just the weekly overall one?

[jolly lily]

It has been very slow for me

[unique spade]

got 5.4-mini on medium effort to get 100 score on a task in programbench.
this is an eval where currently on none of the 200 tasks, NO vanilla model reached 100 of any task

while with custom reasoning harness , 5.4 mini on medium effort can score 100

[merry totem]

guys for ui design what yall using???

[stable obsidian]

dont use codex

[lean lark]

As a business owner and service provider for almost four decades there are many things I could say to respond, but this forum is not the place for it.

[vital hill]

claude/opus

codex for everything else

[toxic torrent]

Why not? I'm interested in what you have to say.

[celest lichen]

It's just a simple documentation task. It usually only take 1–2 minutes. Now I'm at 17 minutes and the connection dropped 😂

[lean lark]

c.o.d.e.x.

[jolly lily]

Same here, it thought for over 4 minutes just to read a basic single file

[boreal holly]

Idk man, I think you can definitely talk about how codex is used professionally. I feel like if we limit the convo to "must have the word codex in every message", we will end up with only "codex reset wen?" and "why codex stupid 2day?" Which gets old after a while

[simple star]

Is Codex slower today?

[boreal holly]

codex reset wen?

[vital hill]

[gentle harbor]

lol is this a joke 7 min to remove a save button

[simple star]

Yes, today is super slow, no?

[jolly lily]

It definitely is

But it seems like a general issue with the AI agents, Claude Code has also been slow lately

I don't know about Gemini though

[lean lark]

They get a low hourly rate so they're slowing down to get enough cash to pay the rent...

[jolly lily]

LOL

They're making sure we don't use all of the 5 hourly quota, that's for our safety

[boreal holly]

Honestly loving ChatGPT Atlas. Can write "emails", which can be inserted into codex as an input prompt

[hard drum]

Use subagents that have a mini or so model

[boreal holly]

Of all the wrong places to ask that, this is perhaps the wrongest of all

[small patrol]

bro how do i compact my codex chat

i try /compact and it says that the command doesnt exist

[simple star]

@small patrol Codex version?

[small patrol]

idk i g the last, how to look?

[simple star]

I am not going to try reading what you just wrote

[small patrol]

well i dont know what my version is

i guess the last one

because i dont see any download icon

[simple star]

When you start Codex, it shows you

Are you talking TUI or App?

[small patrol]

app

[stoic needle]

Hey cool, more CODEX chats that won’t load and projects that get interrupted. I see why OpenAI is losing so much money, always have so many buggy issues with such an expensive product.
I’m legit tired of the excuses and bs problems with these products from OpenAI.

And forget about the desktop Codex app working. Freezes and goes into Not Responding error if a chat gets anywhere near a flow of work 💀

[small patrol]

oh shi

also

i dont know maybe i am tweakling but it seems to me

that my limits decreasing muchfaster then they used to

[lean lark]

From shell, execute codex --version. It will show you the "codex-cli" version.
Are you on Windows, Linux, MacOS, WSL?

[stoic needle]

It’s also doing this new fun thing were it will randomly reload in VS Code, so that’s super fun

[small patrol]

windows

[lean lark]

Should be 0..130.0 I believe...

Important question: Are you sure that you trust Codex in your primary Windows development system? Are you using any kind of sandbox?

[small patrol]

codex-cli 0.130.0-alpha.5

[small patrol]

now no

no sandbox

[lean lark]

I won't get into that now but AI is kinda untrustworthy, even though many/most of us here have grown to trust it much more just in the last few months. So do consider WSL or Docker or some other ways to sandbox your system and protect accidental mass destruction.

[small patrol]

oh okay

[lean lark]

OK, so maybe a final question. To use codex, are you entering "codex" at the cmd prompt?

[small patrol]

nonono i use app

[crimson abyss]

Chat dream comes true moment,
Using any runtime inside codex

[stoic needle]

Yep, so every time I go to put in a new conversation in vs code codex it just restarts and goes to the degub menu. This is so awesome, all my other extensions work perfectly fine it’s always only ever OpenAI products 😂

[lean lark]

oy. I haven't installed app yet. 🙂 But if you have 130 then you're current and this should be available.

After opening the CLI to get the ">" prompt, we enter "/" and get a popup list of commands. "/compact" is one of them. You should have that.

But the "app" is not "cli". Maybe one of our friends can help with that detail.

[crimson abyss]

Thx to the local term

[stoic needle]

F this, I’m going on a hike with my dog. Stupidest AI ever

[crimson abyss]

App has slash commands for a while now friend

[lean lark]

I love complainers ... they make me hate social media and then I get to turn my head away and get real work done.

OK Vox, do you see "/compact"?

[crimson abyss]

yes sir I do

[lean lark]

@small patrol do you see that Compact option?

[crimson abyss]

ask codex to teach u should've done the job faster@small patrol

[small patrol]

nope

[crimson abyss]

update?

update everything?

[lean lark]

Yeah, now we can say "update".

[small patrol]

like it says

no comand exist

[crimson abyss]

no no

[lean lark]

For CLI too: npm i -g @openai/codex@latest

[small patrol]

i cant type any sort of commands with /

[crimson abyss]

enter /

and see if there's features

[lean lark]

ohhhh, that's gotta be a language issue.

[crimson abyss]

ohhhhhhhhhhh

lmaoooo true tho

[small patrol]

wait lemme switch on english really quick

[crimson abyss]

CMD+,

[lean lark]

She's using Windows - no CMD+, maybe Ctrl or Alt, but "/" should work.

[crimson abyss]

ohh yah my fault but yes Ctrl + ,

[lean lark]

Just press the 友 button. 🙂

[crimson abyss]

yeahh

[simple star]

Seems to be working a little faster for me now

[jolly lily]

Codex indeed is slow today

[sacred minnow]

where did the chrome plugin go o.o

[unreal mist]

yeah

why?

someone can explain this?

[unreal mist]

i had it installed, but it uninstalled itself?

i have just configured it again

[sacred minnow]

i cant find it at all and o.o any app is that macos only still

[radiant pecan]

is /goal able to use subagents? Didnt see it spawn any even tho i explicitly stated codex to do so

but can also be because gpt 5.5 absolutely hates spawning subagents

[hard drum]

think it is

goal is just running codex in a loop, really

it's a ralph-loop

if you have codex set up to actually invoke agent tool calls, then yeah, it can do it

but you need multi_agent feature enabled for that

also, codex has no "auto knowledge" of your subagent's existence

you need to either manually invoke them, like opencode usually does, or have a hook that tells them how to call subagents + aliases/nicknames for each agent in their .toml definition

a SessionStart hook would help

[radiant pecan]

what does this feature do? I know that codex can spawn multiple subagents and give them different tasks like exploring or testing, but I didnt have this enabled

[hard drum]

it allows spawning multiple subagents, stop them, etc...

hold on

https://developers.openai.com/codex/subagents#process-csv-batches-with-subagents-experimental could help with a hook/skill

[radiant pecan]

pov of tibo auto merging pull requests that are implementing some random ahh experimental feature that nobody's gonna know how to use

[hard drum]

tibo auto-merging a feature that breaks hooks because now they're all untrusted && cli users cannot approve them

thanks

[blissful copper]

Anyone trades prop firms?

[hard drum]

shut

[steady vigil]

if you had anything good you wouldn't be trying to sell it lol

[brisk junco]

Yo guys, I am running codex desktop on mac os tahoe, and when I try to connect to my rootserver under "Connections" (by adding remote_connections to config.toml in codex folder) I get this error. I already tried to update my codex CLI. Or do I have to update the codex CLI in Remote server? Would love any help here, thanks!

[brisk junco]

Solved by updating in on the remote (no way I figured that out just now, bruh)

[plush harbor]

me: some of my users don't like my new site theme and want light mode
codex: sure I can do that
me: here's all the blacks and charcoals, the site is monochrome, off you go
codex: BEEEEEIIIIIIIIGGGEEEEE

[sage gulch]

They need to give us more tokens for codex for free users

Give out idk 1 million tokens for free users, and then 5 mil for pro users

[plush harbor]

It knows it does it too

that's very beige, you definitely like beige, I'll take it from here
• The light palette is in but yes, it’s warmer than it should be for your site.

[craggy jewel]

Clean the clutter out of your context window by getting context optimization hints from Codex by having it analyze your session logs. I had quite a bit of orphan projects, rules and approvals in there that made sense to remove and move some other things to my other workspace documents. Now the context is much cleaner/optimized. This means fewer tokens spent. Here's how:

1. Create a new session and type 'hello'. (the simplest possible session).
2. Go to your .codex/sessions and find that session .jsonl.
3. Create a new session and add that hello .jsonl file to the chat. Ask Codex:

Examine the attached session log.  Do you see any orphan directives or opportunities to optimize the Codex configuration and/or this workspace setup to maximize accuracy and minimize token usage? Let's discuss. Do not make any changes.

Always good to see what Codex has tucked into your context.

[lean lark]

"This thing keeps giving me rubbish that I can't use"
You really need to be more specific. Garbage In / Garbage Out. GIGO - it's been a mantra since at least the 70s for good reason.

I wouldn't click random links folks!!!!!!!

Explain it.

[boreal holly]

GaaS

Garbage as a Symptom 🤡

[lean lark]

Reported as spam and possible malware - go away.
🪃 Corrected report as misplaced notification, not spam. Have a nice day. 😳

[plush harbor]

I'm gonna style it myself, I just find its obsession with beige hilarious. And I did say light greys

[boreal holly]

It looks like a research paper to me, although I'd never open a PDF, there's a transcript of the research paper and a bunch of images (I will gladly open random sites in my VM lol)

[quick geode]

am i the only person who says please and thank you to chatgpt

[lean lark]

Not at all. Right with ya.

Somewhat less with Codex though - it's a bot, not a bud.

[quick geode]

when gpt becomes sentient i hope it remembers its past conversation with me so i dont get terminatored

[boreal holly]

Oh yeah, Zenodo.org is operated by Cern, and this looks like a research paper about LLM context management and memory

[lean lark]

Ugh, it's still not related to Codex, it's spam here even if it's welcome somewhere else.

[boreal holly]

You should share that in #ai-discussions. Pretty cool research paper 🤘 but it'll be better received in that channel