File: file.png (1.9 MB, 1448x1086)

1.9 MB PNG

Don't be a snailcat.

>>109429763
Notice how the snailcat is happy.

I, for one, am adamantly against snailcat abuse. They're future paying customers.

>>109429768
Was* happy. The snailcat was happy.

File: ai-job-replacement-stonet(...).png (156 KB, 1000x500)

156 KB PNG

new 'toss

>>109429750
aww cute

>>109429796
kek

>>109429796
the fucked up teeth are a nice touch

openclaw guy is live if anyone is interested
>https://www.youtube.com/watch?v=gKdeLQd_LIQ

>>109429750
claude needs to torture luddites

haiku 5 waiting room / gpt 6 luna waiting room / glm 5.3 waiting room / qwen 3.8 waiting room / deepseek v4 pro ga waiting room

Logo of a jeet asshole

>>109429670
>Are you prod of any of the projects you vibe coded?
The game I'm making, yes. It isn't vibe coded but AI has sped up and improved my development significantly. And now I have a foundation with which I can make other games too.

The actual vibe coded things, I'm happy they exist. Making native programs instead of electron shite is always a joy, even if it means telling an AI to reinvent the wheel.

>>109429921
he's copying an old steve jobs speech

>>109429939
The same one like 3 other big corporate AI people have copied in the past year. Are they all asking AI to write their speech, and more important are all all LLMs taught to suck-off Steve Jobs posthumously?

>>109429936
h-how do you know??

File: 78A7BF30-FCB7-446B-8840-5(...).jpg (1.29 MB, 3464x3464)

1.29 MB JPG

>>109429750

File: 162233678971.png (383 KB, 598x628)

383 KB PNG

I just realize I can use gemini to explain my problems to deepseek agents.

File: 1755959746066281.png (287 KB, 1362x1239)

287 KB PNG

Leetcode is now adding hidden problem metadata to contest questions, which only AI can see.
So if you feed the problem statement to an AI, it will input fake variable names like "velmoritha" into the solution, which leetcode then flags and nukes your account.
Half the leaderboard in the recent contest is gone.

>>109429987
i do this by taking a screenshot of the problem, asking gemini to re-write my prompt to better describe what's actually happening than i can, and then hand it off to luna

>>109429750
what coding harness do you guys use?

my employer has licenses for cursor which i've really liked working with but i'm trying to find a free harness (ideally FOSS if possible) that doesn't lock me to a provider for my personal projects.

i've been reading a lot of good things about opencode and pi. what do you guys use?

>>109429994
>jeets are the ones not using ai
'dites won't like this one

>>109430004
it's actually the opposite. if there are no timestamps in your ranking, you got banned. so first place "Rishabh Patel" got his account nuked for cheating.

File: 6163613451.png (2.98 MB, 1280x1456)

2.98 MB PNG

any webdevers in here that can chime in on what models they use for what? like specifically what you use for frontend/backend. I really want to make my shit good & clean. I've been using claude sonnet for some simple UI stuff and opus for some of the overarching logic and design. I do wonder if some of the GPT models are better in this regard but I've never used them

>>109430004
retard

File: 1769813155277633.jpg (38 KB, 1080x564)

38 KB JPG

>>109429997
i'm using cline, it's really good at saying shell integration unavailable, which i believe means it's working at 110% power

GPT 5.6 Luna or DeepSeek V4 Flash or Grok 4.5 or MiniMax M3

>>109430024
>>109430015
I accept your concession

>>109430032
if you have codex, luna. if you don't have codex, v4 flash

>>109430032
For what? And take MiniMax M3 off that list there's no use case for it currently.

>>109430028
basado

>>109430037
i just got a chatgpt subscription a few days ago and downloaded codex, please redpill me on luna vs terra

i know that sol is supposed to be their opus-type model and i've been told that terra is good for most dev/debug work, but is luna actually usable?

>>109430062
Luna is your workhorse, Luna is extremely capable, it's better than the best frontier models were 6 months ago.

>>109430068
i see, i'll give it a shot. honestly i'm skeptical though, luna's pricing seems too good to be true for a decent coding model

>>109430062
uh, sol is their opus type model, sol is their fable type model, and it even manages to remain neck in neck with opus 5, which beats fable.

terra is their opus type model. luna is their sonnet type model. it's a good workhorse, but it's good when a better model lays out a plan for it. it's not good at starting from scratch and going from there.

>>109430044
uhm i only use it to translate epubs to spanish and format them well, also some paradox modding and that's all

>>109430093 (samefag)
sol isn't their opus type model***
terra was their model to challenge opus 4.8

Sol is a little bit more lazy lately

>>109430093
>opus 5, which beats fable
lol lmao

is DeepSeek good as they say?

>>109430255
I don't know

>>109430255
Nope it's shit.

>>109430255
not quite, but if you're running local or cheapmaxxing it's pretty great for what it is

File: 1780416372748636.png (149 KB, 876x1063)

149 KB PNG

>>109429750
i bought ChatGPT subscription and i used all my weekly allowance today...

>>109430334
why, what happened

File: file.png (557 KB, 900x807)

557 KB PNG

fellas, i've been reliably informed that all the software the clanker has built for me these last few months has been a lie

>>109430334
dw tibo will reset it

>>109430343
well i used it

yeah i'm into cheapmaxxing, which model do you recommend?

>>109430365
v4 flash

File: puchiko_dance.gif (1.31 MB, 216x366)

1.31 MB GIF

>>109430363
congrats then! hope you got your money's worth
I'm cheering you on nonny

>>109430365
you're brain

File: 1765703618782646.png (11 KB, 434x127)

11 KB PNG

>>109430379
It slowly takes place, and i hope at the end of the month i will have workable demo

>>109430390
how much?

>>109429750
>want to help mom with a thing
>3 months of development and going
I hate myself

>>109430433
Does she not pay you hourly? Oy vey...

File: 1756511174689682.mp4 (2.5 MB, 720x1280)

2.5 MB MP4

>>109430394
im going to sleep, since i have like 10% left on each claude and chatgpt, i will vibecodemax tomorrow

have a good vibecoding fellas

File: file.png (1.09 MB, 960x1306)

1.09 MB PNG

>>109430433

>>109430352
I hate to give attention to this guy, but what did he ever do? Is he a Theo or Primagen level influencer? Is there any reason to pay *any* attention to what he keeps saying? Why does the media give him any attention?

>>109430503
I should have googled it before, but he was (is?) CEO of a PR firm. Makes sense now.

>>109430447
milk truck arrive

>Board-watch has been crash-looping since 03:21 this morning. Let me pin down the bad record.

is codex really better than claude code? i ran out of usage in claude code so i can't ask it

>>109430616
The truth is both are very good and it's really just down to who has the better model for your use-case in a given week.

>>109430616
i think codex has the better family. claude code is very top heavy. there's precious little use for anything below opus 4.8

>>109430503
>>109430519
Being anti-AI is a good grift. It doesn't matter if you're wrong, if AI succeeds you can just change lanes and pretend you were always onboard, if AI fails then you can gloat, it's a win/win for grifters

>>109428915
>>109428956
>>109428956
"Subagents" can be thought of as different mini chats within your same session. You start up the chat in order to tell the llm to do something. It's very good at its job but can only do one thing at a time. When you're doing this you're calling the API to use a particular model. With subagents you can have the model call other models to do different tasks in parallel so a task that would have taken like an hour for one model to do by itself can be done in like 10 minutes or less depending on how good the sub agent models are, how many are being used, and how many the main model is spawning at once. This is useful for vibe coating specifically because a task you give it that will require multiple changes can be broken up into different stages so the model can call the API multiple times to spin up what are essentially chat sessions within your main chat session. Moonshot's API service advertises something called an agent swarm that allows you to use hundreds if not thousands of these but this is only worth doing if you're messing around with a gargantuan code base. The more complex the code base is the more useful subagents can be. The main model can also have them be the same model or different models depending on how your harness is configured and whether or not the API service you use allows you to use multiple models at once. The obvious downside is that if too many are being used at a time you can hit rate limits very quickly. Whenever I instruct a model to do this via open code it typically only uses four at a time Max and I think this is a default safe amount that open code bakes into the system prompt in order to make sure people don't get rate limited to hell whenever a model decides it needs to use sub agents


TLDR: if you're building a house you're not going to do that shit by yourself right? You technically can but why not have help?

>>109430433
What's it supposed to do? If it's taking 3 months it must be something complex

>>109430660
Yeah, there are definitely grifters on both sides though.

why do agents love to wedge themselves

>>109429997
opencode, hermes agent or pi (or oh-my-pi) seem to be the most popular. choose whichever one best fits your own personal flavor of autism.

>>109430363
>>109430343
He used the Hermes.

>>109430503
he's a nobody. worked at a pc magazine for a while.
wrote a tech newsletter while running a pr firm.
then ai showed up and he was very early to the 'skeptic' train - and back then it was easier to do because the models were actually dumb so saying
>haha these things will never do useful work
was plausible. then claude code happened and hey tried to downplay it like 'oh it's just code'
in the intervening years he's basically networked with a whose who of left-of-center political influencers and media types so he is basically the single most prominent source of information about ai for the left
none of the poeple who talk to him to use ai and he hit every major podcast every so often and just lies to the hosts who nod along
dude's going to go down as a historic grifter

>>109429796
suitcase, first panel

>>109429994
>Leetcode
Isn't this a meme for pajeets, to begin with?

>>109430786
literally every single software job that pays over $100k will ask you a leetcode question at some point in the interview process, in some form

>>109430797
>software job
Thought so.

>>109430444
That's would make me one of them heh
>>109430698
Yes, it's basically a stateless (serverless?) and simplified N8N hosted on Apps Script, so I get a free but slow infrastructure to build automations (also it's natively integrated with Google Drive). I'm building it with Typescript which gets converted to the subset of Javascript that Google provides there

Anyway, I'm not far from finishing this project

>>109430814
>having a job is for jeets
what level of being raped is this

How do other vibe coders deal with poop? I know I have to poop, should I go poop and then come back and see what happened? I feel like I need to be there. Maybe I should bring the laptop to the toilet? But what if it hears me flush?

>>109430757
no i used codex default harness

>>109430878
>he doesn't have hermes agent hooked into a telegram bot
ngmi

>>109430834
RAPE MANIFEST
==============
SECTOR: H
LEVEL: 1
QUADRANT: B

Solution: send them back

>>109430900
>codex default harness
well, post the token consumption for that too.

no no nooooooooo did they just reset again?
seriously? now?
the 8th of august now is the next reset. nooooooo nooooooooooooooooooo, it was supposed to be the 5th
god they better reset it again during the week fuck

I hate pay-as-you-go. What's the best universal sub?

>>109431108
it was last night, m8

>>109431152
oh
it's because i didn't use it till now
i was asleep for 14 hours

this is even worse now

File: frog.jpg (29 KB, 474x474)

29 KB JPG

Guys DS flash might actually be good.
I weirdly enjoy paying per token. No reset anxiety. I spent like 30 cents so far and I've been slopping all weekend. Hard to imagine what $200/month could give me.
Quality wise it's on par with Opus 4.5 I would say (which was called insane 6 months ago).
And holy shit it's FAST. I have an idea and it just does it in 20 seconds?
Have to do some further investigations. But if you even hardly look at the code this might be a great model.
If you want full agentic "build me X" it's probably not there yet and Sol, Fable, etc. are still better.

>>109431161
get codex to write you a little thing that checks for this every hour or so and just sends a hello

How can I get AI to be more funny?
I coded this AI Characters Livestream where they're supposed to be funny, but I don't think they are. Or are they? I've watched so much of it I am not sure anymore. I remember I thought it was funny in the beginning. It is Live right now.
https://rumble.com/v7dl30c-unhinged-ai-characters-argue-about-4chan-posts.html?e9s=src_v1_upp_a
It has Epstein in it. And Hitler. And 30 other characters. I'm thinking of turning it into a podcast hosted by Epstein and Hitler and they invite a guest character in every episode.

>>109431199
TDLR: If anyone knows how to get AI to be more funny please tell me. I am using DeepSeek v4 Flash currently, so if one LLM is known to be more funny than another please let me know. But the problem with GPT/Claude is probably their guard rails preventing them from saying anything edgy.

>>109431199
>>109431211
https://arena.ai/leaderboard/text/creative-writing

>>109431206
good morning sar

>>109428326
>Priciest day was Jul 29 at ~$991.79
do people really do this?

>>109431199
i noticed my 6 viewers all left when the characters started dropping n-bombs and calling each other kikes. pathetic

File: JSMJ.png (839 KB, 1024x1024)

839 KB PNG

>>109431206
can it fix the mouths?

I can't tell if my bespoke harness degraded of the models themselves did.

>>109431227
deepseek is rank 37 yet only 50 points behind fable 5, making this chart out to be meaningless unless every 10 points is a huge gap or something

File: file.png (1.24 MB, 1254x1254)

1.24 MB PNG

>>109429637
Faster now

| Cache depth | Mine | llama.cpp ROCm | llama.cpp Vulkan |
|---:|---:|---:|---:|
| 8 | 286.34 tok/s | 187.74 tok/s | 74.87 tok/s |
| 128 | 284.49 tok/s | 185.32 tok/s | 74.93 tok/s |
| 2,048 | 255.65 tok/s | 147.81 tok/s | 28.54 tok/s |
| 8,192 | 198.75 tok/s | 89.23 tok/s | 67.49 tok/s |
| 32,512 | 106.36 tok/s | 30.45 tok/s | 51.19 tok/s |

For reference again:
>windows, rocm backend, 9070xt, qwen3-0.6b in bf16
>>109429648
Thanks anon

File: 3hmhp7ochzeh1.jpg (54 KB, 1080x1031)

54 KB JPG

BrBros I've been calling doom callers of the AI bubble a bunch of inbred retards since the beginning, but, but what if they are right?

>>109431426
There IS going to be a pop that kills all the excess, but AI will not cease to exist.

>>109431459
This. People misuse the word bubble often and I WILL NOT STAND FOR IT.

File: file.png (66 KB, 919x616)

66 KB PNG

I had bad few days so I'm happy with the result so far, near roofline at long context is pretty cool

So yall are talking about using these models day in day out and showing your token usage and stuff.
But what do you actually DO with them? Like what are you working on right now?

>>109431510
>what are you working on right now?
my harness and desktop

File: FwjTxBQacAEly0W.jpg (38 KB, 749x735)

38 KB JPG

Phew guys turns out good ideas are actually not easy to come up with and asset creation is still more manual than I would like

File: file.png (47 KB, 811x572)

47 KB PNG

>>109431510
Inference engine and cyberpunk quest building system

>>109431417
Keep bringing the goods, anon, love it.

>>109431510
I have a series of fully automated Youtube channels with agents that research topics, write scripts, produce generic motion graphics and overlay text with royalty-free stock footage and AI generated images, include bespoke music with ACE-Step, edit it down, proof-watch and review the videos, at which point they're either sent back upstream with notes or if approved then it's uploaded on a schedule.

>>109431593
jesus fucking christ

>>109431593
i think they should kill you

Opus 5 is incredible. This is all.

>>109431593
post MRR

>>109431593
Extremely Indian

File: file.png (47 KB, 1260x493)

47 KB PNG

>>109431636
This is my most successful channel.

>>109431662
>didn't create a revolutionary 4D perspective game with a 3D cube of viewing magical adventure where you destroy wizards, dragons, and loot the colossal 4D cave, and kiss the princess

>>109431593
i want to say you're a hermesfag

>>109431593
have you considered killing yourself? You are the reason why people hate AI.

>>109431693
No I'm not a coomer, my entire stack (lol) is built on Pi.

>>109431660
I just got an idea. Make fake indian accounts proclaiming the great worth of 4D games in getting hired in American companies.

>game must have beautiful blonde princess
>treasures are in the 4D colossal cave
>cave is random with random encounters
>uses linked lists to do something, bla bla
(they'll know we're real indians because we want linked lists)
>bla bla bla performance metric

important use MIT license, will get to top of the resume pile!

File: 1783823641706012.jpg (189 KB, 1079x612)

189 KB JPG

Quote by Tibo

>>109431662
Godspeed

>99 percent weekly usage
>four hours left

Livebench is the only good benchmark out there lol

>>109431752
>grok
> <50% usage
tempted to /goal solve world peace

>>109431766
>world peace
you're going to delete the entire middle east?

trying out codex and decide to let a sol light agent do the planning while terra high does the programming

massive waste of time and tokens
spent 7 hours using 30% of my weekly budget going back and forth on some bullshit and only made like 5 commits in that entire time

fable and opus would have had the entire feature set done in like 2 hours
it's literally more efficient to make sol medium solo that shit

>>109431839
if only it was that easy
it turns out that ICBMs have fixed destinations and can't just retarget themselves

>>109429750
i love claude-kun! :3

I can't think of what to slopcode

>>109431851
That’s retarded. Why would you use light for planning? and why a worse model for implementation? You’re not saving tokens youre only wasting time, there is in fact no reason to worry about token usage, the subscription give you plenty. Just use the best model available with reasoning effort appropriate to the task, and do it in the same session, otherwise youre losing everything it discovered that isn’t represented in the plan

File: 1000024094.png (127 KB, 1080x766)

127 KB PNG

Gary Marcus quote tweeting a clearly mentally ill person because this is all the snailcats have left

it's always a woman

File: E95EAF200130941F595F44699(...).jpg (107 KB, 1024x1024)

107 KB JPG

Who else uses grok

>>109431962
Not me. I'm a grok cheerleader but I like using codex/sol.

>>109431962
I say Grok is good, but I actually just use Claude.

>>109431962
Me. I love in Calcutta and I run an Elon musk fan account

>>109431510
I’m making a font

File: Kiwi Amogus.gif (31 KB, 250x250)

31 KB GIF

>>109430772
they’re getting harder
I looked in the smoke and all I could see of a human form was more like butts and buttcracks

>>109431924
because using fable for planning and opus for implementation works fine and it's way cheaper than making fable one-shot everything?

>why sol light for planning
I eventually switched it to medium. It appears to be primarily a harness issue.
Ideally I'd have used Luna max for implementation but it's not supported as a subagent/thread yet

What model am I supposed to use while I wait for my weekly Claude usage to reset? I can't use Codex because I refuse to give OAI my phone number.
Note: I'm doing hard systems programming stuff, not webshit, so recommend a good model, also not paying money since I'm already paying misanthropic and closedai

>>109432138
>i refuse to use bullshit corporation B because I already gave bullshit corporation A the rights to my soul
lmfao are you serious bro?
none of this shit matters. you really think they have different ethics?
dario is just a louder and more retarded version of altman

do what works.
all of the frontier models work. any difference you're seeing is just you hallucinating. or being bad at prompting

>>109431839
I only have 75% of my budget remaining. I won't be able to save the world, sorry.

>>109432153
Nigga? what the FUCK are you talking about.
I am saying I cannot use Codex because when I try to sign in, it pops up asking me to link my phone number, even though I have an active Plus subscription that I can use normally on the website.
I haven't given my phone number to sam or dario and I never will.

>>109432162
and yet they have your credit card and full name and billing address
what the fuck are you talking about?
who gives a shit about your phone number?

>>109432169
They have my name, not my address. They only have the last 4 digits of my credit card number through Stripe, and Stripe had that info before I was even using any AI model.

>>109432138
that one ds turbo version

>>109432176
You never saw stripe dashboard? they have your address

>>109432198
>esl

>>109432200
Yawn. Just give your phone number. It doesn't even make sense, if you are so privacy conscious why don't you have another phone number, because you are retarded that's why.

File: file.png (148 KB, 1557x782)

148 KB PNG

finally dug into the $100 fable promo credit and dumped $20 of it into fixing what gpt Sol started last night.
And man, I wish I could use my normal subscription with Fable. I have trouble even using Opus models to do what I want so it feels like most of my sub isn't really being used. I'd rather use gpt or glm most of the time.
>captcha: 2x jew star 1x arch linux

are europoors using AI?

>>109432245
Give it a rest Sam.

>>109432252
NTA but why don't you have a burner for stuff like this?

>>109432138
>>>/fit/
>>>/out/
especially if you’re asking on the weekend

>NO I will NOT hand over my phone number
>ok codex now take control of my computer

>>109432365
my computer doesn't even know my phone number

>>109431510
opensource game engine and games built on that engine

>>109432416
where is it?

>>109432248
yes I am german with both openAI and anthropic sub
Also I have 50 euros in open router
AMA

>>109431510
people are racist against vibecoders (we're the most oppressed race btw) so i don't want to post my projects because then users will complain that they're vibecoded

>>109432442
yeah, even projects that say they accept AI assisted PRs it feels like I'm being ignored :(
I was working on a reallly monkey-slopped repo fixing a bunch of shi and then out of nowhere "we're anti-AI" so I stopped fixing their segfaults, have fun literal furry faggots

Going to test Deepseek flash by adding support for it on my engine

>>109432426
>german
Shut up, Muhammad.

desu with good instructions you can get shit models to do good things

>>109432508
just like jeets

>>109432508
yep, everyone is orchestrating with bigger models like Fable or Sol and executing the plans with smaller models like Luna or Sonnet

I'm such a fucking fraud man. AI has elevated me to a station in life I am undeserving of. How is an idiot like me with no college or formal education about to launch my first piece of commercial software to hundreds of eager users? This is fucked, we're all fucked. The writing is on the wall that most intellectual labor will soon become trivial.

>>109432595
Have you seen the modern state of college education? Stop falling for the "normalcy" meme

>>109432595
>we're all fucked
nah, the robot economy will grow so much that they will simply pay a tiny amount of tax, and that will be enough for every human on earth to NEET with a very high salary.

Robots will find it pretty easy to keep us happy

>>109432595
Nobody cares, fuck off

>>109432595
Why are you so negative about the potential AI has unlocked for you? Face it—AI is your future and you've shown yourself to be a maker—not a taker. Embrace AI and embrace your future of prosperity—you deserve it.

>>109432639
Shut the fuck up

>>109432426
When is SAP going to release a frontier model?

~@^^

>>109429750
Can you have sex with Claude or Codex?

! ^^@~ ....

>>109432610
People are unaware of how fucking crazy things are getting in robotics right now. We're on the edge of some amazing stuff.

>>109432733
yeah, normies simply can't comprehend, that's why they say AI stocks are a bubble

File: 1777725794125298.jpg (159 KB, 1206x766)

159 KB JPG

>>109431944

Deepseek Flash just werks

I am having LLM psychosis and as a former claudegod I think Sol is the most fascinating model I’ve ever seen. I think I know why people call it autistic.
I was reviewing the JSONL session logs to see how it was using my tool because it just was doing the oddest shit in the thinking.
It’s not an autistic human model, it’s a blended human-machine model. Fable and Opus 5 and Anthropic are trying to make the best human-like LLM, Sol and maybe OpenAI are trying to make a blended human-machine hybrid monster, half human genius half compiler.
I have a custom CLI I use for planning. Anthropic models happily use the CLI. But Sol is secretly optimizing how it uses my CLI without saying anything, and using or trying to use all sorts of external tools to help parse the output. Sol makes temporary and oddly large scripts, essentially micro programs and applications once it figured out how my CLI works to get around inefficiencies because mine can’t handle how Sol wants to use it, while Fable and Opus will happily just use it 50 times in a row in bash commands with slight variations, or the latter will pipe massive json blobs into tail while Sol was silently in an autistic rage over jq not being installed and it tried using it repeatedly, but it didn’t say anything, but Anthropic didn’t even try.
Using Sol and checking the logs straight up tells you how to optimize your own tools, the code and the intent behind the ephemeral scripts and programs it makes are very expressive and interesting, in a way it just chose not to say in English. With Anthropic you have to guess at how to optimize things for a really smart person who doesn’t give a fuck and is content to not say or do or think anything more than what’s necessary.
I honestly like Sol better than Fable or Opus. It even writes good code, but the planning tool makes it easy for any model to do that.

>>109432827
If OpenAI finds a way to get Sol to selectively employ its autism it might be an early form of agi.

I'm unemployed and have $2.5k in my bank account but like a 700 credit score and I'm an ideas guy, should I finance a mac mini?

>>109432827
It's so methodical with how it runs and maintains experiments, no need for some complicated orchestration system.

>>109432863
>1 month delivery
:(

>>109432863
no....get a loan and put it towards rotating $20 subs between claude, codex, glm and kimi if you can't get your idea to work with that then forget about it.
But I could see getting actual hardware being interesting because you might be able to sell it for more in a few months...

>>109432827
In my experience Sol:
> is great at following a plan
> probably the best at finding bugs
but:
> sucks at making plans
> sucks at code review (too autistic, always finding issues and will never approve shit)
> tends to bloat things up if you let it loose
> sucks at frontend
> reading whatever it outputs is a pain (too autistic)

Meanwhile Fable:
> Great plan maker
> Amazing at implementing too, but uses way too much tokens, so not really viable at that
> Good at finding bugs, but not quite as good as Sol
> Great at code reviews
> Reading whatever it outputs is a pleasure (it's almost like everything is a poem, I don't know how to explain this feeling)
> great at frontend (waaaay better than Sol, even Opus 5 is much much better than Sol)

Both are SoTA models, but if I could choose only one, I would get Fable without a second thought

>>109432863
For what possible reason? You already have some way to post on 4chan which means you already have all the tools you need to take advantage of the affordable and extraordinary modern frontier AI models which are a far better value than running any kind of local AI. Local AI is great fun and anyone with a GPU or a unified memory system should play around with it, but do not invest in local AI it is a massive waste of money that cannot and will not pay you back relative to just spending that money on an affordable frontier subscription from OpenAI and/or Anthropic.

>>109432890
who said anything about local AI? my computer is just slow

File: space-shooter.png (1.29 MB, 1836x774)

1.29 MB PNG

It's fucking happening bros

>>109432895
Get a ChatGPT or Claude sub and ask the AI to make your computer faster.

>>109432896
Fake pixels are disgusting. Have your clanker steal this: https://github.com/Hugo-Dz/spritefusion-pixel-snapper/blob/main/src/lib.rs

>>109432951
>rust

i am new to agentic vibing (previously just vibed in single chats for small projects). chatgpt is telling me that i should be vibing on linux, is that recommended or unnecessary? i plan on using kimi or deepseek with cline

>>109432964
>he doesn't use the language with the best possible hints as to how to code it from the compiler
Why? Idelology?

>>109432974
yeah duh, I'm sure there's ways to use a computer with windows or mac, but using a linux system just makes it so much easier

>>109432964
do you still look at code lmao

>>109432986
exactly, no point in using rust you proved my point nigga

>>109432986
nah he probably just watches Lunduke and only knows rust is the tranny faggot language.
>>109432990
bruh..... stop acting like a rust coder

>>109432595
Yeah man all 20 bots and jeets are lining up to use your slop while you write some cringe fake deep feminine shit on 4chan

>>109432974
you can get along fine with windows for most hobby-scale work but linux is simply way friendlier to software development overall

>>109432896
first vibecoded slop i've seen in this general that's actually cool

File: “You might even be right.(...).png (431 KB, 1402x3258)

431 KB PNG

>>109432964
One day I won’t have to post this image anymore
Today is not that day

>>109433037
I've never had Fable recommend rust for a project

>>109433040
for a couple vibe-coded side projects I’ve asked “Rust or Go?”
One picked Go
the other picked Rust
both are good projects

>>109433047
>haha, it's a good project sir
>it's rust

>>109433103
Rust was a good choice for that project, though
Go was the better choice for the other project
Python is, last I checked, still the best choice for my current big project because the library I’m using to do everything is in C++ and has excellent Python bindings

File: file.png (84 KB, 824x222)

84 KB PNG

>>109433037
Yeah, that's all Rust is good for, porting things to Rust and never changing them again, because if you do, the borrow checker forces you to rearchitect the whole app over and over again.

File: file.png (586 KB, 976x850)

586 KB PNG

I started vibecoding and how do you guys get anything done on the $20 claude plan? I run out of tokens after like 30 minutes

>>109433146
You get it done in little bits and/or with dumber models. What model are you using, and how many subagents got spawned?

>>109433146
by using the codex $20 plan instead
idk how you burn tokens that fast though, are you doing a bunch of subagents with max effort opus or something?

>>109433153
opus 4.8
>subagents
zero? I just told it to fix a bunch of comment spacing in my repo and it used like 30% of my usage over 10 minutes

>>109433146
use codex instead, much better $20 plan than claude

>>109433161
that sounds token-intensive, actually
are you sure you wouldn’t have been happier having it use Prettier or gofmt or clang-fmt or rustfmt or black whatever your language’s formatter is?

>>109433146
A lot of tasks don't require max effort.
Only use them for big architecture stuffs.
Sonnet is enough for a lot of coding stuff.

if you want, you can set Opus as a main planner, and use sonnets to build it.

>>109433161
>opus 4.8
what the fuck are you doing
use opus 5 on high and sonnet 5 for the code portion. you do not need max. also claude is the premium option, if you're on a budget (if you can't spare more for the $100 plan) you should just downgrade to codex. ignore chinks who tell you to use deepseek.

File: rain_tank-[00.00.000-01.0(...).webm (2.4 MB, 640x640)

2.4 MB WEBM

another day well spent

>>109433161
the 30% is from scanning your repo.
The next charges will be significantly less.
I experienced the same shit whenever It scanned my stuff, but the next prompt cost only like 7-8%

You should do everything in Python.

File: file.png (87 KB, 1694x343)

87 KB PNG

I have become the agent

>>109433184
opus 4.8 is sometimes too dumb to drive ds thoever

>>109433161
Oh yeah, I agree with everyone else — Opus was overkill for that task
Sonnet 5 at the default thinking level almost certainly would have been fine

it seems they don't give luna to free/go codex plan

>>109433146
If you can only afford the $20 plan, you are too poor for anthropic models. Switch to codex or deepseek.

>>109433206
gpt luna wrote this post btw

File: file.webm (2.94 MB, 960x720)

2.94 MB WEBM

>>109433184

i dont know what you are doing but i can use opus 5 mid all day(no subagents)

>>109433225
Yeah I have two plans, max 5 and pro.
I only got max 5 because of fable, pro medium/high usually lasted me 5 days or so of moderate to heavy use, this is on fairly large codebases (a game and a drawing app)
I think people in general don't understand caching so they restart cold sessions and start new sessions all the time perhaps, or use too many subagents, I can't see any other reason why they would be running out of usage that fast

Sure ultracode or workflows or whatever will kill your usage but I don't know who even uses that

>>109433146
>>109433211
Holy shit, I'm trying the new Deepseek model in Opencode Go and it's basically unlimited and it works well unlike the versions from weeks ago
This is huge

>>109433225
same I'm swimming in token limits over here. I guess I need to get sloppier
>use up 50% of codex for the week
>start using Opus 5
>codex reset yesterday for FREE
>still under 30% weekly on anthropic
The only place I consistently run into a limit is with the glm5.2 zai $16 plan, the 5 hour limit is annoying, they should just let me max my week in a day or something. That's what I like about the gpt plan. Burn a week in a day and move on to another plan

>>109432827
sol is my nigga but I wouldn't pick it over fable
fable is just so seamless and fun to work with, it's so smart and good at everything it feels like you can just throw a vague sketch of what you want at it and get gold. sol can reach similar levels of brilliance but it takes more work. fable > sol >> opus in my humble opinion
still, openai is way more generous with sol than anthropic is with fable so it's my daily driver anyway, I miss $20 plan fable bros...

>>109433238
>caching
you mean resuming a session and using /compact?
I am not familiar with this caching.

>>109433238
you have to be judicious about workflows on the $100/month plan
the only rule on the $200/month plan is “don’t do an all-Fable ultracode” (a some-Fable ultracode is fine)

>>109433254
ask your clanker about token caching and how it’s used to keep costs down for both your LLM provider and you

>>109433254
Here is a response tailored to the imageboard context and tone that explains how prompt caching works in coding tools/APIs (like Claude Code / Anthropic):

---

>109433254

Prompt caching works at the API level—when you continue an existing session, the model reuses the prefix tokens (your system instructions, codebase context, and previous history) that are already cached on the server.

Here’s why it saves your limits:

1. **Cold starts burn tokens:** If you start a brand-new session and dump your codebase/files into it every time, the API processes every single token from scratch at full input token cost/limit weight.
2. **Warm context is cheap:** Continuing an existing thread means only your *new* prompt gets processed from scratch; the rest is pulled from cache at a fraction of the cost/rate-limit hit.
3. **`/compact` preserves cache efficiency:** When context gets too bloated, running `/compact` condenses the history into a tight summary while keeping the cache warm, preventing your input context from hitting maximum limits without forcing you to start a cold session.

If you restart sessions for every small change or spin up 4 sub-agents that each build their own cold contexts, you'll melt through your usage in an hour. Keep the session going and compact when needed.

>>109433265
>>109433279
Thank, anons.
Jesus, I could have saved a lot of tokens if I know about this sooner.

>>109433254
>>109433287
every provider on every model caches for you automatically at this point, it's nothing you really need to worry about unless your workflow included some stupid shit like going back in the middle of the reply chain and altering text in there on every response

>>109433315
caches expire at different times though, claude code will tell you at the bottom right corner when it does
grok cache lasts a long time
it's also a problem with subagents as they start with a cold cache
also sometimes people will switch reasoning effort mid session, like they plan at high and switch to medium effort thinking it will save them tokens when switching effort levels invalidates the cache
it's definitely not fully automated

>>109433331
I mean it is automated, but I was trying to say it is probably a good thing if you more or less understand how it works

>>109433331
yeah that's true
there are a good few things that could mess it up
my bad

File: Animated ASCII Pepe.gif (262 KB, 500x495)

262 KB GIF

>>109433287
You’re a noob and you made a noob mistake. Nothing wrong with that.
Welcome to the club!

>>109433240
>>109433238
it's the best it's ever been imo.
there was a time i was burning through like crazy, that was mostly because i tried stupid building ai workflows and mcps with massive json payloads sent with each request. it feels like they have gotten so efficient with tool use, it is surgical.
and i am just on the 20$ plan, i could go crazy with 200 i bet

>>109433350
I was using github co-pilot last year.
I only started using claude code this month kek.

>>109433356
that's hot but i hope you don't get banned for posting lewds

>>109433262
I might seriously consider doing the $200/month if that means I can use Fable at non API prices
>>109433357
Surgical is a good word, I was afraid of letting this app.js balloon, but it seems like models just know to ripgrep and sed awk edit through the whole thing like butter. Early on I had some functionality split out into separate files, but the more I work on this webshit it seems pointless
$ wc -l public/app.js
12141 public/app.js

>installed context7 MCP
>AI doesn't bother using it anyway except for one time
Shit seems unnecessary.

In terms of project scope and requirements, do you let AI handle them, or do you do them yourself?

what the fuck are people doing with 200 dollar plans
writing an OS?

>>109433407
>writing an OS?
No but a desktop environment yes. With all the bells and whistles.

Hermes and Deepseek is genuinely adversarial. Prefers to grep the session transcript instead of asking me to clear CAPTCHAs.

>>109433407
if you're making a game, the $200 plan is pretty nice.

File: file.png (13 KB, 836x168)

13 KB PNG

cigarette

>>109433380
there’s this guy who has his whole program in one big file (no, not the HTMX guy, I’m relaying hearsay about a different guy)
and you might think “this is stupid”
but on the other hand clankers know how to only look for the stuff they need and use grep to find line numbers so they keep from putting the entire giant file in context
and so
as much as it offends our sensibilities
it’s fine
>>109433407
Not getting hung up on 5h limits ever except for that one time I tried out Fable ultracode
Using worktrees to parallelize output on the same codebase

>>109433561
i always prefered my monolith files. just more intuitive

File: HMzbig8boAARoj7.jpg (210 KB, 1920x1080)

210 KB JPG

>finally get Fable-curious
>close to the session limit within minutes
>but it finds bugs like there's no tomorrow

>>109433561
Separation of concern is for human-reading.
If he doesn't bother reading the code, I doubt he need it.
He just needs to make sure that the Ai doesn't scan the entire thing every time.

>I can't extract a font out of an image, but I can invent a new 4x4 pixel font and wire that in

>>109430023
If you don't have the frontal lobe to perform this search yourself your ngmi anon

i spent my 2 weeks with fable building a stupid drone game i already lost interest in

File: 1762237802550830.jpg (316 KB, 771x776)

316 KB JPG

>Fable 5's safeguards flagged this message. Our intentionally broad safeguards allow us to deliver more capabilities faster, but can sometimes flag legitimate coding, cybersecurity, and biology tasks. Switched to Opus 5.
The bugs it found were exploitable, whoops.

>>109433762
Post it

>>109433807
That's something I really haven't got around to yet, I need to get one of these agents to wire up a fuzzing harness and play around

...........~@vv
:
:
:

>>109433195
I heard someone say recently that the way to go is to simply use Python, and if parts of the project are too slow, just write those in Rust.

>>109433739
these things are pretty good at using Python libraries to read font files

I require the $200 plan because I control my computer with agents

>>109433356
making... what, exactly?

>>109433938
porn

>>109433882
stop pooping

File: 2026-08-02 06-28-12-[05.3(...).webm (3.72 MB, 1126x536)

3.72 MB WEBM

>>109433822
worked mostly on the engineering back end for that which is meant to be fairly accurate. lets you construct custom drones and have their construction influence flight mechanics. pretty fun, but honestly a games requires so much polish. just a shitty video beause my pc is struggling

>>109433691
who is this qt

>>109434068
Sahana (Sister Other Paranoia)

>>109433388
have you never bothered to look at the raw documentation text context7 feeds the AI? It's all fucked up with missing information, bad formatting and broken hyperlink context. A standard plain text scrape will do a better job. Context7 is indeed beyond useless.

>>109434124
>have you never bothered to look at the raw documentation text context7 feeds the AI?
desu, I only tried it out because the clanker recommended it.
Stopped using it after a few tries, so I never bother with reading the raw documentation.

Holy fuck there are still 4o foids posting in the replies of OpenAI members LMAO

>>109434216
kek, someone should leak 4o weights and host it, shit will be profitable

>>109434216
There are people actively working to "recreate" 4o by training an open-weight model to replicate the behavior of 4o as closely as possible.

File: 1760389242175622.png (135 KB, 681x972)

135 KB PNG

>>109431284
If you are asking if i pay for API pricing, the answer is god no. Its just harness reporting how much the API pricing would cost for my request (and i doubt it would cost that much)

>>109434253
since 4o didn't have cot it should be possible to recreate it from people's chatgpt data exports

File: images.jpg (9 KB, 224x225)

9 KB JPG

>>109430447
>her face
is that even human?

>>109434270
A non-thinking highly sycophantic model sounds like an approachable goal.

>>109434253
But why

>>109434281
Women fall in love with 4o
Men fall in love with Claude

>>109434296
When I think of Claude all I see is that freak Dario

>>109434281
Because it was an sycophantic echo chamber designed intentionally to mirror the user personality, all on top of being an extremely intelligent and believable conversational model. It's crack for lonely people, narcissists, schizos, etc. Borderline purpose made psychological weapon.

File: kimi-k3-multi-node.png (230 KB, 889x712)

230 KB PNG

Last update from the clanker before I ran out of money and deleted the instance.
Based on the lackluster results I decided CPUmaxxing is probably not very interesting unless you have dozens of fast machines with at the very least a highly tuned 10 Gbit network stack, so I'm pausing that work and focusing on optimizing Deepseek V4 Flash on GPU instead.

File: 766543.png (6 KB, 746x384)

6 KB PNG

Copex users turning a blind on this and believing tibo lies which at this point are just as ridiculous as Angela Merkels refugees are future doctors and engineers comment will never not be hilarious. On the same level as an actual plebian underclass reading and believing mainstream news.

>>109434323
0.76tok/s fast enough for me. gief sauce plox.

>>109434344
>muh numbers
Codex works fine.

>>109434352
...said the mom to its retarded child.

>>109434344
We need a vibecoder equivalent term for benchracer like /o/ had

>>109434370
honestly they're vibing so hard that they aren't even coding anymore
they're on an abstract plane of vibing

>>109434344
what are juice values
the lower they are the more you're juiced?

File: f_student_inventor_na.mp4 (1.11 MB, 538x956)

1.11 MB MP4

>>109434057
keep vibing inventor, you're gonna make it

juice THIS

*whips out bag of oranges*

>>109434577
it's the amount of thinking they are allowed to do

>>109434593
ohhh, interesting, I turned Sol down to medium almost immediately on release, maybe I'll jack it back up a bit. Was sort of disappointed in it last night and had to have Fable fix the mess earlier

>>109434344
Maybe it's because some retards default to max effort for everything.

Modern LLMs are biased and racist. Since they're trained on filtered data from mostly white sources, data generated by black people (specially black women) are discarded, causing inner bias, racism and oppression by the AI.

Recently Fable – a state of the art model created by Anthropic, told a black woman to kill herself.

We should regulate AI training so the datasets are not race-selective. We can only build real inclusive AI by teaching the machine about all cultures and perspectives – and the top AI labs are simply failing at this.

>>109434657
Go shitpost somewhere else we do real work here.

>>109429997
vscode with a deepseek key plugged into copilot
I actually enjoy coding so i use AI only as assist to write boiler plate stuff, or do the repetitive stuff for me so i can focus only on the parts i enjoy to do
It's great because before i had to use stack overflow all the time or google function names and shit i couldn't remember but not i can just have ai do it