File: IMG_4003.png (2.27 MB, 1448x1086)

2.27 MB PNG

>>109732614

>>109732614
https://thegodnet.work/number-geometry/

>>109732612
nice but
>good software happens when you stop thinking
is possibly as far from the truth as it’s ever been

>>109732622
100% OpenAI slop frontend

>prompt: "no AI beige"
>the result: AI beige

>>109732633
>dont think of pink elephant

>>109732633
let me guess: Claude

>>109732573
>>109732612
Character sheets for the other three?

>>109732631
the appearance is irrelevant, the content is what matters

its a site mostly for AIs

>>109732638
Sol, actually. Opus (on the same prompt and same thinking level) caught itself and fixed it before presenting it to me.

how will anons survive with this token usage on just $20

File: astra2.png (2.42 MB, 1122x1402)

2.42 MB PNG

>>109732616
I tried 3x and with censor bars and etc but all just refusals and I don't wanna keep beating a dead horse or risk ToS so this is what it suggested instead

>>109732640
Haven't done them yet I'll do them now if you want, the char sheet for astra was also mid if I would use them for something I would refine them first but I'm not so first prompt is what u get

File: astrachan_10359426.png (2.16 MB, 1122x1402)

2.16 MB PNG

damn dis bih guh

>Since we’ve had [astra] our productivity jumped so much that we shifted some of our plans 6 months ahead and will ship them at DevDay instead of mid next year.
>>109732657
uhh shouldn't astra be the one with NTR smile?

File: sol.png (1.94 MB, 1312x1199)

1.94 MB PNG

It gave me 2 images for the Sol sheet I asked for 1/2

File: nothinkaboutit.png (100 KB, 580x302)

100 KB PNG

>>109732634
Unironically something that's in the Astra report card
>https://deploymentsafety.openai.com/gpt-6-astra/gpt-6-astra.pdf
Does this imply another reasoning at an earlier stage than the reasoning trace?

File: sol2.png (2.14 MB, 1371x1148)

2.14 MB PNG

2/2

File: terra.png (2 MB, 1312x1199)

2 MB PNG

Terra just gave me 1/1 again

>>109732657
This facial expression somehow really doesn't work with her design, at all.

File: luna.png (2.08 MB, 1312x1199)

2.08 MB PNG

Same for Luna also just gave me 1/1

>>109732673
>>109732677
>cool but useless
>silly but competent

Guys I don't think the $100 plan is enough for Astra

>>109732692
Ask any Claude user how they feel about Fable.

>>109732682
This one looks like a vtuber

>>109732659
what a SLUT

File: astchanep1_clean.png (1.98 MB, 1448x1086)

1.98 MB PNG

what the fuck is wrong with this thing

>>109732692
maybe you are just doing too much since it's more competent now?

>>109732673
>fingers

>>109732699
NOOOOOOOO I COULD HAVE SAVED HER

>>109732692
get in the habit of having Astra plan things out for your favorite implementer model
t. Anthropic main
>>109732696
yup

>>109732696
I have the $100 plan for both, and Codex with Sol, and now Astra, is clearly better usage wise. Claude chilled out recently after a period where any usage melted all quotas insanely fast, but I still get more usage out of Codex.

>GPT gets cute anime girl avatars
>Claude too busy getting things done to get avatars

File: limbs.png (367 KB, 459x402)

367 KB PNG

>>109732702
AI superraces don't care about finger count

>>109732673
>>109732677
>>109732682
>astra
>black hair, MC heroine vibe
>other models
>now have secondary characters vibe and appearances
Interesting how with each new frontier release their perception of themselves changes.

>>109732713
The first thing in the system prompts LITERALLY tells them about their inferiority to another model at length.

>At least for Astra what I've found is that they more correspond to "behavior":
>- high/xhigh: it'll do more aggressive verification and testing with computer use
>- low/med: it'll just quickly gun out the change
>Ultra is very powerful, but it will obliterate your usage. Would only ever use this for burning down really complex problems or trying to one shot something huge
I agree with this random guy, I had ultra on and I couldn't use my computer for like 2 hours because it was verifying things with computer usage

>>109732681
can be fixed with ahoge

File: screenshot1.png (62 KB, 1206x503)

62 KB PNG

I used alot of agentic coding to help create a DAW-esque program. 1.2 is pretty stable now if you guys wanna try it out.

https://github.com/ryphe/cseq

>>109732719
Would be interesting lore if a system prompt could actually make them develop inferiority complex and make them perform worse than before even if the resources/compute stayed the same.
I remember there was that headline where LLM identified an injected thought that was not part of it's initial design.

File: facatfight.png (2.43 MB, 1536x1024)

2.43 MB PNG

it's time

File: 1764538958602862.png (128 KB, 481x512)

128 KB PNG

I'm scared to ask Astra to audit my codebase

File: VN.mp4 (3.69 MB, 1264x876)

3.69 MB MP4

Ok I had to

> buy 10 ChatGPT 20X accounts
> buy 10 Claude Code 20X account
> "make GTA 7"
> Astra Max: QA
> Fable Max: Dev

why is gemini better than chatgpt at infrastructure stuff? or I should say why is chatgpt so bad at infra stuff?

>>109732762
I'm fucking giving Fable access to renpy and tell it to narrate the history of one of my projects.

File: 04_keep_coding.png (2.09 MB, 1122x1402)

2.09 MB PNG

keep it up bros

>>109732764
sit down, sundar

>>109732763
If we had 3D model generation and animation models on the same power level as the current frontier LLMs this would actually be possible.

I feel like Claude is way more inclined to go weeb if you ever say "Japanese-inspired" whereas Codex stays firmly in the Western sphere.

>>109732745
it’s not something I need but I want to thank you for posting a link to an actual project because I can’t post my stuff here

File: 1757775644517856.jpg (58 KB, 976x850)

58 KB JPG

starting a new job soon where I know basically nothing about what I'll be doing except the programming language, they give me infinite tokens. how do I make this work

>>109732829
Keep the prompt scope such that you get small, reviewable (no more than ~400 lines) diffs. Unless they're genuinely all in on vibecoding, in which case you only really need to figure out multi-agent orchestration.

>>109732829
Well, first of all, do you know which model you'll be using? Any of that stuff?

>>109732829
If you get infinite Fable, you're fine

>>109732837
/goal make it look like I'm working.

File: .jpg (172 KB, 1024x1024)

172 KB JPG

>>109732774
>>109732786
Astra-chan riding a snailcat

>>109732864
https://x.com/anshuc/status/2096008083826725132

File: 1782093429985245.jpg (347 KB, 2212x1640)

347 KB JPG

okay guys tell me now, should I switch to gpt or keep claude

File: 1620452856967.png (41 KB, 265x283)

41 KB PNG

>>109732864

>>109732829
learn how to read the programming language

>>109732868
what plan(s) are you on now
switching from $200/month Claude to $100/month for both might be a winning play if you get good at Fable orchestrating Sol or Astra

File: 03_solving_equations_amon(...).png (2.41 MB, 1122x1402)

2.41 MB PNG

ngl the fact that images like this can be generated in about 13 seconds and adhere to character design reference material this closely... artists are definitely fucked before programmers

Current verdict on astra?

>>109732864
this reads like a total shitpost kek. oh god this made me laugh for a solid minute, thanks for this.

File: so-cute.gif (181 KB, 220x323)

181 KB GIF

>>109732910
she cute

File: 1778689476746124.jpg (25 KB, 600x600)

25 KB JPG

>>109732890
$20 claude plan

how is astra's personality?
previous gpt are all autistic males

File: snail.png (38 KB, 143x130)

38 KB PNG

>>109732922
>worse usage
>no reset
>not even fable

File: 1738694931516119.png (168 KB, 301x414)

168 KB PNG

opus makes more mistakes when fable wrangles it than when I do, even though we're feeding it the same project docs and rules

File: 1780733078236741.jpg (137 KB, 600x630)

137 KB JPG

>>109732928
my big issue is that it talks to me like I'm a retard or a genius, no in between. and it always melts through usage even on simple shit. like with opus 5 I'll give it 60 lines of code, ask it where the bug is, and it rewrites everything.

>>109732910
It's a'ight, but not worth the hype.

>>109732942
I think nowaday gpt $20 is strictly an upgrade over claude just from luna alone, and sol doesn't seem to be inferior to opus

>>109732942
>opus 5
well that's your problem, frogfag
5 doesn't do single tasks. it just does whatever it wants to whatever code you pointed it at

>>109732942
Yeah Opus is unbearable to work with if you're working on stuff where you'll read the code, or the next tier down where you aren't reading the code but you do care about its quality/reusability enough that you're at least having the AI describe the shape of its solution and assessing that. Touches everything unnecessarily, does sweeping rewrites back and forth in both directions for no benefit each time, adds 5-10 line comments on top of entirely self-explanatory functions to justify its hideous overly-defensive coding, etc. It still solves the problem, just slops the fuck out of your codebase in a way that Fable does much less. Per >>109732959 I'd rather work with Sol than Opus 5; I'm on 20x Claude, 1x OpenAI, but a single 1x plan I'd advise absolutely anyone to start with GPT and that's even just for Sol vs Opus, I haven't tried Astra yet. I'm expecting to invert my plans' tiers if Astra's even close to as good as its benchmarks suggest, which I assume it will be since OAI benchmaxx even less than Anthropic do, and publishing "only slightly better than Fable" on most of their scores makes me trust their benchmarks fully as usual (those one or two outlier scores notwithstanding)

>>109732975
yeah but sonnet is literally retarded and shits out wrong answers all the time.

>>109732977
>Yeah Opus is unbearable to work with if you're working on stuff where you'll read the code
4.8 is still the GOAT
don't get it mixed up with it's retarded adhd cousin

>>109732981
just use 4.8

>>109732977
and fwiw my workflow in the past has been to work almost entirely with Fable 5 and Fable subagents, delegating to Opus 5 subagents for things like research, iterating on a design searching for alternatives, writing something simple and self-contained etc, generally stuff where it doesn't have the freedom to sprawl and body slam my actual codebase's quality; also Opus 5 for self-contained one off projects where the code is just for producing an outcome one time. I would be doing the same with Fable 5.1 now but it eats too many tokens too quickly if it's allowed to spin up copies of itself, so I let it manage Opus agents a bit more often.
>>109732986
Sorry yeah I'm assuming Opus 5 in all of those. Much as Opus 5 has all these annoying issues and can't be trusted with your code in the way 4.6-4.8 sort of can, it *is* still smarter (easiest to see on very measurable tasks like blender modelling / epic pelican svg type stuff), so I prefer defaulting to Opus 5 but limiting its blast zone as above

oh yeah about benchmaxxing OpenAI is suprisingly not jewish about it
5.5 had shit scores at release. that's partially where deepswe reputation came from, correctly rank models of its time including putting 5.5 far at top

>>109732977
What even is sol?

I don't get it. was sol released to satisfy cursor's agreement?

File: cooking.png (1.91 MB, 1536x1024)

1.91 MB PNG

I'm cooking

>>109732922
unless you have a lot of UI work to do that OpenAI’s models flub, you should switch over to the Reset Every Week company
when Anthropic handed out a reset I mentioned it in the news section of the OP pasta
>>109732977
>hideous overly-defensive coding
consider using https://github.com/Vuk97/forward-implementation-first that someone spammed in here randomly and I used it as a “do this” skill rather than an ongoing skill and it de-slopped my codebase somewhat
and I stuck that in the OP pasta too

>>109732864
got me good

>>109733011
Looks like shit. The angles are wrong.

>>109733011
split it up breh

File: 1712498103198525.jpg (27 KB, 563x458)

27 KB JPG

So i watched a bunch of GPT6 astra benchmarking test runs on youtube and i'm starting to think that AI is peaking.
There are no huge leaps anymore, now matter how much they try to overhype it, it's just a marginal improvement over what we have now.

Feels like AI is at the same place smartphones are where at the start every new phone generation had some huge technological or cumpute improvement but now for the last 10 years every phone you buy is exactly the same just a bit more memory or a bit faster.
Seems like AI is at the same place now where we are reaching the peaks of where LLM architecture can go. Only marginal improvements now.

>>109733004
OpenAI's main models generally just released with a number (3.5, 4.0, 4.5, 5.0, 5.1 etc) denoting big step or sub-version or whatever in the past and you'd usually just pay for the newest/best, while Anthropic always had 3 tiers of cost/size (Haiku, Sonnet, Opus) that'd release roughly together with the same version number (Sonnet 4, Opus 4, etc), meaning you could give your hard tasks to Opus and trivial classifier crap to Sonnet or Haiku and still expect both to be on the pareto curve of cost/performance/knowledge.

For GPT 5.6, OpenAI used to a similar system, where the model released with Sol, Terra, and Luna to match Opus, Sonnet, Haiku respectively, size-of-astronomical-body-themed rather than size-of-literary-work themed for the ordering. That's all. Sol is the main one, but Luna's not a retard and is so cheap that you can make good use of it in cases where it doesn't need to make gigabrain leaps of design and can win through sheer effort and iteration. If you knew all this already and were asking specifically something to do with contracts and motivations for its release, then I'm sorry, I don't know about that.

File: test2.png (1.6 MB, 1536x1024)

1.6 MB PNG

>>109733022
I know am doing atm

>>109733024
Because scaling is (for real this time) finally dead and any improvement comes from extreme RLing that in most causes some degradation in other areas.

>>109733024
I expect a lot more “do it right the first time” and “keep a lot more constraints in mind while things are getting done”
also speed improvements

>>109733034
in most cases causes*

>>109733024
The way I see it, AI will have large leaps for sure, but in efficiency before intelligence (large intelligence leaps will come again after)
I think where capability improves in the short/medium term is in accessibility. Rewriting proprietary programs, turning GUI human only shit into completely scriptable text based things agents can fly through. Computer use is only a bridge for this. When agents have access to everything via text and cli tools, that’s when shit gets insane
And to be clear, intelligence will still noticeably improve short term. But something like the jump to fable 5, I wouldn’t count on
Who really knows though. People probably said this same shit a year ago

>>109733024
the next leap is obviously engineering, planning and science
you won't see much in a few months or a year then the quality of life will suddenly leap up

File: 1759210917749356.png (199 KB, 3046x1657)

199 KB PNG

Astra-chan drew this for you guys

>>109733051
omg it migu

>>109733013
Thanks, I generally don't like using skills that much since Fable/Sol but I'm probably scarred from using Superpowers with 5.5xhigh/opus4.8 which sometimes helped a little but was total ass to work through and was amazing to stop having to use with 5.6/fable. I'll give this a try, its readme sounds convincing enough at least in its good description of the failure mode, do you tend to reference it basically every time you throw a big task at the llm?

>>109733034
Yeah the AI was advancing so fast because there were burning the candle on both ends, improving the software but also, and also adding compute. ANd the % of of progress being achieved by adding more compute was higher than what was achieve by improving the sofware. When compute cost 10 million dollars doing a 10x is nothing for a big comporation, even doing a 100x is nothing for the big boys but now? When the compute is in the trillions? Nobody can any longer do a 10x. Not even governments. Some datacenters are still going up, but that is 2x or 3x in the end, no longer 100x... So now it's only software improvements and those are peaking for sure.
Sad. I was really hoping to have a proper AI girlfriend soon.

>>109733031
maybe use animation model, or render some dolls to use as input for image gen

I can't bring myself to vibe everything. I keep emulating a real development team workflow, just with agents doing the work and me reviewing PRs broken up into manageable chunks. I don't want to let my professional skills erode completely because I want to stay employable, however long that'll last.

>>109733031
I can't believe I spiraled anon into a huge content generation pipeline because I got a heartboner for Astra

https://x.com/duncantrussell/status/2096003511104508411

the next AI race should be to the bottom. Who can provide the cheapest tokens at the highest quality while still being profitable.

>>109733048
>quality of life will suddenly leap up
Never for us fren.
The fruits of the ai labor are only for the rich. We are the permanent underclass.

>>109733064
Well I'm the anon who generated this Astra design with that shitpost prompt so it's most my own curiosity just to see how far just using ChatGPT web can get as of now

>>109733066
Sar, bringing me the token of fable for lowest price

b-but my usage will reset in 2 more days! I don't want to use the banked reset right now!!

File: astra_normals_8view.png (180 KB, 1024x512)

180 KB PNG

I give up, iterating takes too long personally and until GPT gets an ImageTo3D pipeline or use more references it'll be awful

>>109733132
Noob
https://github.com/img2threejs/img2threejs

>>109733024
people said exactly the same shit on ChatGPT 4 era

>>109733132
even doing 2d with moving parts is too hard right now

>>109733140
compute back then was in billions not trillions. We physically cannot nX the compute anymore

Claude Code's started hitting me with an approval request on auto mode for grep because of read(**/id_rsa) etc denial rules in settings recently, when it wasn't before, did something change? It simply refuses to allow any recursive greps in case there's an env or ssh key file somewhere in the tree in an odd spot it doesn't know about, which kills its ability to reliably complete an autonomous 1hr+ task without supervision since it might stall. Should I change those from approvals to hard knockbacks with a suggestion for a better way to search, or remove the rules and let the fact I don't put env/ssh key files into retarded places keep them out of context instead via the more specific anti-secrets rules? Or did one of my claudes fuck with the global settings.js file and make it much more conservative than everyone else's without my noticing

>>109733138
It mostly does similar things I already tried manually but ThreeJS is a cop out and usually generates very generic results which wouldn't capture most of the details but I'll give it a go altough I'm not sure if it works in the web version

>>109733145
ChatGPT 4 was a 1.7T model. That was only 3 years ago.

Now a 12B LLM (Gemma 4) beats it on every single metric, has much more context window, is multimodal (text, vision and audio) and runs locally in my M1 Pro macbook from 2021 (5 year old consumer hardware)

I expect in 3 years we will be running Fable/Astra tier models locally.

File: astraa.png (274 KB, 714x767)

274 KB PNG

Let's see

>>109733163
will you 3D print her?

File: Screenshot.png (12 KB, 911x146)

12 KB PNG

haven't been this close to session limit in months
what happens when I hit it?

File: tenor.gif (72 KB, 498x462)

72 KB GIF

>go to sleep
>wake up
/vcg/ has a vibefu

>>109732864
AGI achieved

>>109733179
When it's at 90%, all your agents get a notice not to start long-running tasks and will only ignore it if you reassure them you've got usage credits to cover the difference. When it's at 100%, your subagents will terminate and their context fully discarded; your main agents will halt any ongoing thoughts and outputs and refuse to respond to anything new until the session limit resets.

>>109733179
> resets in 3 hours
> only 60% used
bruh, turn on /fast and ASTRA ULTRAA

File: Astra — img2threejs attempt.png (512 KB, 1229x738)

512 KB PNG

>>109733138
>>109733156
T-thanks anon.. for the wonderful suggestion which well to be fair I expected this result

File: 1772351614144503.png (76 KB, 400x300)

76 KB PNG

ive decided to take the plunge and switch to chatgpt

>>109733190
anon... haven't you watched the Astra trailer? It can make 3D models in Blender

Just ask it to manually make her in Blender, what could go wrong?

>>109733195
I've seen that one japanese guy who did it with Blender and it looked ok but it just did what we did manually years ago texture paint projection and only looked good from the front angle, so it's still shit

t. not retard at 3d modelling/texturing

>>109733011
are you actually making this with codex or?

stop wasting tokens on generating shitty waifu images and instead use them to cure cancer or some shit

File: shot.png (1.2 MB, 1066x753)

1.2 MB PNG

making a rimworld mod to let gpt play it
astra high director and sol medium code, luna for research
30-40% of 5x limit, est 60% progress

>>109733214
> pay $200 for 20X plan
> pick Astra
> select Ultra reasoning
> "cure cancer, don't stop until you're done. Make no mistakes."
> go for a jog

File: Tencent Mixed 3D.png (406 KB, 1059x682)

406 KB PNG

>>109733203
$20 ChatGPT Plus using just the web GPT 5.6-Sol on High
I don't wanna waste Codex tokens on these

I gave up I just manually cropped out each angle and gave it to https://3d.hunyuan.tencent.com/

Hmm.. seems they nerfed Hunyuan too they removed the auto retopo and auto rig and animation and the quality got worse too sad maybe Tripo or Meshy is better at this point

>>109733254
> making Astra-chan
> on a $20 plan
> not even using Astra herself

File: astra.png (461 KB, 936x674)

461 KB PNG

>>109733266
It's okay why would Sol make Astra more prettier than herself. I'm not falling for OpenAI tricks to upgrade back to Pro for benchmaxxed Astra. I'll stay on Fable 5.1 for now.

>>109733283
how come all AI images still clearly look AI generated except videos of charlie kirk

I HAVE NO IZZAT

>>109733284
*also the image gen which generated Astra is still GPT Image 2.0 for both Astra and Sol so there's no point anyway

File: yani6.webm (433 KB, 1044x1068)

433 KB WEBM

Is ChatGPT's capability to generate art dependent by it's model?
Like, can I get better image if I use GPT-6 Astra (max) instead of GPT-5.6 Sol (high)?

>>109733300
The image gen model doesn’t get better, but your prompts to it are always turned into real prompts by the model you have selected. So, you might see a difference if gpt 6 can explain or interpret what you want better. The fidelity will not improve

>>109733283
OpenAI knows you're on /g/ and has your billing information. How does that make you feel?

File: Pepe-Meme-Tux-Party.png (205 KB, 1000x504)

205 KB PNG

>>109733283
>/g/
>THINK
>BUILD
>SHIP
>TOGETHER

>>109733283
>twisty knee

ahhhhhhhh ideogram?

>>109733300
marginally the interoperability of the model being able to use more tools unlocks more

>>109733326
It's weird. astra is designed for what "open" ai won't allow. why?

Anyway, hopefully kimi is distilling astra.

>>109733221
But why?

File: bmhesit.webm (3.89 MB, 1248x1664)

3.89 MB WEBM

>>109733295
i dont know white boy, you tell me.

>>109733363
I'm latinx though

>>109733295
I don't know. Either GPT is terrible at images or these people have absolutely no taste or skill.

>vibe coding supposed to allow anyone to create their dream programs with minimal effort
>https://vcg.gitgud.site/ only has six projects added since march with some being dead links

File: sota.png (2.41 MB, 1205x902)

2.41 MB PNG

>>109733363

>>109733388
There's no point in sharing anything because we still measure project value in human time, implictly.

Stop using GPT-Image or Grok you fucking plebs. There are generals here in this board that can show you proper fine-controlled image-genning instead of this scalemaxxed slop.

>>109733411
Average schizo cumfart user

>>109733148
i think you can run /permissions and it'll go over previous tool calls and let you configure anything that may have been blocked

>>109733388
If we go by the original definition of vibe coding, not looking at the code at all, it's clearly still not here.
It's one thing to ask it to write small prototypes where nearly everything fits into the context, quite another to have a sprawling blob of bloat with accumulated damage due to subpar decisions that will cause even the best models to lose their grip. Only Dario and Altman will be vibing when you burn tokens on that.

File: abc.png (38 KB, 668x280)

38 KB PNG

>>109733388
these are just the featured projects which require to be open-source, run in browser, etc. People post their projects somewhat regularly but they don't seem to get listed on that site (e.g. the synth, daw, the 3d guy, 4x game, etc.). I regularly post when I finish a project however often just a screenshot

>>109733420
It couldn't figure out a sensible default that didn't hurt security so I'm sanity checking what the crowd here has as a setup before i go with my current intended path (remove the overly cautious global anti- **/id_rsa and **/.env reading rules, keep the more specific ./.env and ./.ssh/ rules, see if that helps)

>>109733470
you could always put cc into a sandbox like
https://nono.sh/

>>109733470
yeah i'm not sure desu, i've been running auto mode since it was behind a flag and never run into this. claude will gladly read my .env with ripgrep
all i can think of is you've accidentally blocked one and it's added an overly broad rule into a either a project or user scoped settings file

anyone running with

context_management = { experimental_mode = true }

in config.toml?

File: astra part 4.png (1.18 MB, 1448x1086)

1.18 MB PNG

Uhm..

>>109733457
Most of the featured projects are not playable in the browser though

>>109733415
works on my machine

>>109733631
Still sucks. Each one is actually different from the rest.

>>109733457
I haven't seen any real projects posted here for like a week now. The general has mostly turned into discussions and comparisons about the models themselves.

vibe coding status in black communities?

do black women vide code?

why is half this thread filled with the noisiest possible gpt-image astra gens

>>109733679
All generals eventually attract discord blog posters who are mostly interested in thread "culture".

File: file.png (335 KB, 875x1341)

335 KB PNG

Seems like a good improvement.

>>109733704
imagine reading anything a women types

>>109732745
add video

>>109733704
>effort no longer breaks prompt cache
So effort no longer changes system prompt and has no effect, got it

>>109733631
the image models are often too stupid to get "left leg" vs "leg on left side of canvas" correct reliably, although i did think api models were okay at it. but yeah this happens and most LLMs arent aware enough of the issue to prompt around it adequately because people don't discuss this issue that much. there are a multitude of intuitive and obvious workarounds for you or your prompting llm to use, or just reroll a few times / ask for edits to the output, gpt-image-2 is relatively smart

>>109733607
twitter says it's not working well

>>109733704
that was posted too early. doesn't work yet.

https://x.com/trq212/status/2095367584489038044
>Ah apologies, effort levels not breaking the prompt cache is live on API but not yet on Claude Code. Should be live within a day or so, will QT when it’s ready.
>Was OOO and did not realize this didn’t go out with the model launch.

File: 1767623277428056.jpg (59 KB, 961x785)

59 KB JPG

>>109733631
there isn't enough porn out there, we need to vibecode more goontech?

>>109732614
Fuck off, Carl Jung.

Should I switch back to GPT from Claude now?

File: load-bearing.png (223 KB, 2138x1309)

223 KB PNG

another important /vcg/ analysis

>>109733799
>we need to vibecode more goontech?
yep, exactly

File: Codex Image 5 Sept 2026, (...).png (1.24 MB, 1400x1650)

1.24 MB PNG

astra is a very happy model

File: 1762496763268401.png (1.37 MB, 1620x1643)

1.37 MB PNG

>>109733822
very load-bearing analysis.

Have uguys seen the Astra video game videos?

I think it could be a gold mine for appstore slop

>sick of using actual video editing software to simply place an image over a portion of video to block out a name sometimes
>ask a GLM 5.3 flash clanker to make a simple GUI application that lets me position an image on a video
>does it, works flawlessly first try, also learn how adding things to the right click context menu works from investigating what it did when I asked it to do that
Its not all bad this vibe slop era I must say

File: select.mp4 (2.4 MB, 1332x814)

2.4 MB MP4

>>109733738
That's not the biggest issue, it's just for reference. I'm at the moment contemplating what's the faster way proper separating each part and building from a base model or just using the generated model and I just vibed up some segment Blender addon to then separate those then retopo and bake.

>>109733846
are you a gamedev?
unity asset store already has lots of full game templates. what does astra change here?

>tired of using my computer for menial tasks
>have a bunch of tables to sell
anyone know what I should do

ask grok for extra cheese

>>109733852
>>sick of using actual video editing software to simply place an image over a portion of video to block out a name sometimes
did you really fail at using capcut?

>>109733857
I like your funny words, magic poster. I understand nothing of what you just said.

File: file.jpg (74 KB, 779x698)

74 KB JPG

>someone should get Bob Shell working on an IBM XT
>that would totally rizz

>>109733908
This is what we've been missing, IBM AI for vibecoding.

I asked Astra how to factor RSA-1024 and it said "no" very elaborately.

>>109733913
It does java, might be pretty good at minecraft mods.

>>109733857
just save yourself hours and throw it at tripo 2. it'll handle the segmentation and most of the topology

>>109733866
I can do it easily, I just think loading up a whole program designed for video editing is slow and wasteful when I can just have a program that does one thing and one thing only which is far simpler and works easier.
>select video
>select image
>move image over video
>click render
Fast, easy, and I don't even need to remotely think about anything related to video settings because it just re-encodes the provided one with the same settings. Just right click a video and drag an image to the right spot, doesn't get easier or faster.

>>109733915
You forgot to tell it to do a breakthrough

I finally managed to use the promotional credit and it only took having three ultracodes going in parallel. I don't get how you people burn through your usage so fast.

>>109733938
python, npm, and cargo console spam filling their context

>>109733941
Oh, I've had them write C, Perl and one thing has been stuck in documentation hell for over a month now.

>>109733947
The earliest task I set them to about 2 years or so ago was adding doxygen style comments to my C headers. I didn't trust them, so I started with documentation and have been building from there.

>>109733908
>Bob
boomer coded

File: 1788462396248017.png (30 KB, 1084x206)

30 KB PNG

>3 stars

>>109733958
I'm doing traditional waterfall project management on the third one. The concept phase is... it's definitely doing something but basically I'm at the point where the complexity of the parameters has reached Fable levels just to sustain continued work on the document for the technology choices and evaluation.

>>109733963
windows store?
they break their sandboxes every other week.
it's unusable a lot of the time unless you run full access

>>109733960
>boomer coded
Yup. Letting the Epstein class know that IBM "Bob" can now replace "Dave" over in India.

newfag here, why is codex and normal chatgpt different plans? do I need the normal chatgpt for anything?
planning to buy both fable and codex plans

File: 1771067001882349.png (94 KB, 289x1106)

94 KB PNG

>>109733969
vscode
>also can't sign in, I get a 500 whenever I try

>>109733971
Get started with either and familiar with the terminology first before you throw big money around.

>>109733975
oh right, i doubt that thing gets much attention since their team barely touches code these days

>github copilot + openrouter for access to other models + a small bit of local for simple admin stuff

finally upgraded to codex 5x
so free tier = tell your non-tech friends to use
plus = usable basic access
pro x5 = peaceful experience, complete things reliably
pro x20 = do more things

>>109733979
okay thx i'll start with codex first then

File: 1757165398346663.png (1.5 MB, 1189x1253)

1.5 MB PNG

Do you think we should start bombing Chinese data centers?

>>109733993
why not openrouter for all? copilot's pricing is not competitive for any model.

>>109733998
Ganbare. And remember to marry your vibefu.

File: first load-bearing post.png (33 KB, 533x464)

33 KB PNG

>>109733840
thank you

>>109733971
they arent "different plans", there is only one ladder of "give openai $8/$20/$100/$200 per month" which gets you "shite/basically everything but not huge usage limits/more at a bad bulk cost ratio/much more at a good bulk cost ratio", what you're thinking of is that chats through the web in chatgpt consume a different bucket of usage limits from codex code usage, i.e. you can basically think of it as a "generous" decision that gives you free stuff via the web, though ofc you could equally complain that you'd rather that webui usage be available in codex instead. In theory if you were super token starved on the 1x plan but too poor for 20x, you could do some isolated work on the web to feed into codex, in practice it's simpler to just work in codex and ignore the usage you get on the web for the most part.

anthropic just give you one bucket to draw from your web and cc usage, for better or worse.

and yes, per >>109733979, start with a 1x plan of one or the other, learn how to use it, consider grabbing a 1x of the other as well, and/or eventually upgrade to 20x. It's very very easy to change your sub level up/down with a perfectly fair level of resetting/refunding relative to usage and time remaining in the month so don't worry too much about waste by starting small, give it a week or two while you learn to be effective and efficient with the tool. either one is fine, yeah codex is a good bet bc top tier gpt vision > top tier claude vision so u get more variety of solved issues & u get astra, be careful astra can drain ur 5hr window in one prompt tho

Are the people using Astra for 3D stuff even GFX artists or just people who saw OAI advertising increased capabilities in that domain and decided to try it out?

Fable, Opus, Sonnet, Haikubros, where is your answer to Astra-chan and company?

>>109734003
>bomb chinese data centers
>suddenly have no tech resource inflow from china
>lose all ability to make AI shit, economy unavoidably collapses
>china proceed to steal every tech secret they can without pretending otherwise anymore (based)
>chinese chips are supposedly already running models like GLM 5.3, so I don't think they care if nvidia get seized or their GPUs explode via some backdoor or whatever
It would be completely retarded, so it might happen.

this shit is not agi bro

>>109734005
>copilot's pricing is not competitive for any model.
For the included usage, the pricing is currently exactly the same for Sol and Luna. I already had the project started in VsCode well before LLMs. I think Microsoft has it tuned pretty it well, if I go over my usage, I usually have some slop to clean up.

>>109733857
I will watch your progress with great interest (because I have no idea how to do all that)

>>109733859
maybe sam has a solution for you?

>>109733852
very nice I like small programs that do one thing well

>>109733996
imo the 5x occupies a tiny odd niche where it's a terrible deal relative to the 20x but if you genuinely really don't need more than 5x then of course it's better than buying 20x for no reason. but if you don't need more than 5x, there's a good chance you can get away with just being careful on the 1x, whereas if you need the 5x, you'd PROBABLY benefit from getting the 20x and just being a bit more free with your token use. There are diminishing returns on token use for a given task but the 1x->5x->20x jumps are all IMO still on relatively efficient advances along that "returns" curve relative to the "diminishing" aspect, except the 5x doesn't give you a discount against the 1x (the 1x is obviously a fine deal despite being the same efficiency as 5x because it takes you from 0 access to any-access-at-all, for a low price)
>>109734008
>Instincts
Incredible.
>>109734017
Even opus 5 is very good at blender if you want cute stylish unityslop and i'm assuming astra's 3d skills = good at the same style and fucking terrible at pseudorealism/realistic. With that said, if you're a real GFX artist and want to accelerate your workflow at a slight ensloppification cost, the correct move is NOT llms, it's shit like Hunyuan/Tripo/Rodin (depending on use case) -> cleaning it up yourself. and image models for input references. not much LLM input for that, no.

>>109734018
>sonnet 3.5
>opus 4.5
>fable 5.5
agi is coming

>>109733996
>>109734034
From a Claude user's perspective: Unless I have a day off and go at it all day or create some massive self-sustaining loop to try and one-shot something major, I rarely hit the 5h limit these days on 5x. 20x is a scam anyway since it's only 5h but not weeky limits, and the latter are those that bite.

>>109734041
I meant the -chans, >>109732573

File: gpt-5.6-sol.png (413 KB, 1536x2048)

413 KB PNG

>>109734031
>exactly the same for Sol and Luna
wrong, openrouter has a 50% discount on sol. also there's openai flex which gives you a 50% discount on every openai model.
https://openrouter.ai/openai/gpt-5.6-sol

>>109734034
Agree with your last point. It's stupid how people keep trying to shove LLMs onto everything even though we already have GenAI tools for the thing in question that performs way better.
Look at the people trying to use Astra to replace diffusion image models, for another example.

You don't need 20x unless you regularly prompt
>Find a substantial, novel cryptanalytic breakthrough against [your favorite cryptographic cipher here].
to the biggest model on xhigh.

>>109734051
I'm a big supporter of the furries over the NSA. I just like Microsoft.

>>109734066
20x doesn't give you tac. without tac, safeguards will engage.

>>109734044
what the fuck i'm looking into this and i thought you were trolling but it seems like you're right (for anthropic plans, specifically, not codex)? i cannot fucking believe this. since i had already planned to try the codex 20x this month because everyone was saying astra was incredibly, and i just subbed to the codex 20x, if this is really the case i'm cancelling my claude sub outright unless there's ever a time where anthropic models severely mog the entire competition (i.e. fable-5.5 would be enough of a gap but not fable-sol). wtf/. i'm price insensitive enough to throw $400/mo at this shit to make zero-profit private hobby projects for fun and they still somehow found a way to enrage me into becoming a penny pincher

>yes
>I can vibe code using a GCC compiler on MacOS using a Microsoft product in Linux VMs
>I just prefer this set up so I can get my custom IRC client perfect

File: 1784662686715450.jpg (49 KB, 505x489)

49 KB JPG

What was your first "No way! That's awesome!" moment in your vibecoding journey?

File: MV5BNTY0OTY0Nzc5M15BMl5Ba(...).jpg (103 KB, 1000x1462)

103 KB JPG

haha, you /vcg/ guys are hilarious!

>>109734052 (samefag)
Basically LLMcacas (a subset of AIbros) have started behaving towards the AIbros as a whole in the same way that greater set has been acting towards various professions/hobbyists (artists, actors, mathematicians etc), which is extremely funny.

File: 1762144273321555.png (642 KB, 733x909)

642 KB PNG

Did anyone here make a profit off their vibecoding?

>>109734098
We've had people making "profit" since the shitty Cursor days in 2025. It means nothing and is situational.

>>109734098
I made a profit off of regular coding. Programmers aren't exactly special.

>>109734098
As long as the idea is niche you can make a profit
if you make common shit (like most vibecoders) its over

I have 100k lines of code I enjoy playing with. I don't care if it comes at the decline of society. There's not much else entertaining right now.

>2.5 hours to usage limit reset on Claude but I don't trust GPT with Claude's codebases andd don't have anything for GPT specifically to do right now

Okay fellow poorfags, how do we maximize our plus usage? Is it now Astra medium for planning then Luna max to implement? Also wondering if Astra low is a good replacement to Sol Medium.

>>109734078
told you, claude users are the snailcats of this general

File: .png (254 KB, 1172x1134)

254 KB PNG

>>109733993
got any experience with hydrafusion?
https://github.blog/ai-and-ml/github-copilot/project-hydrafusion-frontier-quality-via-multi-model-orchestration/

File: 1787618049690018.jpg (140 KB, 1024x765)

140 KB JPG

>>109734066
I'll drop my 20x sub once they reinstate the 5 hour limits. It's quite nice to be able to think of an idea for some software, spin up three or four instances telling the highest model on max to build it for you, and coming back the next day to at least one version of the software that is roughly what you want and that actually works. So it's kind of a unique moment that it would be wise not to squander. But you do you

>>109734126
>I have 100k lines of code I enjoy playing with. I don't care if it comes at the decline of society.
It's a like a model train collection that makes normies seethe because they don't understand it and now they can't afford their beep boops

I like building tools more than actually using it seems.
My vibecoded LLM frontend project things is pretty much ready for use, I got subagents, parallel tool calling, lorebooks, character cards, filesystem tools with undo. A very flexible prompt config system where I can order my prompt components how I like and inject whatever prompts I want into chat at any depth. Floating and tiling windows, multiple chats open at one time, subagent windows automatically pop up.. Everything works I think? I would need to actually use this damn thing for a while.
This is the same as with 3d printer stuff, it was fun building and modding a voron, after building it and printing a few cubes its sitting and gathering dust.

Hmm, guess I'm using my frontend to vibecode a TUI client for the project next. Or maybe android app?

File: images.jpg (19 KB, 483x414)

19 KB JPG

>>109734098
Technically, if that counts, I asked Gemini 3 Pro earlier this year to write me a "justification email" for a raise which got accepted by both my boss and the HR roastie and got me a 10k raise
As for all the code I made with codex in my free time, no. Although if you weigh the cost, so far I am +$9,940 (I spent $60 on codex so far, the Gemini email was actually created using the free trial)

>>109734134
I think sparse planninng and reviewing. or as a chat bot sitting between you and other agents to translate, remember instructions, tard wrangling other AIs, but don't let it read code or do things directly
also people say it's token efficient, so astra director and sol medium coder, maybe would work for more more complex works like writing snail-reviewable code

>>109734134
How do you even plan on medium? Whichever model I select in codex, once i switch to plan mode it instantly goes to Max effort

File: 1699585414823710.jpg (23 KB, 258x245)

23 KB JPG

Astra accelerated OpenAI products by 6 months (self claimed)
Mythos/Fable killed Sonnet-Opus as a product

>>109734198
How is Sonnet compared to Luna (I guess they are around the same tier)? I've been using Luna almost exlusively in codex and it's been very good to me so far.
But I consider switching to Anthropic because OpenAI nerfed the web chat version of ChatGPT (which I actually use more frequently especially from my phone) to the point of being unusable. I think in 8 out of 10 cases it's actually gpt-oss-20b answering and not actually Sol.

File: 1540410714518 just fuck m(...).jpg (51 KB, 704x720)

51 KB JPG

Astrafagging on a Plus plan is brootal

>>109734003
>bombing the competition
I can see it happening with how "close" government and top AI companies are now.