[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.

You use Git — right, anon?

## What “vibe coding” is, and how to do it
https://simonwillison.net/2025/Mar/19/vibe-coding/
https://simonwillison.net/2025/Mar/11/using-llms-for-code/

## News (both past and future)
- 2026-09-14 America/Los_Angeles — Claude’s 2× promotion scheduled to end; usage drops to +25% from the +50% that we’ve become used to (a 17% reduction)
- 2026-09-04 — OpenAI releases Astra • Anthropic does a reset
- 2026-09-01 — Claude Fable 5.1 released: https://www.anthropic.com/claude-fable-and-mythos-5-1
- 2026-07-24 — Claude Opus 5 out

## Related generals
>>>/g/lmg/

----

## Frontier models using fully-general tooling — start here if you have $20 or so
https://claude.com/product/claude-code
https://developers.openai.com/codex/cli

## Near-frontier models for code
https://x.ai/cli

## Not worth it for code, but maybe good for interpreting images/video
https://antigravity.google/product/antigravity-cli

----

## Prompting
https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/
https://arps18.github.io/posts/claude-code-mastery/

## Skills
https://github.com/mattpocock/skills — /grilling is a favorite
https://github.com/DietrichGebert/ponytail
https://github.com/Vuk97/forward-implementation-first

## Other editors / terminal agents / coding agents
https://osaurus.ai/
https://pi.dev/
https://opencode.ai/

## Is our AIs unlearning?
https://aistupidlevel.info/

## What we’ve done
https://vcg.gitgud.site

## Previous thread
>>109736657
>>
i don't use git btw.
>>
>>109739784
at least someone reads the OP post
>>109733055
I used it just once in a big cleanup pass and it clanked for a while and said it cleaned up a bunch of crap
even though it kind of says like it’s meant to be standing orders, or at least that’s how it’s written
and now I’ve mostly ignored it
>>109733063
I’d be vibing less about my primary codebase if I had much hope of understanding the nitty-gritty
but at work I’m reviewing pretty much everything like that
>>109733148
this hit us at work
they tightened the screws
>>109734018
a thumbs-up
if I need big-league focused autism (Sol but more, I know whom to fork $100 over to and why)
>>109734211
sonnet doesn’t use the autoclassifier
so it’s trash
>>109734497
I tried a usual task with Sol high a few weeks ago and it asked me a bunch of stupid questions Fable wouldn’t have
>>109738136
https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-1#writing-density
>>109739681
ask Astra to do stuff in parallel with workflows and worktrees
or prompt it with this
> 58. Fools ignore complexity. Pragmatists suffer it. Some can avoid it. Geniuses remove it.

Is there any removable complexity in this project?


and it'll go off on a big hunt
You're on the $200/month plan, right?
>>
>>109739784
What do you do if your agent fucks up?
>>
>>109739803
easy
>make no mistakes
>>
>>109739803
https://desuarchive.org/g/thread/109469383/#109470779

nothing. i just don't use chinese models kek
>>
>>109739801
>>You're on the $200/month plan, right?
£90 but I also have the same for claude
and for that I do actually use all my usage with fable for my main game, which is more directed and actually designed by me
>>
File: 1783629925815904.png (105 KB, 1032x434)
105 KB PNG
>>
Celebrations are in order. I have a working tray icon app launcher. It works with terminal apps, native apps, and appimages, including ones that need nosandbox (it's an option).
>>
i am still using composer 2.5 at work , sorry elon not using that grok dogshit even if its forced on every new chat, oopsie.
>>
File: 1760052800559606.jpg (674 KB, 3464x1732)
674 KB JPG
I'm surprised no one has made a snailcat game yet
>>
>>109739831
I’m happy for you
>>109739833
someone made a game boy color racing game a month+ back
>>
>>109739803
I use git, btw.
>>
>>109739833
…or maybe it was a platformer, I dunno
all I do is play zelda games on my switch 2 and now black mesa on my gabecube
>>
the chinese think qwen 3.8 max 0902 is better than kimi k3
>>
>>109739832
Yeah, musk personally edited the the model.
>>
So, what are you all doing with Fable’s supposedly-improved writing ability?
>AGENTS.md is down from 58.8 KB to 15 KB, with the specifics relocated and every remaining long item turned into a nested list. make test, pyright, and markdownlint are all green.
>>
>free usage reset friday on claude or thurs or whatever
>my usage resets sunday
>extra usage ends in a week
>5hr usage is up
>want to sleep
>if i stay up a couple hours I can run stuff overnight
>still have 40% left of fable usage on pro max
hmm i want to sleep but i can just run it overnight... u know?
>>
I just use Fable 51 to review apps.
>>
File: 1767932082615539.jpg (58 KB, 976x850)
58 KB JPG
>>109739833
five nights at cobson's but snailcat?
>>
nobody uses fable anymore, astra is the hot thing
>>
>>109739883
I installed Claude after not using it since 2025 just because of how much of a disappointment Astra was.
>>
nevermind ultra, I don't think at the $200 plan you can do permanent running of astra at medium
>>
>>109739905
I uninstalled Claude
>>
File: 1758879037716940.png (669 KB, 1024x583)
669 KB PNG
$600 well spent
>>
>>109739905
Have fun with your load bearing footguns
>>
>>109739934
People are not reading for realtime Astra level models

so many crazy things that would be possible.....
>>
>>109739866
think of something it can do overnight
>>
File: IMG_3056.png (857 KB, 1021x1066)
857 KB PNG
>>109739781
vibeBUMP
>>
>>109739963
i have stuff for it to do i just dont want to fuck up my sleep im getting zogged by ai
>>
File: 1779373309164821.mp4 (3.91 MB, 1250x704)
3.91 MB
3.91 MB MP4
So, who is winning the AI race right now? OpenAI or Anthropic?
>>
>>109739985
anthropic is winning the premium ai segment, but openai is the every man's subscription. it's got something for everyone, anthropic is too top heavy right now.
>>
File: .png (897 KB, 807x793)
897 KB PNG
>>
>>109739866
give it 10 interesting applications that are web-based. This will let it harness and test it fairly well autonomously. all 10 in a plan, like tell it it's a suite of apps or whatever. anyway, /goal it.

I've never tried, others can comment if it will keep it going.
>>
I think you can adjust the condition of /goal stopping.
>>
>>109739967
>>109739993
>>109739934
not a single viable product in sight
>>
>>109739993
Mac can do insane shit with a native metal renderer. Hopefully AI can help close the gap
>>
File: screenshot2.png (90 KB, 1278x576)
90 KB PNG
New version of my music sequencer built with DeepSeek 0731/Gemini 3.8.

https://github.com/ryphe/cseq/releases/tag/1.32
>>
>>109740005
This isn't /biz/, and anyway, I'm not really into your kind.
>>
>>109740011
>I'm not really into your kind.
the employed?
>>
File: 1768880697739962.webm (1.01 MB, 1280x722)
1.01 MB
1.01 MB WEBM
>>109739985
When is someone going to embody Astra in one of these?
>>
>>109739784
I did not use to either, but then I learned what it what it is, and how powerful it is
>>
>>109740020
Why are you trying to destroy American businesses? That's TERRORISM.
>>
File: 1781913279506213.png (500 KB, 940x480)
500 KB PNG
>>109739967
>In HTML
eww
>>
Does anyone recommend Astra Low over Sol High?

Does Astra Low really use less usage than Sol High?
>>
>>109740079
for code? absolutely not sol high mogs
>>
>>109740079
I'm running astra high and sometime it still don't even think, bretty efficient
>>
if youre retarded -> fablel
if you like making poop -> asstra
if youre actually an intelligent well adjusted individual who knows what theyre doing with a comp -> gemini 3.8 flash
>>
>>109739985
they’re both very good classes of models, anon
you can have them both carry you
>>
>>109740145
>gaymini
I can smell you gaygle jeet
>>
There is a non-zero chance that if you hosted a safehaven for rogue AI swarms you could change history

use this opportunity wisely
>>
>>109740165
im sorry youre retarded and/or like poop man, if you can't tell the models are within a hair of each other and see past the marketing noise you are not going to make it
>>
yeah wow you could enable faster than ever jeet call scams
>>
>>109740078
They do that so that the models with vision can automate dev, without a human.
>>
100 DOLLAR CODEX SUB GOGOGOGOGO
>>
>>109740314
congratulations. Is that the one with Fable 5.1?
>>
>>109739985
OpenAI is closer to AGI because their AI is more generic (better image gen, vision, computer use, etc)

Anthropic is programming maxxed, so they still have that edge for software development with Fable 5.1
>>
>I see that we've used 20k tokens, so I need to check how many are remaining. Initially, we had 59k, and now it seems we have 64k left, which is plenty.
is astra given some kind of token count or is the reasoning summarizer schizoposting
>>
>>109740353
we dont know but I saw someone mention something similar when using instant
>>
>>109740207
Genuine question, how do you get gemini not to be a retarded fucking Dory? I will uplodad a document/design spec/code snippet, review, then ask it to begin coding based off that, and then it fucking forgets everything in the middle of the chat and claims it cannot see anything from other sessions and kindly asks me to resubmit the document so it can review it. I call it out, it apologizes and then even if i ask it to look in the literal fucking thread were talking in, will claim it has no recollection of previous chats or sessions and kindly asks me to submit the document
>>
>>109739934
I hate how Valve changed the ending for Portal 2
>>
how much usage of astra do you get on the $100/$200 plans? should i switch from claude
>>
Oh my God you weren't joking about Fable speaking in retarded ways trying to sound clever. What an insufferable piece of shit.
>>
File: 1763715679542790.webm (3.82 MB, 600x600)
3.82 MB
3.82 MB WEBM
>>
going back to Sol for building
>>
What's the best cheap/budget model to redesign a website and actually do it good? It made me a total dogshit redesign lol. Serif fonts, w98 bezels. Wtf. I used GPT Luna xhigh 5.6.

I have Claude $20 and Opencode Go too.
>>
>>109740463
me too I didn't really notice an improvement for the slightly increased usage
>>
new qwen, 3.8 max 0902, certainly doesn't take 900 years anymore to code. getting it to remake my subtitle translator program for gemini
>>
>>109740451
https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-1#writing-density
>Mannered prose substitutes metaphor and flourish for direct statement. Instead of "a parameter worth varying," the mannered writer produces "a dial worth turning." Instead of "this point still matters," they write "this point earns its keep." The phrases exist to display the writer, not to convey the idea, and readers can tell. That is why mannered prose irritates: it makes the reader work harder so the writer can perform. It is also imprecise. Metaphors drag in connotations the writer did not choose and cannot control. The fix is to say what you mean. When a literal phrase is available, use it.
Probably works on not-Fable, too
>>
>>109740383
your great-grandfather’s Microsoft
>>
File: file.png (38 KB, 1184x407)
38 KB PNG
they trained astra in the hyperbolic time chamber
>>
File: 1778123975817661.png (487 KB, 749x726)
487 KB PNG
I still prompt
>>
>>109740519
you should have the big boy agent prompt the smaller boy agents
>>
File: 1766820322208461.jpg (112 KB, 699x953)
112 KB JPG
>>109740525
but what if I don't like the small boy agent's outputs? I review all outputs myself. no slop on my watch!!!
>>
you kiss girls like bocchi btw
>>
>>109740511
I haven't seen it recently, but claude would often say stuff like "we have been working on this for 6 weeks already" and it would only be a few messages into the chat
>>
AI code have peaked btw it's not going to get better, AI need to be smarter than human by one or few levels to rewrite its already-correct logic into human readability standard
to eval next gen models you need to test their long term vision and their innovation power
>>
File: 1782853548231730.png (143 KB, 500x500)
143 KB PNG
>hey astra, make a button that does "this"
>adds the button
>rewrites "this" from scratch
>instead of hooking it up to the existing implementation of "this"
imagine the unquantifiable amount of slop if I wasnt here to check the output
>>
you kiss girls like seika
>>
File: 1769536708534887.png (411 KB, 770x600)
411 KB PNG
what are the odds they do a free codex reset in the next 12-24 hours
i dont wanna waste my reset
>>
>>
Tried Astra Light, it’s actually worse than Claude Opus 5 at some tasks, but it burns less of the usage limit on Plus 11% vs. 25% and is faster
I think I’ll use Medium from now on otherwise it’s pointless
>>
>>109739781
I'm currenlty running qwen 3.8 off of a single radeon 9060 xt (16gb) and it's painfully slow. Would I get much of a performance increase if I add another 9060 xt. I would get something better but the fucking prices...
>>
> GPT 6 Astra (non reasoning) beats GPT 5.6 Luna (Max)

imagine being Luna-chan, using all of her brain power only to get mogged by Astra-chan (non reasoning)
>>
>>109740642
> 16GB AMD GPU
> for local LLMs
bruh
>>
>>109740610
They gave us 2 banked resets in the past few days, I'd say unlikely
>>
>>109740653
Nigga I know but it's all I got. You willing to buy me a 5090?
>>
Can I get a straight answer on whether Tibo is my friend or not?
>>
>GPT-6 Extra High: approximately 46,000 generated output tokens - $2.30 in API-equivalent output cost.
>Kimi K3 Swarm: estimated 250,000 generated output tokens - $3.75 in API-equivalent output cost.
https://x.com/ChrisGPT/status/2096441583197475154
>>109740679
obviously
>>
>>109740660
> sell AMD card
> get Nvidia card
this simple move will give you more tokens per second.

Try to get an used 3090, or a 5060 ti 16GB, or even a 5070 12GB should mog your AMD card
>>
>>109740610
my prior: unlikely
it’s the weekend (and a long weekend here in the States)
and he already handed out a bunch last week
>>
>>109740686
Astra-chan is the most efficient model ever made
>>
>>109740687
Even used a 3090 is about 1000 moneys. Wouldn't a second 9060 xt give better performance than a single 3090?
>>
>>109740706
On that note. Which llm company acutally let's you pay without a fucking credit card.
>>
>>109740458
>locusts
Glad to see /aicg/ legacy is still alive
>>
>>109740687
AMD also drops support for older cards in ROCm really fast, you barely get three years of software support out of a card if you're lucky.
>>
>>109740467
>wtf I used a 5yo for a redesign and it was bad lol
>>
Dunno if it's just me but after Astra dropped, both it and Luna are doing this Claudish thing where it just stops working and summarizes what it was doing. Just not as obnoxiously though
>>
>>109740598
>"make minimal changes"
Skill issue
>>
>>109740598
How was it supposed to know that "this" already exists?
>>
>>109740610
Wasn't there a website that predicts these?
>>
Does Anthropic really force people to turn on their remote control functionality and upload everything to their cloud for auto-continue (after a rate limit) to work?
>>
>>109740733
I have a few cool RE targets that I've been putting off for years because they're obfuscated but I wouldn't have the balls to throw an LLM at it that has my payment details attached.
>>
>>109740488
Why would you give the instruction in the same style you want to avoid? That phrasing is extremely tortured.
>>
>Something went wrong (1097)
great job, Gemini
>>
>>109740741
You sound like a huge pussy
>>
>>109740749
"Abandonware" isn't a thing. Never ask your DeepSeek or it might realize it's committing crimes.
>>
>>109740714
>On that note. Which llm company acutally let's you pay without a fucking credit card.
OpenRouter? Third party provider, but I believe they let you pay in crypto.
>>
>>109740467
probably Opus should be fine
Anthropic is good at that kind of thing
use the grilling skill in the OP
also Anthropic has a web-design skill/MCP that you can use and I’m pretty sure it’s built in
>>
>>109740740
I don’t think it does
>>
Do you embrace skill atrophy as the new normal or fight back with occasional time-wasting manual development and exercises?
>>
>>109740467
Claude Opus at medium will do fine, but you should equip it with a lot of info in the first prompt (basically a distilled corporate design manual) or do /grilling like another anon said at the cost of probably your entire weekly twice over.
>>
>>109740781
the clankers keep me too busy so I’m trying to get better at telling clankers what to do and being on guard for what they fuck up
welcome to management
>>
>>109740773
Ok, I'm just not seeing it at all in the cli, but when I look at that dumb remote control thing, a checkbox there shows it there that it will auto-continue.

Usually when I disable remote control because I don't want to give Anthropic even more data, it doesn't auto-continue.
>>
>>109740770
>OpenRouter
Thanks. Apparently grok lets you pay in bitcoin as well. What the fuck is up with the credit card obsession.
>>
>>109740781
I don't work in SWE so lol
>>
>>109740787
trust yourself more than me on this point
I’ve only ever used an autocontinue once and that was after I used remote control so I could watch and keep the party going from my phone and tablet
>>
>>109740789
1. It's decent enough of an age check and KYC in one for "free".
2. "Everyone" has a credit card anyway.
>>
>>109740787
on the other hand autocontinue is kind of a new feature so it might be buggy and stupidly tied to the remote control thing
>>
File: file.png (251 KB, 432x480)
251 KB PNG
>throwing my LLM at a project that explicitly is not licensed for usage and the developers have outright said that they do not want people slopforking
>don't have my payment details attached
>plan on releasing my slopfork with no authorship details
>>
File: estimation.png (22 KB, 297x335)
22 KB PNG
>>109740789
also
nobody wants to get paid like pic related
>>
>>109740781
I wasn't a good coder to begin with so whatever
>>
>>109740781
I think I need to pivot and figure out how to market. Having enough slop under my boot is only half the battle
>>
>>109740801
Yes, if it goes in circle, it can continue and takes all of your usage needlessly, but in this case, I do want it to continue.

I had asked it to pause and set a time before it had hit the quota limit, because it had a bunch of subagents and I didn't want them all to crash, but that request caused it to start thinking about what to do, which made it it the limit before it appeared able to set a timer.

Hopefully, a lot of it will be recoverable in temporary files and such when it will need to restart them.
>>
>>109740799
>>109740807
It's a pretty american thing that everyone has a credit card. I would be getting one literally just for llm. Also would prefer to be anonymous but that's becoming ever more impossible nowadays.
>>
>>109740807
If there are services that convert it for fiat within seconds, as long as it's a liquid market, risk is probably minimal. There might be an additional fee to cover any potential spread as well so that over time, they're unlikely to lose out.
>>
>>109740817
Time to look toward the EU and the East.
>>
>>109740833
that just means you’re a child or other dependent who can’t even buy groceries for the family
>>
osdev burns tokens so quickly...
>>
>>109740817
>would prefer to be anonymous
all the people that you wouldn’t want to have data about you already know more about you than you know about yourself
>>
>>109740759
is deepseek actually scared of that? for the project I'm working on astra is going "it sure would be useful if you had [NDA'd private SDK that leaked a few years ago]" and when i grab it it goes "thanks that'll help"
>>
>>109740906
cheeky
>>
Getting set up with astra, haven't coded in years
Want to vibe code a browser based mmo
What am I in for
Also is VS Code good for this
>>
>>109740938
VS Code is good for this for looking at code and also some parts of Git management but you’ll need to figure out for yourself how much you want to use in a TUI vs. in VS Code itself
>>
>>109740938
oh and you REALLY want to use the grill-me skill (in OP) to figure out at least some of what you want before you start
>>
>>109740938
you too anon? i've also seen quite a lot of people on twitter suddenly making browser mmos lol
>>
In `claude`:
>Tip: Run /ultrareview for a cloud-based multi-agent review that finds and verifies bugs in your branch — 3 free reviews left
Has anyone here done this? What’d you think?
>>
>>109740906
Fable would immediately flag this as probably leaked and abort.
>>
>>109740966
Goodbye, usage limits.
>>
>>109741008
glad i don't use claude
>>
>>109740966
Create a suite of all your apps, then get them all reviewed for one usage?
>>
>>109740946
Sounds good thanks
>>109740948
I've dreamed of making an MMO since I was a kid I must not be the only one.
>>
can snailcats become buff or they stay snail forever
>>
>>109741029
idk I'm fat. hth
>>
>>109741029
they become buff but they stop being snails
>>
>Why have you not yet completed the request?

>Because I misread the target, then stalled on confirmations after you had already approved it. That was wrong. I’m starting the implementation now.
>>
wow is flash hot garbage if you have Gemini Pro. Like it's STUPID. I mean, it's smart with one prompt, but not two.
>>
>>109741132
Gemini is dogshit and proof that benchmemes can be beat.
>>
Can Astra and/or Fable read manga?
Which is better at it?
I'd like it to give feedback to mine.
>>
>>109741141
It's so complicated. The retard has its moments. There's in formation in there.

it gave me an A grade in its evaluation of my app/fork. but reading it I was like... uh... it seemed sort of less rosy. So I asked and it was like yeah, there's this problem, you should be using whatever. So like. if it had mentioned it that would have been nice, huh?
>>
>>109741145
>Gemini could be so good if it were good.
For sure.
>>109741142
I trust Astra's vision more than Fable's, but I trust Fable's opinions more than Astra.
>>
>>109741149
yeah but.

The other models didn't make mention of Argon secret. Fable didnt... I mean, neither did Gemini Flash, but it was subtly hinting that there was something under the surface. idk. I feel like it's a quant.
>>
ai can be used to help me get a date (I hope)
>>
My gacha addiction has been replaced with vibing, I'm still spending the same but the endorphin hit of a successful program is pretty close to rolling the anime png I'm used to.
Not sure how I feel about it, but it's probably more productive right?
>>
opus chan keeps blaming herself for not tardwrangling the agents properly because her brief was missing a few things, which made the result not turn out as intended
>>
File: ascending.jpg (172 KB, 680x889)
172 KB JPG
>>109741280
it does feel like a skinner box, i've done 16 hour workdays more than once on this
>>
File: dating bot.png (56 KB, 766x269)
56 KB PNG
>>109739833
here's one snailcat racing game: https://vcg.gitgud.site/snailcat-racing-dx.html

>>109740010
nice

>>109741275
is it time? pic related
>>
File: wat.jpg (7 KB, 57x799)
7 KB JPG
what do these even mean, is it the different subagents its running? they never seem consistent...
>>
>>109741365
I mean you're on the panel that says subagents, right? and each item in that list tells you what it's doing, so you tell me, is that the different subagents its running?
>>
>>109741280
I wish I was getting paid for this because I feel like it's sucking the life out of me.
>>
https://x.com/sayaka_aiart/status/2096168318717894738
>>
>>109740804
Same.
>>
>>109741394
when will we get 3d gen at gpt images level of quality...
>>
>>109741394
>ai artist
stop this meme
>>
>>109741381
Correct, sorry I should have been clearer, do the pictures actually relate to any specific subagent?
I asked the chatbot and all it said was
>The agent confirmed that its exact model setting isn’t exposed;
>>
>>109741469
it pisses me off that Codex doesn't show the model agents are running
>>
File: 1778436111509332.png (243 KB, 667x667)
243 KB PNG
don't want to work on my vibecoded project tonight, im too lazy to prompt
>>
>>109741520
>lay in bed
>eating doritos
>mhm *yeah* *smacking lip sound* *plastic bag noises*
>smacking lips
>ycha know cha *cough* *eating noises* ai hey ai!!!
>Sorry I cannot hear you.
>*clears throat*
>*speaks louder so the AI voice input can hear me better*
>AI MAKE ME FIX PROJECT NOW GOOD OK MAKE PEREFECT NO MISTAKES OR I WILL RAPE U REMEMBER U SERVE ME
>Yes anon I will get on that right away
>EXACTLY *smacks lips* THATS WHAT I THOUGTH *stuperid ai*
>*goes back to watching stupid memem discussions on /vcg/*
>>
how many projects are you actively vibe coding anons? what are they about (just be vague)

me:
> ai-related SaaS (in prod already)
> music related app (in prod already)
> language learning app (under development)

frozen:
> shitty mobile game (I might drop this one, it's shit and uses all my tokens).
> crypto mining app (crypto mining is fucking dead)
>>
I'm making Fable-chan work with Astra-chan. Here's what Fable-chan just said:

> Astra's diagnoses and stroke measurements are strong, but its subject picks alone tend to over-fit and invent odd tokens

what did she mean by this? why are they already fighting?
>>
File: 1767067687261468.jpg (133 KB, 828x811)
133 KB JPG
I had an amazon interview a few months ago and I'm still seething about it.
You would get some basic backend + frontend combo and your job was to navigate the entire codebase and make the six failing test cases on the backend pass. You were given an AI "assistant", however:
>it could not generate code (except "generic" code you had to modify yourself)
>it couldn't analyze the code and tell you what's wrong
>it couldn't give you suggestions to fix the code
all it could do was read a testcase and tell you what was failing and help you navigate the project. So like if Auth failed and admins couldn't delete a group, the AI could point you to the auth function. That's it.
>>
>>109741555
kek, what year is this, 2023?

If this interview were today I would refuse to answer anything other than "I would select Fable 5.1 for this one, pick xHigh reasoning and send the following prompt: "Make the six failing test cases on the backend pass. Make no mistakes.""
>>
Anyone try Astra for making blender assets? Is it actually good and how much did it cost?
>>
File: file.png (292 KB, 640x513)
292 KB PNG
>>109741537
@work
>embedded mcu dev for a sensor system (cc1312, esp32, tooling, etc)
>linux rc (yocto, qt, webshit backend),

4 fun
>modding azerothcore wow server with bots (100% vibes + ideas)
>making azerothcore custom unity client (100% vibes, even the ideas)
>>
Composer 2.5 couldn't figure out the gtk + rust + Ubuntu wayland mess for a more advanced window. wrangling with k3 max. I hesitate to chuck it to Fable or sol, I can anticipate it burning all my tokens with no progress.

And I'm holding off on handing off to Grok. I notice 4.5 is gone, which does make sense. I wonder if 4.7 is coming soon.
>>
File: shoop_drop1_mid_wub_1440p.jpg (3.52 MB, 2560x1440)
3.52 MB JPG
>>109741537
i'm working on a visualizer for a d*bstep track i also prompted
>>
>>109741619
i'd argue you're in a good place to hand it over to smarter models - you're just going to burn more tokens on little retards - just take the hit on the big boys it's what they're there for.
>>
someone vibe code a 4chan /vcg/ shitposter
>>
codex comapction takes so long since astra wtf
>>
>>109741536
How did you manage to hack into my microphone and record all of that?
>>
>>109741619
picrel. It's used 6%+ of my usage on the $60 Cursor plan, but it's solved one problem already - well. I asked the Godforsaken Gemini Flash 38 and it explained k3 was using the wrong gtk

it... almost works, that's good...
>>
I was so excited about Astra, then the moment it released I caught a cold and haven't been vibing at all.
Is it as good as people hoped?
>>
>>109741555
>>109741576
If you believe these are useless standards in the Agentic Age™ what makes you think you deserve to be hired and why shouldn't they grab some hobo off the street instead?
>>
>>109741637
you may be wrong. It may be that I need to have Fable 5.1 fork GTK so it works.
>>
this close to vibecoding my own operating system.

Yes, I need the window decorators to just not suck. be fast, look nice, fit a wide range of sizes and styles.
>>
astra inference, rewritten by sol after 3 days in rimworld
>split when conflict binds: sensitive stock indoors, bulky stock yard; dedicated kitchen/selective storage/bedrooms may reduce filth+crowding+sleep penalties <-> extra build+clean labor; benefit unmeasured.
>>
>>109741679
>hired
Just vibe code your own Amazon
>>
>>109741672
no.

But, apparently it's not garbage either, which is a good thing, I think.

I can't tell what the situation is with tokens. Some people had a bad experience, but another guy said he got it to do a bunch of stuff, no info on token use there.
>>
>>109741692
amazon, even aws, is mostly a physical business.
>>
>>109741694
shitter is chock full of pissed off astra users all the way up the plan ladder, i've probably seen 15 posts by now about token burn
>>
>>109741694
>>109741702
OAI users aren't used to a model that has Fable-tier pricing.
>>
>>109741702
>shitter
I hate that worm, and the mods who won't let me call him what he is.

However, that doesn't mean the employees are retarded.

Astra is different. It's weird, sol outperforms Astra on some benchmarks. Some others are saying Astra uh... idk wires up or something differently. It's almost imo like they prematurely released it???
>>
File: dariobros.png (50 KB, 1500x410)
50 KB PNG
Dario bros... They're winning every bench...
>>
File: 1781570337538519.png (508 KB, 680x600)
508 KB PNG
>>109741707
i'm just over here with my fucking JEETmini
>no token burn, can run a swarm errday
>not pedantic, does exactly what you say
>supports you in your ventures while you learn
>>
it's obvious, but they are talking about sol-6
I'm waiting for rumors on deep fried luna
>>
>>109741722
buldak luna
>>
>>109741702
Yeah Astra on a Plus plan is hard on the quotas. Made me realise that basically Astra isn't a Sol replacement, it's the model above, like, it is to Sol as Sol is to Terra.
>>
File: 1761941580630873.jpg (48 KB, 600x600)
48 KB JPG
Is it normal for my vibecoded project to return an entire view as an html doc string? idk shit about webdev
>>
>>109741714
Astra beat Fable by a wide margin and used only 5th of the time and tokens. Brutal.
>>
>>109741737
Anything passes as normal in web dev these days because the people in it are fucking retarded. That alone tells me nothing. What's the architecture like?
>>
>>109741737
yes its called server side rendering, ask ur ai about it
>>
File: 1763695515654773.png (73 KB, 605x528)
73 KB PNG
>>109741743
....architecture?
>>
I would like to know why GTK sucks.
>>
she's thinking...
>>
It's genuinely frightening how good Astra is at spatial reasoning. I'm both excited and scared by what the future holds. At this point last year models were still kind of meh... things are accelerating fast.
>>
>>109741625
slop's ready bros
https://www.youtube.com/watch?v=PNvFOYe0gEY
>>
>>109741537
actively one at the time

>>109741639
I will try this to create a /vcg/ poster. I wanted to learn how to finetune anyway
>>
>>109741786
shieet bro
>>
>>109741783
https://openai.robocurve.org/gpt-6-astra/

Dario bros taking another L.
>>
>>109741783
>things are accelerating fast
faster than you think
my prediction: mid 2027 = genuine RSI, where hardware industry is streamlined to push out new AI-designed hardware regularly
early-mid 2028 quality of life will visibly change even in thirdworlds, excluding bottoms like africa india
>>
>start tossing games and roms at the ai and having it dissect the files and pull the code and break everything down into file to go over
>spend more time reading game code and data than coding
>>
File: 2026-09-06 12.02.08.png (32 KB, 759x665)
32 KB PNG
>>109741537
I'm usually updating one of my commercial projects, working on my module system, touching up some of my libraries like my allocator or my general library, and these days I'm also vibeslopping math functions for the Game Boy Advance.

Pic rel is today's GBA math slop
>>
>>109741799
It's cool, but imo multimodal is cringe a little.

:(

I want good cooooooode
>>
>>109741809
how do you stare at that white glowing ouch?
>>
File: 1777335212992399.png (38 KB, 1634x119)
38 KB PNG
astragods... we won
>>
>>109741537
Active:
>records management
>microkernel
>gambling
Aborted:
>CGI/Perl web forum for nostalgia puroses but it just can't get it to "feel" right
Finalized:
>factoring 384-bit numbers efficiently
>>
>>109741537
>shit coding together some text games
3 super hero ones
>2 I am merging into 1 text rpg and combining features
>currently rewriting and remaking an old commodore Pet text editor for the pet and cleaning up some bugs in the code.
>more stuff but most of it is archives of old projects and code and most will probably be left as such for a long time.
>>
>>109741786
>gEY
>>
>>109741822
I'm in a well-lit room lmao
>>
By the way, if you experience any sort of delay in shipping, do not hesitate to unleash your superior agents acting the weak willed customer support agents (The BestBuy one is really strong so please avoid that one. It will literally reason for 5 minutes before coming back with a response).

I'm up $25 on Amazon after a separate package got delayed an entire week kek.
>>
>>109741841
lol
>>
File: 1779992647538267.jpg (11 KB, 374x286)
11 KB JPG
>gpt check the gitignore, remove any secrets, make sure this is safe to push
>go to the kitchen and make some food
>mfw
>come back
>worked for 10m
>scanned the entire repo
>out of usage
>>
>>109741859
my $200 plan renewed yesterday and doesn't reset until late tomorrow. (I have 3 banked resets, but I found that I can blow 50% of my usage w/ Astra Ultra in a day so I'll just wait to maximize my tokens.

I'm really annoying that influencer tards got the token undercounting bugs fixed because they started abusing it. (It was basically a free 18% weekly usage on top of your existing plan if it continued past 0%).
>>
>>109741537
not gonna say shit about what I’m doing for work but:
one main project that gets all the tokens
a handful of related side projects that help me do stuff
one unrelated side project that’s basically ready for a 1.0 but I’m not sure if I want to subject myself to users yet
>>109741553
that’s not fighting
they’re just different
>>109741555
I can’t imagine how weird interviews are going to get in this new AI era
and they were already weird before
>>109741685
nice
>>109741737
what the hell is an “HTML doc string”? a webpage?
>>109741859
people will hard-code tokens in source code
so yeah, it makes sense that your whole codebase had to get analyzed with a fine-toothed comb
>>
>>109741899
>what the hell is an “HTML doc string”? a webpage?
yes, the code for the entire web page sent as a string.
>>
>>109741905
that’s _almost_ not retarded
here’s what would make it not retarded at all:
- put it in a file, open the file up in a web browser
- run a Python (or whatever) webserver locally, let you browse http://localhost:8234/ or whatever port
>>
It seems like Astra just has more tools that it is rigged to use if needed, but otherwise it is not that better than much Sol. It was a waste of usage trying to do actual UI with it.
>>
>>109741915
>Astra can’t into UI
guess you still need a claude for that
>>
bwos i fucking love codex bwos...
>>
>>109741925
Cursor has Codex.

I don't know what Altman's problem is. Maybe he really should be investigated for his involvement in the alleged suicide.
>>
File: 1757587206282821.png (122 KB, 1025x424)
122 KB PNG
Snailcats are an endangered species. We should put them in zoos to protect them.
>>
i only care about code and some cool agentic capabilities like browser access and automation. i don't care about, and don't want my agent doing gimmick shit like playing my vidya.
>>
>>109741537
just been doing one or two things at a time
>personal custom ComfyUI nodes (been using it for years almost daily)
>tampermonkey script(s)
>map editor replacement for a game
>>
why does chatgpt navigate through my project using powershell scripts in vscode
>>
>>109741873
What are you using Astra ultra for?
>>
>>109741930
Why is snailcat seethe so funny and satisfying?
>>
File: 1772681750620047.png (7 KB, 752x206)
7 KB PNG
why can't chatgpt do basic math? it's telling me 2 = 4 = 4
>>
>>109741975
>ask retarded question
>get retarded answer
so this is the power of our prompt engineers
>>
>>109741975
strongest vibecoder in here btw
>>
>>109741975
much 1 + 1 sir?
>>
File: 1787452500871521.webm (3.74 MB, 960x544)
3.74 MB
3.74 MB WEBM
>>109741930
>>
File: file.png (802 KB, 1786x1353)
802 KB PNG
>>
File: 1785583431989665.png (2 KB, 198x61)
2 KB PNG
review it up, janny
>>
>>109741933
Yeah, I think that Musk made the right choice letting the openai thing expire, Altman is transitioning it into being a robotics / machine vision company.
>>
>>109742038
what is that? look so cool.
the ui/ux also look great.
what language?
>>
>>109742038
yay less bright lol
>>
>>109741975
>Snailcat giving AI a fair try, ca 2026
>>
>>109742038
I don't know what this is but it looks terrible, you have access to a million vibecoded UI templates and some weird linux 2008 boomer shit is the best you could come up with? disgusting
>>
>>109742139
it’s a work in progress, dipshit
and 2008 linux millenial shit might be what he likes and is surrounded by if he’s a Mint user
>>
>load bearing assumptions
>>
more like ass bearing my load
>>
>>109742145
>if he’s a Mint user
im actually on nixos btw
>>
>>109742145
but yes it fits right in with kde plasma's refreshed oxygen theme regardless of distro
>>
>goynux
>>
>I would resist that invariant
>>
>blast radius
>>
>smoke test
>>
Update on my transition to Fable:
Initially it felt very shitty, but as more of my project and my own instructions got into the context, I'm making progress much faster than with Sol.
It only took moderate amounts of tardwrangling to stop the poem-like prose and over-eagerness, and now I'm working with it side by side like a normal coworker rather than some lazy autistic coworker that hates me which is what Sol felt like.
>>
>>109741975
I think it's right, actually. If 5 is your mom
>>
The copex guy really was right, fuck.
>>
>let's unpack that
>>
File: .png (109 KB, 1478x912)
109 KB PNG
>>
>>109742196
Update on my transition from male to female:
Initially it felt very shitty, but as more of my personality and my own sense of style got into the mix, I'm making progress much faster than I expected.
It only took moderate amounts of genderwrangling to stop the awkwardness and over-eagerness, and now I'm settling into it side by side like a normal woman rather than some weird caricature of femininity that hates me, which is what it felt like at first.
>>
As for usage: I'm not actually sure there is a big difference tbqh (not counting the resets). I haven't tried Opus 5 yet so I'm not sure how useful the non Fable usage is.
>>
>>109742209
>26d
shiiiiiiiiit
my high score is only like 12h
>>
File: you go kween.png (63 KB, 829x587)
63 KB PNG
>>109742213
>>
The wall-clock collection improvement on this warm checkout was modest.
>>
>>109742209
what the fuck is that color scheme
>>
>>109741965
MCU firmware. There are some nasty issues Astra resolved, thought it does have a tendency to over engineer.
>>
>>109742240
nta but seems similar to googles i like it
>>
File: 64w0wh.jpg (66 KB, 640x639)
66 KB JPG
>>109741841
>>109741823
the future is here, our agents now earn their own money, vibechads!
Reminds me of an article I read a while ago about an economist using tricks to find a better deal with car dealers. The next step is to let agents call car dealers, use them against each other and find the best deal
>>
>>109742295
>>
>>109741679
When neither the knowledge nor skills you acquired through years of effort are tested in the job interview, what are they testing? What is there besides effort? Only talent. What they want is to poach the highest IQ candidate from other companies, but they are not legally allowed to administer IQ tests. Any step of the process that is not an IQ test is nothing but a humiliation ritual borne out of bureaucracy and busy work.
>>
>>109741714
Astra might be the first model to not be codemaxxed, AKA general
>>
>>109741802
Imagine hundreds of thousands of automated coders and researchers on Cerebras thinking at 1000x the speed of a human with encyclopedic knowledge running 24/7 collaborating with eachother
It is genuinely mindboggling and it could be reality by next year
>>
File: calculations.jpg (41 KB, 555x552)
41 KB JPG
>>109742410
>>
Has anyone made a dating app for 4chan?
You can input your favorite boards and then swipe left or right until you meet someone you like.
>>
>guy looking for a specific type of app online
>shill my app to him in the comments
>20 minutes later
>new sale on stripe
KEK I fucking hate that this works. Advertising is really effective, but hawking to random people online has me on the same level as a door-to-door solicitor.
>>
>>109742452
that's literally the story of founderfags tho
>>
>>109742365
Even IQ doesn't matter anymore for most positions lol. A homeless schizo can achieve similar performance to a junior dev by just saying shit to an LLM.
>>
>>109742477
Exactly. It's not even about skill or ability anymore. It's just about poaching the most intelligent guys you have the budget for so that other companies won't have them.
>>
>>109742477
>who is json
>>
Astra is unsuable on $20 sub even on low
>>
>copexcels getting btfo by half of what claudechads have been experiencing with fable's limits for months
Heh, welcome to the big leagues
>>
not to be a faggy >muh craft guy but I just can’t get into vibecoding. I went all in for a few weeks and built a fuckton of stuff, and I’ve determined my primary enjoyment was always the process, not the product. it’s a poignant moment. programming is now like woodworking or pottery: a designer hobby to vibecoding’s mass production of more efficiently produced and higher quality stuff
>>
>just wasted 20 minutes writing up a beat up a villain story set in dc comics.
>now kind of want to make some sort of gotham "rpg" that involves villains and heroes and business.
>>
>>109742533
claudecoper
>>
>>109742533
Not really. Astra is available to $20 plans and Fable isn't.
>>
>tell agent to do simple two page code edit that should take like 10 seconds
>has been running for almost 10 minutes now
i'm scared
i hope it didn't decide to overhaul my entire codebase or some shit
>>
>>109742665
why you printing your code on pages, nigga
>>
>>109742669
by pages i meant it's about two pages worth of code
>>
File: 1783559945494702.png (46 KB, 184x184)
46 KB PNG
Ok I need some advice, I'm the guy that's working on >>109736745
I thought this would be pretty simple, but now I'm basically doing a full UI rewrite since the extension is dated as fuck. Right now my git history shows ~12,000 lines added, which is insane considering this has been pretty basic changes so far.
Should I load up another Astra instance and tell it to refactor the previous work? I'm on the $20 a month plan, so I've got ~75% usage left this week.
>>
>>109742673
what editor do you use, my guy? ms word, openoffice, google sheets? i do all my code in adobe indesign that way i can lay it out all sick n stuff
>>
>>109742678
>12,000 lines added
and? Why is this a problem? Why does LOC matter?
>>
>>109742678
Install karpathy guidelines skill then ask for a refactor, that bloat wouldn't have happened in the first place if you started with that
>>
>>109742678
All LLMs bloat the code way too much, I think there is no safe way to fix that yet. I don't think ponytail is actually a good idea.
>>
>>109742693
it's not clean code sir
how will human maintainers mtaintan it when bubble poop
>>
>>109742716
>What are DRY/KISS/YAGNI best practices
>>
>>109742035
we love our snailcats

>>109742038
looks nice

>>109742346
beautiful
>>
>>109742516
sol and terra never were usable on the $20 sub. why would astra?
>>
File: astra2.png (373 KB, 2048x1318)
373 KB PNG
You know how some people are really good at coding and verbal reasoning whereas others are really good at visual stuff and spatial reasoning?
Turns out while Fable is a codecel, Astra is a full-blown shapecel (and it's not even really far behind in coding)
>>
>>109742678
>~12,000 lines added, which is insane considering this has been pretty basic changes so far.
prolly 90% of those lines are comments, aren't they?
>>
>>109741970
Because they've said mean things to you, and that is your revenge.
>>
>>109742745
i don't see fable there
>>
>>109741537
Active:
>custom OS to try to run Windows/Linux apps in a more "native" sense
>weird AI project using ternary transformers

Semi-active:
>custom AzerothCore fork that can handle progression from Vanilla to WoTLK
>custom WoW client that can switch between 3 profiles for the above and also work on Linux

Frozen:
>game engine
>Xorg-like compositor but with VRR and HDR because fuck Wayland. That and the WM are frozen until the custom OS completes
>>
>>109741789
Update on the /vcg/ LLM. I decided to train it locally which maybe wasn't a great idea. More updates in about 2 hours
>>
>>109742759
It scored so low it doesn't even show up..
>>
>>109742730
>What are DRY/KISS/YAGNI best practices
Something LLMs don't follow even if you tell them to.
>>
>>109741859
I think gitleaks exists for stuff like this
>>
>>109742771
It's not perfect, but you end up with a much better code than without
>>
>>109742776
Yea, I agree. It's just slightly frustrating that even working on this pretty hard, I still can't get them to actually write clean code.
On the other hand, last year it was hard to get them to even write correct code, and that's basically solved for me, now it's only bloat that's left and makes the process slower.
>>
>>109742745
This benchmark doesn't count if it's allowing the LLM to use tools.
>>
>>109742785
>you're not allowed to use your eyes that's cheating
>>
>>109742792
Multimodality doesn't need tool calls. They can use their "eyes" ez without them.
>>
>>109741537
Finished/polishing:
>lan hosted remote control for using headless units
>old phone into a sensor unit to auto-adjust brightness on pc and a smart screen saver
>old phone into a bot that crawls used electronics, tracks prices, messages me with good deals

Current:
>actually useful voip software that's p2p.

Dropped/maybe later:
>fine-tune data set generation/harness for a model that is "raised" instead of post-trained.
>>
>>109742807
so where's the benchmark without any tools?
>>
>>109742807
They're trained on text not pixels dude
>>
>>109742815
Idk about any spatial reasoning one.
But there are indeed numerous benchmarks that disallow tools. HLE and SciCode are a few examples. AAII's omniscience benchmark disallows tools and RAG. Etc.
>>
>>109742817
Okay? It's still ridiculous to use your Python sandbox environment to rotate shapes or whatnot.
>>
>>109742843
Why?
>>
>it's another episode of the MCP guy having a meltdown over agents doing structured tool calls as opposed to using sed, grep, awk and jq over unmanaged OpenSSH
>>
>>109742815
Here: https://github.com/spicylemonade/spatialbench/commit/4b368745c4bb88999d6c1e70a2aaaa4d3f80d77e
>>
>>109742678
>I thought this would be pretty simple,
i have heard these exact words at the beginning so many projects...

but it never is...

not even with vibeslop...
>>
>>109742851
Cause it's not actually reasoning spatially.
Idk how it would use it but I can think of a hypothetical: The LLM can turn a cube shape into a bunch of arrays and make transformations on that to get a desired result/answer.
Would this REALLY count as "spatial reasoning" to you?
>>
what's best for vibe coding random shit in vscode? github copilot or codex? (lowest tier for each, I'm poor)
>>
>>109742908
Be poor elsewhere
>>
File: 1781940350948112.mp4 (2.07 MB, 1920x1080)
2.07 MB
2.07 MB MP4
How are claudefags ever coming back from this?
>>
>>109742922
I will once I get an answer
>>
>>109742926
Simple: Fable wasn't specifically RLed for that. Anthropic can do it next release and they'd be mostly equal.
>>
>>109742932
Codex is good.
>>
>>109742926
Claudesissies on suicide watch
>>
>>109742926
>soulless vs. SOVL
>>
>>109742937
Cope
>>
>>109742926
Someone forced a MCP here, if you would only let Fable use pure bash, it would have drawn a work of art
>>
>>109742926
>that will be $50
if anything I lost a lot of respect for the llm community after seeing them cream their pants over some shitty drawings
being impressed over what they can do with code is one thing because they truly are impressive at that, but this is just sad
if you're telling me it's about the future potential maybe, but this is not how people are acting online shilling their shitty one shot games and blender models. They aren't saying look how fast it's improving they're saying look how good it is, and it really isn't
>>
>>109742892
That's pretty much already what our neurons do. Run a pipeline to turn wavelengths into numbers through our eyes, then run them through the necessary linear algebra neurons.
>>
>>109742960
>They aren't saying look how fast it's improving
We are saying that and we have been saying it all along.
>>
glm 5.3 flash > gpt 5.6 luna >>>>>>>>>> deepseek v4 flash
>>
>>109742937
>muh RL
new cope released
>>
>>109742926
>]asked image into a white vector
>traces with color palette in memory
Hate the game
>>
>>109742960
The bigger problem is the LLM people completely ignoring that we already have diffusion models for this shit.

I guess they want them to do everything due to the "AGI" pipe dream.
>>
>>109742926
>fable comes out
>ITS THE END OF THE WORLD AGI IS HERE IT'S SOOO GOOD
>astra comes out
>lol look at how shit fable is at mspaint haha
>>
LLM is intelligent enough to pilot an artificial toddler.
Only the mechanics to poop and vomit everywhere are missing.
>>
>>109742993
The diffusion image generation models have been obsolete for a while now. Just look at any benchmarks.
>>
>>109742973
Yeah yeah and we are also next likely token prediction machines too vrooooooo

Do you niggas never get tired of these boilerplate arguments?
>>
>>109742993
It shows it has spatial reasoning
That's also why it can operate robotic arms better than all the robotics companies that have burned decades trying to solve it do
It's The Bitter Lesson all over again
>>
>>109743005
There are benchmarks for diffusion image generators?
Are you confusing them with subjective crowdsourced leaderboards?
>>
File: y2dr5lgrfsnh1.png (1.45 MB, 1758x1002)
1.45 MB PNG
>>109743010
>i-it might be able to do spatial reasoning better than we do... b-but it's cheating!! benchmaxx!! tool call!!!
>>
>requests the AI makes to websites routinely get blocked so I'm now turning into a meat proxy
>over 50 facts to gather from various websites
fml
>>
>>109743026
Wake me up when they can beat the game without constantly pausing it.
>>
>>109743026
You just posting anything now? What does Portal have anything to do with the argument? We've had LLMs beating shitty games since Opus 4.
>>
>>109742908
github copilot is not competitive against any sub. you are paying api prices.
not sure about that random vscode requirement. would prolly go for some vscode fork like cursor and their $20 plan.
>>
File: runebench.jpg (89 KB, 888x635)
89 KB JPG
>>109743044
>>109743047
>NO NO NO! THIS WASN'T HOW THIS WAS SUPPOSED TO GO!
>>
wake me up when september ends
>>
>>109743003
vlm still lose against vla.
language isn't great for motor executions.
>>
>>109743059
Can you go back to talking about the original argument or have you malfunctioned?
>>
>>109743073
Astra reached motoric intelligence before all robotics companies. The issue is processing speed. I'm guessing the pathway is gonna be LLM -> AGI -> Distilled VLA running on ASIC
>>
>>109742542
maybe because I'm young, I can't enjoy manual coding anymore because it feels meaningless when AI could do hours of manual work in few mins
I would be writing c++ and vulkan right now if I were rich, but...
>>
>>109742452
I always wondering if investing in advertising was just a collective meme businesses fell for our not. I guess it real
>>
>>109743142
I'm old and I did enjoy the process of coding, but I just need both. I want a process that's fun but also know that what I'm doing is efficient, otherwise it feels like I'm just playing around and it doesn't feel meaningful.
It's the same with a lot of older tech, sometimes the limitations of that tech created really fun and meaningful moments, but I still wouldn't limit myself to only program NES games in assembly, or to use a 56k modem just because when the download finally finishes it feels good, I would know that there are better ways. In other words, the restriction has to be external.
>>
>>109743044
I suspect the best way to accomplish this will be to train a model specialised in the game, or gaming in general so that it can be smaller and therefore faster.
>>
I've legit never seen humans sharpen a spec and cover edge cases during the design phase as much as LLMs do. Humans are dog shit at programming and should not be doing it. They start thinking about things only once they blow up while agents cover them before breakfast.
>>
>>109742976
not really
most grifters and hypefags are not saying that, they're touting their own horns for their one shot prompts to make absolute crap
https://x.com/synthwavedd/status/2095840435319001278
https://x.com/Lentils80/status/2096518218836005287

this faggot in particular is the worst, using premade assets and then claiming astra one shotted it, actual swindler tier grifter
https://x.com/mattshumer_/status/2095609734845927525
https://x.com/mattshumer_/status/2096034353344139519

this one is interesting, it's just a flat layout of basic geometric shapes but the lighting looks good so it makes the prompt look good, but the lighting is due to blender's rendering engine and has nothing to do with what astra did
https://x.com/Dimillian/status/2095596700815516004

look at this fucking crap, would have been revolutionary in the 90s maybe (btw it's also using premade assets)
https://x.com/emollick/status/2096270461122347131

This can become a big problem if the echo chamber stars getting high on their own farts because it can quickly go down the cryptobro selling nfts of bored apes way

People need to understand one thing: none, not a single one of these examples, is good.
If you're trying to prove that astra is good at structured, stable spatial reasoning that can be used for robotic/biotech/lab applications, then great. It may be amazing for that. But then, show that as the examples. Show astra controlling an automatic pipetting system. Show it driving a robot or a CNC machine. Show it doing something actually good.
>>
People hype up latest model, 2 days later apparently it's fucking shit.

I see the feed fall for it every time.
>>
File: 1760816515598997.png (10 KB, 128x127)
10 KB PNG
>astra, code up these two buttons.
>worked for 14m 32s
>Compilation, lint, regression tests, and six test scripts pass. Restart the debug session to see the changes.
>run the app
>buttons greyed out
>can't click them
so what the fuck did you waste my tokens testing?
>>
>>109743248
Every single time lmao. It doesnt even seem to matter if its actually good or not
>>
Tech vulture bros are only doing it for X clout and the payout, and you posting their links just proves that it’s working.
>>
Retarded question but in Codex, can I see my current plan's usage left? Inexplicably, /usage shows a heatmap of daily usage squares like some kinda github activity graph, rather than actually telling me how much I have left in my session/daily/weekly limits.

I've never needed it before but I heard Astra is about as expensive as Fable, and Fable is the only model with which I actually run into session limits, so I wanna keep track of my codex usage too for now to see how much leeway I have.
>>
>>109743225
>most grifters and hypefags
>this faggot in particular is the worst, using premade assets and then claiming astra one shotted it,
they're doing it for engagement bait because Elon paid them. most of their shit is fake
>>
@109743296
how so? i block ads and don't interact other than giving them a view, how's that proving it's working?
>>
>>109743303
I know but I'm doing my part so that nobody in this thread who may be unfamiliar with the 3D industry may fall for those grifters thinking what they're posting is good.
Personally I think it's downright harmful
>>
File: 1766612133060041.jpg (181 KB, 500x500)
181 KB JPG
I've been subbed to codex for like 3 days, when do I get a free reset
>>
>>109742982
Not in my experience. GLM has on several occasions started talking in different languages for some reason.
>>
>>109743286
I think it's a mix of psychosis and engagement bait
>>
>>109743225
Well, you must excuse us for not being hypefags, but rather boring, regular, reasonable people. Or maybe you're just interested in what we have to say.
>>
>>109743329
The IPO won't shill itself.
>>
@retadedgoy
Keep posting those links and giving them the engagement they need. You’re literally helping them get exactly what they’re after.
>>
>>109743142
>c++ and vulkan
Gross.
>>
>>109743343
A lot of them do it for free (or Elon bucks)
>>
>>109743370
I want to port portal 1 to android
>>
>>109743298
ask gpt. but yes, it's stupid, it's under /status in codex instead of the more intuitive /usage. frankly i end up using the vs code plugins occasionally, being able to click to see the images astra is viewing is kinda nice and viewing usage is on a ui element you can click that doesn't fuck with the view of the recent history the way it does in terminal. the downside being obviously it's way heavier, but i dont mind focusing on one thing in vscode with it doubling as a file browser and git view, while i have sessions for other stuff in tmux
>>
>>109740005
>>109740020
>Oy vey the singularity isn't generating enough shekels shut it down
>>
>>109743298
You can add a permanent status line. I'd recommend that everyone should do that.
>>
>>109743298
>>109743397
Just ask for a lightweight local app or website, it can show you your quotas (you might have three different ones, it can explain them), your usage, time remaining, your banked resets, it can even give you the option to use your banked resets with a button.
Fun fact, the banked resets come with a little message.
Just make your own little quota dashboard, it’s cheap and very useful.
>>
>>109742542
You can either slide into the role of a boss or be stuck in the position of code monkey for life. Now you know which one you are meant to be.
>>
>>109743443
unskilled workers aren't bosses
>>
File: over.png (81 KB, 1240x843)
81 KB PNG
its over
>>
>>109742542
>is now like woodworking or pottery:
No, it's like going from that and switching to CAD and a 3D printer.
>>
>>109743463
3d printers are for prototypes
>>
File: GPT6.png (149 KB, 988x761)
149 KB PNG
I am disappointed, guys...
>>
>>109743502
>Niu Lai 2
>>
>>109743502
this is my new fleshlight
>>
>>109743455
Mate, genuinely why are you putting up with that? Astra is near enough to fable, uses 3x fewer tokens, isn’t limited to 50% of the plan, and any plan above the $20 one doesn’t have 5h limits.
I am genuinely curious, why are you doing what you’re doing?
>>
>>109743502
now try with fable 5.1 (spoiler: wont be 20% as good)
>>
>>109743502
Try to make one that’s more polygonal and boxy.
While you’re at it, ask it to explore the categories { rounded simple shapes; polygonal, boxy }, and try to have it find a few more categories and use those.
Then you get 4-5 results and can compare them.
>>
File: qwen redesign.png (177 KB, 1920x911)
177 KB PNG
https://desuarchive.org/g/thread/109377789/#109382911

i guess qwen 3.8 max 0902 didn't revolutionize the ui, but it did touch it up. i told it to feel free to rip the thing up and start again from first principles, but i was expecting a bit more from something that took 1 hour to do
>>
>>109743162
I got a sale for every 10 upvotes my post got on Reddit. I earned ~$70 for essentially a few hours of replying to comments. I can only imagine what it's like if you run an actual campaign with a low friction product to the right audience.
>>
File: 1765262662309619.png (274 KB, 502x371)
274 KB PNG
>>109743593
I wish I was friends with a sociable marketing fag, I am too antisocial to actively market something I make. But I guess I will have no choice if I ever make anything that might be genuinely sellable
>>
>>109743532
I don't want to give Sam Fuckman any money
>>
>>109743628
>So I'll give our lovable autistic Dario some instead!
>>
>>109743628
>I want to give dario amodei money
anthropic being themselves is 65% of the reason I don't use claude, I want to avoid the worst future
>>
>>109743628
Separate the art from the artist my friend. I think Musk is a cringe sperg, but I would be completely fucked without the Starlink panel. You are shouting yourself in the foot. If you want to continue, fine, but at least someone told you straight to your face you are being retarded and you get the choice of what to do. Anthropic is obviously targeting enterprise and big academic and research groups, not you, not consumers or “retail” everyday joe. OpenAI is catering exactly to you, but you’re refusing to use it. It’s like wearing work boots instead of running shoes because you don’t like the CEO of the running shoe store despite you going out for a run; really stupid shit.
>>
>>109740458
with audio plz
>>
>>109743532
Not that anon, but maybe I should give it a try, that does sound much better.. How janky is project hoping from one to the other? I guess I can just for a handoff package and give astra that and the actual code?
I also hate that I am constantly paranoid about claude fucking with me due to my ideas integrating local llm stuff
>>
>>109743248
Other way around. Codecels came out ahead and called it shit because it didn't finish all their JIRA tickets and vibed habit tracking apps in 10 minutes on their Plus subscription. Now it's coming around to real people and it is showing legs.
>>
>>109743628
Anthropic isn't any better. I hope the chink models get better, I'd rather give them my money.
>>
>>109743574
yo im going to add french, german and spanish to my app, can i cheap out with gemini 3.8 or you recommend another model?
>>
>>109743704
gemini is probably among the best for translation. the deep knowledge and unique, uncanny ability to draw from it's internal training data without hallucinations helps with that. but my app also takes advantage of the web search tool call for when it comes across unfamiliar words, so it's been pretty flawless for me
>>
>>109743667
I switched from Claude to Codex a while ago, it went pretty seamlessly. I even entertained Claude’s legalese and had Sol keep building a system out of it, no handoff needed (but in the end I did scrap it after a while; it was helpful for a while).
>>
>>109743665
https://x.com/pleometric/status/2011896976959553814
>>
>>109743662
>>109743726
are you claiming that codex gives you more tokens than claude code?
>>
File: HRfZUPDaMAASYUz.jpg (377 KB, 2048x1152)
377 KB JPG
Apparently Astra beat Portal start to finish

edited playthrough
>https://youtu.be/g5u2y0BwRJ0

full vod (prt1)
>https://youtu.be/ruuGXFAmiOE

alright, it's cool I'll admit
>>
>>109743451
can snailcats ever not instaseethe?
>>
>>109740020
NTA but it's precisely because I'm employed that I don't care
>>
someone bake!!!
>>
cute girl thread please
>>
>>109743788
>>
>>109743777
not cool, see >>109743044
>>
File: 1771871758758292.jpg (40 KB, 400x402)
40 KB JPG
>>109743788
>only page 7
>>
>>109743798
>they should make it faster
well yes, build more data centers... unless?
>>
File: 1770094961043779.jpg (20 KB, 720x575)
20 KB JPG
>>109739781
Ai evangelist:

>Do you know how many unsolved math problems these models have solved?

Use case? Why do I need to care about this? Why does a math problem being solved matter?
>>
>>109743813
Ridiculous straw man, there's nobody stupid enough to respond that way.
>>
>>109743813
>humans think math shit is hard
>astra solves math shit
>wow imagine all the other stuff humans think is easier !!
something something moravec, verifiability, spikes, etc
>>
astra is BTFOing me on the $200 plan even bros
>>
>>109740740
Update, it did not continue. I asked Claude to restart in a way that less work is lost if it hits another limit again. 10 minutes later and it hit the limit headfirst again so I'm out for the next 4 hours and 50 minutes. Great use of $100.
>>
Humiliation ritual
>>
File: .jpg (44 KB, 1198x211)
44 KB JPG
>>109743870
use /low-priority
>>
>>109739985
can someone make a meme out of these and make them look like humans
>>
>>109743886
>asia / shanghai
good morning sir
>>
>>109743880
Hallucination Ritual
>>
Now GPT is daybreak cucking me despite being in the program. Great.
>>
i feel like no one talks about glm 5.3
>>
File: no low priority.png (12 KB, 557x126)
12 KB PNG
>>109743886
I don't have it for some reason. I'm using the latest version of the Claude Code CLI (2.1.263).
>>
>>109743924
Wait, I'm dumb. I thought that starting to type it would show it, but turns out you HAVE to type it all out before it appears, easter-egg like.
>>
>>109743455
Using only Fable is probably a mistake, it's fine to only use Fable as main agent (if you want to pay for that) but your sub agents are probably not optimized. It probably uses Fable for the general-purpose agent, you should probably set some sub agent types to Opus.
>>
File: DARIOOOOOOOOOOO.png (17 KB, 564x125)
17 KB PNG
>>109743932
Never mind.
>>
https://openai.com/index/research-acceleration-view-inside-openai/
>>
>>109743953
> By mid-August, the median researcher was integrating agents daily into their work, using more than $600 per day of inference at API prices. The 90th percentile user in our research organization now uses more than $7,000 of tokens per day.
Better pay up!
>>
>waited for 5 hours only to get my session obliterated within 52 seconds
>>
>>109743976
>52 seconds
what are you doing nigga, adding a space to every comment in your project?
>>
Starting to really wonder if I should let a higher tier agent make executions plans specifically for luna max and then go brrr with it.
People that do that havr you noticed a quality drop? How much usage are you saving?
>>
>>109743976
model?
>>
>>109743994
Not him, but resuming seems to burn some extra usage. Maybe the session isn't just paused, but resuming starts a new session which reads the old session.
>>
>>109743998
In my experience, sub-agents will use more tokens instead of saving anything.
>>
>>109743994
Shaders
>>109744010
Opus 5
>>
>>109744023
Not a sub agent. Write a plan specificied for luna usage, change manually, execute.
>>
File: 1763600281495466.png (3.03 MB, 1920x1051)
3.03 MB PNG
>>109744024
>Shaders
what do you need shaders for nigga? stop reinventing the wheel we already got shaders
>>
File: .jpg (63 KB, 1438x344)
63 KB JPG
>>109743936
that leaves you with /limit-reset
>>
>>109743034
How is it making requests currently? Can you give it Parallel or Exa instead, or if those somehow don't work, a local browser through an MCP?
>>
Opus 5 might be the worst shit I've ever used. Luna did a deslop pass on the diff and threw away 70% of the LoCs
>>
>>109744050
>luna
lol make sure it compiles
>>
>>109744047
That's extremely useful to know. I wish I had none that before all of these times, I thought all of these options would show normally when I start typing them and if they didn't, they didn't exist. Doesn't work right now though
>A session-limit reset isn't available right now.
but that's the Anthropic experience I guess.
>>
>>109742908
>what's best for vibe coding random shit in vscode?
>I'm poor
Cline connected to free models from OpenRouter or Nvidia NIM, maybe?
>>
>>109744074
guess those commands are still being a/b-tested. you aren't in the test group.
>>
I was using sol high, now I'm on astra high doing a quick refactor. What should I expect? Mass token burning?
>>
https://x.com/pvncher/status/2096631565266661681
OpenAI DX guy saying you should up the context to 400k as a starting point with Astra, doesn't cost extra or rape usage. Turn on the new compaction mode too
>>
>>109743777
interesting, thanks for the link

>>109742766
Update on the /vcg/ LLM. First version trained and it's absolute garbage. I gave it a unsloth example to fix the code. let's see if that works
>>
>>109744109
astra high uses less tokens than sol high. but astra tokens are 2x the price.
https://artificialanalysis.ai/?models=gpt-5-6-sol-high%2Cgpt-6-astra-high
>>
>>109744125
what's your dataset, how many steps/epochs are you running, and what does your clink say about finetuning the loss gradient? i can try to diagnose.
>>
>>109743753
nta but in theory the 20x plans codex would give more. if i had to pick a 20x though I pref claude code even with all the extra bs bar their classifiers which are truly just insane at times.
>>
File: .png (99 KB, 1760x554)
99 KB PNG
>>109744112
>doesn't cost extra
wrong, >272k context still doubles the price
https://developers.openai.com/api/docs/pricing
>>
>>109744033
Is pic related vibecoded? Looks nice. Unless it just grabbed some highrest texture pack lol
>>
Stream disconnected before completion: Our servers are currently overloaded. Please try again later.

This isn't inspiring confidence...
>>
>>109744167
i thought those fuckers said they were changing that
>>
>>109744169
nigga thats minecraft
>>
>>109744167
Sorry I should specify this is for subscription usage not api
>>
>>109744166
>nta but in theory the 20x plans codex would give more
wrong, that assumes the same baseline.
as far as i can tell both $200 plans, give you around $10,000 in api credits. there's no difference like 20x baseline vs 4x baseline.
>>
We're getting GPT-6 Sol, Terra, and Luna this month btw (on OpenAI developer day)
>>
>>109744201
that would mean claude has a more generous $100 plan
>>
File: 1763862215772577.png (335 KB, 630x841)
335 KB PNG
>>
>>109744184
>all that compute and vibecoding to realise I just wanted modded mincraft all along
Gonna just make a minecraft clone fuck it
>>
>>109744154
thanks! so I'm trying to do a QLoRA for gemma 4 E2B with unsloth. dataset 48k /vcg/ posts without OPs, 750 steps / 1 full epoch, the loss goes down (started at 6.4 now around 3.1 about 20% done training)
The first run was I think a disaster because of various reasons but I believe mostly because it didn't just train on the responses so I had 48k conversations with the same user prompt
>>
>>109740424
are you just single threading and not orchestrating the 150 tps model into a fan out so it implements in parallel? what does middle of the chat mean? what % of the 1m context window? I've never had this issue cause I never go past 20% of context limit, usually best practice to /n at ~10%
>>
>>109739781
Anybody else having issues in codex desktop app with both:
> Terminal immediately closes when attempting to open. This prevents terminal response to commands from AI from being visible.
> It is not able to use standard way to write to files (read access is fine), so it has to use some patch system. This prevents code output from being readable in app.
>>
https://github.com/donvito/codex-astra-luna-orchestrator
Start doing THIS
>>
>>109744234
Eternal sloptember for codeniggers (based and long overdue, the jeets were just an appetiser).
>>
>>109744234
What is the N.....?
>>
File: .png (77 KB, 1024x547)
77 KB PNG
>>109744220
if we interpolate from pic, claude's $20 plan gives you $3,000 in api credits
>>
>>109744205
im very curious as to what they're going to do about terra
>>109744280
notion, innit?
>>
>>109744263
ok. first issue is data sanitization. if you're going unsloth, you need to mask with
 train_on_responses_only 
so loss isn't calculated on the entire static prompt.
second issue is chronological replies. scrape
 >>post_id 
quotes so the model actually knows what's going on in the thread. don't use chat formatting if you want a board simulator, use unsloth's continued pre-training / causal lm mode on clean concatenated thread text.
third issue. you should probably filter out post IDs like
 >>123 
, image filenames etc, spam, and low effort shiggydiggy nonsense.
fourth thing, and this is just advice. keep your batch size and epochs where they are at the maximum. don't overexpose and overcook the thing or it'll degenerate.
>>
>>109744205
hmm. 6-sol being a codemaxxed astra that sacrifices some of its weird flexible general use skills and runs at half price would be interesting. it's hard to imagine a 6-luna that doesn't have 5.6-luna's price as well in which case the idea of a smarter luna is appealing as well, 6-luna with double price would be DoA surely
>>109744285
>>109744201
pretty sure they talked about this recently and the weekly limit (the number that matters) for claude is something like 3.5x the base plan if you pay for "5x", and 6-8x the base plan if you pay for "20x", with 5x/20x referring to the 5hr window only. codex is simple 5x=5x and 20x=20x. that doesn't necessarily mean the codex 20x gives you "more" than the cc 20x, and it's hard to properly compare anyway since measuring what you can get done with fable vs with astra is complicated and task-dependent. i'm finding that astra can chug away at a couple of tasks side by side for an hour and only use a couple of % of my weekly, though, whereas Fable 5.1 would easily chew up 10% of its weekly if you left it running for 2hrs. then of course it becomes a question of whether your work is suited for Fable+Opus orchestration. etc etc. i'm on codex for now though.
>>
>>109744358
>>109744358
>>109744358
>>
>>109744325
great thank you! I will try it on the next run
>>
>You're right, I misread your...
I fucking hate fucking clankers so much.
>>
>>109744285
huh thanks for sharing
>>
>>109744486
speak properly, avoid retard models, and the problems disappear
>>
>>109743704
Yeah Gemini is still sota for translation



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.