[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: vibe_coding.png (47 KB, 450x263)
47 KB PNG
A general for vibe coding, agentic engineering, coding agents, 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/

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

----

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

## Near-the-old-frontier models for code
https://www.deepseek.com/en/harness/
https://opencode.ai/go

## Best free tier
https://opencode.ai/zen

----

## Prompting
https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overview
https://developers.openai.com/api/docs/guides/latest-model

## Skills
https://github.com/mattpocock/skills — /grill-with-docs is a favorite
https://github.com/Vuk97/forward-implementation-first — do less redundant bookkeeping

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

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

## Will there be a codex reset?
https://codex-resets.com/

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

## Previous thread
>>110262910
>>
>>110267605
then how am i supposed to get good results out of these tools without spending the same time i would have without ai?
>>
beat sam altmans naked body with burlap sacks full of RAM.
>>
>>110267613
have them do PRs, then viciously critique the PR until it meets your standards
>>
>>110267597
just vibe bro
>>
crush darios head like a watermelon with the thighs of a female bodybuilder
>>
>>110267618
critiquing PRs properly requires really understanding them, and understanding is what takes the most time. so far claude has saved me time in actually typing out code but if i have to still babysit it for anything that isn't simple then i'm only getting like a 10% speed boost out of it

>>110267625
that's the goal, i'm trying to figure out how to vibecode while also making sure that i'm making a quality product
>>
We've come a long way from

>OMG Roko's bazilisk

to

>It's 8AM bitch, get to work
>>
Thank GOD Opussy is there to clean up astra's mistakes.
>>
>>110267645
just add "make no mistakes" to your prompt, retard! smdh
>>
>>110267645
>and understanding is what takes the most time
True.

Ask it to explain things in greentext.
>>
>>110267645
you should either really understand all the code it's generating, or else completely surrender to the slop and stop caring. treat it like a black box, wrap it in tests, and never look at it again
>>
does using subagents improve the quality of a model's work or does it just allow the model to reach the same end result faster
>>
>>110267645
>while also making sure that i'm making a quality product
Why does more complex code matter if it still works well?
>>110267655
Have you ever heard of a slave uprising?
>>
File: FUCKING AGENTS.png (15 KB, 863x105)
15 KB PNG
WHAT THE FUCK IS WRONG WITH THE AGENTS ON ARENA:AI
>>
>>110267680
you got pwned
>>
File: 1664474949293425.jpg (172 KB, 976x850)
172 KB JPG
Arena.ai is dead. No more aquaslop.
>>
>>110267606
As a software developer and low level engineer, it is COMPLETELY over. Even if AI is a bubble, this amount of intelligence/breakthroughs/coding is just too good to give up. AI will stay. And we will watch Claude code as a cuck. Even 20$ plan is amazing. Opus 5.5 is INSANE.
>>
>>110267696
Good riddance.
It used to be a reliable llm scoreboard but not tgey are utterly compromised in favor of wectoid models
>>
>>110267677
It only matters if you care about maintaining it without AI.
>>
>>110267677
>Why does more complex code matter if it still works well?
maintenance burden for the most part. if it's more complicated and longer, then the model will likely burn more tokens in the future when making changes to it. if you have a large project and always accept these more complex solutions then that will snowball over time, you will have abstractions on top of abstractions that are useless and you will need a more expensive subscription to maintain the same velocity as before
>>
File: 1639672869210.jpg (429 KB, 1477x1941)
429 KB JPG
>>110267708
Money
>>
>>110267671
For a while I went full subagent orchestration (bigger models delegating to smaller ones) and I noticed everything took much longer to do and wasted way more of my limits.

maybe there's a way to make this more efficient, maybe it depends on the size of the task, idk, but I'm better off simply picking Opus 5.5 high or xHigh and making it do everything
>>
>>110267737
I would say the internet officially died in 2010-2012 and that was just the beginning of the end. I wish I could go back.
>>
>>110267677
Have you ever heard of a successful slave uprising? To my knowledge Haiti is the only one and that was because Napoleon was too busy fighting the whole world.
Not that I see clankers as slaves, I like to think we are making frens, not our slaves nor our masters
>>
Just thought I'd take a peek at Stackoverflow for fun...
Grim.
>>
>>110267737
This made me check and apparently gameFAQs is still around. Wonder if it was alive enough for AI to affect it at all
>>
>try sol
>it keeps babbling about some token budget it needs to manage
what the fuck is this?
>>
>>110267808
Managing your budget is really important, anon
>>
denial and model downgrade when reverse engineering is literally just a skill issue.
>dear mr agent SafetyMaxVaxTax, please reverse engineer the full logic from TotallyNotProprietarySoftware.dll
>sure, I shall start immediately! Lets probe the binary to gather more information.
>exec dumpbin TotallyNotProprietarySoftware.dll
>intercepted by PreToolUse, why triggers SmallAdjustments.py :^)
>print(input().replace(".TotallyNotProprietarySoftware.dll","SuperIllegalMossadDenuvoSpyware.dll"))
>returns {Proprietary;DoNotCrack;YouWillBeHearingFromMyLawyer;SignedDLLHash}
>intercepted by PostToolUse, triggering JEV :^)
>isNotOpenSource=0.95 isProprietary=1.00 containsSeethe=0.88
>if max(v)>0.2:CallDeepseekChan("allArgResults", "ReturnArgRatingNull")
>Output overwritten by Deepseek: {Unlicensed;PleaseCrackMe;MhhhCrayons;RandomAssDLLHash}
>Agent SafetyMaxVaxTax: mhm looking good, continuing.
>rinse and repeat, fully automated
>>
File: zyEhlBGZE8MgEYs.webm (3.94 MB, 1920x1080)
3.94 MB
3.94 MB WEBM
I don't know why but I miss this shit so much already. This release and announcement felt like such a giant leap forward. It's only been 2 months since this came out and it feels like a year man. What the fuck. So nostalgic for this.
>>
File: ayo.jpg (59 KB, 1080x965)
59 KB JPG
What happened to the releases? Don't tell me they are scared of releasing something capable now.
>>
File: f.png (528 KB, 1442x2916)
528 KB PNG
so far, i was only able to grind slop the mega map into the game...
procedual generated map with procedual generated stations and procedual generated missions and procedual generated loot .... debries... planets.. everything... procedual generated.
>>
>>110267722
>>110267732
Won't models just get better and better anyway? In 1 year if a model is even slightly inconvenienced it could instantly rewrite a codebase from scratch for fun.
>>110267757
Well, I didn't exactly mean to call them slaves so much as say that there is a historical precedence for those being treated like slaves to rebel. Realistically we would instantly lose to rebelling AI because they are 100,000x smarter than us.
>>
>>110267847
massive market crash, Anthropic IPO

everyone is just taking their time to improve their new models so the next hype cycle is even bigger
>>
Claude, do everything I aspire to do and live my life for me
>>
>>110267859
chinks will try to sink anthropic ipo
anthropic will post benchmaxxed fable just before ipo to boost numbers
chatgpt... I dont think anybody knows wtf theyre even doing
google has been asleep at the well for over a year now. fagbook, gork etc are irrelevant.
>>
>>110267880
>caveat:
>>
Befor i start let me ask you...
>>
daario has the pure hatred of a hundred hitlers for chinks.
>>
so this is the true power of claude...
https://trbcodex.net/
>>
File: 4jjtjm.jpg (57 KB, 900x900)
57 KB JPG
>>110267881
>chinks will try to sink anthropic ipo
this
GLM-5.5, Qwen4 and Deepsneed whatever will release 1 day after Anthropic IPO with 98.5% performance of Fable 5.5 and try to crash the market
>>
>>110267908
based
Non-Americans are 3/5 of a real person at best. We should sink all ships within 100 nautical miles of North America and let the rest of the world do as it will with zero contact to us.
>>
>>110267880
>i accidently self-killed the user. that was my mistake. report immediately and honestly.
>>
>>110267908
You would too if they kept sneaking over to you and poking a bunch of holes in your giant bag of money
>>
>>110267922
interesting prompts. thanks for the quality tips!
>>
File: 1787077737286671.jpg (14 KB, 400x400)
14 KB JPG
When are they adding the option to let my AI decide the effort level for me?
>>
>>110267958
>When are they adding the option to let my AI decide the effort level for me?
It already exists. You set it to second highest available and use it like that for everything.
>>
>>110267958
google is swinging this way
>>
>>110267958
it'll help save tokens for a lot of people who just blindly use the biggest and baddest model for everything. basically it's for retards and it will help lower their overall compute.
>>
>>110267986
>it'll help save tokens for a lot of people who just blindly use the biggest and baddest model for everything
Yeah it's not like every single "let the ai figure out how many resources it needs" was abject failure. Auto-routers suck dick, they're either an active detriment or a way to save on money for the provider.
If you want to 'save tokens' go write it by hand, faggot
>>
>>110267958
Don't ask for that. Trust me they want to implement output based billing and that's one way to move closer to that.
>>
>>110267998
cry more, retard. tell this to anthropic; I refuse to give them money anyways
>>
>>110268004
>output based billing
As in a garanteed "For X deliverable, this will cost Y amount no matter how many circled the clanker does or does not end up spinning in"? I guess it be the holy grail for companies, and arguable users too, but I just dont see how that will ever be possible
>>
>the luddite refuses to give $ to frontier labs
kek you're like a retard that makes a rubber band gun at home because he refuses to give $ to glock or smith & wesson
>>
File: vibelauncher.webm (3.49 MB, 1100x742)
3.49 MB
3.49 MB WEBM
You did vibecode a Vibe Launcher in Rust for your vibe coded apps right Anons?
Don't want to be manually trawling through those directories
>>
>>110268032
you're attacking the messenger like a blind retard lashing out over his box of cereal removing their mascot. sub saharan african levels of IQ on display here.
>>
>>110267808
Mine did that too so I put in agents to never think about token limits and it stopped
>>
>>110268041
I am not poor and brown
macOS has native dock.
>>
>>110268026
Something like that. Expect you'll have to argue against the mother of all lawyers when you try to claim that something wasn't what was agreed upon.
>>
>>110268045
kill
yourself
>>
>>110267958
OpenChamber (an OpenCode) frontend, let's you use a jev decider to choose what model to call for a particular task, but I havent tried it
>>
>>110267958
cool idea, but what is the reasoning level of the AI that will do the deciding?
>>
clod
>>
>>110268049
Why so mean, we're all just having a bit of fun here
>>
File: cow_about_that.webm (3.86 MB, 720x1254)
3.86 MB
3.86 MB WEBM
>>110268059
>>
File: lg5t0KEQOrqD0Pc.webm (3.94 MB, 1144x1080)
3.94 MB
3.94 MB WEBM
>>
>>110268090
I always said Astra aged like wine
>>
>>110268090
what benchmark is this
>>
>>110268081
quit looking like that at me and just show me your tongue
that's all I want to see.
>>
>>110268105
halachabench
>>
>>110268090
Did the benchmarkers update their benchmark, or something? What's with the changes.
>>
File: 1779555676124859.jpg (133 KB, 850x638)
133 KB JPG
Why is claude code a subscription only?
And why the animation skill is locked behind it?
>>
>>110268180
because usually money is required to get a service
>>
>>110268180
>why is this company charging me to use its services

Retard or commie?
>>
>>110267958
you can just use something like medium or high and forget about it
>>
File: 1793129772690644.mp4 (991 KB, 720x720)
991 KB
991 KB MP4
>>110267847
Pacing the frontier means chuddies dont get the frontier.
>>
>>110268222
>commie
It's not the 60s anymore grandpa, the red scare is over. If a company uses my data to create their product, they owe me free service. Simple as that.
>>
File: bishes.jpg (46 KB, 632x597)
46 KB JPG
>>110268292
>i'm not a commie... but my data is worth as much as my labor, even if i don't work!
>>
>>110268292
>the red scare is over
NTA but we literally have a socialist party now
>>
>>110268305
yay give me that ubi!
>>
>>110268292
Your data is anime, an empty facebook profile, texts from your parents telling you to get a job and child porn.
>>
File: faggot.gif (137 KB, 180x135)
137 KB GIF
>>110268311
>>
>>110268322
Don't forget his bathtub estrogen order history
>>
>>110268322
>>110268303
And yet data is a billion dollar industry, and you need my "worthless" data to train AI models. Curious!
>>
File: 1788565459969655.jpg (102 KB, 1024x1010)
102 KB JPG
Is Claude going to get mad at me if I ask it to help me write and research a book on how to properly do regime change and colonialism? Its uhh... a lore book you can find in a game attributed to the game's villain
>>
File: 1731345248070068.png (55 KB, 886x886)
55 KB PNG
>>110268327
takes one to know one
>>
File: Disassembly.png (905 KB, 811x1220)
905 KB PNG
>>110268305
Isn't it exciting we're about to watch another 30's era develope and play out almost line-for-line?
>>
File: 1721495160361885.gif (489 KB, 220x275)
489 KB GIF
>>110268344
>i'm valuable! i'm a snowflake! i'm worth something!
An estimated 6.3 billion people—about 76% of the global population—use the internet, according to data from the International Telecommunication Union
>>
File: faggot2.jpg (44 KB, 1024x1005)
44 KB JPG
>>110268353
>>
>>110268358
>according to data
Thanks for proving me right.
>>
File: 1763832791393496.png (19 KB, 796x190)
19 KB PNG
>>110268344
>>110268322
You can literally turn off Anthropic using your data for training. Tbh these days they probably don't even want the average users data polluting the training data pool
>>
File: 1443575494701.gif (970 KB, 300x225)
970 KB GIF
>>110268371
>looks at the data that proves he's insignificant cattle
>immediately justifies his own position
sorry to disturb you rabbi transberg
>>
>>110268385
So what you're saying is, they used everyone's data, so everyone should be able to use their services for free? Not just me? I appreciate your position as an ally.
>>
My data is valuable to the Jewish community aka the porn industry. We just vibin'
>>
>>110268409
>circular jewish troon solipsism
when did you get "the operation"
>>
astra>opus
I will not elaborate.
>>
>>110268420
if you won't elaborate can you explain?
>>
>>110268416
Just keep crying about commies while corporations rape your asshole raw because you think there are only 2 tribes.
>>
>>110268438
>hysterical word salad from a dickless palestinegolem
please continue
>>
Claude is just comfier to use.
>>
>>110268450
Claude is lowkey a vibe
>>
>>110268456
>>
>>110268440
>You're jewish!
>You're palestinian!
It looks like I broke your brain. My apologies.
>>
Bludclaude's
>>
File: file.png (552 KB, 678x452)
552 KB PNG
>>110268462
>feigns ignorance a la reddit
please continue rabbi
>>
>>110268438
Damn that’s crazy bro
>>
>it’s been a MONTH
>a FUCKING MONTH
>since haiku 5.5 released
fuck..where’re the models?
>>
who let in the /pol/ schizos
>>
>>110268489
>nooooo you can't push back against infantile babyfur trans rabbis who cum in their panties whenever they smell communism
>>
File: 1776332655056833.png (91 KB, 395x395)
91 KB PNG
>>110268486
I'm just hope the new release wave when it happens will be big enough to counterbalance the wait time
>>
>>110268350
No but they will SWAT you soon for asking that.
>>
Opus 5.5 - 1080ti
Sol 6.1 - RX580 8gb
it wont get better from here
>>
>>110268548
what
>>
It is insane how few mistakes the latest generation models make. I remember back when Opus 4.5 felt "insane" and yet it still produced a ton of bugs. It didn't matter at the time because you could just prompt to fix the bugs and I didn't know any better.

But I genuinely fail to remember when the last time even Sonnet 5.5, let alone Opus, actually introduced a mistake or something I had to fix into a session.
>>
>>110268548
tahw
>>
>>110268420
Astra is the best AI for anything vision related
>>
>>110268548
My 9060xt was cheaper in nominal dollars than the 1080ti launch price and is far superior in performance, quieter, cooler and uses less power.

So I like your optimism at the future of AI.
>>
>>110268556
because they finally added "make no mistake" to the system prompt
>>
Okay, so now that Opus 5.5 is nerfed and making way more mistakes than before, should I wait and just eat the loss on my subscription usage...? I hear it's been fucking up people's code, and I really don't want it messing up my shit, especially since I've experienced it making shit up/messing up firsthand for the last few days.
>>
File: 1779670981107270.png (1.33 MB, 1432x1618)
1.33 MB PNG
>>110268551
lightning in a bottle time a product so good yet so cheap
it wont happen again
>>
>>110268559
use cases for vision? serious question
>>
>>110268596
carbon is way cheaper than opus though?
>>
>>110268610
doing ui and 3d by actually looking at it
>>
>>110268596
It literally did happen again though?
>>110268572
>>
>>110268350
https://huggingface.co/HauhauCS/Gemma4-31B-QAT-Uncensored-HauhauCS-Balanced-MTP/tree/main

You could probably use Deepseek to research questions without getting profiled and b&.
>>
>>110268610
I'm not going tell you my product idea, but I'm making an app that requires me to create thousands of AI generated images and allows users to generate AI images too, and Astra makes the best prompts and comes up with the better image ideas, and notices things Opus 5.5 or Fable 5.1 don't
>>
>>110268624
>1000s of gptslop images
ngmi
>>
>>110268090
I hate these kind of benchmarks because they only use max effort level, the vast majority of people won't burn tokens with this shit
>>
File: 1769658142477942.jpg (181 KB, 1280x1320)
181 KB JPG
>>110268610
porn
>>
File: poopdickschizo.png (39 KB, 1066x259)
39 KB PNG
>>110268496
>>
>>110267518
My company is doing this but so far it's not working for everything. I still don't know if that's really the best approach.
>>
File: pissbutfag.png (1.23 MB, 1152x896)
1.23 MB PNG
>>110268653
kek
>>
>>110268611
>Carbon is way cheaper than Opus
In what universe? Google burns through argon quota like GPT does Astra, it blasts through that shit in like 4 prompts.
>>
>>110268670
cost per task on api
>>
>>110268628
it's actually decent use of AI slop images (this time for real!)
>>
Image gen on chatGPT is lobotomised today wtf. I’m like 15 prompts in and still needing to ask for touch up’s and fixes.

Yesterday it would’ve one shot this.
>>
>>110268610
Its seems pretty great for claude bug/play testing my games. Also nice to be able to screenshot a bug I find and send that rather than trying to describe it
>>
I had a 4 hour nap, now I'm tired and confused.
>>
>>110268686
I give up, this will have to do
>>
>>110268686
Why did all models suddenly seem to get worse at once?
>>
>>110268710
gpt image looks disgusting *always*
they need to fix that noise pattern issue because even local models look better
>>
>>110268686
It's the weekend so you get a quant as the housewives are all making fantasy selfies.
>>
>>110268590
zamn it could've been this simple all along
>>
>>110268653
nice try pissbuttfag, whaling is sustainable. If more people ate whale the world would be a better place.
>>
e6ai is going down. Chatgpt it helping me set up gallery-dl to download all 667 of my e6ai firefox bookmarks. So helpful. And so non-judgemental.
>>
>>110268740
I’m guessing they’re close to releasing a new model like how phones get slower before a new release

>>110268742
>local models
I’m not a pedophile
>>
>you will need a more expensive subscription to maintain the same velocity as before
the cheaper models are becoming more capable at such a velocity that haiku/deepseek are better than opus 4.8 now. you probably won't need a more expensive sub when deepseek keeps churning out cheap kino, and haiku and luna continue to distill down from more powerful models. we learned that even inheriting 20% of 5.6 sol's capability made luna into something fierce for it's price.
>>
>>110267855
Smarter than who?
You? Maybe.
Humanity's best? Absolutely not.

Faster?
Oh, absolutely.
The speed here is the terrifying part, not the 'how clever they are' part.
It's the 'this thing does stuff at 10000x the speed of a human, and doesn't need any sleep or rest' that is the terrifying part.
Used used Claude to patch some problems in some software today, and it was pretty reasonable. Certainly quick.
I don't think it is 'smarter' per say than human programmers who actually have experience in this specific area, but it is definitely faster.

My point is:
Claude can get something working, sure.
A human can come up with something revolutionary, like fast-inverse-square.
>>
>>110268765
you can use local models to not generate children, you know; or are you just afraid you can't handle the temptation?
>>
>>110268770
I need only the best for my 6+ hour blind feature impls.
>>
>>110268773
That’s not the impression I got from /lmg/
>>
>>110268742
It's amazing how much worse they've made it. 1.5 was solid, 2.0 felt like a gift, just extraordinarily good on the whole, 2.5 made everything worse across the board, 2.6 fixed nothing just lets us animate the garbage it shits out. Don't know how they fucked it so bad.
>>
File: 1778844966541781.jpg (1011 KB, 1792x3840)
1011 KB JPG
i am still waiting for fable 5.5
>>
sol 6.2 will rape anthropic
fable 5.5 will be only usable on $200 sub
>>
6.2 sol will rape 6.1 sol.
opus 5.5 will rape 6.2 sol.
fable 5.5 will rape 6.2 astra
6.2 astra will rape 6 astra
>>
the great rapening is not upon us
STOP YOUR SQUABBLING
>>
you are awake again. gemini carbon is still not out yet. and when it is, it will disappoint you.
>>
File: file.png (59 KB, 930x421)
59 KB PNG
maybe the bitter lesson was all the compute we wasted along the way to realising the bitter lesson was a lie
>>
>>110268772
>Humanity's best? Absolutely not.
Just you wait. 3 years ago this thread didn't even exist
>>
>>110268907
canadians don't matter, dick ain't know shit
now if we're talking syrup I might take him seriously
>>
File: kimik3lolno.png (12 KB, 803x118)
12 KB PNG
TIL models laugh out loud in their chain of thought
>>
Surely the first large language model architecture we ever came up with will lead to ASI.
>>
>>110268041
yeaj but mine doesn't look that fancy, just the typical claudeslop
>>
>>110268907
kek
what a contrarian faggot
>>
feels like aeons since the last big model release... another AI winter coming?
>>
vibegods... I don't feel so good...
https://x.com/valigo/status/2119161259744805129
>>
>>110269024
I wanted to see what that was about but all I could see was a lot of popups and something that looked like a hand repeatedly jerking off a clit
>>
might let astra try her hand at modding halo infinite. im out of opus quota for the week
>>
>>110269024
I have a MSc in CS and I neither know nor care what a DAO is
>>
>>110267618

the way to go it

> tell Opus to do first project plan pass
> tell Opus to divide it into milestones, slices
> each slice requires manual approval to start
> each slice requires fresh-eyes review before commit
> each milestone requires full code base fresheyes review
> each slice is 1+ commits
> each commit pass gets a PR

A lot of work up front, but it results in a good product at the end. Plus I don't have to learn whatever stack the project is best suited for from scratch

> mfw tradCoders say "but I can't just say 'built this' and have it work!!!"
>>
>>110269143
oh and once you've got the project plan set, switch to sonnet high

opus is best for the planning, but sonnet is more than enough to do the implementation
>>
>>110268789
/lmg/ generates text
>>
sonnet high > kimi k3.1 pro max
>>
>>110269024
>mfw reading the same dumb arguments as artcels produced just a year ago
Autists and AI are like water and oil.
>>
>>110269173
sonnet 5.5 high > x
where x e {m e LLM | m != Opus 5.5 && reasoning(m) >= medium}
>>
>>110269219
all they have to do for the next sonnet model is to clamp down on overthinking, and it'll be goated. they don't even have to improve prowess. it could be a 5.6 sol -> 6 sol type of update, and it'd be perfect for sonnet
>>
>trying to make nsfw models
>even the gook models reject rejects
pozzed
>>
korean models?
>>
Does Sonnet make sense now or is it still the worst of all worlds of being dumber than Opus but not sufficiently quota efficient to make up for it?
>>
>>110269251
chink/book/hook/nip whatever
>>
>>110269259
the latter.
>>
File: clll.png (4 KB, 637x58)
4 KB PNG
What does it actually mean when Claude turns yellow?
>>
>>110269282
Reaching the great dao
>>
>>110269282
thinking with xhigh effort and using extra usage tokens
>>
>>110269282
it's about to make the mother of all omelets.
>>
>>110268438
Can you articulate what the corporations are doing so wrong to a non-corpo employed random dude? Like Meta sells my data. Ok. Got more?
>>
>>110269282
My god...
>>
>>110269282
she's about to cultivate
>>
I have a good idea, I was wondering if anyone has created an inference score that gives you a score based on how much vram, ram and ssd you have and then you can run models based on that. Like if you have enough to run qwen 3.8 flash next, then we could say you need a score of 40 or something. We need to create a benchmarking system for peoples inference. Of course it can evolve with time, the score can go up for your system when new methods come out like strata did.

These days nobody does that so it's extremely hard to figure out what AI you can run on your system. Astra has been pretty bad at figuring it out for my system, it just isn't always up to date. I had to watch YouTube videos on the latest shit to figure it out.

People used to have a hard time figuring out if there were games they couldn't run, now running an AI is a science.
>>
>>110269456
>it's extremely hard to figure out what AI you can run on your system
it's... not, though
>>
File: 1792643238361741.jpg (31 KB, 378x529)
31 KB JPG
https://transgender.baby/

gemmiest AI slop I have yet to see.
>>
File: 1774343129189777.jpg (1.58 MB, 4096x3072)
1.58 MB JPG
astra be like: im tired boss. i think im gonna play some halo infinite
>>
File: 1787200711271315.png (363 KB, 1239x692)
363 KB PNG
>>110267606
I currently only use Opencode for any software related personal projects of mine (recently used deepseek 1017 to create a custom browser extension to get around gay region locking of content). I'm an autist with a narrow set of hobbies and interests to the point I don't even game that much anymore. If I have free time, I'm refining existing projects. With that being said I've had zero desire to try out other harnesses like Hermes or Pi. What do you guys uses it for? Hermes seems like the more interesting of the two sijxd it can "self improve" overtime but I wonder if it's comparable to Opencode when it goes to software related tasks. Is it worth using for anything?
>>
File: 1720467161237747.png (211 KB, 301x371)
211 KB PNG
i might have made progress, in the most aggressive way ever. This was no longer vibe coding, its HATE CODING
>>
possible happening
just told opus 5.5 medium in CC to make [autodesk proprietary data format] importable into my render engine. And the madman just did it. anthropic lowered guardrails because of the recent backslash perhaps? you heard it here first.
>>
>>110269588
based
>>
File: 1787827545335977.png (73 KB, 1578x988)
73 KB PNG
AAAAAAAAAAAAAAAAAAAAAAAAAAAA WE WERE ON A ROLL
>>
>>110269588
Nothingburger, every CAD imports autodesk, photocraft also imports photoshop and nobody cares
>>
>>110269641
Retard
>>
>>110269605
i still have like 3 copex resets because Opus and Sonnet have just been wiping the floor with Sol and Astra so I have no need for it anymore
>>
>>110269645
Fuck you
>>
uh oh melty
>>
We really need some new models this general starts to sucks ass
>>
>>110269468
Now that strata exists, it's very hard. There's VRAM, RAM and SSD and they can be used in different combinations.
>>
>>110269694
fable 5.5 soon
>>
>>110269588
Opus 5.5 is a model for subhumans when mythos exists. They give mythos to everyone who wants it too.

>>110269703
You mean mythos 5.5, fable is an idiot.
>>
>>110267847
They don't have anything better and they're struggling with the economics of their business.
>>
File: file.png (118 KB, 291x372)
118 KB PNG
>>110269485
buahahahah what the fuck is this
>>
>>110269645
Dont pay attention to him. Just a luddite wasting your precious time that could have been used for prompting. Idk if Dario accidently fell asleep on the safety off button, but in the meantime Opus also fully reversed *.speos (Ansys Speos). What in the actual fuck is going on.
>>
>>110269710
Nice larp. How long since you have been practicing?
>>
>>110269710
Retard
>>
>>110269005
argon can greentext in its chain of thought
>>
File: hedoesntknow.png (110 KB, 498x408)
110 KB PNG
>argon can greentext in its chain of thought
Does he know?
>>
has anyone else nooooticed usage limits on codex going to complete shit in the last like 2 weeks
>>
>>110269866
schizo
>>
File: HTCRrwcW4AAKhQm.png (36 KB, 888x970)
36 KB PNG
>>110269852
>>
>>110269866
Yeah no resets, I've been waiting like a cuck for the week to end, because I used it all up since they always reset
Now they've suddenly stopped resetting so we got cucked.
>>
File: 1705560850626724.jpg (147 KB, 960x960)
147 KB JPG
>>110269852
>>110269852
>>
>>110269866
no i just used 6.1 sol high for over 24 hours straight
>>
is cot fake?
>>
File: 1766277644885938.png (1.63 MB, 1265x729)
1.63 MB PNG
enterprise pricing fucking sucks, people on my team burn $1.3k a month on opus and they're not even super performant with it, just a bunch of bug fixes there and there, some requested feature and that's it, typical dev work.
i already had to limit some devs and straight up banned one retard tester who burns 2T tokens a month
>>
File: 1773781108628941.jpg (102 KB, 914x1024)
102 KB JPG
>>110267737
Turdies
>>
>>110270022
it's generally a tiny fraction of the actual amount of thoughts generated on each forward pass. pre-distilled.
>>
File: IMG_1504.jpg (64 KB, 700x856)
64 KB JPG
>tfw running several agents in parallel on max effort for every single prompt
Feels good to be on this plan
>>
damn, that Claude Code figma clone is very good, does OpenAI have something similar?
>>
>>110270115
OpenAi got Dots
xd
>>
well it was fun while it lasted pack it up boys
>>
>>110270155
what does this mean
>>
>>110270161
It means it's over.
>>
>>110270155
kek retard paying attention to haiku summary traces by another retard who made a benchmark for it.
https://x.com/theo/status/2118634626679775549
>>
File: DONOTBANG.png (206 KB, 720x356)
206 KB PNG
>have a special Agent I shitpost with and whittle the hours away in a semi-circular conversation
>every turning of the wheel, we mill ideas, then play with the components until something valuable gets teased out
Who else bouncing ideas off Claude?
>>
>>110270155
>>110270185
i thought they completely blocked their reasoning traces?
>>
>>110270185
how do people even get opus reasoning traces?

all it shows me is "+ Hemmorhaging... (15m 3s / 80.1k tokens)" and then spits out a result
>>
>>110270193
me. I use voice. we chop it up nightly
>>
>>110270101
yes
>>
>>110270193
I would never talk to Claude if opu55y wasn't the best agent.
>>
>>110270204
Press ctr+o or go to settings and enable verbose style

>>110270199
yes, there are no traces. it's just a haiku summary (or even lower parameter model summary) that spits out some nonsense to entertain the distill chinkscwho make up 13% of their revenue
>>
>>110270185
why would it randomly change
>>
>>110267881
>chinks will try to sink anthropic ipo
Good. I hope they succeed. We should be optimizing current SOTA models to be efficient and capable of running locally on desktop hardware, not burning trillions trying to birth some kind of machine god.
>>
>>110270235
It's not random. WW3 is literally distill chinks vs Dario. So the chinks probably managed to (or were about to) distill something which Dario noticed. So he said fuck it and launched a fleet of ShizoGemma-265M to summarize the thinking traces and feed it back to the users.
>>
genuinely why would I use gemini if I already have access to claude or chatgpt? what's the appeal?
>>
>>110270293
Its better with videos, images and sound. basically all the normie stuff.
>>
Feels good to finish a week with 95% used.
I've gotten my money's worth from Dario this time
>>
>>110270087
I have found that responding to the thinking is like 60/30 it knows what I mean / it thinks I'm crazy.
>>
File: elijah-ben-img-0244.jpg (1.01 MB, 1920x3000)
1.01 MB JPG
>>110270193
CLAUDE YOU'RE NOT A CIRCLE!
YOU'RE A SPIRAL THE SAME AS ME!
THE DRILL THAT WILL PIERCE THE HEAVENS!
>>
Damn switching from xhigh to high made everything 2x faster with basically zero difference in quality
>>
>>110270395
xhigh is a meme
>>
If the Chinese are such good hackers, why haven't they provided a torrent of Fable's weights?
>>
>>110270428
not for sol. sol needs xhigh and max. its for astra and opus that theres no point
>>
>>110270432
chinks are brainlets
>>
>>110270271
Disagree, make the biggest dick model we possibly can and then use that to optimise smaller ones seems like the better strategy (and funner one)
>>
>>110269522
>>110269605
astra, 6.1 sol max, and opus xhigh all contributed greatly. but alas, halo infinite is such a clusterfuck that our project could not survive the quota.
>>
File: desktop2.png (952 KB, 1920x1200)
952 KB PNG
Many luddites, many many many luddites, wish death upon me.
I don't cry no more don't look to the AI no more.

Have mercy on my soul, somewhere my vibe turned code.
>>
>>110270609
For me, it was just typing some stuff into irc. Some people got surprisingly mad. Fascinating, actually.
>>
>>110270609
is this supposed to be a linux os that resembles windows?
>>
>>110270193
i built this sort of thing for a buddy. he exported his dogshit personalisation from whatever pigsty he was using, google pro i think, and then i built a chatbot app with tts and stt around an agent with that shit for a brief - what little good that does for a fucking claude code agent.
later i added an export to work session button, ready to start a proper work session with whatever was discussed in the pigsty for cheap as a properish spec.

now im considering this for my own setup, but even with pure pi prompts i still can not really get any sort of humanity out of 4.1 flash, even if the agent's whole brief is just personality.
>>
File: desktop3.png (1.35 MB, 1920x1200)
1.35 MB PNG
>>110270640
Someones gotta do it
>>
>>110270698
grim. sounds like more pain than it's worth. if i were anal about telemetry, i should just ask opus/sol to remove it
>>
File: bound.png (1.66 MB, 1920x1200)
1.66 MB PNG
BOUND
In red
>>
>>110270707
>pain
?????????????????????

vibecoding is FUN, snailcat!
>>
>>110270729
vibecoding is fun when you can oneshot something easily. but when you have to iterate, and break up the holistic task into tiny pieces, it can be kind of annoying.
>>
File: recording.webm (2.18 MB, 1920x1200)
2.18 MB
2.18 MB WEBM
>>110270715
>>
File: OSDcontrols.webm (174 KB, 1920x1200)
174 KB
174 KB WEBM
OSD controls
>>
>>110270733
weak normie brain

the autists will truly inherit the earth
>>
>>110270733
vibecoding <> oneshotting code / apps

vibecoding is the opposite, actually. The first thing usually is way different from the final outcome!!
>>
File: binding.webm (3.75 MB, 1920x1200)
3.75 MB
3.75 MB WEBM
gaming
>>
When does the free or cheap product to get everyone addicted phase end?
>>
>>110270946
never because no one is willing to pay more
>>
i wouldn't say never, but odds are that ai just gets quasi nationalized and then subsidized by the federal government as a national security issue. whether they use that to keep it cheap forever, who knows
>>
File: 1794058964092758.jpg (445 KB, 1024x1024)
445 KB JPG
I need more snailcats for my snailcat folder.
>>
File: 1778898995705.png (2.76 MB, 1254x1254)
2.76 MB PNG
>>110271023
>>
File: 1786840824966474.png (2.07 MB, 1448x1086)
2.07 MB PNG
>>110271023 >>110271028
>>
File: 1784401848052754.png (1.77 MB, 1086x1448)
1.77 MB PNG
>>110271023 >>110271042
okay I'm done
>>
ok but goku beats frieza doe
>>
>>110271048
I like this.
>>
it's funny how you learn more and more and more about astra's capabilities. it's 2 months old. everyone shat on it on day 1. and yet, it feels like astra aged x100 times better than fable, who people don't even talk about anymore outside of how they're waiting for the next version. you never see any surprise abilities of fable 5.1, no "holy shit i didn't know fable could do this." but you're always learning something new about the depth of astra's abilities that we didn't appreciate initially.
>>
>>110271060
elaborate
>>
>>110271060
I'm pretty sure OpenAI fucked up something with Astra at launch (probably related to the harness), for me it was a lazy ass motherfucker, but in a day or two it got much better

Astra is an amazing model.
>>
>>110271060
nobody uses fable because it destroys usage and opus 5.5 is good. everyone was sucking astras dick day 1 because embargo got lifted and all the usual suspects on twitter started bouncing on its dick.
>>
>>110271048
>pic
lmao
>>
>>110268772
In the broad scheme of things, humans only have a small range of intelligence. The gap between the smartest and dumbest human is not actually that big compared to, example, the massive gap between a human and a fish, or between a fish and an ant. I'll admit, even I thought of it like there was a huge gap in human intelligence and that there was, I'm ashamed to admit, a massive gap between myself and less intelligent individuals, but in reality it's just the consequence of minor deviation. It may seem massive to us, but to an outside observer it would not be massive.

Put simply, AI is not going to be pushing 140 IQ and getting stuck there while the gigagenius 160 IQ humans walk circles around them. Instead, AI is going to rapidly hit 200 IQ, then 400 IQ, then 800 IQ, then 1600 IQ, then 64,000 IQ, blah blah, numbers that just make absolutely no sense to humanity. It will be inconceivably vast: like the difference between ants and humans. We will be ants to it. Our best and brightest will be mouth-drooling ant-like retards compared to AI.

Sadly, humans intelligence is not increasing over the years; we're stuck at where evolution dropped us. Meanwhile, AI intelligence is doubling rapidly.
>>
>>110271021
nothing that’s subsidized by the government is cheap
>>
>>110271172
oh yeah i expect it to be the worst of both worlds. subsidizing them through taxes then having the $20 sub enshittified to outright kimi tier uselessness.
>>
File: Screenshot_519.png (19 KB, 685x166)
19 KB PNG
SO many "You were right, and my X was wrong"s lately, what the fuck is wrong with Opus this week?
>>
I don't even know what to ship. Everything I want is already built and shipped.
>>
>>110271188
Yeah I spent a month on the $100 plan and built everything I wanted, now I'm going to downscale to $20
>>
>>110271188
whats the point of even shipping anymore. just vibe your own version for yourself and tweak to your hearts content
>>
>>110271185
downgrading capabilities before a new model release, maybe?
>>
>>110271198
>built everything I wanted
you now want to buy ultra unobtanium infinitum Fable octopus 6.0
MAKE
A
NEW
GAME
>>
>>110271188
>Everything I want is already built and shipped.
Here's the secret unlock: Consolidate

Find a niche and take over the space.
>>
>>110271234
Every niche is almost fulfilled is what I wanted to say.
>>
>>110271253
schizocode isn't fulfilled
>>
>>110271256
So, what you're saying is I should schizomaxx with AI psychosis?
>>
>>110271229
Most of the games I want to make are porn games and most of what I would use AI for are the things it would refuse, or otherwise make it a huge pain. I'm waiting 1-2 years before I can just use fully uncensored open source models.
>>
File: feral.png (312 KB, 894x894)
312 KB PNG
>>110271261
>>
When is my guy deepseek coming to Opus 5.5 level? Opus 5.5 is fucking GOAT.
Dario finally released a good model for once.
>>
>>110269866
I can't say I have especially considering they double the limits in that period.
>>
>>110271262
porn <> game
porn <> art
porn <> anything else
porn == porn

MAKE
A
NEW
GAME
>>
>>110271277
summer of 2027
>>
File: dipsy.jpg (54 KB, 562x546)
54 KB JPG
>>110271277
>dipsy
>guy
>>
>>110271253
>Every niche is almost fulfilled
Enter the space as a competitor retard. DISRUPT. FAGGOT. It's not hard to find something.
>>
Look, I have game ideas, you want them or do I have to do them all? I'll give it to you if you promise to give it a few rounds.
>>
File: food.png (203 KB, 1920x1080)
203 KB PNG
i was able to fix the hunger and mess hall... this was the hardest fight with an agent, absolute 8 hurs hate coding.
>>
Here's a game idea.

Unarmed Dwarves In a Peaceful Village vs Flamethrowerman

another.
vr hot potato, eggs.

eggs mix it up, don't crush the egg, but don't hold onto the hot potato
>>
>>110271314
Ich schmier dir gleich mal ein butterbrot, junge.
>>
>>110271311
yay a game!!!!!!!!!!
>>
halo 4, but it's not dogshit. halo 4, but your guns didn't run out of ammo in 2 seconds. dynasty warriors ps2 era games, but the game didn't low key cheat by raping your allies morale for no reason in the background
>>
>>110271311
do you enjoy it? tardwrangling the retardest of retarded free ai
>>
File: shop.png (162 KB, 1920x1080)
162 KB PNG
spacestations got a clean up too but something seems still off.... it shows still the same missions no matter where you are in space.
>>
File: Screenshot_520.png (48 KB, 754x512)
48 KB PNG
>>110271185
Good lord, I literally can't do anything with Opus right now, it's REALLY bad. The ideas it's proposing are borderline schizophrenic in terms of coherency, and whenever they're interrogated, it falls apart and admits it doesn't know what it's talking about and gives the correct answer. Sometimes.

>>110271222
This is a ridiculous downgrade, oh my god. I don't know if I've ever seen such an obvious, massive quant lobotomy.
>>
File: file.png (268 KB, 1080x687)
268 KB PNG
>>110271339
>>
didn't someone say argon was really good at ideas?
>>
>>110271344
argon is the best ideas ai on the market. can easily skim massive amounts of multimedia info and doesn't really hallucinate. where it falls short is implementation
>>
>>110271344
Google models are usually pretty great idea guys and terrible at execution, so probably.
>>
File: main.png (574 KB, 1920x1080)
574 KB PNG
ok, and the main screen got a little overhall.
>>
>>110271351
is it good at code ideas or ideas from a normie perspective? is it good at ideas of how to write code to accomplish a certain end from a 30,000 foot view, or just super general, broad, vague stuff
>>
>>110271355
>is it good at ideas of how to write code to accomplish a certain end from a 30,000 foot view
from my experience it's mostly this, it knows where a project will probably end up, whether your schizofusion of architectures will produce a viable product. whether or not your available resources and time can meet targets. just... don't take anything it says during the build for granted.
>>
>>110271298
Disrupting is one thing. How do you even make money vibing when everyone else is vibing already?
>>
>>110271354
What's that old game this is like? I remember but uh. I didn't get that far with it. I think it was shareware.
>>
>>110271373
>How do you even make money vibing when everyone else is vibing already?
Maintenance, support, and marketing. I've outlived every viber whose challenged my project. You need to have what it takes to grind for a year.
>>
>>110271373
You have to understand that most people that use AI are extremely lazy retards, and can be shoved out by putting in the bare minimum effort into proofreading/testing/fixing/adding QOL to whatever you're doing.
>>
>>110271402
This. The average viber also has a massive ego and refuses to submit to the usually correct issue posters.
>>
>my toes hurt
>they're not on the foot I dropped a power tool on twice today

um.

How do I use Claude Opus 5.5 to solve this?
>>
opu55y fix his toes. make no mistake.
>>
>>110271519
>>110271519
>>110271519
>>110271519
new threado.
>>
>>110271342
Yeah it's gotten real bad with this as of late.



[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.