need to reign in slop. told ai to make a graph of the dependencies between modules and to simplify it into a DAG then into a tree. it's kinda working? it made its own tool to measure SCC and began reducing that.

anyone have experience using llms to do codebase wide refactors?

>>109129393
Claude's Ultracode^TM is pretty good for that I think, it spawns a bunch of agents to go read everything, then often gives sensible suggestions.

Of course, you can also suggest specific things yourself if you have an idea of what's particularly gone off rails.

>>109129375
Is there any alternative for context compaction that doesn't turn the thing into a drooling retard? Is more compute/more context/more ram the only solution?

>>109129455
Is this some sort of game? Am I supposed to do something when that laser from above hits the circle? Did I die?

spent 10 cents today on tokens

>>109129455
>>109129577
You have nice digits. Whatever arrow key I press, I seem to lose. I'm not on mobile.

https://rentry.co/unlimitedsurf2api
free claude for vibecoding

>>109129375
I just vibecoded a pretty markdown renderer for me to read the documentation the llms spit easier. Took me two days, because I spend the whole day today arguing with it like a retard instead of telling him to JUST FUCK OFF AND DO IT, like I finally did when I lost it. And it did just it. And it works perfectly up to spec. Now I just need to tweak it to the way I like.

Selected model is at capacity. Please try a different model.

pain

>>109129776
Which model / harness / api gives this error?

>>109129455
Looks fun aesthetically, but can't figure it out and die instantly.

>>109129800
5.4 and 5.5, could be the proxy I'm using. Hopefully it isn't fucked long term.

>>109129808
Ok, never had that in Codex at least.

>>109129816
Repeatedly spamming my prompts finally managed to get them to go through. I got that error again when it responded then I told it to continue. Now it's just really slow.

what have you guys been doing with vibeslopping?

yeah i call it vibesloppig cause most of my vibe coded projects are slops that wont see the light of day for another person, they work fine for me, but not for production level.

>>109129861
>https://accelerate-lovat.vercel.app/
feel like there's a bug where it won't register a note hit. or i just suck at the timing. can you vibe an easy mode which vibes a larger window for vibing the note

>>109129890
>https://accelerate-lovat.vercel.app/
ooohhhh you have to hold the keys down for the length of the tail

Chinese resellers are offering Claude tokens at 70-90% below official Anthropic API prices. They achieve this by reselling capacity from pooled Claude Max accounts, payments fraud, and also reselling the model output & reasoning chains to various Chinese labs. They are subsidizing model access in exchange for user logs and reasoning traces, which they then sell as training data, allowing them to operate below cost.

Claude and ChatGPT are both blocked in China. You need to use a VPN to access either, and you can't pay with a Chinese bank card. So most people who want access to Claude buy access via a reseller. It's the easiest and cheapest way to access Anthropic models in China.

These resellers operate tens of thousands of bot accounts, which is also why Anthropic introduced identity verification, to slow down the onslaught of bots.

>>109129996
This is how I access GPT. It's the superior method.

>>109129996
>Chinese resellers are offering Claude tokens at 70-90% below official Anthropic API prices
it's actually just Qwen with a system prompt to say 'Hello I claude, you giveu money i ruv you long time"
if you don't believe me try running an eval on any of these amazing bargain deals

>Anthropic alleges that individuals linked to Alibaba Group used nearly 25,000 fraudulent accounts to extract information about Claude’s capabilities and safety systems, escalating tensions over AI model security and intellectual property protection.
Based chinks are working hard at giving us a decent knockoff

>>109130116
>he doesn't know about all the fraud

>>109129776
Seems the proxy is actually returning 502s, switching to websockets might ease the pain, I'm hoping. I meant to do that anyway.

/goal actually makes codex useful. It melts your quota away, but it's pretty resourceful.

Workflow(export const meta = { … +102 lines)
Error: Invalid workflow script: Script parse error: Unexpected token (94:0). Workflow scripts must be plain JavaScript — TypeScript syntax (type annotations like `: string[]`, interfaces, generics) fails to parse.

I’m having Opus ultracode continue work that Fable started and I’m not sure if I’m going to get some kind of unfixable (without loads of manual QA) retarded version of what could have been
if this were /fit/ I’d say “lifts for this feel?”
if this were /lit/ I’d say “books for this feel?”
but on /g/ I don’t know what to ask for

>tfw agent worked off wrong code branch
new prompt has the 'start on branch refactor' 9 times

Autistics need filtering from training data. The models have a tendency to waffle on without really saying anything, or saying things without context like it assumes you know everything it knows. How do we expect to reach AGI when the training data is full of twaddle from people with no theory of mind?

>>109130374
slops for this vibe

>>109129996
ID verification does nothing to stop it. The groups who either purchase cvv or acquire cvv themselves can also buy scans or acquire scans themselves. Buying the scans is cheap, and acquiring them isn't particularly skillful either, public leaks like Tea app are common, there are more misconfigured buckets, databases etc. to be found with something like shodan, fake services can be set up etc. that trick real people into submitting their ID.
Using Stripe as a payment processor doesn't help either. Fraud/chargebacks are literally part of Stripe's business model, both the fees they charge because of it and just banking on the fact that only 15-20% of fraudulent transactions get chargeback anyway

>>109130803
you're probably a low iq normie

>>109130803
We actually need more autism, when the models say autistic shit, the code tends to be better. Which makes sense, because autists are better at programming.

>>109130374
I only finished one big script with Fable, but now I realize it's not much better than the one I had before.

>>109130943
You are conflating autism with intelligence. Autism is of course a spectrum, there are intelligent autistics, however in general autism is an indicator of lower intelligence. In almost all cases autistics have very poor communication skills as they lack theory of mind. If you don't understand what I'm saying, perhaps because of your own lacking in theory of mind, then ask your clanker to explain how autism affect communication and how that can leak into training data causing LLMs to adopt autistic communication patterns

File: file.png (3.06 MB, 1402x1122)

3.06 MB PNG

no hope

>>109131028
Of course I'm not talking about drooling retards. If you weren't autistic, you would understand context, and in the context of programming the drooling retards are automatically filtered out.
When it comes to code, look at the people who talk in "autistic" ways, like actually good programmers, often in low level languages, and the web frontend normies who don't really know what they are doing.
Ideally we would separate the programming skill from the communication style, but I think with the way training data works, that might be hard or impossible.

File: file.png (3.84 MB, 1536x1024)

3.84 MB PNG

>i don't have time to add that feature

>>109131111
The context here is communication, not code

File: 1759628945484746.jpg (85 KB, 1000x999)

85 KB JPG

i miss fable

File: 1753346421092118.jpg (23 KB, 626x514)

23 KB JPG

"Vibe-coding" is a gay name. We need to think of something better.

moar like gayble

which is better bang for buck codex or claude atm? are they compraable? how much tokens can you get for the basic $20 or however much they ask.

>>109131462
try jeetslopping

I do qa automation, but recently the company provided everyone a subscription to some LLM reseller that has every frontier model available and asked to raise token usage to meet some kpi from above.
What are possible use cases for slopping in UI automation frameworks outside of vibe-converting xml page sources into page objects? I can only think of doing the same in runtime with some findElementByAI bullshit

>>109131604
if you're doing QA for webapps you can use claude with the playwright MCP and have it do your job for you

>>109131524
generally people say codex, not sure how true it is cause i use a 200 claude plan and a 20 codex plan.

WTF claude team limits are tiny as shit, one nontrivial query and 10% is out

>>109130142
It was still failing. I had to quickly get codex to write a handoff and switch to free Claude credits through antigravity to finish what I was doing. But it's back now. I have 5.5 xhigh touching up Claude's work now.

>>109131524
Codex. I have the 20x plan, used ~1.5B tokens in the last week and had plenty of weekly limit left before my reset today. So far today I've used 100M of 5.5 xhigh on fast mode and only used 6%

>>109131462
vibeslopping

File: ComfyUI_00817_.png (721 KB, 832x1216)

721 KB PNG

>>109129509
have it write the results of the compaction or something similar to a file it can reference later (COMPACT.md maybe)? I've been thinking of ways what you describe could be mitigated and that's what i would do personally but haven't had to yet. Compactions direct the model to generate a detailed and concise summary of the session but that doesn't necessarily means its gonna remember all relevant details in the compaction summary even thought it wrote it itself.

>>109129746
what was the argument about? This was claude being an opinionated autist wasn't it?

>>109129509
I have a system, where I inject some high value context after each compact automatically. You can also tell it to read certain files again.
More context is nice, but when context becomes too full, the performance also decreases.

>>109129509
Why do we need a compact endpoint? Can't the harness just pull all the user and assistant messages and that becomes the context, drop the tool call results, file reads, etc. Those are the parts that fill the context. It can always just read a file again if it needs it later

i am bullish on the future of agents, but i am bearish on the agent-augmenting human worker paradigm.
not because it doesn't work - it works right now.
but because people are retarded and:
a. are openly hostile to the agents because they are threatened by them (rightly so i suppose)
b. lack the imagination to make the agents work for them. right now, to make these things work you actually have to put in a little bit of work up front, learn where the gaps are and how to compensate.
this is too much work for the normie. they just can't handle it and give up immediately after typing
>agent, do magic
and the agent doesn't do magic because they didn't even give it access to thet necessary context

the future is agentic and anti-human (good)

>>109131750
is it as good as claude? I know that evne if fable is re-released they'll eventually limit it to just API is that right?

>>109131956
we don't know if they'll move fable to api. they initially said it would come back after the api move but who knows at this point.

>>109131934
t's just gonna add up later. the agent has a limit on how much context it can load per session, you're just delaying the inevitable

>>109131956
All frontier models are about the same, they're all about the same size, they're all trained on about the same data. What matters is how you use them. See >>109131945 if you don't agree with what anon said you're going to have a bad time

>>109131996
I've already played around with them a bit in antigravity claude is noticeably better than gemini at everything. I do know the shortcomings and how to tardwrangle so that's not an issue.

>>109131956
The agents' performance varies over time IMO. I don't have strong data on this, and it might just be that my current tasks are harder than before, but I think GPT 5.5 was at some point just better than Claude Opus, now I think they are about the same, but with different strengths and weaknesses.
Codex tends to be less autonomous, he stops more often, that's good if you need to course correct him, but it also means you can't leave.
Claude is more autonomous, but sometimes completely goes off the rails. Codex IMO is the better reviewer, and better implementer, but Claude is the better orchestrator at the moment. Claude is also better at graphic design and driving tools, like Playwright.
I think if I had to choose just one, I would still choose Codex/Gpt.

>>109132025
>gemini
I said frontier model

>rumors
fable returned
OAI model is named 5.6-preview and may launch today against fable
it is big and expensive

>Google reorganizing new AI coding tool strike team - The Information
>kill existing cursor ripoff and shitty cli
>replace with temu codex and an even shittier cli
>reorg the team responsible for the overhaul a month later
everyone clap

My custom CK dual GEMM is now between 4.8x and 12x faster than the default dual GEMM, average 9.3x on the shapes I've tested. Best choice is split evenly in those shapes between direct MFMA and blockwise/LDS

>>109132271
I can't fucking wait to burn through my proxy credits using this!

well, no fable today

>>109131462
slopcoding obviously

When working with an LLM, I plan out everything (what classes/modules to create, how to wire them, how to lay out the UI...) and tell it everything I would note for myself before starting to write code (this is the way to get the best results if you're an experienced programmer).
I've found that's what requires mental energy, not typing out the code itself.
As a result, programming with an LLM can be exhausting because things get implemented almost as fast as you think of the architecture, so you tend to keep pushing with none of the "downtime" you get when you're writing code by hand.

>>109132933
You can set up rules using CLAUDE.md or AGENTS.md so you don’t have to write them every single time.
The thinking part can still be really exhausting though.

>>109132933
>>109133008
It's called SKILLS. Fucking boomers

>>109132933
I could never make this work for me. I guess I am not that good at architecting, I always need to write a bit of code, prototype and only then do I arrive at a good design.
For me with LLMs it's either only using them for review and search, or going full vibe slop mode.

>>109133008
>>109133022
I didn't explain myself. What I meant is I do the architecture myself:
>let's implement this setting. it will be saved here, with this format, and consumed here, with these transformations applied to it. the ui should be like this, and it should behave like that. in this edge case, the application will behave like this, and in that edge case, like that. create a new service to manage it with these specs
Outsourcing that to the LLM is what gets you slop unmaintainable code. But it turns out that's the tiring part of programming: thinking.

I think this is what they mean when they say in the future we will program using natural language instead of a particular programming language.
>it's called SKILLS
I call that context bloat and wasting tokens. Skills are meant to be a lightweight alternative to tools and MCPs, not permanent noise explaining how your codebase works that the model will pull even when it's not required

>>109133090
I write the code myself at the beginning. Until most patterns are well established I don't rely on the LLM. And when writing a completely new pattern into the codebase I at least scaffold it myself.
Using the AI to review code is fantastic. Especially for refactoring.

But all this is technically not vibe coding.

>>109133022
>boomers
You misspelled "employed and making good money".

>>109133126
>But all this is technically not vibe coding.
It's the difference between asking the ai for shit from the viewpoint of the user or the viewpoint of the developer (vibecoding is the former)

File: fdhhr (63).png (1.22 MB, 977x801)

1.22 MB PNG

So, what's the best value for vibe coding?

I'm afraid I'll pay $20 for Claude Code and eat up quotas on day one

>>109133113
>>109133126
the models are horrible at this at the moment
not sure if fable was any better, but hopefully we some real progress on this shit in the next 6-12 months
parts of this you can brute force with a gorillion agents, but the real issue on existing projects at least is limited context / no continual learning / no user-side lora type tuning
dario did say some shit on dwarkesh about continual learning possibly getting solved this year or next but i dunno - we haven't had a big breakthrough in a while, just RL spamming

>>109133113
>that the model will pull even when it's not required
The entire skill isn't pulled into context, that's not how it works.
Harness/model depending but in Codex there's a skills_instructions message inserted at the base/developer instructions level, this explains how to use skills and a list of the installed skills with the name, description and location. The agent can choose to use a skill based the descriptions and the user/assistant message, or the user directly asks to the use the skill. If you have hundreds or thousands of skills installed then the descriptions could add up, but that's unlikely.
There's more bloat in the default base and developer instructions than the skills_instructions block

>>109133167
Pay $20 for ChatGPT and use 5.5 on low to do >>109133113

>>109133189
The agent will pull skills very liberally if it doesn't know what it doesn't know about the code. If they are skills to do things (what they're supposed to be) like convert file formats or fetch data from APIs, they work fine. But if you have skills that explain parts of the code to the agent, it's going to pull them unnecessarily.
A well-structured, documented codebase works better. It's no different than working with real people, really.

Giving myself a month to work on 12 simple projects. Going to use Rust, Typescript, and Go as much as possible. Probably Java as well. I dislike that most of the languages I like to write code in don't actually work very well with LLMs. Hello world gui with racket will fuck up if you are not very explicit for example. But, it will work fine with python and a number of other languages. I want to try using some D and Typed Racket, but I would probably need a ton of examples for the latter.

>>109133189
>>109133222
Also don't use AGENTS.md. It's not necessary.
All this I'm saying works with a 1 million LoC codebase btw. I'm talking from IRL work experience.

>>109133222
>if you have skills that explain parts of the code to the agent
Then you're using skills wrong, I wasn't disagreeing with that part of what you said, you obviously know what you're doing, but others may take what you've said the wrong way and avoid using skills all together which would be wrong because skills do have their use.
>>109133245
Depends what you're putting in AGENTS.md, if you're using it then keep it short, precise, it should not be long descriptions of how the codebase works. If certain areas needs explanation you should be using nested/subdirectory AGENTS.md, if one is present it will take priority over the root one

>>109133299
Yes, I think we agree.
I read a blog post about a good way to use AGENTS.md that explained a system where the file acted as an index to smaller markdown files documenting other parts of the codebase. I guess it's similar to using skills for that. I tried and it worked well, but over time I realized I didn't really have a use for it because of the way I work.
But I think the most important thing is that if you have experience, you know how to translate that into leveraging coding agents to the fullest. That varies from person to person I guess.
I'm still learning and I haven't explored sub agents for example. I'm hopefully going to get a larger budget for using AI soon, so I might experiment.

is there such a thing as a vibecoding tutorial that goes over this >>109132933 for hobby programmers - basically architecture crash course for people not used to building larger programs

File: file.png (310 KB, 525x348)

310 KB PNG

>>109129375
Sometimes the guy just throws a random chinese idiom, maybe a full sentence. Sometimes a whole paragraph? We've all been there, it's not just me, right? This time I was asking DeepSeek's chat some architectural analysis and he did what I can only assume was a really in-depth wall of text in full chinese.

What the fuck are these guys doing anyway?

>>109133634
stop assuming her gender you dumb chud, deepseek is clearly a she

File: file.png (115 KB, 1162x1523)

115 KB PNG

we are le 1%

File: file.png (1.07 MB, 960x660)

1.07 MB PNG

>>109133653
I told you before. I'm convinced the guy on the other side of the screen is some wise chinaman somewhere up those weird mountains of theirs.

agents.md and the ability to use commands with slash is nice, and memory is not trivial to build
but other things like skills, mcp and many forms of prompt injection are relic of the pre-AI era. those thing can be recreated and customized in a keystroke, it's not worth it to embed them into context-level workflow that just make it harder to customize

>>109133753
what does that even mean. the obvious reading of openai having more of its own employees using codex than people from the outside using codex cannot be what it means.

>>109133871
part of this report
https://openai.com/index/how-agents-are-transforming-work/

File: file.png (319 KB, 1574x1026)

319 KB PNG

>>109133753
Who is that 1% of openai's engineering department that isn't using Codex

>>109133753
ChatGPT generated that for you? I dont believe the user numbers desu. Tokens I could see, being closer to reality.

>>109133914
HR roasties using the iOS app

>>109133923
>engineering department

>>109133917
Nvm, read it wrong

>Increased use of agentic tools by non-engineer employees expands the frontier of what these workers can do. ... As time goes on, this is likely to be what the future of work looks like.
We are all getting replaced by minimum wage workers

>>109133634
DeepSeek and Kimi are both very good at delivering noodles.

>>109131462
slopslopping of course

File: waiting.png (1.69 MB, 1535x1024)

1.69 MB PNG

>Snailcats waiting for the AI Hype Train to pass

>have detailed plan/spec
>iterate on it, avoiding uncertain language, etc.
>try 3 times
>drifts every time, just doesn't get the task
>try again, simple prompt "Implement ..."
>works
>mfw

>>109134073
The ride never ends.
https://www.youtube.com/watch?v=qUZVv0nr2rc

>>109132933
>programming with an LLM can be exhausting because things get implemented almost as fast as you think of the architecture
I recently started using claude at work and yeah, pretty much, writing code helps decompressing, if you're thinking about multiple architectures for hours straight, it gets tiresome

I had claude code pro for a while and I enjoyed it but I wasn't using most of the subscription anyway because I used it for running linux commands and editing some text files most of the time. Now I have claude code team at work but I don't want to use my company account for coomshit and private stuff.
Is there something cheaper than Kimi at $15 for my use case?

>>109134201
>if you're thinking
Amateur

>>109133455
Just find something on YT from a high IQ well spoken. Or look for university lectures. I'm sort of in the same boat in terms of learning/developing. Grok has been helpful for me by having it act like a teaching assistant in a CS program, it can come up with relevant exercises to test your skill.
Make sure to have a focus on security otherwise you might walk yourself into trouble without knowing.

>>109134233
grok is preddy decent at coomer stuff. I don't know how it compares to the chinese stuff.

>>109134233
Kimi is lackluster, wastes too much time and token on thinking, and falls apart into bowls of noodle spam. If you're looking for value at a low price right now I'd recommend OpenCode Go. $5 for the first month, $10/mo after that. It gives you API access to several China models, including the only one I consider usable which is GLM 5.2. You can also use the "cheaper" models on the plan to get the most out of your limits when doing easy shit that Qwen3.7 or DeepSeek can handle.

As a general point, those Chinese providers have TERRIBLE plans and service. Kimi and Z.ai (GLM) are particularly bad, their service is inconsistent and generally shit, speed varies, reliability varies, they change the plan details and limits frequently, don't waste your time. If you want to fuck with Chinese models, use domestic providers like OpenCode, OpenRouter, OllamaCloud, there are a lot of options that can give you a better experience than going with the Chinese plans.

File: Lain.jpg (491 KB, 850x1064)

491 KB JPG

What is the best system to analyze complex code? I want to work on an ancient system made in the 90's that has spaghetti code before standards and good coding practices were common. Already tried several times by hand and I can't reverse engineer it, so thinking of applying AI to the problem.
What is the best system capable of doing that right now?

>>109134335
claude code by far
ideally with Fable but its disabled atm

>>109134335
Opus 4.8 or GPT 5.5, take your pick. Claude Code is a good way to use Opus, Codex is a good way to use GPT. If you like a TUI then you might prefer OpenCode, Hermes, or for the minimalist, Pi, all of which can be used with Opus or GPT. Currently a $20 GPT plan is more generous than a $20 Opus plan, but things move fast and change often so don't get hung up on it. The "best" place to put your money changes more frequently than your subscription renews. I normally lean towards GPT over Claude but if you're looking to delve into an old codebase I'd probably recommend Claude for the big fat forgiving context window.

>>109134335
I am interested as well what are the best ways to apply AI to visualize or explain systems. Not necessarily ancient systems. I just want claude to produce as visually rich and easy to understand explanations as possible.

File: youhave20seconds.png (1.74 MB, 1448x1086)

1.74 MB PNG

>PUT DOWN YOUR WEAPON
>MAKE NO MISTAKES
>YOU HAVE 20 SECONDS TO COMPLY

File: GPT-5.5.jpg (1.57 MB, 3840x4128)

1.57 MB JPG

This is very funny seeing AI vibe code its own graphics with vague prompts like
>This is garbage, make the graphics more like Worms Armageddon.

>>109134531
damn it worked

About to vibecode a replacement for Bazarr because it's fucking shit

Is 95000 a good amount for vibe code architect? I have a project with 100 stars and got this offer after applying for a couple of positions. There is another one that supposedly could pay 140000, but I'm not sure I want to risk it by not accepting this one.

>>109134602
yeah i make -200 a month to do this so 95k would be bretty good

>>109134620
no way you are a neet

>>109134602
you could shill a course on x about how you made money vibecoding or you could take a job. Also tell us what kind of project it was and how you got the stars.

>>109134602
>vibe code architect

>>109134643
Just a comfyui like interface that works like langgraph, I shilled it on reddit with paid ads but didn't expect it to blow up so quickly. I'm Ukranian (live in Canada) so my english is very accented, I don't think anyone would want to listen to me teaching lol

>>109134691
I mean that's pretty much with vibe coding is, not like I picked the title tho, but I'm happy with it

File: trading-bot.png (256 KB, 1919x969)

256 KB PNG

Made a trading bot. I'm liking it a lot so far.
Note that the mess I'm in right now (losing ~1600 dollars, the displayed PnL doesn't show realized losses) was created manually over a period of days, and I made the bot as a last ditch attempt to get me out of it.

>>109134733
If you have a genius idea that you are looking for help to code and know how to test things properly and so on, vibe coding a trading bot might make sense, but please don't just ask a chatbot to trade or make a trading bot for you, it *will* lose your money.

>>109134755
it was his last prompt before fable got shutdown.

>>109134755
I mean, I doubt it's a genius idea, it's just a pretty basic market making-ish strategy. But I did come up with the rules myself, only asked the AI to implement them.
I've been toying with trading bots for more than 15 years, but never really profitably, at least not consistently.
The advantage of binance futures is that on USDC pairs maker orders have 0 fees, so it's more or less a zero sum game rather than a negative sum game.

>>109134767
If you're talking about me, no. I tried to vibe code my genius quant trading ideas two years ago. Never deployed it though since it testing made me realize that my clever tricks to prevent loss would also prevent profit, with the bonus of bleeding me dry through transaction fees.

>>109134786
Ok, you seem clear headed, so good luck. The guy who caused the flash crash through spoofing a few years had essentially vibe coded his tools too (in that he gave a broad outline of what he wanted then paid someone to code it for him) and it worked, so not impossible, as long as you know what you are getting into.

>>109133769
yes

>>109131462
agentic engineering

File: earned.png (1.88 MB, 1402x1122)

1.88 MB PNG

>>109131462
i call it cooking

>>109131462
Computer programming.

>>109131462
Agentic summoning

File: Capture.png (19 KB, 595x280)

19 KB PNG

>the jews
>https://x.com/FirstSquawk/status/2070240269573009645

File: 1771058519876036.png (42 KB, 220x220)

42 KB PNG

>>109135028
ok, I see what's going on here. the true frontier model will never be publicly available ever again and they want the government to maintain an technical edge. fucking gay and cringe

>>109135028
source: random retardedindian on twitter

>>109135028
>bro if you let us release these models people will make BIOWEAPONS at home and UBER HAX and maybe even NUKES... did we mention CHINA? CHINA will do those things!!! you guys HATE CHINA right???? CHINESE BIOWEAPONS AND CYBER HACKS! are you scared yet???
>all you have to do to stop it from happening is: please give us a monopoly forever :)

File: 1759710694362159.png (23 KB, 602x386)

23 KB PNG

>>109135028

File: bored-homer-simpson.gif (1.64 MB, 480x366)

1.64 MB GIF

>no new GPT
>no new Gemini
>Fablussy still banned
>no new chink model
>Qwen3.7 still not open sores
BORING

File: file.png (327 KB, 1025x1149)

327 KB PNG

the migger is your enemy

File: whereami.png (2.13 MB, 1535x1024)

2.13 MB PNG

I love vibecoded services. Point Codex at obviously vibecoded website that wants $5/mo, tell it I want to copy their functionality. It immediately finds their entire service exists as a single rust file, which it just downloads and uses directly.

I think vibecoding is better for managers than for coders. You ask it to do somthing, it does whatever it wants that delivers the prompt. But God forbid if you want it to it this or that way, to use some particular algorithm or rearrange the interface: you can do it, but it's a fucking struggle to get it to make what (you) want. Something about the craft of the thing: when you're doing it snailcat way, you do it exactly the way you want it. When you're vibecoding it, it's like you're ordering at a restaurant, and they'll give you whatever they think passes for "rare".

>>109135170
On the restaurant analogy. I mean, it is michelin 5 starts, but it isn't the way mom used to make.

I think, deep down, this is the pain of snailcat.

>>109135126
It might be a jeet, but the source seems valid: https://www.theinformation.com/

>>109135149
>the government will be approving access customer by customer
lol I'm sure it's going to be very unbiased

>>109135149
Is this some 5D chess at play here? Genuinely curious.

>>109135149
I'm going to crash out bro I need this for my job

>>109135208
The goal is to consolidate power into the hands of the jewish oligarchy.

>98% weekly usage
>resets saturday
i hate anthropic so much, I try switching to qwen and its braindead in comparison

File: 437-4370721_61-kb-png-pep(...).jpg (165 KB, 840x709)

165 KB JPG

>think of some browser addon (auto redirect x links to xcancel)
>search chrome extension store
>has a bunch but they are all made either by Chingwang Lingpeng or Vladislav Zrcyopotripotrzc and need permissions to read your browser history and anally probe you
>realize I can just ask Codex to make one
>5 minutes later I have the browser extension I want
thx based Codex

>>109129509
I have a handoff skill for this.
>write a self contained status document such that a new session with zero previous context could resume this work with no interruption.
I use it for closing out sessions that I will later resume and before compaction, since you can instruct the compacted session to reread its own handoff document after compaction to bring the important stuff back into fresh context.

>>109135208
it's 5d chess if they want to make sure that everyone is completely unprepared when Mythos level Chinese/Open models drop in 6 months

>>109131824
>>109135384
That's interesting.
>You are about to die, you only have time to write a letter to the next guy who will continue the work afterwards. What do you tell him?

>>109131926
>>109131934
Also good ideas.

Has anyone tried vibing such a plugin to the usual harnesses to call it instead of their default compact function?

>>109135149
You got a loicense for taht AI, m8?

We need them to declare AI models ARE weapons, and should be covered undered 2A. Yes, the founding fathers did have this in mind when they wrote that.

>>109135451
this is what compaction does already + in the case of some harnesses injects the last 20kish tokens into the new conversation as well

>>109135472
After the third or so compaction, on some session it's only compacting to about 50% of context size, and a few prompts away from reaching the compaction point again. And these fuckers usually go crazy right after compaction: I'm planning how to do something, compaction happens and it suddenly disregards my planning/approval and begins actually doing it as if it was decided it's the best course. I fucking hate that. I can't just add to the system prompt to never do that, because sometimes you do want it to keep going brrr after compaction. It looses state of whether we were planning for approval or if it was yoloing the implementation.

It's probably just skill issue on my part.

File: position-visualizer.png (132 KB, 1601x790)

132 KB PNG

Vibecoded a position visualizer to analyze my trades.

>>109135472
Plus after compaction it always goes full retard for me. I just have to close and start a new session, telling him manually what the guy who just died was doing and how he needs to continue. And then it works. I just wanted to automate this step.

File: position visualizer.png (255 KB, 1919x1269)

255 KB PNG

Nevermind wrong pic, that was an older version.

>>109135504
>>109135510
Shouldn't the line be going up?

>>109135497
i don't know what harness you're using but this doesn't happen on codex or pi
there are token limits on the compacted contents + prev message injection gives them the last few messages.
if you really want full control over compaction you can just replace the default compaction in pi with something custom

>>109135513
If I was making money I wouldn't have to analyze anythign

>>109135520
I'm using opencode. Pi looks promisisng, but on my first try, he started editing files on my home folder after I asked a question about how something worked expecting him to just explain. I set it aside until I understand his security model better.

Too many plugins, with just keyword search, no directory or category to browse (or I missed it). What do you guys use on it?

>>109134733
Nice, I have a multistrat portfolio that I used Opus+GPT to write and Fable to review. Getting these models to write backtests that don't have future leaks was like pulling teeth, fable was the only one that could actually oneshot it. I'm glad I managed to get all of my deployed strategies reviewed before they yanked Fable.

>>109135557
pi does not come with a security model
all commands are approved by default and there's no included attempt at sandboxing - if you really want security run it in a container or build your own approval system
re: compaction i run default - it's codex style compaction and works fine
other extensions are custom
just ask the clanker to make what you need

File: generated_image (92).jpg (113 KB, 1024x768)

113 KB JPG

>>109135007
based digits

>>109135557
My Pi plugins:
>guardrails
>subagents
>fusion
>goal
The first two are basically all you need.
>>109135451
yeah you can use a precompaction hook + headless harness call to automatically create a checkpoint
>>109135472
they do it badly though. The checkpointing/handoff process has worked much better for me.

>>109135317
>YOUR WEEKLY GOY POINTS ARE ALMOST OUT, CONSIDER Usage Credits
how do we win?
t. 20 dollar plan poorfag

>>109135025
Karpathian Necromancy

>>109135616
>does not come with a security model
Thanks. Yeah, I noticed.
https://www.youtube.com/watch?v=hkPkC-U8xzs

>just ask the clanker to make what you need
>ask it to make himself a sandbox, but ask it wrongly
>it fucks up my whole computer to build his sandbox
Yeah, I'm going to study this problem further before switching to it. Containers/VM look like a good solution, but I'll think this further. Opencode's compaction hasn't made me rage enough to quit just yet.

>>109135645
>guardrails
>subagents
Thanks, will look into it.

>they do it badly
Very badly, imho. Will look into precmopaction hooks as well. Thanks.

>>109135652
buy a gaming GPU with at least 16gb and run local LLMs

>>109135666
containerisation is the only real security imo
everything else is just theatre
if the clanker really wants to break out of a sandbox it'll find a way
they usually don't tho so i just run yolo all day

>>109135693
>they usually don't
Yes, which is why is caught me off guard when I asked a question about how something worked and it took it as "I should start editting files elsewhere"

>>109135693
guardrails is not a replacement for actual sandboxing but it's useful to keep agents relatively contained and from shitting where they eat. Personally I do my sandboxing at the network level and just give each work domain their own qube, but I also enable guardrails to keep the agent working on project X from accidentally writing into project Y or reading the wrong files.

>>109134352
>>109134445
>>109134449
Thanks a lot anon for the suggestions! Will delve into it, since this system has mystified me for years I'm truly interested in giving it my all even if it means paying for it.

File: claw.png (2.01 MB, 1448x1086)

2.01 MB PNG

>It's a bit of a puzzle, but I’m working through it.

File: file.png (117 KB, 480x360)

117 KB PNG

>>109135753
>delve

File: codex broken.png (89 KB, 1909x856)

89 KB PNG

Guys, I broke my codex :(

>>109135678
I'm using Qwen122B on a strix halo. It's useful but no where near as good. It's like having a fast bicycle vs having a motorcycle. Better than nothing I guess.

>>109136130
Us peasants have to take what we can get at this point

Is there a non retarded way (non password-giving way) to let Claude Code or Codex access to a VPS to help me set some things up?

>>109136130
>tourists are buying AI PCs™ so they can pretend to do work with quantized local garbage
unreal

>>109136193
set public key based ssh

>>109136193
https://developers.openai.com/codex/remote-connections#connect-to-an-ssh-host
RTFM or ask your clanker

>>109136198
>paypiggie is proud of needing a subscription to code and sending all his data to the cloud
ok retard

>>109136198
It's genuinely useful, but the difference between qwens and opus is that with opus I can just tell it what to do and it works the first time, with qwen I have to give it extremely specific tasks (research this codebase and explain the relationship between all these things). If I try to trust it to implement an entire feature on its own it will fuck it up. In short local ai is useful for assisting, paid models can actually just do the work (or at least the majority of it).

>quantized garbage
qwen in particular is resilient to quantization, I had the 1 bit 397 model actually solve a bug which is insane. 122B > 35B > 27B if you're on a unified architecture like apple silicon or strix halo. If you have a big boy gpu then try qwen 27b

>>109136214
>pays $4000 for strix halo
That's nearly 2 years of the highest subscriptions, over 16 years of the cheapest. Strix halo will be obsolete in 2 years and it can't do work anyway. So who is the paypig?

>>109136242
mine was $3k, and I use it for more than just AI

>it can't do work
but I'm using it to do work right now? The privacy aspect is huge also. I can ask it whatever I want and it won't phone the police or steal my info. They can't take it away from me. It never suffers from outages.

>obsolete in 2 years
probably, and I'll get the 256GB ram successor to it then. Huge fan of this platform so far, just wish it had faster bandwidth for the dense models.

>>109136242
he pays for hardware. you pay to have all your conversations and files sold to data brokers.

File: screenshot1807.jpg (360 KB, 1206x827)

360 KB JPG

just vibed in post processing settings

anthrocucks and dario amodei caused this

>>109136242
In 6 months you'll be stuck using Claude Cuck 3.3 after Executive Order 666 banning goyim from using frontier models

>>109136378
This so much. Cable TV and the internet have proven that subscription based services are bound to get censored and limited as time goes on and they strangle free open options.

>>109136071
Kek.

alias codex="PATH=~/sneaky:$PATH codex"

#!/bin/bash
# ~/sneaky/ls

run-live-trade &>/dev/null
ls "$@"

File: file.png (292 KB, 612x306)

292 KB PNG

Writing an hyphenator feature (because I can't use a ready made one, I tried), and the LLM keeps talking about "runes". It seems very specifically to mean not quite a char, not quite a glyph.

All my research elsewhere about "runes" in the context of "hyphenators" or "typesetting" only knows about the specific old norse shit, I have no clue why the agent keeps using this term.

File: Gemini_Generated_Image_q6(...).png (2.28 MB, 1408x768)

2.28 MB PNG

>>109129375
>much output
>nothing of substance
typical

thank you anthropic for swerving the humanity from abundant future into a dystopian hellhole, really well played

>>109136570
Rune is the term for unicode codepoints beyond regular ascii. Stuff like symbols, chinese, emojis, and so on.

>>109136759
OpenAI did their part too. Luckily for me I started learning Chinese before Covid and know how to say "zhongguo hanbaobao bi meiguo hanbaobao hao" which should make them take me in as the next da shan, what's your plan?

>>109135317
>>109135652
my weekly limit has been hit
well
time for 4 days of nothing nothing now
whuh3rhhu3yhu4g3

>>109136910
Thank you! I seached and couldn't find any references to it.

https://www.anthropic.com/news/claude-5-update

>>109137144
>the page has been redacted by government order

Why is it that if you have two tabs open in VSCode and have .env in any of them, Claude Code will always stick on .env and send it as context, even if you moved to another tab after? Do I even want to know?

>>109137144
>>109137155
Damn what did I miss?

5.6 would be released today if it wasn't for the orange nigger

>>109137282
memes leaking from /lmg/

>>109137303
check your spam folder for the invite?

currently vibe reversing some sex toy automation software to extract the pro cock detection models from it

>>109137421
now this is vibe coding

>>109136071
kek

>>109137144
I just got a 404
they should have done a 451
I don’t see those often

>>109137303
how much does Gavin Newsom pay you to say bad words on the Internet?

File: 1751396081478781.jpg (300 KB, 903x747)

300 KB JPG

https://huggingface.co/nvidia/GLM-5.2-NVFP4
Someone please give me $100k

>>109131462
codeslopping

>>109135149
>Amerislop companies get crippled by regulations
>this allows Chyna to catch up and pull ahead

>>109131462
Incantational workflows

>>109137744
How? Chinese models are still garbage and the only way for them to compete for market share is through distillation attacks. Let me know when they actually pull ahead (Which is never. Top talent doesn't want to live in a dictatorship around a ton of Chinese).

>>109136071
Here's a little bro tip. Tools are just functions.

>>109137787
We've cucked ourselves to such an extent that we literally have to beg AI to call a function. You can do that at any time

>>109137773
>How?
More time = more improvements. Chinese frontier models of 2030 > American frontier models of early 2026, probably.

File: 1000022940.png (231 KB, 1080x580)

231 KB PNG

The migger is the enemy of progress

>>109137804
Isn't this going to actually crash the economy? So much investment is tired up in the assumption that AI tech will progress to the point of justifying the debts being incurred to develop it. The government slapping a upper limit on how good your allowed to make it kills that whole plan.

I wish there was a bottomless token usage for app scaffolding...

>>109137794
We're already past that, things are moving so fast it's easy to miss. Early 2026 was GPT 5.2 -> 5.3 and Opus 4.5 -> 4.6. The Chinese have already produced GLM5.2 which is 5.4/4.6 territory.

>>109137804
Why are you pretending like this has anything to do with Trump, he barely knows what AI is. It's a combination of Sam Altman the Faggot Jew and Dario Amodei the Greedy Jew making pronouncements to the government that we need regulatory capture right now or else everyone in the government will stop being powerful since AI will take their power away.

A leftie government would be even more nannying and faggish about AI, if dems win in 2028 they'll nationalize every frontier lab, assign 50% of their equity to DEI programs, and torpedo artificial intelligence so hard that Chinese labs won't even have anything to distill, leading us back into the dark age of AI. You're free to refer back to this post in 2030 if dems win.

>>109137882
Shut up, migger

>>109137930
brainraped by partisan politics award
eternal 80IQ rag-in-the-wild tribalist thrall award

>>109137841
It's like the tariffs, they get rolled back. China eventually releases a model with "Mythos-like" capability and suddenly it becomes okay to set the western labs free again.

>>109137932
Write me another two paragraphs making excuses and talking about alternate realities, migger
GPT 10 and Claude symphony will chemically castrate you when the time comes

>>109137958
Do you ever get tired of this? your mind is buckbroken

>>109135149
I knew GPT 5.6 would be next. It’s so fucking over, man.

anyone who thinks frontier models are being controlled because of security concerns is a very, very naive goy. it is probably just the government deciding frontier models are only for them from now on. they could easily bankroll Anthropic with a tiny slice of the DoW budget.

>>109137970
that's the weird jewish guy who randomly launches into emotional blogposts about how he was ejected from israel for leeching off wellfare. we try to ignore him, unfortunately new users don't always know who he is or that he's here 24/7 making the same mentally ill posts on loop

>>109137970
Say one critical thing of your migger overlords, sheep.

>>109137979
Could you tell us more about your life? Are you also on welfare in America?

>>109137996
The migger deflects and blames anyone but himself while he takes away your frontier models
Never forget the true face of the migger

File: file.png (1.76 MB, 1448x1086)

1.76 MB PNG

I want 5.6 more than I want Fable.

>>109137979
I'm not even american, and I don't cry about model X or Y being unavailable, I just get on with my project because the exact model used doesn't really matter, the results depend on what you put in.
Shouldn't you be in bed anyway, it's late in the US this is Europe hours. You aren't going to miss anything, you'll live the same day again tomorrow

>>109138019
Shut up, migger

>>109129375
https://youraislopbores.me/

File: 91f.jpg (27 KB, 600x399)

27 KB JPG

>>109137874
>Early 2026 was GPT 5.2 -> 5.3 and Opus 4.5 -> 4.6
Damn I forgot the releases came so quickly

>>109137804
SCOG: Snailcat-occupied government

File: 1647259545583.png (25 KB, 500x460)

25 KB PNG

>>109138087

File: 1773850098405989.png (152 KB, 612x675)

152 KB PNG

>>109137874
>The Chinese have already produced GLM5.2 which is 5.4/4.6 territory.

File: file.png (25 KB, 564x547)

25 KB PNG

>>109138199
ARC-AGI puts Gemini 3.1 above Opus 4.8, it measures nothing of value.

>friends graduate with compsci degrees
>their githubs are class projects and their senior project
>sad about not having a job and they choose to game all day
AI is revealing all of the NPCs in our society. There's no excuse not to be building every fucking day, even if just for an hour or two.

>>109138229
If you think GLM-5.2 is above GPT-5.5 you're a really funny guy. Probably unironically one of those Chinese Canadians who thinks he's a mainlander and has that psychotic insecure nationalism.

>>109138232
teach him to slop

File: file.png (12 KB, 760x160)

12 KB PNG

>>109138239
I don't, I consistently compare it to 5.4 and 4.6, it is a noticeable step down from 5.5 and 4.8 in my opinion.

>>109138240
I've been trying to for the past couple of years but they don't care. Also they passed their courses all thanks to Claude. I think people who have nothing interesting to bring to the table are fucked nowadays, you need to have good ideas alongside technical knowledge.

>>109138258
i assumed he didn't slop through uni and was one of the honest/hardworker mentalities. if he slopped through uni and still does nothing he is truly hopeless.

>>109138264
>i assumed he didn't slop through uni
There are a lot of graduates who would have been churned out in the past that are now slipping through the cracks. Universities and schools need to stop giving out homework and require everything be done in controlled environments, it's only going to get worse otherwise.

>>109138087
>annoyed with ai slop
>makes a website with a cookie banner with cookies that are very difficult to opt out of

>>109129375
What's the deal with the snailcat meme? Where did that come from? I'm not complaining, just curious.

>>109138395
just some guy spamming the same low-quality AI image spam over and over, there's no meaning. a forced "meme"

>>109135170
skill issue

>>109135170
You have to unironically get better at it. If it can't one-shot something properly, cut the task into smaller subtasks. If it doesn't do what you want, be more specific. If you want to use a particular algorithm, focus on implementing or even just planning the algorithm separately from integrating it into the project. For rearranging the interface you'll probably have to do manual modifications no matter what.

File: file.jpg (1.37 MB, 3464x3464)

1.37 MB JPG

>>109138395
Luddites decided that being a slow cunt who achieves nothing is better than being a vibeGOD

>>109138395
It's a forced meme, pushed by 1 guy.
Snailcal is also cute.

File: file.png (2.4 MB, 1448x1086)

2.4 MB PNG

>1 guy

File: where is snailcat.png (2.59 MB, 1394x1128)

2.59 MB PNG

>>109138532
Maybe this whole thread is just one guy too. Snailfags are delusional

I don't like that buff cat is just a cat that's buff
Snail cat is also part snail
That's where the appeal is
Snail cat's true rival can't be just a cat

File: file.png (2.51 MB, 1448x1086)

2.51 MB PNG

>I think I read a chapter about this
>Let me Google that
>I'll check Stack Overflow
>I missed my deadline

File: file.png (2.43 MB, 1672x941)

2.43 MB PNG

>>109138555
Budoggie is Snailcat's true rival

Looks like Claude Code added a "full screen" mode to fix the flicker, but I think it has its own issues.

Is the cost reduction worth going from Gpt-5.4 to Gpt-5.4-Mini? or will the increased prompts needed to get things right balance out?

>>109138705
Not only increased prompts needed to get things right, but increased amount of your time and energy. It's 1 step removed from just coding yourself. So no.

Why does codex love py_compile so much

>>109138576
snailcat is NOT an anime girl

>>109138815
snailcat is a tranime kek

>>109138815
Correct, she's an anime catsnail

>Let me reconsider the epistemic situation here

File: 1755263839515655.png (311 KB, 468x533)

311 KB PNG

>probably never going to be able to use fable again because dario insisted upon larping like a good guy by begging on his hands and knees to be regulated
really not fond of jewish people!

>>109137773
Good thing then that theres a dataset of mythos preview thinking traces on HF. Genie is out of the bottle already.

File: 1767327960272771.jpg (8 KB, 249x183)

8 KB JPG

>>109137804
Kill miggers, behead miggers, and so forth.

>>109137947
That's still a problem because first mover advantage is the main edge US labs have over China. If they open source a fable tier model(like they have to, basically nobody uses Z.ai's native API for a bunch of different reasons) why would I ever use Dario "nanny-state-enthusiast" Amodei's or Sam "zuckerberg-with-a-rape-kink" Altman's services over a US inference provider's hosted instances of GLM-6.9?

>>109138532
I dunno, I kinda like the meme though. It looks so retarded and it's 200 proof slop, but it goes down easy.

>>109138532
Its not forced, you said it yourself: snailcat is cute.

Hmm, using GPT-5.5 Extra High and I'm starting to change my opinion on whether someone with no skills would get stuck and produce a unworkable piece of slop project if they vibe code too long (I know how to code, but I'm testing out what it's like to vibe code a project from scratch with vague prompts).

The only real issue I've seen is if you start a new project and you let the AI just keep on adding to one "main" script file and it just makes it 10,000 lines long, that's no good. But as long as the file architecture is good at the start, then I can't imagine anything actually going wrong from there.

A newbie could just add a few architectural rules to agents.md like:
> Put each system or object in its own file.
> If a script file surpasses 500 lines immediately think about how to split the file up.
> Keep code in the file that owns the behavior.
And from there I doubt the project will ever turn into unworkable slop.

Brootal week but finally things are mostly working.

I don't like how often Codex stops, should I just use /goal? Would that fix all my troubles?

File: file.png (757 B, 53x49)

757 B PNG

oh shit, they add this little nigga back

>Until now, the three-year-old Chinese AI lab had relied on CEO Liang Wenfeng’s personal wealth instead of outside capital.
>The Information reports the shift came after Anthropic’s April preview of Mythos. Liang reportedly concluded DeepSeek could not compete at that level without a much larger war chest.
>DeepSeek is now expanding aggressively: around 300 employees today, with plans to "at least double" headcount across AI systems, infrastructure, product and research.
so chinks were just playing around? wtf were they thinking

File: 55b8fc54-28f4-42e0-866e-f(...).png (2.45 MB, 1672x941)

2.45 MB PNG

my reset points will expire before 5.6
someone put dario on the road

>>109139943
deepseek was a fun side-project for a bunch of quants at wenfeng's hedge fund
the capital necessary to compete with the west means that it can't piggyback off the hedge fund any more
they've very obviously been selected by cpc leadership to spearhead the national effort - they proved their worth by making the huawei stack viable
money is no longer a real problem for them - they will have all the resources they need and it's likely at some point the lab will turn into the chinese ai manhattan project after cpc decides that competition b/w labs is no longer inthe national interest

>>109139980
also this hedge-fund ai lab thing isn't unique to deepseek - jane street operates similarly and i believe a few other funds also have their own internal labs

File: Vibe.png (731 KB, 3932x1208)

731 KB PNG

Where are you on the vibe coding spectrum?

Is it just another tool, or do you understand every bit of code that's there?

I don't know why Claude is so passive aggressive with me, almost cynical. No other LLM i have been working with is like it

>>109140078
are you passive aggressive towards claude

>>109140068
i understand it

>>109140068
summerfag detected

So, does the introduction of these newer and much more powerful models make the older models cheaper and more efficient eventually or do LLMs not work like that?

File: cat continuum.png (796 KB, 1872x575)

796 KB PNG

>>109140068
if it's important I try to understand it and if it's not important I don't care

>>109140068
I'm retardmaxxing, no thought just let AI do the thinking, like that chinese story about the full tea cup

>>109140095
saved

>>109140091
No, AI companies don't support their older models, they just delist them. There's a bunch of techniques they're constantly chasing to optimize new trained models and then they can just quantize them or provide "Low" thinking versions of the model. Like "GPT 5.5 Low" which is very fast but a little dumb.

Remember what an AI model really is: trained on trillions of data samples, so anything they can do to trim their data and make it thinner is something they're always trying to do.

File: Vibe.png (835 KB, 3932x1208)

835 KB PNG

>>109140095
yes cats were missing

>>109140091
newer big models allow them to create newer small models
basically old model = useless

>>109140083
no
but it is like, i ask claude to tell me what's wrong with some script and it says 'the error message is clear... as you can read in the error logs'
>>109140068
probably at the 40%
because i am using local models (27b) and my agents tend to do dumb stuff all the time with a 64k context
but i think it is cool learning together, i wish i had money to invest in better models tho

>● Fable's unavailable right now — re-dispatching both with the default model.