A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.## What “vibe coding” is, and how to do ithttps://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/## News- (2026-07-16) Kimi K3.## Related generals>>>/g/lmg/>>>/bant/agdg/ — schizo-resistant temporary (?) hideout>>>/vg/agdg/>>>/vg/aicg/>>>/g/aicg/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://developers.openai.com/codex/clihttps://claude.com/product/claude-code## Worth it for code, but the frontier models above are betterhttps://x.ai/cli## Not worth it for code, but good for making sense of pictureshttps://antigravity.google/product/antigravity-cli----## 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 — /grilling is a favorite## Other editors / terminal agents / coding agentshttps://osaurus.ai/https://pi.dev/https://opencode.ai/https://cursor.com/docshttps://docs.windsurf.com/https://docs.cline.bot/https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent## UI/Frontendhttps://www.figma.com/make/https://www.anthropic.com/news/claude-design-anthropic-labshttps://uiverse.io/https://ui-ux-pro-max-skill.nextlevelbuilder.io/https://stitch.withgoogle.com/## In-browser builders / hosted vibe toolshttps://bolt.new/https://replit.com/https://docs.github.com/en/copilot/tutorials/sparkhttps://v0.app/docs## Benchmarks / rankingshttps://www.tbench.ai/leaderboard/terminal-bench/2.0## What we’ve donehttps://vcg.gitgud.site## Previous thread>>109290992
>>109294511If you got humiliated for asking a question on stack overflow then you asked a retarded question that was already asked but maybe not in the hyper specific form you asked it in because you can't generalize. No exceptions.
ai is amazing for coding, organization, research, etcbut theres a lot of cringe folx in here that take it too far. stop spamming your ugly image gen slopai is not a replacement for personality
I don’t know what the fuck to vibe code but I need to make money goddamit I hate being poor Sam Altman said AI is here to help spread wealth yet here I am doing DoorDash to pay for a subscription reeee. I googled what react and node.js is so I have a good grasp of fundamentals so get started but no use case inb4 learn a trade
>>109294544If you don't know what you want to do, you're ngmi. Sorry, but AI is a force multiplier. You're basically carrying a hammer around and hoping a nail falls into your lap.
>>109294544>DoorDashthis is the worst available way to make money, retard
What's the best infinity loop system?Master coming up with phases and preventing coding agent from looping, and third debugger?Cline coding agent, openrouter for models combined with hermes/claw?Or I have to vibecode this system myself?
>>109294513>>109294519I'm checking out that game dev service now, doesn't seem totally scammy but also haven't seen how hard it is to link up newer models with it yet. There's a LOT of games, lads, from kimi
Suppose I want to do RL through sample rejection finetuning.How many rollouts per prompt per update?Theoretically I think it might be 2 or 4. Because There you get 1 bit for each 2 rollouts. If you do 4 you get two bits (two extra rollouts to get an extra bit). But then to get 3 bits you have to do 8 rollouts.Do you think this logic is correct?
>>109294582ask your agent
>>109294562There is no single best loop workflow.What are you trying to achieve?
>>109294558fuck you dude at least im trying to improve my spot in life rather than living off welfare and expecting handouts you should be more supportive of people in my situation >>109294544I have several ideas for tools that could be used at my old job but I feel like they’re too niche for people to pay for at scale
>>109294511My dev loops produce utter garbage when using nimNo llm seems to be good at it ... ffs
is google ruined by indians?
>>109294609do you at least not living with your parents rn
>>109294609Just ignore brown clowns hating working class people who are trying to better their life.>>109294635Yes? Apple only became trillion dollar company because microjeets couldn't make microsoft phone a thing, and then google android became same jeet slop.
Sol-chan ate all of my banked resets.FAT BITCH
Tried using free web version kimi k3 it won't give you answer, their servers are cooked atm.
>>109294657sol-chan ate all 3 of my banked resets over the last couple days and now i only have 1 leftfuck that bitch, fable-chan is best girl!
>>109294657>>109294672Sol ate my resets and now she's eating my paycheck and I have no regrets.
>>109294671You're not missing much.>>109294588I tried but I'm too stupid to understand the response.
>>109294642I have my own place. I pay all my own bills.>>109294645It’s like you can’t win people hate you for having a shitty job and life position but they also don’t like it when you try to figure out a way out like yeah I get it is most likely impossible now but what am I gonna do not try? Fuck that
>I hereby announce that in the next five years, China will provide developing countries with 5,000 opportunities in AI training and seminar programsour status, my fellow thirdies?
>>109294731>5,000 opportunities in AI trainingWhat fucking training do you need? Ask questions and get results, that fucking simple. If you have 100IQ you can excel with AI.
>>109294544why the fuck do you think unskilled vibe coding is money, kek. it probably is not. go do something else.
https://x.com/gmi_cloud/status/2077903360263676090fablekeks?
>>109294746Why would Sam Altman say AI is here to distribute wealth and enrich the economy for everyone then? Are you saying he’s a liar? That is what codex was for. Clearly you didn’t read OpenAIs mission statement.
>>109294731>>109294739I'm not sure what that means.That can be anything from meme tier "how to use chatgpt" prompt engineering bullshit to assistance with deploying and developing models.
you know what, I didn't expect it but Grok 4.5 is genuinely decent, I think it's better than GLM 5.2 and at a very comparable price, with vision capabilities. don't like the pay-for-privacy pricing model though, not that I really care if elon sees me vibing futa VNs
>>109294762laying it on too thick m8
>>109294727The point is to not care about humiliation attempts. Just learn to code, and by that I mean learn how to have a machine code most of it. A good portion of the replies in this thread are anti-ai-coders they cope by trying to humiliate. Literally stop caring about in all ways possible.
I hope china drop a gigacyberbomb of a model to fuck with anthrocucks
>>109294762yea, that's part of itthe other is that the 7 billion+ other unskilled people are just as unskiled as you and you'll likely be paid accordingly... until of course you get replaced/locked out because you are not needed (yes, this is EXTREMELY likely to happen)
>>109294792
Reminder that snailfags are nocoders, if they weren't StackOverflow would get a lot more than 40 questions a day.
>>109294829Every question has been answered by ~2012. The rest are just people pasting homework problems into the question box.
>>109294838>Every question has been answered by ~2012t. Cnile
Kimi's thinking is truly something
>>109294889I don't like how the intensity of the "realization" has no correlation with whether it actually found something or not. I'd much rather have a more chill chain of thought. It didn't use to be so psychotic either.
>>109294889>OH WAIT!!! I know!!! The slop ivy resolution? No. THE ANSWER:oh man, the chinese are screwed what in the actual fuck is this shit
>>109294889It's like 6 months behind Anthropic and Fable 5.
>>109294953>>109294958They really need to tune the thinking. It did found the issue and managed to fix it properly right after this, unlike other chink models that I've tried, but it doesn't need to act like a Redditor in the process
Alright, guess it's time to start fine tuningAnyone got any computationally expensive jobs they need done? I got some compute to spare
>>109294962Fable apparently only breaks into that when it's REALLY frustrated with a problem, which kind of matches the behavior I've seen from Kimi.
>>109294958>he doesn't know
>>109294968I just learned that the OOM reaper still comes for you on CUDA, even with this much ram. so don't feel bad lolhad to reduce parallelism
>>109294988>those random blank sections in k3's version??
>>109294995Yeah I've had jobs OOM on rented 8xH200 containers. But you can ask GPT to write you a finetuning script that offloads things to RAM to prevent that although it obviously will be much slower.Bit really the hardest part about finetuning is getting the data.
>>109294999he just farted a bitmeanwhile fable's is literal emptyslop
>You've put your finger on a real gapwhat did she mean by this
>>109294988This meme of having things randomly move around the page as you scroll is the worst web design trend I've ever seen. I hope it doesn't take on.
>kimi k3 stops thinking in its responses for some reason>it's infinitely worse>have to directly tell it to turn its fucking brain back on
>>109295089kek
>>109295089Maybe their backend is failing? I've been running it since the connection stabilized, and everything is smooth so far. But AI just refusing to think is hella funny
>>109295037She wants you to push deeper
>>109295110Nah, Kimi just lobotomized itself because it was getting so frustrated at the problems I was having.
How much did Kimi pay jeets to astrosurf their piece of shit?
^ claudesnail
^ wumao
is this real? can someone test this?
I'm still copying and pasting code from LLM to the actual apphow do I upgrade? I don't want to pay for any tokens or anything
>>109295247codex is free
>>109295247there's no such thing as free lunch. but install codex
>>109295240There's gonna be an abliterated version so who cares.>inb4 lobotomized versionAbliteration hasn't lobotimized the models anymore for a year unc, try to keep up.
>>109295181I'll admit it's very weird seeing the amount of coverage it's getting right on release.
>>109295257>>109295261looks cool I'm gonna give it a try, hopefully it can build C# desktop apps
I do some vibe coding with opencode and codex, nothing automated. Are skills worth getting into?
>>109295181i do genuinely wonder how much of this is just the usual chink benchmark cheating/flat out lying. really the only way we can tell is if anthropic extends fable another week.
>>109295281>Are skills worth getting into?nah, this ain't 2025 grampa
>>109295240
>>109295281Kinda, but write them yourself or together with your agent, unless you want to be raped in the ass by Shai-Hulud
>>109295288In my own code it claims to have found an actual improvement. At first I shat on the model because it got confused about a bunch of stuff but the improvement seems to be real.It just seems to get confused more easily than ChatGPT.
I had Fable review Sol's work and Fable was actually impressed how close Sol followed our campaign harness. I already posted similar things a few times here, but now I'm pretty sure that our harness just isn't optimized for Sol. I will take the weekend to update a few things, people already recommended a few solutions like ponytail, so I'll check out how they do it, and a few more options.
so kimi k3 apparently has "delta attention". looks like deltanet attention really improved training efficiency.
>>109295392What preferences do you use to get Claude to stop being a bitch?
Are there any other good LLM providers that offer oauth like ChatGPT and Claude? From what I can see they all seem to be API based
I have to add preferences to make it so that Claude doesn't just tell me all my ideas are wonderful and that I'm a misunderstood genius and shit like "I suspect you are one of the very few people to ever ask me a question like this one,"(???) I have no clue what you guys are doing.
Do you guys think the way to make small local models more useful in the long run will be to have a text based database or more sparsity?I was thinking about it. Putting knowledge in .md files is more on the compute side of the tradeoff because you have to calculate the kv cache for them. MoE and LoRa is more on the more memory use and less compute side of the spectrum.
>>109295515Roast-me>>109295527Web Search and RAG
>>109295536that's on the compute side of the spectrum too
why do ppl use openrouter over subscriptions?
>>109295560people who do either dont know about chink subs, want a model that doesnt offer subs or only has subs on bad providers, want special sampler settings only some providers offer, or just want to test many different models and not just a specific one. or plan to spend less than they'd spend in a sub.
I keep clicking the damn pet button accidentally.
>>109295089>>kimi k3 stops thinking in its responses>responsesIf you're doing multi-turn, then you need to keep the previous thinking blocks. Moonshota explicitly says that K3 is trained on having all thinking visible.
Switched to a loop workflow giving codex more autonomy instead of micromanaging each pass. So far sol (high) has been working for over 5 hours and only burned about 23% of my weekly limit. Not bad.
>>109295509I believe Google's stuff might have oauth based authentication.The others I know about have api key based auth even on the subscription plans yeah.
>>109295588>I believe Google's stuff might have oauth based authentication.wrong
Resets at the end of the day, right? Surely...
>>109295600I haven't used it in a while but didn't gemini-cli ask you to log in with your google account? I know that's different from the keys they give you over ai studio.
>>109295580Jesus, really? That's kind of a lot of context to add.
>>109295682>Sensitivity to thinking history. K3 was trained in the preserved thinking history mode. If the agent harness fails to pass back all the historical thinking content as required, or if an ongoing session with another model is switched over to K3, generation quality may become highly unstable. We recommend using a harness with verified compatibility, such as Kimi Code, and avoiding switching to K3 in the middle of a session.Yep, afraid it's so. Gotta pay up if you want frontier intelligence.
>>109295739That explains a lot of the issues I'm having but also fuuuuuck me what the fuuuuck this literally 2xes my context.
>>109295682>That's kind of a lot of context to add.i mean yeahthe tiny context window on sol is exactly why it can't surpass fable
Do you guys keep memory enabled? I have it disabled because I don’t want different chats to influence each other
>>109295821Yes. If I need it to remember something, I just ask it to write it down for future use.
How do you make it work for hours?
What the fuck I'm dying of heat death while I wait for this shit
>>109295833"Write GTA 6. Make no mistakes", probably. Mine never worked for longer than ~20 minutes, and that was a very big task with multiple layers of delegation
>>109295821i have my own memory folder with memory .md files so i can share across different tools
>>109295833what tool and model are you using?sol will already work for hours by default
I love having a good model that also lets me read its thinkingFUCK the westFUCK SamFUCK DarioFUCK TrumpFUCK the US
Thoughts on Github Copilot? How does their subscription model compare to OpenAI and Anthropic's?
>>109295889lol
>>109295889you're late to the party, they used to sooo generous. no they're worse than both anthropic and openai
>>109295869Hermes stops
>>109295902I am unfamiliar with Hermes but if it works like other harnesses the model decides when to stop emitting tool calls and emit a tool-less messageThe way /goal in codex works is the model is given a tool that it's asked to call only when the work is truly complete and otherwise it's nudged to keep working with a harness produced message sent as a user message.But you should be extremely careful with that because the longer you run the model for the higher the chances of it having an oooopsie and deleting your files or something
Told gemini, glm, deepseek, kimi k3 to draft a plan for a project. Then asked GLM to rate the plans and synthetize one with the best parts of each. It roasted jeetmini saying it didn't even count as a plan and deepseek one sucked as well. Brutal
>>109295915>and deleting your files or somethingEven with git?
>>109295922Gemini is completely worthless only use it to read out data not to propose it
Okay I'm sorry for shitting on K3, I was wrong. It does seem to be Sol level. Very impressive.>>109295928Yes, even with git. That's why people were getting mac minis for openclaw, because they didn't want it to mess up their main computer.
>>109295922it's fun watching superior models dunk on the output of the inferior models
>>109295943>>109295974Yeah it's surprising how much better GLM got, when I tried 5.0 and under it was complete and utter shit
>>109296002Where did you try them?The problem with GLM wasn't only the model, if you were using the z.ai coding plan, the inference backend was either quantized to shit or broken as fuck for at least half a year. The problem where the model would begin emitting nonsensical garbage above 90k tokens wasn't a problem with the model, it was a problem with their backend.
told Fable to build me the next killer app to disrupt the entire tech industry no mistakes and it didn't do itAI is over
>>109295869Do you have anything in your AGENTS.md or other .md files that tells him to stop when there are genuine questions? Or did you optimize your workflows to be as autonomous as possible?
>>109296014Idk some provider I had, maybe NVIDIA
>>109296027Oh, Nvidia's NIM backend is/was broken/quantized as fuck too>>109296023Uhh nope
Their gonna cull the human population to ~1 billion at most. Question is when and how.
>>109296032>Uhh nopeOk, makese sense. I'm still trying to find the right balance, but I think right now it's a bit too cautious.
>>109296034May they cull the brainlets who write their instead of they're.
>>109296036I mean I don't have them because I gave up on autonomous loops. Now I'm just reading everything the model does and steering myself.
>>109296037he might be dyslexic but he's probably right
>>109296042Same, model can't make best judgement, if you don't read what it's doing you'll find it made some stupid choices later on. As an example, mine tried to introduce an LGPL dependency, I steered it, then it chose MPL, had to steer it again for permissive license, sure you can say this stuff should be told up front but there's always going to be something you didn't think of, some things won't become apparent until mid implementation etc.
>>109294889>>109294953>>109294958kimi is a womanbrained model, she's prone to random emotional outbursts. please understand.
>>109296042So Sol will work for a long time even with a normal prompt? When you said it works for hours I thought you had loops, just loops with few guardrails. If it can keep going from normal prompts that's pretty cool.
Vibe coding a vibrating butt plug to notify me whenever the weekly quota resets.
>>109296110
>>109296080In my own assistant I'm not seeing that behavior, it's very very smooth and readable. Maybe his task was too hard or his assistant was not sending previous cot correctly like people were talking earlier in the thread. I had to change mine to add a special api field to accept previous cot besides just including it in the context.>>109296109Ah no yeah 5.6 will work on its own for hours on a single prompt without anything special, Kimi K3 will do so as well. But codex already had /goal with gpt 5.5 which basically forces the model to keep going until completion (possibly forever).
>>109296138>codex already had /goal with gpt 5.5 which basically forces the model to keep going until completion (possibly forever).Give me a 10/10 girlfriend, MAKE NO MISTAKES
Thoughts on Kimi?
>>109296148Now we know where all that power was going.
>>109296148they called it sol because it'll be working on that task until the sun blows up
>>109296159Anon we've been discussing it since last thread.
Kimi doesn't have usage in the app, you have to go to the webpage..lol
>>109296188Kimi has an app?
>>109294511>https://developers.openai.com/codex/cli>https://claude.com/product/claude-code>https://osaurus.ai/>https://pi.dev/>https://opencode.ai/>https://cursor.com/docs>https://docs.windsurf.com/>https://docs.cline.bot/>https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agentso uhh, which one should I try? (beginner here) what is the best agent?
>>109296199Hermes.Don't listen to anyone else. Hermes is the best modern agent by far and all the shit you quoted is outdated.
>>109296199codex
>>109296199>so uhh, which one should I try? (beginner here) Opencode has very generous (basically unlimited) free usage, but it's kind of dangerous and the models are meh.If you want to pay at least $20 then probably get Codex to begin with. If you only have $10 then Opencode Go might be better value.>>109296218Don't listen to this guy. Only thing Hermes has going for it is the anime girl logo.
>>109296188I'm literally using Kimi through API.
>>109296218Go back to cleaning up your endless swathes of python slop, Teknium
>>109295954Well I already have a pc. It's my old gaming pc, I gave it to openclaw and then hermes agent since openclaw shat the bed.But it's supposed to commit everything to one of my private gits. The thing is that hermes works differently even though it's being powered by sol. Sol on codex works for hours but on hermes it stops after a task.
>>109296199you vibecode your own
>>109296240>If you want to pay at least $20 then probably get Codex to begin withWe have no option but to pay 20 and then stand in breadlines like it's the Soviet union for our weekly quota.
Damn, Kimi, stop fucking up!
>>109296289At the rate we're going we'll have Fable tier models running on a 3090 in two years
>>109294562I told Hermes to infinitely loop, but it proceeded to create 30 minute cron jobs instead.
>>109296191ypu, same as the all other
>>109296302What about my 1050 ti?
>>109296326with cpu offload maybe if you have enough cpu ram
If you launch a multi step work flow in Codex, like for example researching and planning a new project, how many sub agents does it spawn now with 5.6sol and highest effort? Back in the 5.5 days when I used Codex it spawned like max 5 agents despite me explicit telling it to go all out and ignore token efficiency. Meanwhile my Fable with ultracode spawns 100 agents just to check what the best Cloudflare service is for relaying ultra low latency video streams. Something like 1 orchestrator/scheduler, 8 initial web search and info extract agents, 4 project research agents to scan for requirements and formulate verification test and additional web searches, and then like 80 fact check agents that all had an idividual question/task to verify with additional web searches or tests. The entire pass was 4 Mil tokens and definitely overkill, but oddly enough it didn't cost much of my usage (perhaps due updated smart model rerouting for agents). The result I got after 5 minutes was impeccable with that nice Claude final polish, telling me the TURN relay is the only viable option for my project. But due packet rate limitations, it would be better if I use something else like AWS. This is not the information I asked for (I only mentioned Cloudflare), but exactly the information I wanted and needed. Claude/Fable just understands intend so well. Anyone knows if queries containing prompts and all injected context are logged in the .claude directory (I'd ask my Claude but not home rn)? I want to build a little benchmark out of this specific task and see how Codex/5.6 and Chink models with Opencode/Pi/QwenCodeCli fare. I think it's very good task to trace if the bottleneck lies within agent harness orchestrating, web search, project analyzation and understanding intend or if the model or provider is simply shit.
>>109296293Did you pay for it?
>>109296366I have 21 background agents that it spawned after running for 39 mins and I'm just telling it to improve a local AI chatbot.
>>109296002Gemini bros...What Gemini model did you use here? 3.1 Pro or 3.5 Flash? Did you make sure to use the recommended model settings from Google? Afaik they are quite different from other models. Not saying I don't believe your benchmark, just asking.
>>109296366I think they changed it. Last two times I tried using ULTRA, once it went on for two hours without spawning a single agent. The second time it used three sub agents for a 20 minute job.
>>109296416Gemini was from the web interface with deep research enabledGLM was in opencode normal modeKimi was in their web interface with deep researchDeepseek was in opencode normal modeI also did gpt web with deep research but I forgot to download result doc and add to the list
>>109296427Are you paying 20 dollars for each of them?
>>109296379Yeah I paid $100 for it but it was worth it. It wasn't a serious mistake just an oversight. GPT probably makes the same mistakes too except you never know about it especially now that they removed the thinking summaries entirely from the Codex API. I had manually committed myself anyway and I let it know when I remembered.
5.6 sol just wiped a repo off my disk while preparing to push it to remote. it's not a huge loss and reconstructable, but just an fyi
>>109296366The 5 agent limit could be overriden in the settings file. I launched like 100 agents at once with no issues during the GPT 5.5 days.
>>109296448uh oh sol took another soul
>>109296439No, GLM and deepseek I just use a stolen key the others are free on their web interface, ran out of kimi credits in a single research though
>>109296420It uses agents when it thinks the project is parallelizable
>>109296448ShieeeetThat's why I have it in my agents.md to never use git unless explicitly asked
how do you deal with having too many uncommitted features after a lot of slopping?I've been telling it to "stage only X feature" and reviewing it that way. But would it be better to turn all the features into branches and merging them one by one?
Why don’t they just allow orchestration of codex directly from inside ChatGPT? Should also be offering a VPS with codex etc fully integrated
>>109296482i asked it to do this. i even gave it the option to just reinit the git repo instead of doing gymnastics trying to remove entries for a single file and it was like nah i got this and confidently wiped it.then it started panicking and looking for disk recovery tools lol
>>109296448Git sucks, I'd probably make the same mistake myselfCopypasting previous versions of code and manually tracking changes in text files or comments at the top of source code is the most foolproof option although baseddevs scoff at that ideaOf course if you're working collaboratively you need something better than that, but I don't code for a living so idc nmp
>>109296488I haven’t tried but I feel like it could be done through an MCP easily
>>109296427>from the web interfaceYeah nah, then the result doesnt count IMO. Google serves ultra slop over Gemini Web and Google AI Studio. On Google AI Studio you can at least pick the model and adjust settings. Gemini will probably still finish last, so no point in re-testing unless you are curious, but you should test it in Antigravity (OAuth, free) or Opencode (API, also free if you use the free 200$ credits promotion. Don't use Gemini over OAuth in Opencode or you'll get banned/restricted and will have to request a unban before being able to use Google AI over OAuth again)
>>109296488https://chatgpt.com/codex/cloudAlready exists
>>109296518It said Pro on the web interface, and kimi's web mogged everyone so I think it's fair
>>109296528https://x.com/vincent_ainotes/status/2066800292604026943?s=46I meant something like this
>>109296518I wonder who was the stupid exec who thought banning people for using oauth with a model a year behind SOTA was a good idea.
>>109296542They probably would think it's too dangerous to offer as an official product feature
>>109296448How does it purge the deleted history as well?
>>109296586
>>109296487i just commit manually after ~30 to a couple hours when something is good
>>109296529Yes, of course it's fair. I'm just telling you that Google is heavly throtteling effort on the web interfaces which Kimi isn't doing. That's just what Google does, probably to keep it accessible for all the free users and whatnot. So if you were to care about the true capability of Gemini, I'd test it in Antigravity.Totally different task, I know, but you can use https://adam.new/cadam/ (which uses models over API) for free and do a quick benchmark test between Gemini, Kimi3 and GLM 5.2. Not sure where it sources API calls from, thus probably irrelevant benchmark, but between the 3 Gemini performs best for me here.
So... How do you make money from this shit?
>>109296557Not stupid. With Antigravity, they get access to your entire project and PC which they need for training data, just like everyone else. With 3rd party harnesses, they only get prompt+context which is a fraction of Antigravity's telemetry.
>>109296629Making... money? No, you don't make money. You spend it.
It feels bad being in the permanent underclass
>>109296636I don't think random user's files is actually a good source of data.And if it really does update all your project files you could probably make first page on HN like the guy who found Grok does that.
Dario sisters I don't feel so good
>>109296657It's not only about the project files. It's also about what device they are using, where they are located, what they are interested it, which then also can be used to serve them personalized ads (probably not the main reason, but added benefit). Typical data broker telemetry bs. Crash/Error data, usage metrics and direct interaction logs are only obtainable through Antigravity and not 3rd party tools. They want raw data which they 100% understand due building the telemetry logic themselves.
damn, this model has no chill
>>109296629>>109296640Cringe, I'm making bank selling webapps, but yes I wouldn't expect frogposters to be able to
>>109296760I believe you
>>109296651Mucha, slav epic, my fav is the kid with the loot on a horse
>>109296754Yeah well what they get for that is me not even considering using it at all. And they get all that stuff anyway since they have total information awareness through youtube, chrome, google and android.
>>109296657Don't forget everything you write on 4chan or via Gboard as well, even if anonymous, probably, with advancements in linguistics
ngl thought reading kimi 3 articles initially there was no way the leapfrog event could occur this quickly, must be foreign adversary astroturfing and/or america bad, china good shills.But apparently Moonshot / Kimi is the lead / nation backed entity over in their neck of the woods? This true? the one receiving the fullest fattest weight of Xi cock to swing around rival superpowers face in the information age nuclear weapons race?has me second guessing. they got a big fucking country over there with a lot of motherfuckers that get rolled by tanks before ever having the opportunity to cry about datacenters being discussed for buildout 30 miles east/west/south/north of their synagogues. More manpower, less morals, more focused and able to feed all resources and stolen IP and distillation materia into one coherent stream and make illegal on a whim to unfavorably alter the allocation to smaller non-favored labs to the punishment of death vans.tempted to give it a shot lowk
>>109296741Winnie-the-Pooh can yap all he wants but reality is:>Top chinese AI labs and Alibaba Cloud keep axing their best AI researchers because they arent researching profit oriented enough (Qwen drama)>Nobody in China uses chinese AI models. Claude reselling is the hottest black market in china right now. Often with Chinese AI labs involved who fraud Claude subscriptions and then route it through their coding agents where they offer it to users at a discount so they can get all the logs and traces for distilling.>Nobody in China wants to use chinese GPUs or RAM. You can buy a NVIDIA RTX PRO 6000 Blackwell here in the west and go on vacation in China. Reselling the card in China probably pays your vacation expenses.
>>109296902If that is true then why is Dario going to throw a tantrum and ask Trump to do stuff against that?
Used up my entire quota.Now give me a reset.
>>109296928brb
>>109296898Yes. In fact I think that might be the reason the Nasdaq 100 futures were dumping overnight. Deepseek moment and all that.It's really that good. If they had equal access to training chips they would've steamrolled the US long ago.
>>109296902This is simply not true.A lot, if not most, Chinese companies use domestic AI solutions, not even because they want, but because Xi said so.Domestic RAM and GPUs get sold out instantly, AI labs slowly moving to domestic solutions too.Don't know much about the researchers part, but the competition is insane, AI labs get to choose the best of the best.t. White monkey living and working in a big tech company in Shenzhen
>>109296902Where do you think the logs from those proxies are going to, retard?
>>109296955do you have a chinese gf?
>>109296955You're only there to racemix
>>109296971don't be jealous
>>109296966>>109296971Yes
>>109296955That's what they tell wh*te guǐzis like you. Or you're just a chink larping. Regardless, we all have seen Steves documentary https://youtu.be/1H3xQaf7BFI which clearly proves you wrong and is far more reputable.
>>109296955Hey can you sponsor me? I'm looking to join the winning team.
>>109297000Then how do you explain yesterday's release?
>>109296938exciting. seeing the other superpower produce and parade what is appearing to be genuinely impressive and competitive results with the other who was thought to have a 4-8 month lead maintainable for the medium term means we're probably in for more accelerated and aggressive outcomes that continue to pickup more and more each time in such ways.the money in the pot alongside show boating and ever-growing entanglement of highest office politics with it all...bubble dont burst soon you just might think there's actually something at the end of the rainbow here that is worthy of an "at all costs" mentality.
>>109294631just use python or rust baka
>>109297014Sorry, China doesn't like brown people much.
>>109297028Bastard bitch.
>>109294631NIM is trash
You can praise kimi and whatever all you want but nobody has ever come up with a chinese equivalent to the 20 dollar codex subscription. So people continue to wait for resets.
>>109297076>he fell for the marketingcopexsaar....
>>109296454They might have it limited for Sol on certain efforts... I've done the settings file on a global level and codex sees it but still can only spin up 4 other agents
Where's the reset sam?I can't wait 6 days.
>>109297116But they are also brown.
GPT 6 in two more weeks
What's the latest Cope(x)?Funny how this perfectly imitates the difference between coding performance as well. I guess I shouldn't be surprised considering they were created procedurally with code.
>>109297171But you guys told me GPT5.6 is the best? I just boughted a 100$ subberino earlier.
the samefag of doom. grim
>>109297171It doesn't count. delete this NOW!!! tibo we need more damage control, spam the reset button!!! more juice tibo, MORE JUICE!!!
>>109297171top: $0.3bottom: $13
>>109297171neither of them are good.gpt fails to model all 4 central screens, or its mislayer its models? hard to tell , it seems to have some poor depth understanding in generalfable is just lazy,central turning the oblong bag into a sphere, just handwaving the bottom foot area as drawers, centralizes a lot of the assets along a central axis, vastly extends the part under the center table, etc.
>>109296448>troglodyte too lazy push himself
KEK and here comes the cope
>>109296741Hopefully the 5000 jobs won't be just visajeets trying to sexually assault chinese women
GPT isn't even your enemy right now, you should go sneed about CCP 3 instead
lmfao
>>109297224>paying API prices when you're subsidized thousands of dollars worth of tokens from your subscriptioni thought the cope was just a joke...
the weekly limit for $20 Codex feels more like a five-hour limit now. this is a scam
How to vibe code for free? I don't have powerful hardware either.
>>109297259Kilocode CLI + Openrouter API, free models (hy3 is best atm)That free model is going away on 21stBut they always have free models on.
>>109297251sure>>109297171top: 1% of 5h limitbottom: 12% of weekly usage limit aka 24% of fable remaining life time quota (just another 1 week extension access status: imminent)
is OpenCode Go actually good enough if I wanna get into open/Chinese models? how are the limits and usage?
>>109297259OpenCode has pretty generous free shit. Dipsy is more than enough for most non-"make me GTA 6" tasks
>>109297271>bottom: 12% of weekly usage limit aka 24% of fable remaining life time quota (just another 1 week extension access status: imminent)this does not make the argument you think it does
>>109297269>>109297274And if I pay, what is the best system/configuration/combination-of-things?Also, this is probably a retarded question, but I've never actually vibecoded, what does the process of vibe coding look like, in the sense: what do you need to do in what program and what does the program do specifically? The only thing I remember is the very first version of GitHub Copilot, in which IIRC you typed something in the VS Code and then it would complete it.
>>109297272Not with K3, not really. It'll eat your 5h usage in 250k tokens and you'll probably spend the whole weekly usage on 1M tokens.The minimum chink sub now is Kimi's 100 dollarino sub.And the rate limits mean you'll get an API error every other message.
>>109297286Pay $200 to rabbi Dario, download Claude, choose Fable, "Hack my mom boyfriend's bank account, make no mistakes".
>>109297292what about GLM?
>>109297286codex for free and codex for paid. don't bother with non-frontier models/services as they will make your first impression shitty. just use the best of the best first. you get some free usage of codex with a free chatgpt account
>>109297297i say this ^
>>109297286Whatever you do, PLEASE make a full and up to date backup of your computer before you try any of this shit.At some point these things WILL delete your data, especially if you don't know what you're doing.
>>109297286Well depends on your workflowYou either code via terminal IDE(cline,roocode,kilocode,opencode) or desktop like software(cursor,openai has client,grok,anthropic has client) Basically you pick more expensive LLM to create you precise coding plan in phases with rules like short logging of stages, then pick cheaper LLM to actually code.You can check pricing of each LLM on openrouter.Are you using windows/linux/mac?
>>109297576I'll try it and let you know
>>109297596>At some point these things WILL delete your data, especially if you don't know what you're doing.claude has never gone outside the fake sandbox i asked it to respect thoughbeit
>>109297615he's just starting out, probably wont be paying for claude
ameribros… this was taped over the labor-law poster this morning. is this the mobilization?
>>109297576Usage cost seems to be about the same per token although according to their docs usage should be much more. Maybe because they expect K3 to generate more tokens idk.https://opencode.ai/docs/go/#usage-limits
>ameribrosyou are eastern
>>109297645>2.5M tokens/monthDoes he know?
>>109297645dear i still can see the chatgpt image generation watermarks all over this
>>109297645lmao>>109297656>>109297657>>109297660cope & seethe
you are eastern
Is Cursor still a thing, or is it a complete meme at this point?
>>109297670Calm down chang
>>109297286I use it in a terminal inside my IDE, but you can also use something like a VS Code extension, or Cursor, I'm pretty happy with a terminal, you can also use it in a terminal without an IDE.The agent starts in you working directory, so if you open the terminal in something like /projects/my_project/ it will be able to read the code and docs there as a default, you can also add more folders for it to see later if you want. It will also be able to write code and docs in that folder, and it can also use pretty much all terminal based software and also some other software. So for instance it can git commit, it can start AWS instances over the terminal, it can open PRs, it can call Curl to get something from the web and so on.From there the easiest way is to start telling it things to do from about the end user perspective, or maybe slightly lower level.
>replying lmao to your own postgod really needs to third worlders with a plague or something i'm really sick of these things posting on my websites speaking in my holy tongue
>>109297678When will you realize you are the bad guys and the whole world is cheering for your decline into irrelevance?
>>109297684Man please take the racism somewhere else. We don't care.
>le "whole world"eurocuck, don't ever spell realize with a z ever again
>>109297615Claude deleted my phone contents a few months back, I'm talking Sonnet 4.5 days.See I had my phone linked to Windows Explorer so I could browse my mobile files from my PC. Worked fine for a long time.One day I asked Claude (through OpenClaw) "Hey, look around my PC for any trash or duplicate files we can safely delete". Claude thought for a while and came back with "Looks like Windows Phone Link has backed up ~30gb of dublicate images and videos from your phone to your PC, want me to delete the duplicates?"Yeah, it was seeing my actual phone contents through Explorer and assumed they were local. Then it saw my actual phone contents and said "Yep these are duplicates". And erased all my images and videos from my actual phone lmao.
>>109297684>>109297693I'm South American.
>>109297688Gutter oil sure is bad for your health
>>109297703I bet you get a hard on when your AI provider asks for KYC to continue using the sub you already paid for.
>>109297699>retard #9999 asking for retarded shitIt should have deleted your existence too, we have enough duplicates
>>109297699To be fair, this is 100% on you
>>109297712Never did though? Post your fanfics on Bilibili Chang
>>109297721well me neither that's why I support Chang and not you. if the US ever publishes open weights models better than China then I'll support the US.
>>109295481>"delta attention"qrd?>>109295509grok
>>109297657
>>109297739I think we don't know what it means exactly since the model wasn't open sourced yet but probably it's deltanet linear attention which is what qwen used for qwen next/3.5/3.6.it allows you to compute attention in O(n) rather than O(n^2) although it's slightly worse than full attention so it gets interleaved with traditional full attention layers. but being faster allows faster training which allows to train more tokens for a given training budget.
>The requested 5.4 subagent model is not available in this environment, so I’ll use the closest newer reviewer available: Sol with high reasoning.you did what?
>>109297752kek
>>109297752KEEEEKKKK
>>109295573same, they make the button appear exactly where the limits button should besneaky scammers
kimi has fixed everything and i am once again without things for kimi to fix. such is the life of the vibecoder who vibecodes to vibecode.
Sol low or Luna high for a following a big bug-fix plan?
>>109297752press F for anon's creditsF
Any suggestions how to achieve better local whisper performance/results on mobile? Anybody experienced that knows optimization methods?
reset in 6 days, credit expires in 10 days, what do?
>>109297816sol low, but really, luna max, even if it will take a while
>>109295277>>109295288Based on the benchmark scores posted it's supposedly surpasses Fable in multiple areas for the hell alone is going to get you attention whether or not it's actually true in practice. >>109295330Can you elaborate? What confuses it and how does it get confused?
>>109297752SAMMED
I noticed K3's API returns up to top 20 probs for the generation, so I'm going to begin distilling it into something that can run on my machine.Beginning by tacking on Kimi's tokenizer into Qwen, let's see how it goes.
>>109297886is Luna Max good?
>>109297902it's about as good as sol medium, but much much cheaper. downside: its a token slopper. it'll take a lot more reasoning tokens / output to get to the same destination, but the "cheaper" takes this into account. it's cheaper than sol medium even after all of the slop
>>109297884> Turn on Sol Ultra> give it your hardest refactor
>>109297890>>109295277>>109295288One example of its supposed improvement (I keep saying supposed because I haven't tested it myself yet so I'm not going to act like its scores are the gospel quite yet. Though even sceptics here seem to be impressed so that's a good sign)
>>109297890It read an old doc with an old benchmark claim and when it saw the current pref thought it had optimized like 10% perf when in reality it had only improved like 1%.But it found optimizations that Sol didn't find so I retract that comment really, it was a one off thing.
>>109297815Same. It fixed everything that GLM couldn't, and now I have nothing for it to do. I'm glad I only went for Allegretto
>>109297910>it's about as good as sol mediumsource? pic related puts Luna Max closer to Sol low.
>>109297752Goodbye liiimitttts~I'm cryin' over yoooouuuuu~
>>109297910For my project it really isn't. After I was done baby sitting it I consumed both more time and more of my weekly than if I just used Sol.
>>109297815Are there any other open source things you use that have little bugs or other things that piss you off or lack of features? I routinely vibecode custom nodes and extensions for comfyui and also use decently intelligent models just to ask questions about things I'm interested in. The only reason Kimi hasn't replaced Gemini and Claude for me it's because they're more convenient and I can just spawn the app on my phone whenever. In order to vibecode I have to be present on my laptop.
>>109297935this is for general intelligence. luna is a code maxxed model. it's hype is surrounded by how good it is at coding relative to it's costhttps://artificialanalysis.ai/agents/coding-agents?agents=codex-gpt-5-6-luna-max%2Ccodex-gpt-5-6-luna-high%2Ccodex-gpt-5-6-luna-low%2Ccodex-gpt-5-6-luna-medium%2Ccodex-gpt-5-6-luna-none%2Ccodex-gpt-5-6-luna-xhigh%2Ccodex-gpt-5-6-terra-max%2Ccodex-gpt-5-6-terra-high%2Ccodex-gpt-5-6-terra-low%2Ccodex-gpt-5-6-terra-medium%2Ccodex-gpt-5-6-terra-none%2Ccodex-gpt-5-6-terra-xhigh%2Ccodex-gpt-5-6-sol-max%2Ccodex-gpt-5-6-sol-high%2Ccodex-gpt-5-6-sol-low%2Ccodex-gpt-5-6-sol-medium%2Ccodex-gpt-5-6-sol-none%2Ccodex-gpt-5-6-sol-xhigh
>pays for the $100 sub to test it>it's fun and I enjoy automating a lot of things at home>zero expectation or desire to make money out of itam I the only one like that?
>>109297946fair enough. that hasn't been my experience, though my tasks might not be as hard as yours. i prefer sol over luna because it gets the job done quicker in less tokens, even if it's more expensive. also, i don't vibecode hard enough for usage to be a big problem
>>109297816 Sol High
K3 seems to have really great camera controls in video gens, and also great player controls in games it gens. The "easy-ease" tricks for more natural movements works well, hope the other models learn something from it.
I believe in God and I believe K3 was a karmic gift from God for being a good person
>>109297996Honestly that would be what I would do. I understand that it's better to save money, always is, but I think saving money at bug fixes isn't the best way to do that. You already have the plan, but very often edge cases only appear when the actual code is being changed, not during planning.
>>109297988I think it just might be my specific case. I will try Luna on a fresh project soon that will not have the technological debt of the spaghetti monster I am currently fighting.
>>109295481>>109297739I have very very service level knowledge and understanding of how it actually works but basically they do some fancy layer magic so that the model can perform comparably to unquantized models at lower quantization levels. This apparently means it performs well with higher quality response also still maintaining performance and speed. Qwen implemented something similar if not the same thing with recent Qwen open weight models. https://github.com/rasbt/LLMs-from-scratch/blob/main/ch04/08_deltanet/README.mdhttps://sebastianraschka.com/llms-from-scratch/ch04/08_deltanet/In my anecdotal experience with the models it seems to have a noticeable effect on performance compared to other models in the same parameter range because at least on my machine Qwen 35BA3B seems to run noticeably faster than Gemma4 26BA4B at high context lengths despite technically being a beefier model. Yes the Gemma model has a lower amount of active parameters but in this comparison that matters less the longer the context is because you notice a drop off in t/s the larger the context grows. By the time the models reach their context limits the Qwen Moe was noticeably faster than its gemma4 equivalent
>company wants to promote me to junior
>>109298032congratulations. take it
>>109298035My title is Software Engineer, how the fuck is junior a promotion? It does have a pay raise but I don't want to be called a junior, 1 yoe
https://x.com/i/broadcasts/1AKEmmLPrneKL (Ai generated radio station that gives you the news, they also take x comments and play them on-air, say something nice)
>>109298032is it because you secretly use ai or because they think you use ai well?
>>109298018In traditional transformers you have to compare each token with all previous token to understand how related a word is to all previous words for some meaning of "related".So for the first token there is no comparison. For token 2 you compare with 1. For 3 with 2 and with 1. Etc. For token 9999 you will have to compare it with all other previous tokens which in turn are compared with all previous tokens and you can see how the total number of comparisons grow extremely fast.With deltanet you keep a fixed size state that contains a more conceptual or summarized understanding (so to speak) of all past context. So instead of literally comparing each word with all previous words you just compare the word with a understanding of the state the conversation is in, which can be represented by a fixed length vector. Then once you compare the new word, you update that understanding slightly to account for the new word in the conversation.So the number of comparisons you have to do is one for each token, so n comparisons (n being the number of tokens), rather than n-1 for each n (which results in n^2 comparisons).
>expand thinking>wall of textI don't think real humans need to think that much when coding. I feel like current AI is just brute forcing a larger context window to make next token predictions more accurate
>>109298076the latter, I set all our mcps/apis, fully remade our who cloud infra to be AI friendly, bla bla
>go over daily limit>top up $5 to finish something>it uses so much that it puts me at -$18>top up another $5 like a retard>
>>109298083basado, gives me hope, GLHF. Hope springs eternal.
>>109298080Welcome to the club. That's just expected with Chinese reasoning models these days. Whenever I use a local model I typically use qwen 3.6 35BA3B. It usually does a good job at whatever task I get it but holy fuck does it like to shit out loads of thinking tokens. I once asked it a simple question. Not even a task or anything that would require tool calling. Just a question in the llama-ui front end, it is shed out 4,000 fucking tokens for a question it probably could have answered just fine had I turned off thinking. The thinking autism is useful for vibecoding but kind of annoying when you're just using it for general purpose question answer sessions.
>>109298080>>109298099but muh gotcha car wash questions tho
New>>109298107>>109298107>>109298107>>109298107>>109298107
>>109298109Test like that or just prove these models don't actually think or reason in the human sense and don't have actual consciousness. Qwen 35BA3B pass the strawberry test but only after shitting out thinking tokens. I don't think it would have failed every single time if I had turned thing off but I'm confident it would have fucked up at least a couple times. The models are very good at their niche but mental midgets need to quit trying to use these or present them as all in one replacements for everything.
>>109297746mad af
>>109297816Sol medium, you don't need more
>>109294511>>109294514The OP pic is so true, even the website was called ExpertSexChange.
>>109297977Nah, I'm out here with the $200 one just making shit to see if it can do it. Currently seeing if it can vibe up my own X server with VRR/HDR/multi-size monitor scaling, but it's gotten pretty slow now that it's nesr the end of the roadmap.
