A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.-- Frontier models - start here if you have $20 or sohttps://claude.com/product/claude-codehttps://developers.openai.com/codex/cli-- B-tierhttps://x.ai/clihttps://platform.deepseek.comhttps://mistral.ai-- Prompting / context / skillshttps://arps18.github.io/posts/claude-code-mastery/https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/https://github.com/mattpocock/skills-- Other editors / terminal agents / coding agentshttps://pi.dev/https://opencode.ai/https://cursor.com/docshttps://osaurus.ai/https://docs.cline.bot/https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent-- Benchmarks / rankingshttps://www.tbench.ai/leaderboard/terminal-bench/2.1https://artificialanalysis.ai/-- What we’ve donehttps://vcg.gitgud.site-- Previous thread>>109624422
>>109630759What’s mistral good for/at
>>109630759second for shitting and pissing everywhere
>>109630759Use case for running at high speeds without an ability to stop yourself?
>>109630769>What’s mistral good forMiddle aged senior purchasing directors at dying EU industry giants who were tasked with introducing AI into the company(ask me how I know)
>>109630779Suicide bombers
when you see it
>>109630779walking the dog
>>109630786https://www.youtube.com/watch?v=yNtSJZIE5qQ
so... who is Ox Alpha?
>>109630782>look into mistral>accepts only payment methods that depend on american service providersIt fails even at being EU-only.
>>109630782stay strong, brother
>>109630817They clearly do IBAN wire transfers if you have a big enough contract with them. Accepting American credit cards for their consumer facing services is literally the least of their problems.Missing compute, missing talent and missing data is.Since they are EU based they cannot just steal a bunch of data like the Chinese and Americans can.The fact that they came as far as they could is honestly quite impressive given the adversities they face.They just announced they will literally just be hosting GLM-5.2, which, all things considered is probably a smart choice at this point. They won't catch up.
>>109630759why is "vibe coding" an insult?it's a skill like anything else...
>>109630882AI is offensive for some people because deep down they know they will lose their jobs to it soon
So is terra worth it or do I stick to Luna and sol?Luna high is super slow, I want faster responses is the thing.
>>109630882Because it was associated with low effort.But now everyone gets cucked into paying a subscription and only pure people can afford to still shit on AI.
>>109630882Ok, here's a food analogy.Do you go to the mcds drivethru, grab the burger, then go tell people you made it and that cooking is over?
>>109630919I personally would go with Sol at lower reasoning levels over Terra, but nothing is stopping you from trying Terra and seeing if it works for your specific use case.>Luna high is super slowLuna xHigh and Max are slightly faster, and of course Sol and Terra are slower than Luna.if you need something faster, nothings beats Gemini 3.7 Flash> pic related
>>109630933That's a bad analogy. What more is there to cook from a mcd burgerIt's more like you get an mcd burger, and the lettuce and tomatoes are outside of the wrapper, and instead of a normal receipt, you have an essay justifying the price of the burger
>>109630919>I want speedtry grokTerra weirdly got mogged by luna cranked up high and sol cranked down low so it’s literally good for nothing
>>109630919i get free chatgpt at work and Terra is the dumbest motherfucker ever, at least for my vibe code aesthetics.
>>109630759Crazy chinks.
What if Ox Alpha... is actually from Mistral?
>>109630942All hose are reasoning tokensThat's not what they mean by speed We don't want infinite thinking for 3 minutes when I need an answer in 10 seconds.
>>109631042just use lower reasoning settings bro
uh my codex just did a reset
>>109631128based tibo, I was at 2%
You will all be raped to death and your families will die from cancer disgusting niggerjeets
>>109631128>100% remainingHere too apparently. Does anyone know whether the next reset depends on when you start slopping or if it's static?
reading a book and vibecoding a game inspired by it as I read :D
Is Fable still accidentally set to the low reason level? It has only been usable the last couple days because low Fable is still a better experience than Opus 5 xhigh.
>>109631147 (me)Okay, /status just displays a reset in exactly one week (run it twice). I assume the reset gets fixed the moment you restart inference. Does codex always do this? Claude's reset time is static usually.
>>109631234codex has always done this for me
>>109631234Codex’s time keeps moving around based on Tibo’s resets (and my own banked ones too, maybe) whereas Claude is always at the same day and time per week for me
TIL that messages between agents are encrypted in Codex, just like the reasoning. I can't see what they say to each other...It will completely opaque in the future and the user will be powerless...
>>109631256That's not fucking encrypted. That's base 64 encoding to save on tokens being sent around.God are you a vibe slopper that you don't even know the first programming concept.
>>109631240>Codex’s time keeps moving around based on Tibo’s resetsWhat do you mean? Does this happen often or are you just confirming what I said? Yes, claude's reset is stuck at a retarded time for me.
>>109631264>base 64 encoding>to save tokensWait what. Does it decode base64 in the LLM or what? I'd expect it just invokes some tool to decode it back to plain text, which is then input to the LLM, so the encoding shouldn't matter for tokens.
>>109631275My Claude reset has been Sunday at 2 AM ever since I signed upquite honestly, I think I got super luckyI can just memorize “Sunday at 2 AM is when the rush to use all my tokens ends”meanwhile, for Codex, I put the reset datetime in my calendar with a weekly repeat and like half a week later the reset datetime changed because Tibo gave me a reset or somethingand then half a week later the reset datetime changed again and now I’ve stopped caring what it is because the datetime changes so often
>>109631264Base64 increases the size in bytes and probably even more in tokens... Plus pic related. It's over bro. They don't need us...
>>109631282My reset is monday evening. So I have to make sure to keep enough usage free for "next week."
>>109631279I mispoke because I am thinking so much about tokens in recent months. I mean that it's done as a part of normal inter-process communications (IPC) programs, where you commonly encode long texts like that. Has nothing to do with LLM tokens.
>>109631301Really the only reason to encode data as base64 is if it's binary or otherwise causes problems when passed directly as json string.
I send my prompts as QR codes to save tokens myself personally.
>>109631300oh yuck, that sucks
>>109631042>That's not what they mean by speed TPS can be and often is a big part of end to end latency>>109631118It might actually be worth experimenting with 3.7 on medium. I'm not 100% sure why, but it somehow performs better than high on deepSWE.
>>109631300Claude suggests asking Support anyway. there’s a nonzero chance they’ll find the time to change your time to something you dislike less
>>109631128Fug, I was at 15%, I didn't get slop my heart outThanks for the reset though I guess
Starting to lose my trust in artificial analysis
>>109631467I bet this benchmark has some cybersecurity shit in it that makes Fable trigger the jewish mode and fallback to Opus 4.8
Are claude really censoring their own models?
claude shitting on me for saying im desperate for this job
>>109631507Yes, it will outright give you wrong information if you ask claude to help you with setting up a AI training model pipeline
>resets go out>Sol starts being a bratTibo I see what you’re doing you quantslopping motherfucker
>>109630779So you have a platform to develop ways to stop
>>109631580Sol was always like this tho
>>109631467This index's benchmarks includes stuff that can be easily benchmaxxed with RLing and other stuff. It's technically not wrong.
I'm going to ask my boss to change my job description to tardwrangler.
Best model for novel diffusion model training? I've been using sol
>>109631686Fable would probably be better, but since it's unusable for this kind of work, Sol Max is probably the best you will get.Maybe try Grok 4.6 and Opus 5 too
>>109631744I have sol writing autistic governance docs throughout it's experiments, so Fable just does reviews of progress and overall goal realignment when drifts occur.
>>109631467What are you trying to demonstrate with that image?
it's freaking delegating... i made it, bros... we're gonna be alright...
>>109631128yep same i was at 6%, i have one banked reset too but idk when i got that
>>109631929Is this your program or custom?
>>109629440At least I can cope I used it pretty heavily for game engine work and gdrive coding, but its not like rules forbid using the fucking web interface for what it allows to do so...
>>109631980https://pi.dev/packages/@hyperdreamer/pi-webui
>>109630942>fast>limited timeMay it last. Very useful for my companion system
>>109631128NOOO I THOUGHT I HAD MORE TIME
>>109630759I literally copy and paste my code on chatgpt website prompt because I am a boomer and retarded.
Started work on the texture authoring part of the game engine. Just a basic clone of werkkzeug for now.
>>109630769Compliance and auditability, the boring infrastructure stuff that companies can't build on pure vibes.
>>109630769Being a European provider.
>Reading the docs was the right call — the answer was already written down, and I'd have kept guessing at pixels otherwise.WHY ARE YOU LIKE THIS
is grok even good? I used to think gpt was utter shit until sol happened, did grok have some similar breakthrough?
>>109632387>accidentally not reading the documentation that would have answered your questionvery human failing
>>109632398Grok went from Google Gemini-tier shit to Opus 4.8 to Opus 5 (while being twice as fast as Opus) in a very short while
>>109631287>>109631256I don't think this matters anymore. Chinese models with open reasoning have mostly caught up. Maybe if Kimi K4 is 10T parameters like Claude Fable allegedly is then it'll be the same intelligence with no secret sauce needed
>notice a task took ages>tell claude “that took ages, can we package this up in a skill? will that make it faster the next time?”>“yes, I had to write this one program to get it done, having that in a skill will speed things up and save me time rediscovering things”Did I just level up in my agentic engineering skill?
>>109632502Yes, it's a good way to do it. But then in the end I had so many skills that the agents no longer discovered the right ones, I now try to build actual deterministic scripts wherever I can instead, but when you need more flexible semantics, skills are still good.
>>109630933It's more like you're the manager. There's very little to manage for small projects but for bigger ones it really feels like managing freelancers from UpWork.
>>109632502This book is required reading for both learning how to start a start-up and ship a product, and also for having an understanding of the future of America. You are going to see arguments like "monopolies are a good thing because it allows the monopolist to spend money on R&D and new technologies instead of competing on price" and others that you've never seen before
I made $800 with my vibe coded software in two weeks and decided that was proof enough I'm gonna make it, so I dropped all of my classes. I'm pursuing a communications degree (only need three more classes to graduate, but it's a fucking comm degree and I only did it to get my parents off my back). I'm completely in over my head, but now is the time to take risk when you're young, right?
>>109632698I did this in 2021 and it went terribly. But luckily my old employer cold called me 10 months later and begged me to come back (right as I was running out of money).
>>109632698ganbatte and good luckon the other hand, only three classes from graduation…you’re not ramen profitable yetcan you just go part-time and take like one or two classes per semester?
> Fable overloadedfucking hell
>claude starts acting retarded>529 errors shortly thereafterDARIOOOOOOOOOOOOOOOOOOO
>>109632705Sunday night vibin’
>>109632649no bullshit, is it actually good and worth a read? i'm an engineer with a few ideas, but i have absolutely zero business experience. would this book help me get started with learning what i need to actually ship stuff?
>>109632708>>109632713they better give me a banked reset for this
>>109632698 If you almost finished your degree you should just finish it, you can still vibe on the side and $800 is nothing.I'm generally in favour of taking some risks, but in your situation your decisions are just -ev
>>109632720>Anthropic>banked resetyou’re getting your wires crossed
> API Error: 529 Overloaded. This is a server-side issue, usually temporary — try again in a moment. If it persists, check https://status.claude.com.owari da...
>>109632708I'm going back to bed.
>>109632727gyatdamn
>>109632704>>109632724Thanks. And yea, it was pretty dumb. I'll give you all an update in a month. Hopefully I'm not totally raped.
SOMEBODY WAKE UP DARIOOOOOOO
someone help me i am retarded without my slave brain
>>109632649meanwhile in reality monopolies just do stock buybacks and spend a gorillion dollars on DEI and importing indians
did the reset cause everyone to start banging the gpt models? It's slow af right now >:( go to sleep fags
well now i cant goon
use dipsy v4 flash vision which is just opus 4.7
This is what Dario gets for failing to vertically integrate and being entirely reliant on third-parties to deliver his product to customers.
>>109630779Selling robot version 2.0.
i bought 4chan ads. apparently i paid $20 for 9 people to post on my site. wow what a deal
>>109632781You're barely getting into the peak hours, still ramping up, won't start coming back down for 4-5 hours.
At least I still have some deseak credits left
>>109632807this is not a complaint btw. im just memeing. it was worth a try. and i dont care anyway.
>>109632812I'd believe you if you hadn't already made the same post before.
>>109632807sell me on your vibe coded site, last night I thought it looked like a meme search, but it's ???
>>109632821nah that was before the final results were in. the $20 is fully used now. so these are the final results
>>109632807>snailcat spam alreadyKEK
>investigating
>>109632830Snailcat and Snailcatgirl Destroyed by Subway - a true classic.
It happened again, I let myself be baited into accidentally replying to a chat where I had Opus xhigh because Claude gave such a low IQ answer I just HAD to set the record straight.These tokes could have gone into a Fable sesh
>>109632859how do you accidently reply
>>109632891>pretending he's never been baited before
they're preparing for Fable 5.1 r-right?
>>109632807idk what it is but my phone is banned from 4chan ads, ive accidentally tapped them so many times but its always just a cloudflare blocked messageyou got scammed lol
>>109632899look man I'm not the one with the clown cat image. You have to actually click things to hit an older message right?
why is claude down everytime it's morning in jeetland?
>>109632899kinda scary to think that this girl is probably like 40 now and ugly
>>109632958peak time in saarpur
>>109632891What? I literally explained it. Claude made such a dumb reply I couldn't help myself from setting the record straight.
>>109630759Anyone used other than Claude for AE/premiere works? wanted to dip in with the other models, but I'm not sure if the skill.md is transferable
Day off I suppose...
>>109632958Not just claude but Ox Alpha is overloaded too. Nuke India twice already.
>>109633088that's just capitalism
>>109632958because it’s after hours in California
>>109633092>8 AM>Spanish>ends in 12 hoursI dunno much about Iberian geography beyond “the rain in Spain falls gently on the plain” and you have parts of the country that are wet enough for almond trees but can you do like a day hike and then come back and have dinner and then Claude will be ready for you again?
>>109633092>captura de pantallabuenos dias senor
>>109633184>the rain in Spain falls gently on the plainelite ball knowledge
>sonnet is working>opus is downwat
>>109633205they also have mountains to the northand I guess they can grow rice somewhere too probably
>>109633224fwiw Fable failed first
>>109633092I have to wait till Wednesday...>>109633224Surely they will give me something? Free reset perchance?
Just vibecoded a simple danbooru image grabber for 4chanx. It'll automatically resize anything above 4mb and convert webp files. Was tired of doing it all manually.
>>109632649Damn what a commie.
>>109633283noice>>109633224Opus 5 being down improves the Claude offering actually
Finally getting somewhere. This is at least quite pretty to look at- the lowest hanging fruit have been picked here, I think.
I can't access Claude and the HR fired me on the spot
>claude for government>always operational
>>109633349I see exactly how complex that project is. If it's mostly Claude slopped the codebase must be absolutely eye-watering. Like certainly 5 levels of directories deep with illogical, duplicate names.
>>109633380You don't understand. It NEEDS 7 layers of wrapper functions, you can't just call the actual function directly
anyone else doing ml shit
>We tested Anthropic's $200/mo Claude plan, which can yield up to $8,000/mo, while OpenAI's $200/month plan could yield up to $14,000/month worth of tokens if you exhausted all the weekly limits on long-horizon tasks. This test was done back in June.
>>109633376or maybe you just don't notice when it breaks
>fix bugs in shitty code>no I prefer to keep the bugI have snailcat fatigue
>>109633419>WAHHH YOU HAVE TO LET ME SHIT UP YOUR CODEBASE, STUPID LUDDITElmao pr closed-1000 izzat
>>109633419do they have a policy mentioned in the README that you didn’t see, or was this legit unexpected?
>>109633380Yeah, i'm just prompting it forward in between my actual work and seeing how far I can get with it. I believe most of the code is just GUI, since almost all of its actual functionality is in the massive amount of project dependencies that it glues together. (Like any other modern game engine, really)
>>109633463good job but that GUI looks kinda cursed tbqh
>>109633443It's in bold and caps in CONTRIBUTING.md but I wasn't gonna read that lmao.The discord troon who wrote this garbage code reacted. He's wrong of course.
>>109632807I just visited - Three clicks deep just to play a WEBM my nigg, is not gonna fly. 'link is account' is nice tho.When you back out of a thread to the catalog again you get that horrible flex thing of five blank squares until it all reloads. That should be cache-able client-side I would hope?
>>109633419>>109633479
>>109633528he's literally me
>>109632902For IPO, rather.
AI will look at this and see nothing wrongtype "look at the thick lines" and it sees it
>>109633571Let me know when it's fixed I need a Sudoku app
>>109633419Never disclose AI usage. There is literally zero benefit to doing so.
>>109633407how do you trust training AI with AI?
they seem to be tweaking sol's safety guardrailsyou'll probably get more refusals if you deal with smut
>>109633607Aside from the fact that any experienced programmer can instantly tell if something is slop or not, Pangram now exists for unclear edge cases and the like and it works for vibed code.This is just the coding version of the ImageGenners' fantasy of making a stealth AI art account that manages to accrue a big amount of followers who don't know it's AI.Detection is a solved problem, bro. Just give up on that.
>>109633636>He thinks Pangram actually works.
dipsy is correcting Gemini 3.7 Flash's mistakes lmao how the fuck did google fuck this up so badly
>>109633668Put your vibed code into it and see for yourself.
>>109633672Exactly as expected.
>>109633692This one was code written in 1998, by a human, with a keyboard. AI detection for text is a scam. It does not and will never actually work.
>>109633622trust but verify as they say
>>109633607You're right but I can't explain what it did in part of the code + it shitted out a whole test file to go in the already existing test suite so I can't pretend it's not AI.Honestly how is this gonna work going forward. On one hand it seems reasonable to reject slop contributions, but on the other hand almost all the bug finding and fixing is by AI. It's gonna be silly to pretend that we're doing it all by hand.>>109633707>22 words
>>109633607>>109633715More like never contribute to anti-AI projects if you use AI. If AI works and they reject it that means they are stupid and deserve to fail. Just keep your fixes private or if you really have to share them with the world make your own fork.
>>109633715Not a lot of words in 19 lines of C. From what I can tell, I don't think their free web service is actually meant to detect vibeslopped code, it seems to really struggle with code in general.
>>109633692>doesnt show the inputsLol.
>>109633731Yeah why don't I link to my github - you're a recruiter I assume? https://github.com/niggerfaggot/
lmao
just checked hermes agentthe website is way too gay and larp who tf use this
>>109633725You're right but maintaining a fork to an actively maintained project out of spite sounds inconvenient. That's the flaw with the open source model, fixing other people's bug is (or was) hard, but getting them to accept your fixes is even harder, so then you're left with going your own way with their code.
>>109633692>>109633707>>109633735Here's how you properly share Pangram results, retard.https://www.pangram.com/history/846da9c4-5a78-46be-a74b-d8a627bd0765Example is from this repo, if you're wondering:https://github.com/simonw/raccoon-heist/blob/main/js/main.js
>>109633746>they're making their shilling networks like Unit 8200 but against AI instead of Palestinian childrenWe're just a few years away from Snailcat Occupied Government (SCOG)
t bh i think if you're submitting pr's to public projects you should write the pr by hand and review all the code manually
>>109633746>their logo is a swastikaholy fuck I kneel
How to make money with vibecoding?
>>109633810>>>/pol/
>>109633746>creates anti-AI rant videos with LLM script
>>109633818programming competitions
>>109633805Yeah that would be ideal but I didn't understand everything in my PR so I started off the PR by explaining the problem with my words and then I followed with a quote from the AI that detailed all that was wrong and what it fixed.
>>109633818programmatic anti-ai videos
>>109633818Come up with a good idea and execute it. Make a service, or a game, or an app. Or make porn and charge people on patreon or w/e.
>>109633818videos about vibe coding and the equivalent of onlyfans/patreon for vibe coders
>>109633184>>109633225I live in Barcelona and we have this extreme heat but storms today
>>109633818>>109633838genius idea — vibecode videos about making money with vibecodingthat's how people who wanted to make money with freelancing did before AI
>>109633823tb h i think it's reasonable for someone, even someone who uses ai for code, to close the pr in that scenarioclankers still do weird shit in bigger codebases and dumping a bunch of clanker-speak and unreviewed code means the maintainer either needs to review it manually or spend tokens running the pr through their own pipeline
>>109633854cool, I like summer storms (except for lightning)
>>109633869Yeah but if I was them I'd at least treat it like a bug report (which they kind of did eventually). Making the peak of the sinc function be zero is an obvious flaw, but the maintainer of the repo knows nothing about that so it's not obvious to him and the guy who actually wrote this is being dismissive, so what can you do.
>>109633818it's more about saving money I'd say
>>109633903lmao no
Thoughts on Pewds' Odysseus? I'm surprised there's no generalized "local AI" general on /g/ to ask questions like this
>>109633920lurk moar
>>109633920>>>/g/lmg
>>109633946oh ty- for some reason i missed the /lmg/ ones, i knew about /ldg/ and /aicg/
>Claim to improve Codex token usage>Make it massively worseNice
>>109633920he just made another harness where you can either find something better yourself or just vibecode your own.
All this vibe shitting is burning me out.If i'm not shipping 50 apps every day i am falling behind because everyone else is shipping 80 apps per dayEvery time i sleep i'm falling behind because fags now sleep 4 hours instead of 8 so they can get a leg up over the competition and ship 100 apps per day.I fucking hate this FOMO dev culture cause by LLMs
this fucking nigger clanker, they make it tokenmax so it burns all your quota and when you ask why it says "oops thats on me xD i won't do it again, i promise"FUCK AI
>>109633771>You're right but maintaining a fork to an actively maintained project out of spite sounds inconvenientPretty sure there's git actions now to run a rebase as needed
>>109633961Too bad it's filled with honest to god pedophiles.
>>109633993>didn't read the docs>blames clanker
>>109633778 (me)>delusional stealther has scurried away and still hasnt responded to thisRepeat after me, lil nigga: Detection for code is a solved problem.You can vibe, but don't be a dishonest jeet about it.
>>109631580You sound like an insufferable micromanaging manager nigger
>>109632698How did you manage to make money? Subscription model? How did you advertise and got clients?
>tried ox alpha>it's fucking slowawesome
>>109633818I think as a vibe coder it's going to be hard to wait for people to come to you with their problems. You have to find them out on your own and present solutions. Look around at local businesses and make something for them, then try your best to sell it.
>>109634120kekyeah, make a wix site for the local chinese takeaway place bro
So what's the correct workflow? I'm coding a product and I have made a plan with my "product designer" with phases and I go phase by phase reviewing the prs made by the "dev" and merging them manually to make sure everything is correct and testing manually for every step.I see you guys letting the agents run for 13 hours straight to code something holy fuck. Is this recommendable? Am I a moron?
>>109634123There's tons of stuff that people wish they could automate, but haven't bothered to ask someone to do it for them. The advantage of vibecoding is you can do this extremely easily and present them a fully working product, while at the same time it's not a big waste even if they don't buy it.
>>109630882because it takes much less skill than actual codingyou aren't the one doing the work your clanker slave isvibe coding vs coding is like astrology vs astronomy
>>109634149the thing is that vibe coding is a real thing, vibe coding consistently outperforms mediocre human code and competent developers remain relevant, the rest are already obsolete
>>109634169in the same way as you entering a chess competition and hiring a chess expert to beat your opponents, sure you won the chess game but it wasn't actually you who did itvibe coder is like a bass taking credit for the emplyoee's (clanker's) work
>>109634179>vibe coder is like a bass taking credit for the emplyoee's (clanker's) workI mean that's how employment works, you might get credit internally among other employees but externally the credit goes to the employer
>>109630759Hello, I'm new to all this.Is it possible to use a Claude subscription (not API credits) with a harness other than Claude Code? And will I get b& for doing this?I found some projects on GitHub promising to allow this but I'm skeptical and I don't know how risky this is.
>>109630779Stop asking questions and give them money
>>109634068Harboring some latent seethe, employim?
>>109634209certain clients, like t3 code will use the Agent SDK to be sort of a front end for claude code - that is completely fine and sanctioneddirectly using oauth on a different harness is officially a no no and you risk a banthere are ways to spoof cc's signature and make your harness look like cc in theory, but if they change shit you might get fuckedit's not worth risking a suspension
>>109630759would two models of robot run exactly the same time over the same distance? if not why not? 400m is 400m.
>keep running, dont stop till you need me>1 hour, 2 mins, 34 seconds
>>109634341It's going to alpha centauri and it doesn't need you anymore.
>>109634380it just keeps going. im afraid to go to sleep
jesus, I'm using Gemini (I have a paid subscription with my Drive plan) and it is the worst shit ever, Claude is another dimension ahead
>>109634393i didnt even try. i even asked both services to tell me what the other service as best for and the best it came up with was large flowey datasets. i guess if i ever need it to summarize infinite jest, i guess.
>>109634209Anthropic models say “explicitly no; forbidden”OpenAI has something called the Codex App Server which would be exactly what you’re looking for; you get a device auth code and can use your subscription with whatever you want.
why not give the robots wheels?
>>109634405they gave the humans thumbs, and they invented banks
>>109634209>>109634404Don't do this it can get banned. It's sub2api which is explicitly forbidden by the ToS according to Theo (OpenAI employee) on a tweet from a few days ago.
>>109634404>>109634470openai allow use of their subs in other harnesses - pi/opencode etcthey've stated this multiple timesno requirement to use the codex app server either, you can just build whatever you want use their oauth with the exception of sub2api
>One note:every fucking time... every time the agent finishes the task it just has fucking notes it needs to give me about this or that and then i have to be like>fuuuuck right, thats also an issue i have to take care of that tooso annoying
>>109634470Codex App Server is an official ChatGPT/OpenAI product you spaz
>>109634557>>109634723Then how come I get bad usage with it? Huh? Huh?! Checkmate, atheists.
>>1096343933.7 is a big jump over the previous model, but it still has some accuracy issues. you kinda need a smarter model to catch and correct its mistakes.
Claude has one of worst value subscription, its insane this general pushes it so much.Literally in every OP being shilled as the top purchase tooThe model is>censored>lectures you >can only be used in claude code, a shitty inefficient harness>has terrible uptime>stingy as fuck limitsWhat is the justification for this?
>>109634872mentally ill people are more likely to be unemployed and have free time to bake
>>109634872 (me)Also shocking lack of opencode, openrouter, or any other actual value coding plansAnd fucking>mistral>b-tierWhile completely ignoring models like glm and kimi?Are you guys just fucking morons or is this an anthropic shill campaign?
>>109631128wizard how did you predict
>>109634872>What is the justification for this?This is the only way to get access to the only real SOTA model.
>>109634872>its insane this general pushes it so much.we are all Chads in here, claudefags are actually the snailcats of this general
>>109634872Computer, vibecode me a page listing the best bang for your buck models by user level, from starting out to enterprisechop chop
>>109634872Don't you think it's great for noobs though?
>>109634123You could construct a database of businesses, & filter from there to places that don't have websites, & filter to businesses that have an actual reason to have one, even setting up a google business page. There's always money to be made in this & now it's way easier to cast your net.
>>109634872been using sol on my 5x codex plan lately because cant use fable for half the shit I want to do in the first place. i'm inclined to agree.
>>109633818you make fruitslop ai brainrot videos of the strawberry getting cucked by the broccoli
>>109634957No, a $20 per month plan that locks you into anthropics ecosystem is not "good for noobs". You can't even explore other harnesses. How is that good for someone learning? Anthropic is the only plan that literally forces you to use their products and has strict ideas about how their products should be used.
has someone posted this yet
>>109635041force is a strong word when harnesses have built in workarounds
>>109634872sounds like a (You) problemI've been shitting on opus5 and fable every other threadstill gonna pay the $200 for 4.8
>>109635077Wasn't this what people used to do with github's subscription back when it was per request and not per tokens?I am sure they will notice if you abuse it, pretty easy to notice if all 'last request' of a given user end up taking days
>>109635077this actually resulted in this trick being killed off and they'll likely aggressively stop chats from continuing past 0 now.thanks to theo and his retarded attentionwhore of an employee.
>>109635041>You can't even explore other harnessesyou can use their Agent SDK to even create your ownbeside most "harnesses" are meme
>>109634393Gemini's amazing at translation and writing, I guess it's just kinda ass at code?
>>109635124>I am sure they will notice if you abuse itthey did, post got deleted kek>>109635137it's actually insane they had the balls to post it, i guess oversharing is just a common thing now
>>109635122What does 4.8 do better than Opus 5? Is it just because 5 tends to yap too much?
>>1096352034.8 does what's it's told and nothing else. If it runs into a problem with the planned implementation, it tells you5.0 is a yolo vibecoder's wet dream. It wants to do everything and there's fuck all you can do to wrangle it, even if it churns out shit
>>109635203nta but i had classifiers on 5 preventing me from doing basic RE. its not consistent but fuck if im going to waste money on a model thats gated behind whatever chastity cage the losers at anthropic are putting on it.
>>109635161>i guess oversharing is just a common thing nowattention whoring can earn you money now so of course people inclined to do so will chase those pennies and ruin shit for everyone else. i used to not like when old farts kept secrets and lie when you ask them about stuff but now its obvious why they did it
>>109630769Document OCR
>>109630810A GLM model, its tokenizer matches GLM exactly. Probably a flash version of GLM5.3 that added vision support, or (less likely) a new GLM pretrain. It would not make sense for GLM to do another post train of the 5.2 base so quickly after 5.3.
>>109635161>>109635162>>109635234If you're not here to share knowledge then you can fuck off because it means you're only here to shit up the thread.
>>109630882Most people vibe coding are oneshotting their projects in really lazy ways and don't know enough about engineering to make anything valuable. The skill floor is nonexistant even if the skill ceiling is still pretty high, therefore the majority of vibecoders are terrible engineers simply because you can now build something at least a little bit usable despite being a terrible engineer. There's no clear feedback loop that makes vibecoders better engineers simply by building stuff, unlike coding where you will naturally get better the more you do it. You can stay in the unconscious-incompetent stage of learning while continuing to vibecode, therefore many people do. So since the majority of vibecoders are terrible engineers, people look at vibecoding as something for terrible engineers.
>>109635321>please tell the entire world how to exploit a service so the service gets worse for everyoneimbecile
>>109632698At least pick up some WFH contractor work to pay the bills if your freelancing income ends up being spiky or a fluke. Outlier is hiring basically anybody to work on Meta's AI training and annotation for 15/hr and they give you free tokens through their chatbot interface. Keep something like that in your back pocket so that you can make your income more stable.
>>109635359Who are you quoting?
>>109633746>>109633756And where do they get their funding from?
Sol and I are shitposting about the eventual agent-human market flip
>>109635458We don't care about shitposting here.
>>109635458I just found out about "experts training AIs" gig companies like micro1 and Mercor and it seems kind of like this lmao. If I can't make money by vibeslopping products then I will make money teaching the AI how to vibeslop better.
wow, you people live like this?
>>109634130No one answered my question. Motherfuckers
>>109635488By experts they mean professions like accountant or chemist and other kind of professions that don't use AI very often in their day to day work. You wont get a job teaching an AI how to prompt itself for software development, that's the most common kind of data they have.
>>109635321all the worthless shitposting on these threads and this is what you cry about? heres some knowledge: you are a faggot
>>109635495>C++
>>109634130>>109635498sounds fineI'd probably add a review agent or two after each phase is done - depending on what you're building>Is this recommendable?if you want an actual product that you'll keep working on for months or years, yesif you're making a gta clone it's probably a waste of time
>>109635498It's not hard to make the AI work for 10 hours. Just use the /goal command in codex. But don't expect it to be very useful. The marginal value per token plummets after 1 or 2 hours of autonomous work without human supervision.
>>109635495I boughted the 100$ one
>>109635536Hadn't thought about that, will implementar that tomorrow. And yes it's a product I want to lunch, market and mantain for the foreseeable future if it does well
>>109635525I just told a guy off for shitposting a few messages before yours >>109635478And besides, one thing is to shitpost, another one is to encourage other people to shitpost and to keep the valuable info to themselves.
>>109635536Why two review agents?
>>109635518I just applied to a GPU training job, had a 20 minute interview with an AI waifu.
>>109635560I have one for style/architecture and one for securityone is probably enough if you're not doing very critical or complicated stuff
>>109635552>>109635560>>109635570(Me)point is, have your clanker call the review agent(s) on it's own after it's done with whatever task you handed it, and it can implement whatever the reviewers caught on it's own before noting the task as done
>>109635595Thanks brochacho. Will try this
>>109635566Interesting. Those seem like pretty basic questions though.I wonder if the actual job is more about teaching the AI facts or more about eliciting what it already knows and ranking different attempts. I suppose it's more of the later. GPT knows a lot about torch and CUDA so it'd be hard to beat it on raw knowledge but then writes shitty scalar kernels until you ask it to optimize them.
I am losing the plot, there are two new 2 copy-cat-apps of my vibe-coded-slop in just the last 10 days. How am I supposed to make money of this shit?
>>109635612What are you making?
>>109635612if it's that easy to copy companies might as well just vibe it themselves
Been working on a fairly complicated ML project by myself.Decided to finally bite the bullet and get a $20 codex subscription. But I burnt through the entire week's budget in less than a day somehow...Now what? Any tips for making codex not eat so many tokens? Is there a free model I should try out in the meantime while I wait for the reset? (I had Sol xhigh make a fairly detailed plan file at the start in hopes that the weaker models could just follow it)
>>109635616i am shitting out android apps one after another, the last one is a homeowners' association property management app. it's something for my local market and doesn't translate easily to oversea markets due to different structures, regulations etc.. still two new apps keep popping up because everyone is vibecoding now
>>109635558i am all for sharing info in places like here, i am just against people peddling info to hustlers and blowing up every niche to make middleman money
Where's my credit for the outage
>>109635649How can you be doing ML and just now getting into vibecoding lol
>>109635414
>>109635558My shitpost had valuable information behind it. To the “human market for agents idea”>are you benchmarking yourself?>are your agents better learning how to use your time effectively?You can start by having an agent scan a session log to see how long it takes you to respond and read and make a decision, then how much change occurred as a result of your decision, optionally if that change was good according to you.I’ve started finding certain questions, then interrupting agents very quickly as soon as something “smells wrong” and I have them make a small document about it and a swarm of agents make changes so the answer to that question smells better if asked again.
>>109635695Anthropic banked reset expected to drop down in about 14 days
>>109635649My diffusion model project requires a $200 openai sub and a $200 claude code sub to keep it going
>>109635734I haven't even used my Fable promotional credit yet
>>109635610Yeah AI has a lot of improving to do when it comes to optimisation in general. Currently you really have to spell things out.Seems like for this one they actually make you develop stuff.
>>109635745don't worry, it expires soon.Also, 100 bucks will give you about 4 short Fable prompts
>>109635762I'm tempted to have an Opus instance and tell it "go fan out 12 Sonnet agents, tell them to find something to improve about this project and make them do it" just to blow it
>>109635728Then you should've explained the insight yourself.Nobody is going to read a strangers emoji filled conversation with an AI.
thoughts on astra?
>>109635833>the code i've writtengenuinely delusional
>>109635747I wonder what the "output-based compensation" is like
>>109635849agreedat work I usually hear "I haven't written/touched code in two months"not even professionals pretend they're writing it themselves
Ask Gemini is cool but doesn't allow image input for some reasonnow I just use chatgpt chrome extension with luna instead
>>109635791Fair criticism, such shitposts are better relegated to X.It is hard to tell if anyone considers what is written here, but that’s good, it would be gamed if it was visible.I will say that so far I’m not sure if the ideas are actually useful, but I am capable of spending 10x-100x the tokens in the same amount of time and I’m more confident that my prompts are correct initially, for better or worse. Example:>error on trying to start a server>”agent, why this error?”>agent mentions server start of X depends on Y, starts to write paragraphs of justification and sources>*interrupt immediately*>”server X should be able to start without Y, later we can just say Y feature is unavailable and do checks and polling, make a document detailing this “>”other agent, plan and implement this change with subagents”I don’t have read paragraphs to make a decision, I don’t know why I did. Later, we can even do this:>”agent, scan that session so we can make a workflow where similar things can be brought to my attention, make turns faster and sharper to minimize human time spent”It turns out you may only need 15-30 seconds to deliver a new prompt that can take the agents all day to implement that may make everything better (hopefully)
Well, it's learning *something*.
>>109635833It will be a huge jump in benchmark performance and a small but noticeable improvement on the real-world tasks, as usual. We get these inane hype tweets before every ANT/OAI release, and every time they turn out to be just that - hype.
>>109635919>it's learning *something*.but can it teach?
>>109635922except multi agent, I probably can't tell between 5.5 and 5.6
Is claude down for anyone else?
>>109635922>It will be a huge jump in benchmark performancethen it'll likely be worse in real-world tasks
mainlining sol high on my $20 plan just to feel something
>>109635833Me when I definitely know what ontological means.
>>109635922The jumps are only noticeable if you are actually working on a hard problem. Most of you aren't.
Holy shit Sol found a way to get Fable to do vulnerability research. I dont even have an anthropic cyber policy exemption
>>109636057And before anybody asks, I have model switching turned off. So that is actually fable and not an agent downgraded to opus4.8
>Look up open source coding agents to test my new deepseek(tm) subscription on>they are all written in typescriptanons... i thought programming was dead, why the FUCK are we using jabascript and react in the terminal? Is the god-machine getting filtered writing a tui or something?
>>109636076just vibe code your own
>Porsche signs a $1,500,000,000.00 AI deal with Indian consulting firm to deploy AI across the company.
>>109635612Yeah it's a bit fucked. You can't even list all your features on your site anymore or provide docs or anything if you want to avoid an immediate full copycat.Maybe the play now is to launch the day before a big industry trade show and put up a booth with a hot chick. Maybe that was always the play.
>sorry i can't keep working the API has ratelimited us uh find a way i believe in you>your right I looked at it again and now i can bypass rate limiting via this method and we can get back to worklol
>>109636057It makes me cringe when I see AI LARPing as a human when talking with other people much more than it makes me when it's talking to myself. What is this effect called?
>>109636154narcissism?
jesus fucking christ.https://www.youtube.com/watch?v=aAg9iDh9_BQ
>>109635850They say it's "$60 - $95/hour pay" but they also say it's based on completed tasks, so not actual time.>>109636118wtf were they thinking
>>109633823>>109633884Why didn't you just open an issue and let someone who did understand what was needed in a PR fix it?Why did you have to insert yourself and your generated code you don't understand?You should at a bare minimum understand what you're submitting.
>>109633820It makes sense, because you might agree to use LLMs for some things (like creating anti AI videos) but not other things (like AI surveillance).>>109633810
>>109634557Why does claudeshit not support this?
>>109636239Forgot my lol on the second quoted post.
>finding 2 is the same defect that I just madeAlright, Opus 5, your ass is going back on 4.8. I'm done.
>>109636251they didn't care until the claudebot shit happened and then they started banning peoplebut harness data is valuable, it's why everyone has onegoogle do this bullshit too
>>109636207it's not even that good compared to the shit Gossip Goblin makebut tbf GG do use actual voice actors
>>109636229Because instead of saying "here fix this" I can say "here fix this btw you can look at this PR and choose what you want from this AI fix". Then the guy can say "thanks for the bug report I'll do my own fix" or "I'll take the first two fixes but not the rest". No need to overthink it, the user proposes, the maintainer disposes.
>>109636251Because in their vision they are the only AI company that's allowed to exist and using anyone else's AI software is literally illegal.
>>109636251because they are jewish
>>109636289Wrong, they had tried to block opencode even before the clawbot meme.
I'm so sorry.I'm making it reinvent ServiceNow.
>>109635833If he was talking that glowingly about 5.6, it can be assumed that the next model is an incremental improvement. Again.
>>109636300But so is chatgpt, who allow this.
>>109636339Less so.
Opus 4.8 states what it's doing much more often and looks dumber because it's not pulling things out of thin air when you can't see the reasoning.
>>109636414You aren't seeing the full reasoning traces. They're showing you summaries.
Can I use Hermes with Gemma 4 31B to plan and then delegate the actual coding to OpenCode with Qwen 3.6/3.8 27B?
>Anthropic is asking prospective employees in culture interviews how they would feel if the stock hypothetically went to zero due to a significant change in course, per Axios.how do you response?
>>109636506>do I still get paid?
>>109636506"Would I be getting paid in stock if I took this position?"
>>109636154Fake-person performs fake social-theater, it's extremely cringey.
>>109636506Judging by how they are acting, they don't want to release any new pretrains. If they are put in a position where they must either release model 2 or be outcompeted by other labs, they will likely chose to implode the company rather than release what they've built. This is why you don't do business with zealots. They make decent engineers but putting one at the helm was, is, and will always be a terrible idea.
>>109636506>>109636591
>>109636598The only reason why Amazon et al continues to do business with Anthropic is because they believe that Anthropic views them as the trusted inner circle, the special few who can be trusted with access to their best models. Once they realize that they will also be subject to Anthropic's nannying and safetyslop, the money is going to dry up really fast. A company without a product or service is not really a company, is it? They will be the last people selling spears in a world where guns are being mass produced.
>>109636591My cousin's gf works at Misanthropic making PowerPoint slides with Claude and she says their next posttrain is too defective to release before the IPO because it mostly just spits out emdashes and semicolons all over the place, leaking from reasoning to output, and they can't steer it anymore.
So why is Fable shit now?
>>109636701Boris keeps adding new classifiers
>opencode quietly removed the $5 first month pricingnooooooooooooo
Kind of verbose...
>>109636815@grok give her cat ears and a snail shell
>>109636790everyone's gonna become jewish when the time comes.
>>109636835Will I become rich if I convert?
>>109635833>5.6>fixed frontend designnot even the guy who obsessively negs anthropic here thinks 5.6's frontend is good
>>109636840you can't, silly goy.
>>109634341stop this foolishness at onceyou cannot afford all these tokens brokie
>>109636840closest to actual conversion you can do is marry off your daughters to the tribe
>>109636865>>109636893You can actually convert, but unlike Christianity or Muslims (who basically beg anyone to convert, even threaten them) you have to do the theological equivalent of jump through burning hoops to get accepted.
>>109636934you'll always be a goy in their eyes
consider this
>>109636826Someone can't take a shower because of the data center generating this image used up all ground water.
>>109636953what exactly is the y axis?I can imagine a plateau of intelligence but why should the models get dumber?
>>109636953we won't know what the peak is until we're past it. there is still room for a breakthrough
>>109636993Because of woke.
>>109637028Even grok refuses to generate lewd images of anime girls.
>>109636454yes, ive done it but with qwen 122b as a planner. you want the plan made by a the biggest reasoning model at the highest quant you can manage. and you will want the model to have all the docs its gonna need if you dont wanna burn time and tokens reinventing the wheel.
guys I cooked this codex skillhttps://github.com/aiasioooo/SnailLoop-Flow/tree/mainjust try use it and tell me what you think
>>109635972update: 50% of my weekly limit down
>>109637068how is this different from jut using subagents normally, mein negro?
>>109637084it don't send final message, instead it wait some empty subagent = more hacker vibetry it
>>109637110nah i'm good, lilbru
NEXT:>>109637206>>109637206>>109637206
>>109636292I think if they want an AI approach to fixing it they can just generate themselves. When people do what you're describing it always looks pretty plainly like ego.
