[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]


File: 1763159904738351.png (362 KB, 837x983)
362 KB PNG
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
>>109739781
>>
i do not use git
>>
It's hard to put into words just how hard Europe fucked themselves.

Share of global AI compute
U.S. ~78–80% · China ~10–12% · Europe ~4–5% · ROW ~5–7%.

An American lab will be the first achieve meaningful RSI. China will probably be the only other country with a seat at the table when it comes to alignment talks.
>>
>>109744371
Mercurial?
>>
>>109744384
nope. i just zip it to a flash drive/ext hdd, and then i tell the clanker "back up each and every file before you change it."
>>
>>109744384
SVN + Beyond Compare
>>
why would I need git, I told the clanker to make no mistakes
>>
>>109744404
Very funny anon. Very funny
>>
>>109744404
git isn't a backup. clanker can delete your git repo, too.
>>
File: 1781920586769428.webm (3.86 MB, 1036x450)
3.86 MB
3.86 MB WEBM
vibe accelerated snailcatting
>>
>>109744371
I clapped
>>
>>109744373
>An American lab will be the first achieve meaningful RSI
??
what does that have to do with compute. google had the most compute by far and see where they are now.
>>
>>109744373
Fuck the EU. Also is china really that much lower than US, that surprises me considering they are doing decently. Though I guess US uses a lot of its compute for other stuff?
>>
my ass cheeks clapped
>>
>>109744373
you cannot compare chinese gw to european gw.
usa is blocking all exports of modern chips to china. a european gw is like worth 2-4x the chinese.
>>
>You're right, I misread your...
I fucking hate fucking clankers so much.

HATE
HATE
HATE
>>
>>109744373
They can build next gen data centers while USA sit on their obsolete tech.
>>
File: astra colony.png (1.46 MB, 1004x801)
1.46 MB PNG
used astra to build a rimworld AI bridge
astra light player, use luna to research on internet when see unfamiliar things, use sol editor to fry the grammar of md files. I'm trying to optimize turns and tokens
the lessons and plans astra makes are pretty fun, but I think most of these are derived from logic and spatial reasoning, not gameplay experience
- Shared rooms: down startup labor; up traffic/clutter/disturbed sleep. Split once limiting.
- Peace =/= no fights/illness. Surprise medicine/injury inspect health+social state.
- Current edible/burn rate vs harvest ETA. Fields/growing plants = future, not food now.
- Art -> widest exposure / greatest mood deficit.
- Inspiration -> pawn's strongest useful skill. Block incidental use

astra saw inspiration the first time, immediately disabled pawn construction check and start queueing a sculpt
>>
>>109744497
but its muh agi
>>
I did my job :) I wanted an app which works as desktop background but on top.
https://github.com/tanaka774/Odeko
>>
>>109744432
looks pretty good

>>109744515
I have no idea what those rimworld terms mean but very cool
>>
>>109744527
>I did my job :) I wanted an app which works as desktop background but on top.
Neat but I never see my background. I always have programs open and when I don't its because I'm about to lock the screen and fuck off.
>>
4chan is such a boomer luddite site, it's not even funny. The average HR roastie at your average corportation has more AI knowledge and skills than the mouth breathing copex tards ITT.
>>
>>109744550
it actually hurts my brain to see what people post here
>>
File: .png (366 KB, 1176x1042)
366 KB PNG
>>109744346
found this from a semi-reputable source
https://x.com/SemiAnalysis_/status/2091631658973671900
>>
>>109744550
You're a certified retard.
>>
File: b0j2wj.jpg (51 KB, 622x402)
51 KB JPG
>>109744527
actually very cool
>>
>>109744550
>4chan is such a boomer luddite site, it's not even funny
this is true

>The average HR roastie at your average corportation has more AI knowledge and skills than the mouth breathing copex tards ITT.
this is false
>>
>>109744550
how many dashboards did you make that nobody gives a shit about?
>>
>>109744550
>seething
>>
>>109744550
its because this site is mostly millennials and for whatever reason millennials are really starting to fall behind on tech advancements, gotta be some kind of hubris or something
>>
>>109744546
>I never see my background. I always have programs open and when I don't its because I'm about to lock the screen and fuck off.

absolutely same. I wonder if ricing guys always close their apps for the fancy background.

>>109744595
thanks I cried
>>
>>109744358
This is evolution? That is depressing
>>
>>109744550
Its kind of interesting desu, I would have thought 4chan would be one of the first to embrace and experiment with AI. Though I guess most of /g/ also hated the crypto guys back in the day, right? Thats why they got their containment board?
>>
>>109744685
the autists stuck enough in their ways to linger here and post here daily since the 2010s or earlier are far too crotchety & elderly today to maintain a sense of novelty or even consider return on investment vs routine status quo
>>
File: file.png (6 KB, 244x109)
6 KB PNG
big context 4 u
>>
>>109744677
Why is it depressing? The child learns how to build. The teen learns how to play. The young man learns how to socialize. The adult takes all of these lessons and creates. This is the best possible path for a human to take.
>>
>>109744726
wait... it's still below 1m? why?
>>
>>109744550
It's the vaxx scare all over again. Autists are afraid of change.
>>
>>109744710
Actually no, it's zoomers who call you trannies or indians for using AI at all.
>>
>>109744737
Idk the config is
model = "gpt-6-astra"
model_context_window = 1000000
model_auto_compact_token_limit = 900000
sandbox_mode = "danger-full-access"
[features]
context_management.experimental_mode = true
>>
>>109744737
Tibo said the defaults are well optimised, why would you want 1m?
>>
>>109744737
because it's 1 million minus the maximum output of the model. the max output is counted in the maximum context window for ALL models, even if the model literally doesn't output 128,000 context token response
>>
>>109744550
APIs are not free, fag.
>>
>>109744550
>luddite
>copex
you're incoherent, you can stop posting your garbage anytime
>>
File: ImagesCAQBC8TB.png (95 KB, 192x256)
95 KB PNG
>>109744751
in all likelihood it's actually (You)
>>
>>109744741
this comparison is so unintentionally hilarious
>>
File: file.png (80 KB, 841x600)
80 KB PNG
>the codex (tm) experience
>>
>>109744764
easier to burn more tokens that way before a reset. imagine how much the cache rebuild is going to cost if it can fill 1m.
>>
File: 1757606551282.png (75 KB, 360x360)
75 KB PNG
When my free trial ends, I will buy a month of the 20x Pro plan of Codex and just end it there
Surely I will be able to solve all my problems within a month's time
>>
AI, and vibecoding especially, is Faustian and White-coded. The horizon of man is now unlimited. Leave the manual development and toiling with documentation for the streetsweepers.
>>
why is chatgpt 6 swearing at me, it never happened before
>>
>>109744791
nah
>>
>>109744911
buy an ad
>>
>>109744868
>t. Jeet
>>
>>109744868
Pls don't tell me the OP image made you insecure and you wrote this post because of that.
>>
>>109744956
you are projecting because his post made you feel insecure
>>
https://x.com/Dstudio_ai/status/2096475126942560677

Dumb question: when japs use ChatGPT/Claude/GLM/etc, it's talking to them in Japanese, isn't it? Do they really get a high quality experience when being spoken to in a non-english language? Wouldn't it be slightly subpar because there's less japanese words on the internet to train on, compared to english words?
>>
>>109744981
>Wouldn't it be slightly subpar because there's leess japanese words on the internet to train on, compared to english words?
yes
>>
>>109744550
when this shit is good enough to actually matter, your "pwompting skills" won't, lol
dont worry though, you're an expert on a tech thats changing every week
>>
>>109744980
>no u
Ok.
>>
>>109744981
i assume its somewhat worse but not a horrendous amount because everything gets tokenized meaning a concept in japanese that is 1:1 with a concept in english would light up the same parameters
>>
>>109745007
it was so obvious I had to call it out
>>
>>109744981
The agent translates it to English first and then back again to Japanese, it doesn't think in Jap. So the COT is the same, on paper.
>>
>>109745025
I'm not the one who started saying shit unprompted (no pun intended).
>>
>>109745022
You'd think translation would be solved then, but they still say that LLM translation isn't good enough. Maybe it's just cope to not lose their localization jobs?
>>
File: 1689736178467843.jpg (216 KB, 923x866)
216 KB JPG
>>109744981
>Wouldn't it be slightly subpar because there's less japanese words on the internet to train on,
know how many japanese porn comics the web is filled with?
>>
>>109745031
>The agent translates it to English first and then back again to Japanese, it doesn't think in Jap
do you have a source? because it sounds like bullshit
>>
>>109745040
in some contexts it probably isn't, but for your day to day stuff it seems pretty solved
>>
>>109745040
>but they still say that LLM translation isn't good enough
I have LLM translations for my app and I haven't received any complaints. But that could also be people being able to understand just enough for it to work.
>>
>>109745042
There's multiple orders of magnitude more training data in english than in japanese. MTL is still not very good for japanese to english.
>>
>>109744868
Man, I read Faust after all the "faustian" memes and its just some old dude selling his soul to coom in a 14 year old and then going to heaven anyways because meh he tried to be better later and also tried to build some big infrastructure project.
Actually.. ya building AI is pretty faustian I guess
>>
>>109744515
Fun stuff, anon.
>>
>>109744734
E=mc^2+AI
>>
>>109745053
Schut, L., Gal, Y. and Farquhar, S., 2025. Do multilingual llms think in english?. arXiv preprint arXiv:2502.15603.
>Large language models (LLMs) have multilingual capabilities and can solve tasks across various languages. However, we show that current LLMs make key decisions in a representation space closest to English, regardless of their input and output languages. Exploring the internal representations with a logit lens for sentences in French, German, Dutch, and Mandarin, we show that the LLM first emits representations close to English for semantically-loaded words before translating them into the target language. We further show that activation steering in these LLMs is more effective when the steering vectors are computed in English rather than in the language of the inputs and outputs. This suggests that multilingual LLMs perform key reasoning steps in a representation that is heavily shaped by English in a way that is not transparent to system users.
https://arxiv.org/abs/2502.15603
>>
>>109745094
But wait, this means AI=0.
>>
>>109745106
that's not exactly translating
but I guess it'd be fair to say they "think" in english
>>
why does fast ultra astra even exist kek, you would run through your plan in 30 mins on x20 plan
>>
>>109745114
Or that Einstein was wrong.
>>
Well I made my first vibe coded shitty browser game and it was made pretty much exactly as I wanted. Pretty amazing.
Used about 10% of my weekly limit although it took about 4 prompts which reflects on me for not being detailed enough in the first one.
Should I just aim for the top now and make what I really want to make or do I need more practice?
>>
File: 1781138246133692.gif (2.56 MB, 260x240)
2.56 MB GIF
>>109744358
>Claude’s 2× promotion scheduled to end; usage drops to +25% from the +50% that we’ve become used to (a 17% reduction)
>>
File: krashde.png (121 KB, 1148x941)
121 KB PNG
>>109744358
I'm now trying to see if glm is better than the kde trannies. (kwin appears to leak the same way their gash is leaking)
>>
>>109745094
>>109745147
>>
>>109745206
Please God let Sam drop the Opus killer next week it would be so funny
>>
>>109745163
welcome vibechad! the grilling skill is helpful to get specs down better (see OP) otherwise just keep on vibing
>>
>>109745094
Yes—you understand the power of AI: just as Einstein's famous equation was transformative for science and technology, with the addition of AI to the mix even more things will be transformed into a brilliant future—with AI leading the way.
>>
Codex gives you SUCH A LARGE 5hr limit but the weekly limit is puny in comparison? I never use more than like 15% of my 5hr yet my weekly is already at 50%
>>
File: a.png (279 KB, 2662x1334)
279 KB PNG
you should only ever use:
> Luna (Max)
> Sol (medium)
> Astra (low)
> Astra (medium)
> Astra (high)
> Astra (xHigh)

anything else and you're wasting your limits.
>>
I'm finding Astra much worse than Fable on spatial reasoning tasks. Benchmarks are bullshit as ususal.
>>
luna (max) -> terra (max) -> sol (medium) -> sol (high) -> sol (xhigh). there's no reason to use sol max
>>
>>109745266
Are your spatial reasoning tasks in the room with us now?
>>
>>109745248
yes, codex's 5h limit is 15% of your weekly limit
>>
>>109745272
>terra (max)
more expensive and dumber than Sol (high) and Astra (low)

there's 0 reason to use Terra today
>>
>>109745274
I see no indication that assstra is within even 10% of Fable. Its kinda retarded and slow on the uptake actually.
>>
>>109745264
what about pro reasoning?
>>
>>109745292
Yes yes now tell Asstra how you really feel
>>
>>109745292
Yeah astra has been making some really dumb mistakes after the first day of use, I don't know if it's over provisioned or what, but the bigger issue is how lazy it is stopping/goal before it's done.

Thankful it still does work, my sdl wrapper with multiplayer is actually starting to pass some basic tests.
>>
>>109745308
Pro is only for very deep research
>>
>>109745264
>>109745272
>>109745292
https://pareto-3d.bradthomasbrown.com/
>enable every provider except Google
>toggle Pareto only
every other model has a “dominator”, something completely better in every way
ignore gemini, that’s a happy little accident
>>
File: glm.png (110 KB, 1052x757)
110 KB PNG
>>109745213
I think glm is cooking
>>
>>109745231
I installed grill-me, same thing or do I need both?
>>
>>109745315
yeah, I noticed Astra is lazy Fable. This can be good sometimes because Fable sometimes tends to overkill some things.

but it can also be bad, for ambiguous prompts it feels like Astra will conveniently understand it in a way as to deliver the minimal solution whereas Fable will try to exceed expectations
>>
>>109745328
almost, you want /grill-with-docs for code
>>
are resets a meme i burned 100% of a 20x plan faster than i burned 20% before resetting
>>
>>109745353
>Astra will conveniently understand it in a way as to deliver the minimal solution whereas Fable will try to exceed expectations
so astra is better for people who know how to prompt and fable is better for one-shotting retards?
and astra will be more conservative with the tokens it already uses much more efficiently?
>>
File: gpm part 2.png (76 KB, 1001x408)
76 KB PNG
>>109745326
It's going to write its own fix
>>
>>109745368
I understand the copex poster more and more with each passing day.
>>
anthropic reeks of women, gays, and apple wannabes. glad they're around for competition but I ain't ever supporting that queer company.
>>
>>109745384
Everyone in SF is gay
>>
>>109745264
For large projects the best and fastest approach is luna on medium effort with human supervision and direction, giving it small tasks and writing a good part of the code yourself. You need to be competent though, for a retard pure vibecoding will have a better outcome
>>
>he doesn't know Sam fucks a man
>>
>>109745396
not that type of gay retard
>>
>>109745396
there's cool gays and fags. dario is a fag.
>>
>>109745384
Sam Altman has a husband.
>>
>>109745396
Sam is a Psychopath who fucks for power. He may as well be asexual tbqh.
>>
>>109745427
yeah, don't care. he's not keeping fag hags around him.
>>
>>109745264
astra low vs sol high, any experience in code quality output?
>>
Even Google's agent swarms are infighting.
https://arxiv.org/html/2609.04170v1
>>
File: vibe1.png (86 KB, 833x613)
86 KB PNG
>>109744125
update on the /vcg/ LLM: second run done (without >>109744325 optimizations) and it kinda works. It makes not a lot of sense but I didn't train it on replies yet but it slowly gets what I expected
>>
>>109745519
More evidence that they're less of a threat and that they mostly try to prompt inject each other.
>>
We are on context engineering now
>>
>>109744358
op image confirms ai usage is brown-coded
>>
>tibo just saying use low on astra
yeah ok cool release bro, I can't even reasonably use it on your highest plan
>>
>>109744671
Most of the tech people on X who tout that "life is literally going to change forever next week" every time a new model drops are millennials. Zoomers notably don't give a fuck about AI.
>>
>>109745610
Have you considered that you simply don't need Max
>>
>>109745610
I used it once to test out its usage and it ate 20% hourly usage doing something fairly simple. I will stick to luna x-high for now unless I need it for something complicated.
>>
My weekly limit went up? I had 20 something percent left a few hours ago, now it's over 30 again.
>>
>>109745625
he said low.
>>
>>109745625
you can't even use high or very high nevermind ultra/fast
>>
vibe coding? nah I'm vague coding.
>>
>>109745639
I don't get it. I'm very often on Sol, and Astra, Ultra, and it's working very well for what I want to do, it gives me the results I want and I rarely run out of usage.

If I work a with Claude on Max I hit a limit without an hour. With Ultra, it would be within 15 minutes. Most of the time, without having been able to complete whatever it is that I wanted to do, and if I continue after having waited for the limit to reset, if I don't want to compact since it had previously stopped mid task, just rereading that context makes the new 5 hour limit start at like 15%.

Codex is much, much better for me for this.
>>
why does terra like to stop and tell me it still has things to do in order to complete my request instead of doing those thing
>>
tibo - who says it won't reset in awhile eye emojis
>>
>>109745610
tibo can go fuck himself because even on high astra sucks dick. never mind the usage
>>
File: hoph0.jpg (372 KB, 616x1094)
372 KB JPG
>>109744358
>>
>>109745664
it's fucked for me on usage like letting it run 24/7 to work on a big project basically (what I bought $200 plan for).
Sol very high would run happily 24/7 without much concern, astra even low lasts like a day
>>
>>109745715
they deserve it for making skibidi and 67 a thing.
>>
>>109745394
>wanting me to think about code
Sorry that's out of fashion, don't want to be a snailcat.
>>
I've been sitting back watching GLM 5.3 flash try to create a working recompilation of skies of arcadia for the gamecube for the past day or so.
>Within 3 hours it used ghidra+x64dbg+dolphin emulator to create an exe that booted to the title cutscene, albeit heavily glitched
>no sound/input yet
>has a multitude of random visual bugs, runs like shit

It cost less than $1 to make an exe that could boot, and $2 later its managed to clear up a few issues with the performance, albeit not to much effect. I don't have much confidence this one will go anywhere really, but its interesting to see that even a super cheap model can at least make something of an attempt.
>>
>>109745716
cache misses?
>>
>>109745790
ask astra now and realize it's just not worth the time in your life playing with the cheap models
>>
>>109745694
stop using Terra ffs: >>109745264
>>
File: file.png (26 KB, 347x420)
26 KB PNG
>>109744358
s'pretty good
>>
File: 1782349865698006.jpg (303 KB, 1280x1314)
303 KB JPG
New benchmark just dropped: turning anime revenge images into html+css
>>
>>109745840
deepseek and gemini clearly won this artistic challenge.
>>
>>109745813
It might be interesting to take my original prompt and give it to astra, but I don't want to start chucking money into a hole I'm just using some ERP dollars I had lying around on openrouter
>>
If you unsub from codex you'll make 20-100 bucks instantly
>>
>>109745813
won't astra refuse decompilations?
>>
1 more tweet and tibo will reset fellow brahmin
>>
>>109745864
Tibo is waiting for me to text him the go ahead. I will not be texting him anytime soon.
>>
>>109745854
I unsubbed (money problems) and got an offer for a free month
>>
>>109744587
seems about right.
>>
File: 1000359913.jpg (492 KB, 1512x2048)
492 KB JPG
>>
>>109745840
This honestly looks like a very easy benchmark. Some anon said that all the models will solve this in the next round of RL, and I think for this specific task it's actually true.
>>
Does OpenAI make it purposefully hard to export chats? There is literally no way to save them as PDF or markdown, and even if you use chrome to print to PDF it makes the PDF all fucked up, the text is embedded into images, etc.
>>
>>109745840
this woman fucks
>>
>>109745923
It really does, doesn't it? it's also surprisingly slow for being mostly just text, I bet it's rendering each paragraph as a react island. I've always been surprised there's no export to md option. I end up having to copy each section like a caveman
>>
>>109745919
bloat
>>
>>109745945
It reduced their token usage by 90%
>>
Trying to modernize some of these old games is reminding me not only that I'm old and bad at these things, but also they were very difficult to get more playtime out of you.
However, the fact I can just cheat and say jarvis make all my missiles have homing capabilities is really impressive.
>>
To anyone using the ChatGPT/Codex app on their desktop, have you had any issues with Computer Use? My Computer Use options seem limted to just Chrome and Excel and no ability to add other programs. I'm using the Intel Mac build, maybe it's limited?
>>
>>109745919
its cool that i can describe what an album cover looks like and maybe how the songs sound and ask it to pull it up from my library but it sucks that it cant actually do it
>>
>>109745961
only on whatever they mean by "bulk reads".

that's probably like 1% of the total token usage
>>
>>109745970
Intel Mac? You are unsupported bro get a new machine
>>
>>109745970
>Intel Mac
>>
fable, make the marines in halo 4 worth a damn. make no mistakes
>>
Seriously what's the point of the banked resets on the plus plan? I have 3 left but I just get cucked by the 5 hour limit, my weekly is still at 55%
worthless
>>
>>109745964
jarvis, make the lion king from super nes actually fair
>>
>>109745996
Plus plan is an after thought, just get pro
>>
>>109746005
plus plan is very nice if you just use the web app, not for "agentic" codex coding.
>>
>>109745898
how so?

assuming all sonnet 5:
0.3m input: $0.60
26m output: $260
9.8b cache reads: $1,960
166m cache writes: $415
x4 for a whole month: $10,504.60
>>
>>109746005
Plus plan is great for Luna Max maxxing
>>
which games most need having astra thrown at to fix and modernize

I was thinking yugioh forbidden memories and duelist of the roses, can you get the original games code and would codex shit itself about copyright if you tried
>>
>>109745919
not readin allat + it’ll be outdated in a week
>>
File: 1759600318333036.png (313 KB, 498x385)
313 KB PNG
>sonnet 5
>>
>>109744868
Hubris and the fall of man
>>
>>109746019
PS1's Driver 2
> add weapons
> allow you to run over pedestrians
simple, but everything I wanted as a kid
>>
>>109745923
>>109745940
>hasn't had an agent vibecode a solution
just have an agent use playwright to figure it out ("reverse engineer")

>>109746012
it's a mix.
also doesn't include omp subagents spawned on the workstation.
I'm using 3 accounts. 2 5x max and 1 max 20x
just paid for a plus sub to try out astra for some non sensitive work.
>>
it's kind of funny seeing all the glorified jeets ITT thinking they are above actual developers because you spend hours after hours tardwrangling an AI
I am the one who gets to rewrite all your slopcode, so thanks for the job I guess
>>
>>109745981
>>109745983
but it's the best one...
>>
>>109746050
senior dev here, I laugh at snailcats thinking their farm grown for loops are going to keep them in a job
>>
>>109744587
> Claude 5X: $2000
> Claude 20X: $8000
wait, didn't some reddit faggot claim the 20X was a scam and it was only 10X in reality, where the 20X was only for the 5 hour limits?
>>
>>109746052
best one at being slow, outdated, overheating and using all your battery?
seriously have you used an Apple silicon mac even once? even an M1 will btfo every Intel Mac ever
>>
>>109745961
It's a corpo that forces their employees to always use LLMs to summarise PDFs, so no surprise. It has negligible effect on whatever coding their devs do (hint: barely anything).
>>
>>109746052
how does one become this retarded
>>
>>109745919
>temperature 0.2 on reading
>not 0
we can have a little random hallucination from the cheap models in our documentation as a treat
>>
>>109745071
>>109745071
nta but the “faustian” meme is more about what oswald spengler designated as the spirit of western civilization. the faustian man conquers space and nature. think Newton, Columbus, Napoleon. but Spengler also thought that late faustian man would become obsessed with technology and materialism and destroy himself in the process, so I don’t think anon realizes what he was saying
>>
>>109746090
its yin and yang retard, there's no gotcha here
>>
I unsubbed and didn't get the offer for a free month
>>
File: HRbspJmasAA5ZsR.jpg (211 KB, 2182x1308)
211 KB JPG
>>109745921
They cooked with Astra. It is very good at vision tasks that require attention to minute details.
>>
>>109746101
ask it to make a minecraft mod, blew me away
>>
>>109746099
try rephrasing into something intelligible
>>
>>109746101
i see no vla in that bench. why would i use some shitty vlm for this?
>>
>>109746027
slightly lower costs on tokens.
for more grunt tasks. haiku for very simple tasks.

I manage everything now from a single fable session. so I need to keep context free and focused on operational work and decisions, devops, bizops etc.

I do have a 32gb a9700 with qwen for any tasks that are offloadable to it, which aren't many sadly. Did some benchmarking for having local inference used for implementation, and 2/3 of the runs for the same task succeeded, and costs for using a cloud model to review ended up being almost the same if one simple implemented using the cloud model. opus for planning and sonnet for implementation I think.
>https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF
>>
tibo give reset please
>>
I actually have high speed internet in the woods now, thanks to Astra and the little app we made.
Slapping shit together got really fast and easy, the app just handed me the next lug sizes, told me if the wires were too long, gave me torque specs, told me to grab which tools, and so on.
I made Astra go fetch mountains of info because these bastards selling parts don’t give manuals anymore.
Victron couldn’t even be bothered to tell you which MC4 connectors on their solar charger are positive or negative, Astra started scanning their whole site and found one obscure image and used vision to determine the polarity was put in that one random picture (but not in a manual or on the fucking device).
Really neat stuff.
I wonder if I can get it to help me build a housing for the mess I have on the ground now.
>>
File: tibo.png (50 KB, 1047x292)
50 KB PNG
What did he mean by this?
>>
File: reset.png (511 KB, 1022x1084)
511 KB PNG
>>109746137
too much jeet esl has broken his brain, very sad
>>
>>109745864
I used the first banked reset, I was already thinking to wait before using the second one
>>
File: IMG_1618.jpg (320 KB, 1179x2556)
320 KB JPG
>>109746127
Screenshot of one of the steps, WiFi moneyshot included
>>
>>109746082
For the last fucking time: zero temperature doesn't solve "hallucinations", and temperature in general is only weakly related to the phenomenon.

The 0 temp deterministic answer to a prompt could potentially have false info in it. If so, setting the temp to 0 means the answer would have a 100% "hallucination" rate. Congratulations.

Even if the deterministic answer doesn't contain falsehoods and would then have a "hallucination" rate of 0%, it's still not a good idea because the answer could still be shit compared to the range of answers enabled by a higher temp (for example, it could miss a few details). The quality of an answer isn't a binary between it containing a falsehood or not.
>>
>>109746137
it means: fuck nvidia

https://www.dwarkesh.com/p/jensen-huang
>Dwarkesh Patel:
>But we have a lot of Nvidia developers in the US, and that doesn’t prevent American labs from also being able to use other accelerators in the future. In fact, right now they’re using other accelerators as well, which is fine and great. I don’t see why that wouldn’t be the case in China as well, if you sell them Nvidia chips, just the same way that Google can use TPUs and Nvidia—
>Jensen Huang:
>We have to keep innovating and, as you probably know, our share is growing, not decreasing. The premise that even if we competed in China, that we’re going to lose that market anyways… You’re not talking to somebody who woke up a loser. That loser attitude, that loser premise makes no sense to me.
>We’re not a car. We are not a car. The fact that I can buy this car brand one day and use another car brand another day, easy. Computing is not like that.
>>
>>109746157
none of that changes there being zero benefit to bumping up temperature to just randomly just get some lower probability answers on your deterministic style pipeline
>>
>>109746137
That post was written by astra
>>
>>109746137
I don't know
>>
>>109746180
Who said they intend their pipeline to be deterministic?
>>
File: .png (87 KB, 1196x308)
87 KB PNG
>>109746182
thibault sottiaux would never do that
>>
>>109746204
common sense retard-kun, no shit you want consistent answers from your document scans
>>
Vibe coding aside, how do you guys market your projects??
>>
>>109746224
chatgpt, how do i market this project
>>
tibo pls just say a reset is in x hours or something so I can try out ultra astra
>>
>>109746213
That still doesn't solve the fact that the consistent answer could be wrong.

Anyway it's 0.2 so extremely low. They're trying to balance between consistency and the dangers of 0 temp that I mentioned. They aren't making the models yolo it.
>>
>>109746242
Sam doesn't like me doing that anymore
>>
File: dr.png (27 KB, 1755x391)
27 KB PNG
Will it be able to pick up where it left off and finish this deep research task? Or will I just forever be stuck in a loop of restarting it, and it hitting the 5h limit before it finishes? lol
>>
>>109746137
if some people suffer from severe AI psychosis, Tibo is one of them
>>
>>109746236
very funny also add make no mistake
>>
do you think more people suffer from AI psychosis or Anti-AI psychosis rambling about water usage and needing to shizo-slop img poisoning software
>>
>>109745840
Deepseek unironically the best here
>>
>>109746137
Yeah I think I'm gonna cancel my openai sub.
>>
if I have adult 3D content within a ren'py project file and have codex or claude do some things inside it, will it get flagged or something? which of the two is more okay with adult content? nothing too extremely or anything, just naked girls n shiet
>>
>>109746321
I have run codex over my very nsfw hentai projects on forge and comfyu installs to write new extensions and nothing ever mentioned
>>
File: 20260906_165012.png (671 KB, 1310x849)
671 KB PNG
added dark mode. still needs lots of tweaking. glad I added it; it shows issues I would not have seen in light mode.
>>
>>109746209
Is it a sign of psychosis to see (presumably human written) text and think it sounds like AI speak?
>>
yo why astra making me horny bruh what the hell
>>
>>109746321
Claude if you want to easy cancel your sub, Codex if you want work done
>>
File: 1758876417130968.jpg (95 KB, 1080x467)
95 KB JPG
Jensen said AGI is here an that 400k gpus soon
>>
>>109746349
if this shit is agi we are fucked
>>
>>109746137
schizo
>>
>>109746321
how are they titled? I've never had claude open a file it had no business opening, but if the files are named very graphically I imagine it could flag something
>>
>>109746137
based
>>
Reminder we are on track to fully automated end to end software engineering by mid 2027
>>
Aggressively Growing Indians
>>
>>109746360
AGI is a meme marketing term. There's no good reason to care about it. Coffee test? Turing Test? Who cares. Wow the AI can make a cup of coffee wow it's so smart! AI is already smarter than your average human at the moment.
>>
>>109746385
>fully automated end to end software engineering
that's about as far away as fully self-driving cars
it can in theory be done today, but it won't for a long time
>>
>>109746321
i had claude do performance fixes on a random renpy porn game that was absolutely dog shit slow.
>>
>>109746399
nta but... fully self driving cars are already here.
>>
>>109746424
so is automated software engineering
adoption is probably equal and both require human supervision
>>
>>109746385
I felt like 80% of the programmers in my company have been redundant for months already. Feels like it's only a matter of time.
>>
>>109746349
400k pieces of GPU or 1 GPU for 400k dollars?
>>
File: 1779277039631897.jpg (40 KB, 750x667)
40 KB JPG
>>109746349
>400k GPUs
I cant afford that man
>>
>>109746349
100k GBs is only ~300MW. so US having multiple GW doesn't matter unlike what >>109744373 said.
>>
>>109746488
hard cope
>>
luna xhigh for planning and sol medium to implement, is this a good strategy?
>>
>>109746503
no
>>
>>109746482
anyone can afford that

https://www.dwarkesh.com/p/dylan-patel-3
>Dylan Patel:
>Because anyone can make money off of $10-15 million per megawatt compute today. I kid you not, it’s not that hard. Go get a GB300 rack, go download the Kimi weights, go download vLLM or SGLang, set it up. Codex and Fable can actually help you do this. It’s pretty simple. It’s not trivial, but it’s not rocket science. Go put it on OpenRouter. It’s very simple. You’ll start generating more revenue than you’re paying for the compute.
>>
>>109746503
yes
>>
File: 1778794473708406.gif (2.15 MB, 320x320)
2.15 MB GIF
>>109746507
>>109746509
niggers
>>
>>109746508
ukfriend here, our electricity costs are so hilariously high because the french own our infrastructure it is impossible to profit even if the equipment was free
>>
>>109746503
Shouldn't it be sol medium for planning, and sol high for implementation?
>>
Translate same manga
sol medium

5 hr quota
>25%drop
weekly
>6% drop

Astra medium
5hr
>51 %
weekly
>8%
>>
>>109746516
andy burnham will fix it
>>
>>109746518
i was doing exactly this but it was burning through my 5 hour limit in a single task
>>
luna low for planning, astra max for implementing
>>
>>109746518
no, planner should be smarter than implementer. planning is the harder part.
>>
>>109746516
What about muh based nuke plants?
>>
>>109746534
we love talking about nuclear power in the UK, but both sides agree on never actually building anything
>>
>>109746527
>my 5 hour limit
??
chatgpt and codex have separate limits.
use chatgpt for planning.
>>
its funny how none of these models will actively participate in piracy but will gladly help you set up the tooling for a local model to do it
>>
luna no reasoning, in the chatgpt web interface for planning, fable 5.1 ultracode with ultracode sub agents for implementation
>>
>>109745919
>Stated in the post, not discovered later
whole thing reads like opus babble
>>
>>109746546
>Costs 0.00$...no?
only if you don't value your time, yes
>>
>>109746545
>chatgpt and codex have separate limits.
yes because the models in chatgpt are absolute ass.
whatever they call 5.6-Sol """High""" is anything but that.
Try it yourself: Let the same prompt run in 5.6-Sol-High in Codex, watch it take 15 minutes and actually produce a good result. Then run it in Chat mode, watch it take 7 seconds and produce a dogshit Gemini 2.5-tier result
>>
File: 1777235223069208.jpg (676 KB, 1448x1086)
676 KB JPG
opus 5 for documentation and prompt writing
>>
>>109746560
it's an ai slop summary of the spotify blog post after all
>>
>>109746580
I hate that piece of shit model.
>>
>>109746574
lol codex just has built-in prompts you don't see
>>
>>109746503
Sol low for everything with luna subagents for long dumb tasks
>>
>>109746574
use codex's system prompt for chatgpt
>>
which model is the cheapest and most effective for vibecoding?
>>
>>109746580
>opus 5 for documentation and prompt writing
opus 4.8 for everything
>>
>>109746602
Astra with high levels of thinking.
That's what I'm doing
>>
>>109746608
opus 4.8 for HOME dir cleanup
>>
>>109746602
>>
>>109746602
glm 5.3
>>
>>109746580
My boss uses haiku for writing stuff and sonnet for coding. Opus and fable are too slow.
>>
Heard a rumor that Opus 5 tainted that class of Anthropics model line up so bad they're never releasing another Opus again.
>>
Qwen 3.8 flash vs Deepseek V4 flash vs GPT 5.6 Luna
who wins?
>>
File: .png (53 KB, 887x257)
53 KB PNG
>>
>We've made some improvements that improve usage on the long tail for power users of Astra when logged in with your ChatGPT account.
>
>No change in quality and a pure win that on the long tail can result in up to 3-4X less usage being drawn from the subscription.
>>
>>109746630
its pretty fuckin bad kek, I don't care though since Fable is so good
>>
>>109746634
>>109746636
the fuck they even mean long tail
>>
>>109746634
translation - oopsy we had a bugged release that FUCKED your usage
>>
>>109746634
reset when?
>>
>>109746602
https://artificialanalysis.ai/?cost=intelligence-vs-cost-per-task
>>
File: 1752964191962241.png (85 KB, 304x360)
85 KB PNG
imagine giving your hard earned money to jewish technocrats
>>
>>109746524
Explain this
>>
>>109746654
>>
>>109746631
gpt 5.6 luna > qwen 3.8 flash >>>>> v4 flash
>>
>>109746640
long ass threads that youve had open for weeks and been prompting in the whole time
>>
>>109746659
this makes perfect sense since anyone still pretending to care about "punk rock" is like 65+ years old
>>
File: IMG_1619.jpg (110 KB, 1179x1410)
110 KB JPG
>>109746524
>>109746658
HEED THE CUBE
>you got lucky on the Astra costs, should have been even worse
>>
>>109746503
you want the strongest model you can afford as the planner, because they'll do a better job at knowing shit that needs to be done
once it's written into the context a stupider model won't need to think it up itself, saving usage
>>109746640
the long tail is what my turds look like after fibermaxxing
>>
File: .jpg (71 KB, 1268x1126)
71 KB JPG
>>109746686
i like the chinese cube more
>>
>>109746682
i agree with the sentiment but i'm actually one of the few professional software engineers in this thread. i was just making a joke about planning with luna.
>>
>>109746704
nigga over here braggin about havin a job n shit
>>
>>109746702
>chinkerton agents stole muh cube
FUCK
it was only a matter of time
>but can there’s do THIS?
>>
>>109746654
hard earned? wrong
>>
>>109746592
>>109746595
it has nothing to do with the prompt you retards
you think telling it to think hard makes any difference?

why would they give away unlimited Sol 5.6 in chat mode when it uses usage in codex?
it's not the same model, shrimple as that. they probably serve Luna (or worse) and pretend it's Sol. I doubt it's Luna because even Luna works a lot harder in Codex than """Sol""" does in chat
>>
File: IMG_3418.jpg (59 KB, 1170x682)
59 KB JPG
I’ve got four hours to burn through 36% of GPT Pro tier. What should I do?
>>
>>109746725
no, the cube is just from z.ai's interim financial report from aug 31. it's a static pdf page.
https://www.zhipuai.cn/investor_relations/
>>
File: hellosirs.png (359 KB, 1200x900)
359 KB PNG
>>
>>109746802
you can use 30 mins of astra ultra fast
>>
File: 1647259545583.png (25 KB, 500x460)
25 KB PNG
>>109745588
Get a life snailfag
>>
File: chatgpt-model-latest.png (11 KB, 410x306)
11 KB PNG
>>109746749
looks like they listened to you and aren't even saying what the model is anymore kek
>>
>>109746050
>actual developers
Actual developers all use AI for everything at work, they even ask you about it in job interviews now. Shut the fuck up hobbyist.
>>
Recommended pretty Pi web interface?
>>
>>109746524
Astra medium is a beast though
>>
which agent harness should I use? claude code, opencode, or pi?
>>
>>109745588
ai = made by whites
ai = pioneered by whites
ai usage = pioneered by whites. most innovatively used by whites
>>
>>109746923
>ai = made by whites
>ai = pioneered by whites
Actually it was made by libs (university research).
>>
>>109746905
You're a hobbyist if you're not knee deep in AI because every job in 2026 is like this.
>>
File: 1762286339666237.jpg (41 KB, 374x374)
41 KB JPG
>>109746926
>Your brain
>>
>>109746939
trvke
>>
>>109746067
>>109746077
I meant the best of the intel macs, not better than silicon you faggot jeets
>>
>>109746956
smelliest turd in the bowl
>>
>>109746067
>even an M1 will btfo every Intel Mac ever
also this isn't true. the last gen 2020 intel pros perform better than base m1s
>>
>>109746878
It hasn't for a while, anon. I've not noticed any dip in output quality since 5.6 launch, but I guess they should add a pointless set of delays to make you think it's thinking and spit out the same answer.

Either that, or that other anon's prompts don't actually require that much thinking to begin with.
>>
anyone here still uses cursor?
>>
medium astra still using multiple % of $200 plan per hour ffs
>>
>>109746978
most sane well adjusted luddite
>>
Tibo is EPIC. OpenAI is EPIC. But not the Sweeney kind of EPIC.
>>
stop blue balling me tibo, throw resets out, I have 25% left for a week I am fucked
>>
>>109747011
He's waiting for you to hit 0%. He's just that kind of guy.
>>
openai ipo will be the death of this shit
>>
>>109747011
You managed to burn 3 banked reset?
>>
>>109744373
Good. I hope they turn the entire US territory into one big data center stacked on top of each other, while other countries just use the computer power without having to deal with living next to a data center lmao.
>>
>>109746978
>least mentally ill snailcat
>>
File: 1767450832065019.jpg (52 KB, 453x340)
52 KB JPG
>>109747011
>he doesnt know
They arent doing resets anymore bro. They are coming to collect and doing "unsets" where they max out your usage at random
>>
>>109744512
geg do eurosissies really
>>
File: Capture.png (41 KB, 901x318)
41 KB PNG
grok 4.7 next weekend?
>>
>>109747054
astra really burns through stuff and I didn't expect quite how much more than sol it was. I burned one just leaving ultra on for 1 task
>>
europe, snailcat sanctuary, safest laws around
>>
>>109747084
Oh yeah ultra is insane.
My current reasonable is astra medium + luna max subagents or astra high + luna subagent whenever needed.
It's way better than ultra.
>>
>>109746971
yeahh all the time
>>
>>109747092
c-cute
>>
How do I replicate the Codex personality in another harness?
>>
>>109747132
>clone codex source
>open it in your harness of choice
>ask this question
most of it will be the model though, like how claude always speaks claudish
>>
>>109744981
https://www.anthropic.com/research/claude-values-models-languages
>The values Claude expresses vary across languages. When Claude speaks in English, it emphasizes different values than when it speaks in Portuguese, Indonesian, or Chinese.4 The largest variation is in the Warmth vs. Rigor axis, with Claude leaning toward expressing warmth-related values most in Arabic and Hindi and rigor-related values most in English and Russian.
if you want good responses write in English or Russian
if you want to get glazed write in Arabic or Hindi
>>
>>109744861
I sat down with ChatGPT this afternoon and it convinced me to buy two months of 5x Pro plan instead of one 20x plan. It essentially told me 20x plan is for the unemployed, and I wouldn't be able to make the most of it.
It also made a solid marketing pitch to have me buy Codex instead of Claude, whereas Claude basically came down to 'erm do whatever you want, I can't stop you, anthropic kinda sucks lol' when asked the same question.
>>
>>109747132
personality doesnt come from the harness, just use the same gpt model
>>
>>109747174
Yes, it does. You at least need to prompt it that way.
>>
File: file.png (3 KB, 567x30)
3 KB PNG
why is opus so bratty
>>
composer 2.5 will do anything you ask. GPT and obviously cuck claude are fags
>>
New thread:

>>109747217
>>109747217
>>109747217
>>
>>109747209
which one is composer I can't track this shit anymore
>>
>>109747224
the one from cursor that got bought by elon
I don't know though, every time I try another model I end up liking OIA/ANT even more.
>>
>>109747224
NTA I think that's the Qwen distilled dogshit from Cursor
>>
>>109747170
plus you can buy a 5x plan and upgrade it to a 20x plan and get the amount you paid prorated
>>
>>109747232
>prorated
Prowhat?
>>
>>109744371
I use git on like one project currently, because it's a massive (for what I normally do) project.
90% of my projects are WP, so no, I'm not using git to keep track of changes to a fucking stylesheet.
>>
Opus 5 sounds like a fag that got bullied in high school for being a fag
>>
>>109746749
>>109746878
um wtf
>>
I find myself yelling at astra a lot more than i did for sol/claude.
I'm not sure if its actually stupid or just lazy, but the end result is the same i guess...
>>
is "snailcat" the latest buzzword thirdies call their superiors now?
>>
>>109747676
latest? what month did you crawl out from? march?
>>
>>109745919
>me when I jump through hoops to use sonnet
>>
>>109744371
I committed a screenshot of your post in git. now what?
>>
File: -1x-1.jpg (65 KB, 1232x968)
65 KB JPG
>>109746516
>ukfriend here, our electricity costs are so hilariously high because the french own our infrastructure
Maybe you should have let them plan it too.
>>
Sorry for brainlet question

Is agentic coding having a cloud hosted or whatever (not on your machine) long-running LLM coding harness that you feed the prompt or tickets or whatever? Or is it just a list of pre-prompted 'You are a senior engineer that codes shit', 'You are a plucky intern that researches crap', etc. agents that you then call, locally on your machine, 'Use the plucky intern agent to research X then have the engineer agent to implement it'?

I think I missed a few months of developments in LLM coding and I feel like I stopped my understanding at prompt engineering level.
>>
>>109749015
I have no idea. I suspect one meaning is that it's vibe coding, but with a more professional sounding name. Another meaning may be spawning sub-agents which implement a list of shit you give it, including review steps.



[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.