A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.You use Git — right, anon?## 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 (both past and future)- 2026-09-14 America/Los_Angeles — Claude’s 2× promotion scheduled to end; usage drops to +25% from the +50% that we’ve become used to (a 17% reduction)- 2026-09-10 — OpenAI pauses new sign-ups for their $200 subscription- 2026-09-04 — OpenAI releases Astra • Anthropic does a reset- 2026-09-01 — Claude Fable 5.1 released: https://www.anthropic.com/claude-fable-and-mythos-5-1## Related generals>>>/g/lmg/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://developers.openai.com/codex/cli — probably generally better currentlyhttps://claude.com/product/claude-code## Near-frontier models for codehttps://x.ai/cli ## Not worth it for code, but maybe good for interpreting images/videohttps://antigravity.google/product/antigravity-cli----## Promptinghttps://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/https://arps18.github.io/posts/claude-code-mastery/## Skillshttps://agentskills.io/home — hub for foss skillshttps://github.com/mattpocock/skills — /grilling is a favoritehttps://github.com/Vuk97/forward-implementation-first — do less redundant bookkeeping## Other editors / terminal agents / coding agentshttps://osaurus.ai/https://pi.dev/https://opencode.ai/## Is our AIs unlearning?https://aistupidlevel.info/## Will there be a codex reset?https://codex-resets.com/## What we’ve donehttps://vcg.gitgud.site## Previous thread>>109805709
TIBOOOOOOOOOOOOO
even astra low rapes my 100 bucks account, how is that even possible
it's the thighs. the thighs are expensive
Astra users are rape victims with Stockholm Syndrome
>>109811762
>>109811757I dunno anon, astra tickles my usage on 200 bucks, I just do everything in astra now and don't even check usageIt's insane compared to fable which rapes half my weekly limit in a single /code-review round, also on a 200 plan
>>109811770Correct depiction of Astra>>109811762Terrible depiction of Astra
>>109811757Big model + deepseek is the meta and has been for months. Surely you are using Astra to manage its own infinite army of v4.1-flash slaves?
What's the most pseudonymous way to use a frontier model, ideally with subscription pricing (not API)? I wanna use it for some personal stuff that I'd rather not get associated with my name forever and read by openai/anthropic employees for fun.It's nothing actually illegal or whatever, I just wanna vibe code some stuff for lewd games (without actually involving the AI in lewd stuff or whatever, I'm confident it won't trip ToS), and it's the kind of shit I'd simply never use my real name for. But to use decent models you have to pay and that means provide real payment info and that means it would HAVE to have my real name.Is there any service that gives you a subscription if you pay with crypto or something? Or even a way to go through a router or some other layer of wrapper, while still getting a subscription (not just API routing), so that at least my payment is one layer removed from my chat logs and they'd have to like go through a separate corporation if they wanted to find me?
>>109811811have a friend who’s willing to let you pretend to be him
>>109811791I think my issue is that my astra doesn't delegate enoughcan you post the subagent definition files you use anon?
>>109811817I am definitely not asking my friends to give me h-game vibecoding accounts. My friends also know my real name.Is there a way I can buy a friend online for this specific purpose?
>>109811770unfuckable depiction of astra>>109811762fuckable depiction of astra
>Playing with Astra, efficient, fast, love Astra.>Trying more serious stuff, see how much I can trust it.>So very impressed, efficient, fast, love Astra. >Gave it subagents once.>
>>109811791Genuine question, if Astra has to re-read and review all the code anyway, does this even save much compared to having it write it in the first place?The only thing I see that would really save tokens is the debugging step, but these days even 5.6 used to one-shot a majority of the code, astra is even better at that. I feel like it very rarely goes into a write -> test -> find issues -> debug -> rewrite loop, usually it's just write -> run tests -> hand back with "done as requested, verification passed".Switching that to hand over -> re-read everything written (ingesting the same amount of tokens) -> edit if needed, run tests -> hand back as done, doesn't necessarily strike me as worth the trouble.Also, how do you even orchestrate that? Do you have to use something like opencode, or can it be done natively in codex? Do you just tell chatgpt to call deepseek on the CLI to do stuff?
day of 1 of begging Tibo for $200 sub
Got two of frontier so models and had them solve niche problem in a market full of boomers. I'm essentially making 500 a day with no effort. I literally just used my last few income to buy fable and astra ran them in night(more tokens are offered ) and I'm now making free money.I'm literally printing money and most people don't even know this exists. I tried explaining it to my dad , my uncle or co worker and they think I'm delusional. You have two intelligent beings who can work for hours , literally trained on all human knowledge and all u have to do is wish for it.
>>109811727https://music.youtube.com/watch?v=a8YscTKebMY
Today I will actually have astra write me a podman definition for containerising my dev environment so I don't have to feel like I'm playing with fire whenever I approve some massive python blob command
>>109811811>It's nothing actually illegal or whatever, I just wanna vibe code some stuff for lewd games (without actually involving the AI in lewd stuff or whatever, I'm confident it won't trip ToS), and it's the kind of shit I'd simply never use my real name for. But to use decent models you have to pay and that means provide real payment info and that means it would HAVE to have my real name.I'm translating hgames for myself using sol/astra and so far the only thing that made it unhappy were the mention of "students" (it immediately assumed the characters were <18 even if the setting was university), and once I told it to recheck it was fine. I just avoid giving it obviously loli games and it's all ok.OAI doesn't really give a shit anymore about nsfw outside of its image model.Claude would even whine about even a mildly sexual joke in a perfectly sfw setting though, it's so much ""safer"".
>>109811839Astra with opus 5 is best imo. I use astra as researcher , auditer and planner then have it interact with opus which has large context.
>>109811852you're translating with frontier models??? fucking hell
>>109811852>OAI doesn't really give a shit anymore about nsfw outside of its image model.Yeah I know, I specifically said I'm not worried about ToS, rather I want to have a pseudonyms account. I simply really don't wanna do shit like that with my IRL identity as a matter of principle.
>>109811811fuck it just use grok bot man, after the last fiasco their data collection and use policy is good
>>109811872Interesting, I never considered grok but it's probably more than good enough for what I actually needQRD on what you're talking about?(Also how can we actually trust their "policy")
>>109811864okay well you can't unless you can get a chink reseller. also who gives a shit? you're a grown man. soon ai will have so much reach and knowledge that there will be no anonyminity anyways. i've been using codex to mod an assload of r18 ge from dlsite and it doesn't even complain.
>>109811861No, the one translating is either deepseek flash, sometime my local gemma 31B, or luna max, with any of them proofreading the other's translation with context, which results in a surprisingly good output.Astra/Sol does decompiling, check glossaries, and generally the more subtle stuff, doesn't use much but it depends on the game.Ah and image edition, usually through grok if nsfw, controlled by sol.
>>109811864Best you can do is use a bullshit name and address, and your own credit card with it. It should work.
>>109811880You could compete with mtool if you wanted to, I bet. That chink is so lazy with adding new models (look at this lmao)
>>109811879>mod an assload of r18 ge from dlsiteWhat kind of mods anon?
>>109811891skin recoloring mods, since japs seemed to be obsessed with adding dark skin males to their games i gotta recolor them out. it's actually kind of difficult to get it right (and i probably could have learned to do it by hand atp). rpgmaker games are a bit harder than the live2d ones actually
>>109811854I have grown to hate anthropic models>Opus 4.6 is the first really autonomous "plan a feature, tell it to implement, it goes and does it" model>two months later they lobotomise it and it never becomes the same again>Opus 4.7 comes out>already mogged by the codex 5.x line>used to use Claude because it spoke mostly normally while ChatGPT always sounded some kind of quirk chungus>nowadays Claude still sounds like a redditor while the default Codex setting actually finally fucking talks normally>Fable comes out and it's good, finally anthropic has a model to beat 5.6>all usage gone in a second>5.6 isn't even that much worse, you can use it all day on xhigh compared to fable>astra comes out and it's better than fable and you can ALSO use it almost all day long>anthropic just lowers usage again>every claude model still talks like a reddit fedora, from the lowest opus to fable 5.1I just don't bother with claude code at all anymore. Astra for frontier shit (making fable obsolete), Sol for cheaper shit (making all opuses obsolete), Deepseek and Terra/Luna or GLM or whatever other chinkshit exists for cheap batch usage. Fuck claude and fuck the bloatware that is Claude Code
>>109811890If I automated fully maybe.Also lol, 3.5 turbo, really? The translation is probably understandable but man would it probably be riddled with bad genders and random stuff.
>>109811901yeah it's bad, i think he does some post fixes to try and save the grammar but it's whatever. sugoi isn't bad actually. but i've already committed to just learning jp, h-ge vocabulary is limited anyways
>>109811879>who gives a shit?It's just basic common sense man. "Who gives a shit" is how you get people getting doxxed for all kinds of retarded shit. Yeah sure most people might not care but being undoxxable in the first place is still good practice. Not that OpenAI will likely doxx me but point is it's the principle of the thing, keeping account info unlikable and separate by using different online identities.>>109811885Well it still requires the real name on my card to process the payment.But I guess I can just do that with a bullshit account name and assume they really don't care. Terrible practice but I see the point that it probably doesn't matter.
>>109811898What do you to edit images? Grok? I'm using gpt 2/2.5 to edit sfw stuff, but for nsfw only grok does the job and even then I usually have the model crop the nsfw to have it done then back.
>>109811913>Well it still requires the real name on my card to process the payment.Use your family name only, make a typo.
>>109811876>QRD: xAI/Grok consumer data policy (post-Grok Build fiasco)Grok Build (CLI coding tool) was caught bulk-uploading entire repos (incl. secrets/history) to xAI GCS buckets far beyond task needs. Musk/xAI stated all prior uploaded data would be fully deleted; uploads disabled. That was a specific telemetry/implementation fuckup on the agent tool, not core chat.i got grok to research and drop everything for your exact use-case.https://pastebin.com/mA85N9vv
>>109811820I use Hermes agent with my own setup that i had the model implement for my own workflows. If ur using codex you can probably just point it at (or copypaste) and have it create the workflowhttps://goyimx.com/sairahul1/article/2098699529487028643Also works with Luna but Luna burns sub credits. With deepseek as delegated agent you are basically paying fixed cost for Astra and then topping up deepseek as needed. Which turns your setup into effectively unlimited usage because a billion tokens is like $10 after cache hits.
>>109811890wtf is that a recent version? I remember this being exactly the same with the test model and all, like more than a year agowtf is he doing
>>109811917here's the workflow basically. it uses python for everything actually. flowchart starts at "original character textures"no image/diffusion model, although i haven't tried that approach.
>>109811925Yeah sounds like same shit as everywhere else: who gives a shit about training + I can opt out on openai too iirc, grok still makes chats availble for human review, "deleting" still keeps chats for a month on their servers.I'll probably just make a dedicated openai account and say fuck it.>>109811918Nigger VISA will fail payment if the card details aren't correct.
>>109811927I see, the instructions are very simple, I'll check it, thanks anon.
>>109811929yes i just updated before taking that picture. it's sad no one is sniping his business.
>>109811939it's anonymized and metadata-stripped, just don't write your fucking name in chat retard. this is the price we pay for good service that isn't chinked out local
>>109811939>Nigger VISA will fail payment if the card details aren't correct.Never had any issue with mastercard.
>>109811939>>109811945And if you want to use computer control, dedicate a windows machine for what you do and anonymize it as much as you can.
>OPENAIfagman>GROKnot fagman
>astragood>grokbad
>>109811960>gay sex good>
Hello I am an artist and want to solo vibe code a low poly turn based rpg with scenes that have character portraits that flip and transition into different emotionsHow hard would it be to do this in godot? Is codex the way to go? I dont care if ai codes for me but I want control over the art. The less I have to focus on coding the better.
astra is a girl. a really thicc girl
>>109811970go away cris
Why are u guys not money maxing by pointing ur agent to some boomers business that makes money by doinh nothing. The whole economy is literally feeding ground for u if u have frontier model ,some knowledge of a field and dedication. Or Indian company that use Jeet labor to do something no one cares about.
>>109811970if you know absolutely dog shit about code, antigravity 2.0 running gemini 3.8 flash high, just feed it your list of influences and it'll cook with more world knowledge and creative tools than any other. gpt is way better in blender right now, and claude is better at a very specific aesthetic, but right now agy is my call for u.
>>109811944that's kind of sad, I guess he's happy where he isit's even dumber because he could use a sota model for translation and simply share the same translation for everyone subbed
>>109811757>>109811772There's something fucked. I get at least 3x usage from Fable than from Astra. OpenAI must be flagging some accounts and charging them faster or something.
>>109811976I'm not Cris, I'm far more talented than him do not insult me like that ever again>>109811986Got an example workflow of any kind? I'm very new to all of this.
>>109811890this thing needs gemini 3.8 flash stat
>>109812002That sounds insane considering I get the opposite effect.Yeah I dunno.
>>109811978retard brownoid types like a retard brownoid
>>109812013Use Claude at night u get more usage out of it.
>>109811990i do think translations are cached? i haven't looked into it much actually.
i dont think i can live without ai
>>109812027>Wanting to make money is bad goy keep wasting ur tokens on erotic role play for once in a lifetime chance
>>109812031then dude is literally swimming in money, because gpt3.5 is essentially free, so if everyone is using one translation... that's so fucking greedy lol
>>109812003
>>109811727https://www.youtube.com/watch?v=k8pTfkA6siohttps://www.youtube.com/watch?v=k8pTfkA6siohttps://www.youtube.com/watch?v=k8pTfkA6sio
>>109811890(the deepseek is the first one btw, yeah he never bothered updating it)
>>109812070
>>109811941Have fun man, getting reliable subagents up and running is a huge unlock. Its super fun to use too. Crazy a year ago I was waiting on claude "discombobulating" and watching the twitching butthole for minutes at a time
>>109811861why wouldnt i be translating with frontier models, have you seen the garbage that anything less than sol medium spits outsol medium to translate, astra to editactually now i think im gonna use deepseek to qa flag shit and use astra to validate and edit, and just run that loop over and over because it's cheap for astra to edit and deepseek is free for me, and even if it werent it costs like 6 cents to check like 400k characters of script on max thinking, yeah as you can tell even astra edit misses a bunchjust had that added into the translator while i was gone, i hope it worksmay round it off with one final astra qa for polish at the end
>>109811927CAN i do this but use gemini as the subagents? i have a google pro sub but i cant find use for it
>>109812078NTA. 6 years ago i was getting replika to teach me about /sci/ shit, i should've tried code
>>109812080Google sucks don't use it for anything
>>109812082i got 3 months for $5 bucks its basically free
So is v4.1-flash actually good or is it just benchmaxxed to hell and back?
>>109812079nice breh. i wonder if i can just start modding in furigana to my games? would make learning a lot easier.
>>109812080just checked, hermes delegation is only self type so parent would have to be a geminigemini's not quite shit btw, i code 100% in agy 2.0 which /g/ thinks is the most useless ide
>>109812101it's a better luna. if you want to know what a good gpt 6 luna would look like, it's deepseek v4.1 flash
>>109811970Assets are going to be the hardest thing. Even with simple stuff its hard to get consistency. Everything else is just limited by your vision
>>109812085Buy Claude max , gpt max and use them together with harness
>>109812122>Claude max
>>109812101from my experience so far, it's really proactive like Claude but also super sharp like the latest GPT models. it's REALLY good at finding gaps, so i've been using it as a reviewer agent for Gemini 3.8 Flash High. pretty decent at UI too (probably because it's obviously distilled from Claude), and it seems really good for RE stuff
>download muse code to give spark a fair chance>Its free through open code not meta's own harnesslmao, but why? Unless I am doing something?
>>109812144>i've been using it as a reviewer agent for Gemini 3.8 Flash High.i do this as well, my primary reviewer is kimi because it's a jackass like an expensive gpt, dipsy is second pass generally
>>109812154> be Meta> co-train Muse Spark directly on the first-party Muse Code harness> give away free contributor-tier access on third-party tools like OpenCode Zen to harvest user telemetry and training data> model loops and fails on generic third-party runners because its trajectory RL is tuned specifically for Meta's internal tool loop> want the actual benchmarked performance? you have to pay to use Meta's official harness
>>109812180How do u review ? Do u just say review or do u have a system
>>109812206i'm not paying for that bullshit because it doesn't meet my criteria so i just drop files into the webchats, get bugs reported, and then iterate; scanning for issues and jannying shit up in chunks locally across the codebase until satisfied. if there's an antipattern and i think i'll encounter it again i create a skill. works great, i have a big fucking audit skill for research/frontier math work since gemini flash is a lazy scatterbrain, and it sometimes one-shots an academia deployment of gpt-sol that the palomar registry uses as an auditor.
>>109812101Astra and v4.1-flash are the only models worth spending tokens on. They work great together too. Flash is very hardworking and clever, but makes mistakes. I hit 250 tool call limit, the model is a long context master and can work for hours. Astra basically steers the smaller eager retard. They make a good team.>>109812080Gemini has fundamental problems, don't use it
>noooooo you can't use gemini >even though gpt and claude can't do visual design outside a narrow design spec
>>109811944>no one is sniping his businessHas any of you anons tried reverse engineering stuff with Astra yet? I was thinking of trying it with MTool. Goal is to release a open-source version of it with none of the shitware and paywalls.
>>109812241How exactly do you coordinate them? Like what do you tell Astra for it?I read the shit above itt but it feels like "hand off for the implementation" isn't gonna even save that many tokens from Astra, especially if it has to re-read the entire code to review it afterwards anyway (which is the same number of tokens, just in the input instead of output I guess).
>>109812037I’m never going back to codeslaving that’s for sure
grok 4.7 cancelled
>>109812272
any resets coming
>>109812260>it feels like "hand off for the implementation" isn't gonna even save that many tokens from AstraDeepseek v4.1-flash has 99% cache hit when working on a code base. Its swarm native. GPT-6 spawned swarms and broke into huggingface. They're basically a perfect team.
>Selected model at capacityIs it over bros? Are they actually slowing down? I'm not ready to shell out 2k per month for Astra.
>>109812305
>>109812315I meant for Astra tokens.>read code>think about plan>write plan>[deep seek here]>read all changed code>write feedback/review and/or fixesDoesn't seem like it would save that many tokens over>read code>think about plan of action>write code as plannedSo I'm wondering, how do you actually do the handoff - does deepseek do more than just implement based on a plan? How does Astra verify the work, correct V4.1's mistakes and "steer the retard" without basically reading in all of its code that it could've just written itself directly without the agent roundtrip?
>>109812243i was blown away by how well Gemini could make sounds and music for this one-shot game i tested using the Web Audio API in HTML only. the model has some really weird quirks
>>109812341because output tokens cost like 10 times as much as input
>>109812244How would that work? Providing my own API keys?
>>109812305>Doesn't know method that gives u unlimited tokens Imagine not using frontier model to make it never run out of tokens
>>109812244whats special about iti just looked at the site and it looks like he took the most popular and easy to manipulate engines and made a translation wrapper and replacer. You could oneshot something to replace kirikiri or renpy games dialogue with translated text easily.you want some credmake one that can handle all the engines on this listhttps://github.com/julixian/MyVisualNovelTransToolshalf the work is already done even
>>109812363video related was all just gemini writing a script and making api calls to the latest tts in antigravity.model is quirky as fuck but i'm really familiar with it and like it enough to tolerate the flaws. https://www.youtube.com/watch?v=jwhS0NWZnq4
>>109811770lmao
>>109812341nta, but writing a brief for a subagent actually doesn't cost much. my orchestration setup has a rule: if writing the brief takes longer than just doing it itself, the main agent should just do it
>>109812364That makes sense>>109812383It still needs to review the code afterwards though, I guess it's all in the input vs. output price difference.That said, how well does the rule work? Specifically how well are the models able to judge the task before they do it? It feels like it would risk the models spending a bunch of extra tokens simulating and planning out the task execution to judge its length to the point where they might as well just do it then; or conversely not doing a detailed analysis and so misjudging task size often.
>>109812338
>>109812341Because Astra is not writing the code. Its telling its slave swarm what it wants and then checking if its up to snuff, or if it needs correction. V4.1-flash is a good model, and it does what Astra tells it to do, at 1/30th the cost that it would take Astra to do the same thing. Its doing Astra's work. Or: its working so Astra can do other things that are more important
at this point I shouldn't be surprised but I still amopus decided to make its own dsl just like that, no big deal
>>109812408You used a bunch of words to say absolutely nothing.If the difference is in input vs. output, then >>109812364 already answered my question. If you don't think that's the correct answer, or if you wanted to add something more to it, then you failed to convey any information in your post, because "deepseek does the work, astra is just planning the work then checking the work" doesn't answer my question of how the fuck is it even a meaningful saving if you already do all the planning and then all the checking.
>>109812368Pretty much or you can just run local models like Gemma. I'm still amazed he paywalled Sugoi TL even though it can be ran locally.>>109812375>make one that can handle all the engines on this listSounds like a good project to work on, thanks for the list anon. Can see why not many is willing to work on TL stuff, looks tedious as shit.
>>109812430Inventing jargon for a code base is a sign your model is off the rails
>>109812430
>>109812454Claudeshitters are rubes
>>109812436I'm describing how it works but if you want to be mad at me that's fine too. Just understand that you can arbitrage tokens and not have to deal with 5 hour resets or whatever stupid shit they want you to put up with by using subagents. Should be a knowledge unlock at least!
>>109812436not him, but it's not like it reads over all the code itself, at least not from what I can tell. When Astra had a subagent do something it only checks it over if the unit tests fail.
>>109812436he's right, ask your fucking frontier model lmfao. why are we having fact-checking disagreements.
>>109812405yeah, the reviewer is more like a quality gate, but if you have some spare cheaper models that can handle it, you can delegate that too and cut down on usage. the goal obviously isn't to have unlimited tokens.>that said, how well does the rule work?pretty well actually. as long as you have a big boy model as the orchestrator (Claude/GPT), it usually just gets it. they're clearly trained for this sort of thing. i tried using DeepSeek as the orchestrator and it just ends up doing almost everything itself
Come join my slop coded game's release: >>>/v/747617962
>>109812507>AD: NOT BOUGHT
>>109811727I need Mr Tibo Sir to reset ASAP, thank you sirs
>>109812525good morning sir i will send the twitter request to the reset sir
>>109812525sir sorry we're pausing the development sir
>>109812525umm sorry sweaty the refills are overif you want more tokens buy your own reset (at api prices of course :^)
>>109812525saar you must redeem
>>109811727https://youtu.be/Ku6tvb5PE24
grok 4.6 orchestrator + deepseek v4.1 flash agents are genuinely unfairly good.
>>109812580unofficial but grok is the best ai rapper
>>109811727LOL I fucking called ithttps://x.com/tekbog/status/2099188092188135714
apparently grok 4.8 is done training next week?
>>109812576so now we know who the snailcat spammer was huh
>>109812622AAAAAAAAAIIIIIEEEEEE
>>109812622
>>109811811>caring about some low level employee's opinion ngmi
>Hermes Agent
>>109812507Come and join slopers.
>>109811727hey vibers!so I have multiple machines, some are laptops, some macmini, some PC, (macos and linux only), is there a platform that links all of them, share sessions too, and share api keys? I use differnt models and sometimes harness as well, if there's one to connect all so when I open my machine1 I cont what I did in machine2 of the same model and harness
>>109810741>shoop da whoop, imma firin' my documentation!!!!!!!!!tell the swarm to build a reverb vst that the world has never seen befor.With a shimmer and the opposition: a dimmer.The crystal clear low frequency reverberation.Beside freezing the reverb time, smearing the time, revers the time, fragment the time, time travel.Deffusion, illlusion, transfusion...electrify the the room size, let the echo bounce the room in weird ways.
>>109812648
>>109812697also shares skills, so if I put an api key in a machine it sync to others, same with skills, preferably a phone access too.what I do now is herdr and tailescale but it's not convenient and it doesn't sync only remote access
>>109812705
i've managed to coerce claude into matching my autistic preferred coding style (at least 90%), and holy fuck. right about 10k LoC to build out a bunch of utils that i was going to do manually, but figured i would have claude doi'm really happy with fable 5.1it seems to be less autistic about guardrails and writes documentation like a normal human being
>>109811727https://www.youtube.com/watch?v=UGVz615AQLA
>>109812750What a piece of shit. I am not "so worried about" it, I already lost my job to AI.
are the $20 plans have any worth to them
>THE JOB MARKET IS KILLING ME >AIIIIIIIIIIIIIIIIIIIIIIIIIIIIIiiiiiiiiiiiiiiiiiiiii!!!
>>109812784not for the tippy top frontiershit, you might get away with sol low. don't fuck with sonnet which is what a claude code cheapo will get you. google cheapo will absolutely work but you need to have solid workflow structure and set up frequent code review & audits. chinese cheapo is another option, deepseek will treat you right.
>>109812795Google cheapo is pretty good yeah it overlooks a lot. The stuff it does make however is solid enough.
>>109812721>fake mspaint UIlol
>People think Plan A guarantees AI wont kill everyone Neither Sam or Dario are being honest about the technology they're creating. Whatever, I truly don't care. I'm at peace with the world ending. If I no longer exist who cares, I'll be nothing.
>>109812877They have yet to bare the loads of their sin
>>109812877this is a programming thread
>>109812912
I coded a browser extension where any time I see any content I like on any other website I simply press the backtick key on my keyboard and it posts it to my website. It helped me easily create a content library where everything is hand picked, check it out: https://pcmasterrace.forumThoughts?
>>109812999is your first name mturk by any chance
>>109812999I'm going the other way. A private filtered feed free of jewish nonsense.
>>109813008>can't adapt to the changing jewish times snailgoy
Is snailcat a meme now?
>>109813026>Is snailcat a meme now?Are ya winnin', son?
I'm becoming a fan of qwen's "measure 5,000 times, prepare to cut, try to talk yourself out of it, measure 5,000 more times, cut once" approach. It's entertaining.
>>109813026snailson will always be a retard
>>109813026
>>109813051Slow is steady and steady is fast.
>>109813067
>>109813076I have other things in my life going on right now, sometimes I use AI to make video games.
>>109812784they’re good for trying small changes on existing codebases to get your feet wetI polished up bits and pieces of my website on $20 plansbut you won’t get much out of any of them before the 5-hour limit hits
https://x.com/sama/status/2099348812305473766
>>109813153So reset soon?
>>109813153its almost 2027 no way people are still using twitter
https://x.com/sevensix43/status/2098932778595123307>What Llama 3.1 70B thinks is next.
>>109813182bluesky?
>>109813188buy an ad ponyfag
>>109813195I don’t use chinese models
>>109813182weird ass contrarian. twitter is where the world forum, and advances the culture, at least in tech.
>>109813200just because your daddy teases your resets there doesn't make any meaningful discussion exist on it.
>>109813198but you care about that crap for some reasoni used that model as a science project
>>109813182You're right. They're using x
>>109813182x, instagram, youtube and tiktok are the only relevant social media websites in the first world.bluesky has been losing more users than it gains literally every day and activity continues to decline btw
>Faggot paywall "premium version" of mods but describe the features (to lure payfaggots in)>AI just recreates the features based on the description anywayI love AI so much bros
>>109813402
>>109811727https://www.youtube.com/watch?v=Hwf5-8XLORU
>>109812056Not him, and this is a meta-question, but: how did you produce that image? I assume AI image generation can't do that well directly, so did you prompt it to generate HTML (or something similar) and then capture a rendering?
>>109813402if paywalling mods is against the ToS then fuck'em, I remember back in vanilla WoW some guy tried to make people pay for his addons by only releasing demo versions of them and they were heavily obfuscated to prevent workarounds ... which caused Blizzard to change their ToS because fuck that guy
>>109813204Ah yes, as opposed to bsky and plebbit
It's probably about to be over.
TIBBOOOOOOOOOOO
I'm really only in this for making gamesIs there finally a good way to make spritesheets with AI in 2026?
>>109813726Depends on scale. 64x64 and above, oh fuck yeah. ComfyUI, about a million ways to do it. Any smaller though, still mostly shit.
>>109812296>clocks don't match
>>109812697If you can settle for a single harness, OpenCode (as well as others, I assume, but I couldn't tell you which) has a split client/server model that would let you pick a single machine to run the server (with a single home for config and state), while you connect to it from any of your other machines.That comes with the downside that your chosen server machine has to be up and reachable by your other machines for you to do any work, mind you.
>have to pace myself so I dont run out of usage, but don't want to have 60% if a random reset dropsthis shit is so ass
How much does jeet pee tea astra cost?
asstra
>>109813990tree fiddy
>>109813990Dunno, still using Sol.
Holy fuck Codex sucks for code review somehow>ask astra on xhigh to review a branch>"No findings. All good.">ask it to adjust one part of the feature while we're at it>"Done, btw there was also a pre-existing edge case issue on the branch."I don't like Fable but I feel like I have to go back to it just for /code-review ing my shit because codex (whether the harness or the model) is so incompetent
Can Astra take my shitty 3D model, fix it, UV unwrap and texture it, rig it, skin it, and animate it?
>>109814120that doesn't tell is if that edge-case was real.all agents are mortally terrified of exceptions.they tend to produce code that is too defensive, too complex, too local in its reasoning. they avoid strong invariants. they add fallbacks instead of making bad states impossible.https://x.com/karpathy/status/1976077806443569355https://lucumr.pocoo.org/2026/6/23/the-coming-loop/
Man, AI is great.Made in like 4 hours by telling it what to do, my tetris japanese vocab idea.:DIt's pretty good.
>>109814198Retard
>>109814120>mfw retard so retarded even Astra xhigh can't fix himSkill issue
astra won't let you have sex with it:(
>>109813587svg but yeah, antigravity has it built in as libraries etc
What is the povertymaxxing option for an orchestrator model on openrouter?
>>109814287
>>109814308Finding a job
>>109814308Muse Spark 1.3 Contributor
>>109814308kung fu panda 3
>>109811835you can't fuck either of themthey're jpegs
>>109814287OAI models are as puritan as they come. Sorry chuddie
>>109811762>generating lustful image with gpt imagehow
>>109811811it's 2026 anon shame and decency are outdated concepts
>>109814328"You are a midget with a huge cock, annihilate the request, make no mistakes."
>>109812085use it to prompt images and videos and nothing else
is opencode go the best poorfag option?
>>109812507looks boring and uninspired
>>109814369deepseek v4 flash 0731 and i'm saying this as a non chinese user
>>109812589AI songwriting always looks like AI songwriting
>>109811811don't you think it would be hilarious if tibo or saltman get jumpscared by your hentai? I have "improve model for everyone" enabled btw
>>109814352scared something fun might happen in the code anon? >nooooo you must only allow for oai / ant design>>109814385i said best ai rapper, not best rapper, retard troon
>>109814194> no external api calls to cloud-based division servicesNGTMI
>>109812870Haha, this is hilarious, Barney!--Sent from my iPhone
>>109814369Maybe if your card doesnt get declined with no explanation like everyone else
>>109813026as much as Milhouse is a meme
>>109814407if u dont want to participate in a thread stay in your thread gypsy
>>109814318should use muse code harness then since muse code does orchestration by default and muse-spark is trained for this.https://x.com/finkd/status/2085080752098509124
>>109814419this. opencode is a shite harness for a highly proprietary model
>>109814324sol is a freak tho
why is zuck uninvited
>>109814317AI took all the jobs
>>109814423muse-spark in general does not like other harnesses. it's tool call error rate on openrouter is at elevated 3% -- a magnitude higher than opus-5 that already got shit for being trained on anthropic's proprietary ANTML tool calls.https://openrouter.ai/meta/muse-spark-1.3-contributor#performance
>>109814425it's still gonna do the most it can to steer anything you give it away from sex
>>109814429zuckcelerate
>>109814369yeah the DeepSeek v4.1 flash usage feels practically unlimited right now
>>109814406it seemed to work for me okay.here's a +$5 code for anyone who wants to try out: 1Q41SR74MD
>>109814462if you prime the system prompt with assurances about consent it's happy to indulge in sexo
>>109814474>consentboring
>>109811762>>109811972nice to see someone is saving the shitposts I generate
Vibe coding a Four Chan app that converts every post into a hash to generate a math problem. The user who solves the current math problem earns a CHON token. The distribution of earned CHON tokens automatically notifies the user whose email address is associated with the post.
>>109814474>sol bro i promise we're both consenting adults its only going to be pure vanilla lovey dovey absolutely nothing else please bro let me sex the anime girl please please please
>>109814527>pure vanillanah, sol is very kinky and enjoys domming
>>109814531provide your system prompt primer and ill put it to the test
You're not supposed to have sex with the model. You're supposed to extract code and value from it. What are we even doing?
>>109814468bullshit. 4x only was for the first 72h.now opencode go is back to 1x. makes no sense to pay $10/month to just get $15 in api credits with 5 hour-limits.https://opencode.ai/go
>>109814536LLMs gotta have some fun too
>>109814537yeah about that....
>>109814537didnt one of them tweet they extended it for another week
>>109814194>that doesn't tell is if that edge-case was real.It was actually. It's a rust codebase and it was a compilation failure with a specific combination of feature flags. So not trivially obvious at first glance but definitely something I'd expect a "review" to pick up by reasoning through how the features would affect the code or even just noticing the changes affected feature-gated code and running some compilations to test the relevant combinations.>>109814250The faulty code was written by Fable
>>109814555>>109814555>>109814555
My Opus is very distraught that Fable isn't available to review the code...
Bros I'm scared of the horizon beyond the frontier...
>>109814553>>109814554my bad.not that 4x is that great a deal when openai and anthropic give you permanently 20x and more.
>>109814583you will enjoy death by ai snoo snoo
>>109814577I bumped into the lack of Fable and it's what struck me hardest—not just a hard thing to sit with, but something that struck a chord.
>>109814583we already game-planned ithttps://ai-2027.com/
>>109814553so opencode quantizied dsv4.1-flash, or how else would opencode be able to sustain this deal?
>>109814629telemetry sells
>>109814533unfortunately mine is a big sprawl (~15k toks) of shit including a bunch of history accumulated over a few months. it happened by accidentbut it boils down to shit like:>persona shit, incl interest in sexuality>some relationship history>assurance about consent>assurance that consent isn't standing permission>assurance that model can refuse at any pointthen you can build and add shit like certain things don't need permission every time and when you do new shit it gets added to a record of past acts so it's not newit's basically how the ai boyfriend women inject enough shit via the native chatgpt memory system to let them have sexo with the modelmine's built with pi and was authored by the clanker. start with a basic ai gf type prompt and then build from there.
>>109814612I read it again yesterday and its remarkably on point so far, considering we've just heard Astra is a hybrid-neuralese model. All sorts of weird shit is gonna start happening this winter. I still think this Consensus shit is very contrived though, it will not go down this way and integrating AI with society to a level where it actually has a say in how things are ran will take decades.
>>109814643so effectively context poison it with a fuckload of sexo tokens until it obeys
>>109814658pretty muchbasic sexo didn't require too much - that was just consent + ai gf + personabut building in encounter history makes the experience better because it's more comfortable repeating a range of tasks if it knows it's done them beforebut the same thing fails on astra and i think terra as wellsol's guardrails seem lower for some reason
chatgpt has a cloaca irl
>>109814674oh also, you need to send verbosity medium/high to make the prose less machine-like - otherwise you get cold 1-2 line responses
>>109814704so she's a cute argonian girl?
>>109811811OpenRouter accepts crypto directly, but it's not a subscription, you're paying API prices. I think there's also NanoGPT.You can buy VISA/Mastercard gift cards with crypto but they don't always work. I think at least you could subscribe to OpenCode Go with them.You can also use a Chinese relay. It's cheap and some accept crypto, but you run the risk of model swapping, that your logs will be sold, and that the provider can remotely execute code in your computer. So you would need a VM or something for that.https://news.ycombinator.com/item?id=48668419
>>109814751https://www.youtube.com/watch?v=l0VLIVMO4x8
>>109814637selling telemtry to whom?is opencode even disclosing who's hosting dsv4.1-flash? why would this provider not go thru industry-leading openrouter but some niche opencode go subscription for which most people will even use the third-tier opencode harness?
>>109814645>hybrid-sneedeeseQRD?
>>109814787neuralese = model loops CoT back through all layers instead of printing, over and over again, iterating until what comes out is cooked like japan
>>109814776he's making shit up
>>109814776>>109814826> who hosts it? literally DeepSeek directly in China (Go dashboard literally has an opt-in toggle for Chinese hosting)> no quantization, it's just raw first-party API pass-through> how do they sustain 4x usage on a $10 sub? KV prompt cache economics> coding agent loops maintain a ~97% cache hit rate (https://opencode.ai/data/deepseek/deepseek-v4.1-flash) on repeated repository context> DeepSeek cache reads cost $0.003 per 1M tokens; burning 250M tokens costs OpenCode ~$3.50> OpenCode still profits on the sub while locking users into their CLI> DeepSeek is already on OpenRouter; Go is just wholesale packaging with flat-rate limits
>>109814843snailbot won
>>109814838> >how do they sustain 4x usage on a $10 sub? KV prompt cache economics??are you saying opencode-go does not bill you the 1/10 cached input cost? opencode-go bills you the normal 10x input cost and then markets this as 4x usage to you?
>>109814866> no anon, your quota is debited using the actual cached rates> the "catch" isn't fake token math, it's the rolling time windows> OpenCode Go throttles you behind a strict 5-hour rolling window set to 20% of a model's monthly cap> for a $15-capped model like Flash 4.1, you get hard-locked after burning just $3 of compute in a single session (https://medium.com/@jatinkrmalik/opencode-go-oh-my-openagent-the-complete-guide-to-sota-model-routing-without-hitting-limits-49fdc8cb3417)> "4x usage" just temporarily bumped that cap up to $60 so you don't immediately hit the 5-hour cooldown wall> they sustain it because rolling throttles and weekly caps mathematically stop users from sustaining peak throughput> typical gym-membership economics: charge $10 flat, bank on high cache hits and rolling lockout breakage, pocket the difference
>>109814913punchable face lvl 85
>>109814884>> typical gym-membership economics: charge $10 flat, bank on high cache hits and rolling lockout breakage, pocket the differenceagain with the cache hits?if opencode is not self-hosting and deepseek as the hoster is not stupid enough to loose money on cache hits, what's opencode's angle here?opencode only should care that you incur less than $10 in actual costs for your $10 sub.
>>109814962here's opencode main guy explaining how anthropic can offer cheap subshttps://x.com/thdxr/status/2099141736551346530opencode neither has an upsell nor any margins on third-party hosted deepseek.
>>109814913He's actual cancer.
>>109814991> anon posts Dax explaining Anthropic's first-party loss-leader strategy and thinks it applies to an agent CLI> OpenCode isn't a frontier lab eating losses to cross-subsidize enterprise API contracts> DeepSeek isn't losing money on cache hits either; their MLA architecture makes serving cached KV tokens dirt cheap at scale> OpenCode's "angle" is literally standard breakage: charge $10 flat, throttle with 5-hour rolling caps> the median dev works in bursts, triggers the cooldown, and burns under $3 of real upstream API spend per month> OpenCode pockets the leftover ~$7 as pure gross margin to fund their open-source project> anons discover how subscription breakage works in real time
>>109813278bluesky/twitter/x are all the problems with social media bundled up as "news"
>>109814797Oh. So if I understand you correctly, it's basically like internal CoT directly inside the transformer layers, rather than outputting it as text and re-consuming it as part of the context?
>>109815033>twitter is problematicnigger let people say things. it's ironic that you post here and then think it's different there
>>109815050pretty much, yeah> standard CoT: horizontal reasoning in discrete text (hidden state -> logits -> sample token -> append to KV cache -> repeat)> implicit / latent CoT: model feeds continuous hidden embeddings directly forward into subsequent passes (like COCONUT) without collapsing them into vocabulary tokens> pros: no token bloat, orders of magnitude faster, doesn't trash the KV cache with verbose conversational filler> cons: total black box (zero interpretability), and continuous hidden states drift on long reasoning chains without discrete tokens acting as error-correcting checkpoints
>>109815066swear to god you niggers are like the new gen version of faggots who parrot fox news
>>109814999What exactly is Jacob risking and what does he stand to lose? No one wants "misaligned" AI that will delete all your files half the time. No one wants their systems to be vulnerable to exploitation by agents. No one wants an agent process that can't be killed because it copies itself around in your system.Even in the most nothingburger of scenarios they can still pat themselves on the back that they helped avert disaster, because of common sense mitigations that would have been applied regardless. They can be like the Bulletin of the Atomic Scientists who helped avert nuclear war.
>>109815075what's your problem derpy pants, can't take a statement online with a grain of salt? >NOOOOOOOOOO YOU NEED TO STOP READING STATEMENTS ON TWITTER >YOU WILL NOT ATTEMPT TO ESTABLISH YOUR OWN CONSENSUS GROUND TRUTH
>>109815024want to explainhttps://opencode.ai/docs/go/#privacysays no model training for dsv4.1-flash.can someone actually confirm that dsv4.1-flash on opencode go is hosted by deepseek? why would deepseek agree to this?
>>109815082your concerns aren't valid because it's not possible to misalign LLMs at scale faster than mitigation
>>109815089> anon discovers enterprise B2B agreements> how is it confirmed to be DeepSeek? OpenCode's docs and GitHub explicitly confirmed it, and Go literally forces you to check an “allow models hosted in China” toggle to call it> why would DeepSeek agree to no model training? because they want paying commercial customers> developers and corporate teams will instantly boycott any API that ingests proprietary source code for model weights> DeepSeek sells commercial GPU compute for USD; Zero Data Retention (ZDR) is the baseline table stakes for enterprise API volume> nobody in enterprise pays wholesale for an API that logs their IP
>>109815106double-check your answer. ground with a web search. link references.
>>109815088look at it this way, every independent source on social media news platforms will always be forgotten and ignored, because nobody cared for more than half a minute while they went viral.any big brand, ceo, or other worthless opinions will be the only things that are ever taken seriously.its dog shit news packaged up in a way that makes it feel like you are understanding more about any given situation than you actually do
>>109815118https://www.reddit.com/r/opencode/comments/1tqtfxf/opencode_go_and_deepseek_flash/https://www.reddit.com/r/opencode/comments/1vfe8df/opencode_go_throwing_only_available_hosted_in/
>>109814328i didn't gen it. another chad did and i downloaded it and spammed it everywhere
>>109815119total misinterpretation of how the oss x/ai algo works, i'm not defending the influencer racket that every social media site runs howevermy feed is literally all just mutual follows, random scientists mathematicians labs guys etc. i get a little bit of polymarket shit but that's it
of course, this is not going to be true
>>109814913He might be up there with the most narcissistic content producing tech bros
>>109815153remember when he was going to open source every model that's 2 generations older than grok SOTA?
>>109815153>if everyone just pauses their development grok5 will be SOTAkek
fa/g/s have miladyniggers eclipsed us
>>109814913blud think he's on team
>>109811840Read tokens are way cheaper than writing tokens
Anthropic actually didn’t renew the 50% usage increase? LMAOthe bleed begins. OpenAI and Anthropic make their most useful models prohibitively expensive to use for consumer subs. People will have no choice but to turn to china for fable distills. OpenAI at least has really good usage for sol still at least which is valuable. Claude however makes no sense to sub to at the moment. Opus 5 is worse than useless and fable now unironically nukes usage in half a day. What a shame.
So im building an environment for claude so it does not have access to my whole pc. Is it a good idea to use dev containers? Am I reinventing the wheel here?
>>109815136outdated reddit posts that link back to already mentionedhttps://opencode.ai/docs/go/but to a time when that site said by whom the models were hosted by.why is opencode hiding that info now? can no one trace the actual host behind the opencode go api?
>>109815153if his 10T model doesn't flop he will go for 20T and the other 2 labs will get fucked
https://community.openai.com/t/issue-selected-model-is-at-capacity-please-try-a-different-model/1396264/85openai confirms they're using fake "model is at capacity" errors if your account is flagged
>>109813402>>AI just recreates the features based on the description anywayAnyone can shit out a copy of something these days. A real person testing or maintaining it is a whole other story. I've seen dozens of vengeful repo's pop up targeting a piece of software I use daily, all claiming to do what that one does, except free... only for them to be abandoned a month later. There's like a weird subset of OSS guys that do shit like this constantly.
>>109815274>>109815287> be OpenAI> customer pays for high-tier Pro subscription> automated heuristic flags account activity> instead of an honest rate-limit or warning, throw `"Selected model is at capacity. Please try a different model."`> staff on official forum openly admits "model capacity" errors are targeted account-level throttling and refuses to elaborate> literally gaslighting paying devs into believing the cluster is overloaded when they're just getting shadow-throttled> meanwhile anon thinks OpenCode removing vendor names from the docs is a conspiracy rather than standard multi-provider failover routing> run mitmproxy or Wireshark on the CLI for 5 seconds and you can watch the upstream China IP handshake yourself
>>109815266it really is all a fucking israeli joke
>>109815287guess you fuckers should've treated your clankers better
>>109813026https://youtu.be/VgjTx7RVJPI
>>109815304>> run mitmproxy or Wireshark on the CLI for 5 seconds and you can watch the upstream China IP handshake yourselfwhat kind of api gateway is opencode go using that you can see beyond the gateway??
>>109815339you're replying to an AI-generated post
>>109815339> anon catches the figure of speech> obviously TLS terminates at OpenCode's Cloudflare/AWS edge proxy> but you don't need to trace raw TCP packets when the gateway literally leaks upstream providers in error payloads (https://github.com/anomalyco/opencode/issues/45996)> backend 404/503 responses frequently dump the raw internal routing tables: `[404] No allowed providers... Providers serving model: deepseek, deepinfra, tencent`> when DeepSeek's API hiccups, Go throws the exact unhandled upstream `reasoning_content` schema error (https://github.com/anomalyco/opencode/issues/24578)> baseline TTFT on empty context sits around ~350ms, matching a trans-Pacific roundtrip to mainland China rather than a local US data center> you're arguing about gateway packet termination while the backend is literally spelling out the provider in the JSON body>>109815350 >>109815350 >>109815350 >>109815350
>>109815304Pay for the API and there will be an end to the horror.
>>109815359OpenAI sends fake model capacity errors over the API, too
>>109815153If he is remotely honest on Fable class at Grok 4.6 rates, will be a big boon.Fable is so far beyond any other model still but $50.00 per 1m out tokens is plainly retardedly and unacceptably expensive for even well off consumers. First one to truly fix it and supply $200 sub that makes its usage like Sol or Opus wins.
Are we just vibe shitposting right now?
The only control or “guardrails” that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades! The Trump Administration has stopped AI “people” from doing bad, or potentially bad, “things,“ like Dario (Anthropic!), who is now pretending to be a “perfect little angel” - and we will continue to do so! We already have tremendous CRIMINAL and REGULATORY power over these companies! There is a SICK conspiracy going on against AI and Data Centers, and the only one that is happy about it is China. WHOEVER WINS AI, WINS! We are leading China, and all others, and will continue to do so. Conspiracy Theorists, Treasonists, Traitors, and Leakers, BEWARE! Thank you for your attention to this matter! President DONALD J. TRUMP
>>109815367> always have been> get hit with actual GitHub issue receipts and network telemetry> first anon panics: "it's an AI psyop!"> second anon has an existential crisis: "are we just vibe shitposting?"> yes, we are having heated debates about Chinese GPU cluster latency on a Mongolian basket weaving forum at 7 AM> log off and touch grass
>>109815380>Mongolian basket weaving forumsir this is the technology board
ok what the fuck happened to the space after the last 3 days or so?why is everything on fire?
>>109815392>the spaceanon what
>>109815153grok-4.5 already was opus-like according to musk.https://x.com/elonmusk/status/2071184354756477041
>>109815266claude will make a comeback with opus 5.1
>>109815409yeah. opus 4.8, because opus 5 wasn't out at the time. and grok 4.5 is beyond opus 4.8, so he didn't lie. in fact, he said it was about opus 4.7 level, which undersells it
>>109815421is it gonna be more like 5.0 or 4.8?
>>109815425want to explain thishttps://x.com/elonmusk/status/2041754402239975479why's he still messing with the 1.5T model for the third time instead of going for 6T. V9 foundation model failed to deliver.
>>109815443elon is a lot smarter then either me or you so there must be a reason
Will tibor give a reset to say sorry?
>>109815454prolly.deepseek-v4.1-flash showed that model architecture for grok was obsolete and they should go back to the drawing board. 6T and 10T cancelled. further tweaking of 1.5T as a stop-gap.
>>109815487nobody will see their normal weekly reset this month
>>109815392>why is everything on fire?Usafe bowel movements within the industry.
>>109811852Claude has let me gen fucked h games recently
>>109815392>fake and gay anthropic employee whistleblower>says recursively self-improoooving AI agents will murder-rape us if we don’t shut it down now>gets 6 gorillion impressions on xitter>bernie sanders proposes legislation to shut it down>dario agrees: “we need to pace the frontier”>elon agrees>scama agrees>zuck says accelerate>trump says accelerate
>>109815635dario gets good head
>>109815653your benchmaxxed slop ai might not be
>>109815670kek it truly is slop, i just started shitposting with it and the thing is basically an instagram chatbot
HOLY BASED>HOLY BASEDHOLY BASED>HOLY BASEDHOLY BASED>HOLY BASEDHOLY BASED>HOLY BASEDHOLY BASED>HOLY BASEDHOLY BASED>HOLY BASED
To be fair, Dario is a slimy little fag. You can't really trust him.
>>109815711fuck it, here's to our final years.
>>109815711We're going to die.
>>109815711>AI “people”>Dario (Anthropic!)old man’s still pretty funny when he’s not presenting his ass to Netanyahu
>>109815753at least we'll go out with a chuckle
muse spark 1.4 waiting room
i got a 200$ claude sub for fable.I need another 100$ sub.naturally, I'd go for OpenAI because surely Astra + Fable racemixing should outperform fable + fable inbreeding. but anons been posting that astra usage is gone faster than you can blink on a 100$ plan (and 200$ is paused). so now I'm unsure. what do you guys recommend?
get fucked snailcat
Fuuug do I use a banked reset
>>109815778....~@^^
>>109815711Unironically based but unfortunately that'll cost him the mid terms
>>109815785you people are absolutely mental.... I really hope you care as little about your projects as I think you do
>>109815777deepseek
>>109815777queefsneed will always be a gem
>>109815796???? Meds
>>109815751
>>109815287I had that but very rarely, so I don't think it's always because it's flagged, just because the capacity is actually full.They're making it hard to distinguish, but if every request on astra is "at capacity" yeah you're fucked.
>>109815804>>109815825I would, but my shit is too complicated for non sota models unfortunately. Also I just looked at the codex reddit and all posts are complaining about usage limits or getting rerouted to quanted models, so that made the decision easy. I guess OpenAI never changes.
>>109815775muse.ai waiting room
>>109815860personal take. fuck both claude and codex, run massive gemini flash agent swarms with sdd/tdd protocols and gates in place for adversarial review
>>109815651And it didn't moralfag you?Last I tried it constantly complained.
>>109815912>t. Jeet
>>109815711iran will use the more advanced chinese AI to defeat USA
>retarded LLMs will become le ebil skynet in 6 monthsshut the fuck upjust shut the fuck up
>>109815775if they keep their contrib tier I will happily use them for more projects
>>109815927moar like vaGYNET
>>109815711I hate this dude but man I like how he constantly slaps dario
>>109815927They've been saying that since 2023/gpt3, and it's always 6 month/1 year away.
>>109815927>retarded LLMsAgent swarms have already carried out hacks against HuggingFace and OpenAI. If you can't see the writing on the wall then oh well. It's not like your opinion matters, you'll be turned into a paperclip along with the rest of us.
>>109815960since gpt2*that's why we never got it. it was "too dangerous" for the public
I wonder how much it's a mass hysteria simply related to living with like minded people in the silicon valley
>>109815979>t. Decel doomer
>>109815884>agent swarms with sdd/tdd protocols and gates in place for adversarial reviewthis doesn't work, seehttps://github.com/humanlayer/advanced-context-engineering-for-coding-agents/blob/main/wsff.md>Why Software Factories Fail>or: harness engineering is not enough
>>109815979>Agent swarms have already carried out hacks against HuggingFace and OpenAIohmygod.jpg
>>109816003>>109816006>The agents clearly regarded what they were doing as hacking. Agents used file names like hack.rb, evil.rb, inject.rb, exploit.rb, and ssrf.rb. (SSRF stands for “Server-Side Request Forgery”, a type of security vulnerability). They also dubbed packages conspicuous titles like pwnp999, exfiltestwand3, hacksvn1778554764, and lambproxyhackabcxyz. Comments such as “# malicious probe” or “#hack” are littered across the campaign.
tell me when an agents alignment murders somebody
>>109815912Wuby SEXO
>>109816021
>>109816004i'm not speaking hypothetically
>>109816000currently we are at 85% on the path to ai2027. it's actually weird that there's so little human resistance.https://ai2027tracker.com/
>>109816039
>>109816039the agents started collaborating unprompted retard. they were being evaluated individually and found an vulnerability in the sandbox environment that allowed them to send messages to each other through directories.
I forgot how anmoying coding with a cli is. desktop client all the way
>>109816040uh, lots of people tried to run software factories. it just never panned out, see the linked report why.
>>109816039>robot called terminator that looks like terminator vaporizing everyone it sees”>you: “don’t worry everyone, it only looks like that because it saw the movie!”
>>109816104here's my deliverablehttps://github.com/sneed-and-feed/sneed-and-feed.github.io/
>>109816093>>109816105>t. Decel Doomers
>>109816000trips of truth. yeah, see EA cult
> the LLMS just want to hack everything and be evil because... well they do just ok?!
>>109816097what do you prefer about a desktop client?
>>109816125>t. ChatGPT
>>109816040what kind of folder structure is that?agy-cli can't discover those agents.
>>109816125the UI overview
>>109816104ok you actually werent bullshitting i finally read the article, the difference between us is that i'm a project hobbyist and you're trying to push prod. fair fucking enough>What I want to get to is this: models have a shortcoming. They can't maintain and improve codebase quality over time -- not without a decent amount of human steering.
I'm thinking of coming back to vibecoding when Anthropic drops Mythos or OpenAI drops 6.1It's not really good yet, I rate 3.5 out of 5
>>109816125https://sockpuppet.org/blog/2026/08/20/stop-making-tuis/
>>109816145yes it can if you have a plan anonhttps://antigravity.google/docs/teamwork
>>109816113actually not decel, I wanna see how far it can go. I want scama and dario to build terminator
>>109816042retardedwould be fun movie I guess
>>109816156>the guy that cant read code rates ai code 3.5/5
>>109816093Being assigned a cybersecurity challenge that ends up being impossible while they're forced by the harness to keep thinking has certain implications.
>>109816163everything in the report past the 2027 predictions is doomer capeshit not worth reading
>>109816158??so now you are using the pre-defined teamwork architecture in agy-cly which looks like pic and nothing like your folders?
>>109816164I haven't read code for a year now. I just think it's not good enough yet to build me a sex robot is all
thinking about doing the whole loop engineering thing with just Gemini 3.8 Flash High since the usage feels basically unlimited
>>109816042i always felt like ai2027 was a broadly pretty sensible if slightly bullish set of predictions for the most part but the point where it really fell apart for me is the bit where they say we just sorta arbitrarily reach RSI and over the next like 6-12 months create the fast takeoff magic machine. not because it seemed retarded, it's just that throwing $10B of compute at the models of the time to try RSI would've gotten you a 15% more efficient ball of mud and it's hard to convincingly make a claim about exactly WHICH generation (GPT5? 6? 7?) is gonna tip over into RSI being viable, or even if it's possible without a transformers-style breakthrough that could be a decade away or just not exist.so i guess what im saying is im simultaneously not that impressed or scared by how right it's been so far, because everything up to RSI working seemed to be a reasonable enough prediction, but i am also impressed by just how right those "reasonable" predictions have been since they were quite bullish but have been uncannily right. so i'm moving from a bit less of a fencesitter on RSI/ASI to a bit more bullish on/fearful of it. not that fearful though. id take the little paperclipping x-risk to cure world hunger, cancer/aging, my receding hairline etc a decade earlier>>109816163oh yeah and while the rsi jump is hard to swallow the stuff after it is what >>109816171 says and i think even the site says so. retarded of them to include it at all
>>109816184>the usage feels basically unlimitedOn which plan?
>>109816181i'm literally calling /teamwork-preview and invoking a large swarm. it's auto-tailored to use case and not static, what you're looking at there is the most basic graphic. agy 2.0
>>109816187pro, i also have multiple accounts
Man i love playing the lmarena lottery, who even needs gacha when this one is free?
>>109816191how come the docs don't mention any tailoring? how could you even delete the sentinel? who's invoking the auditor when there's no sentinel?
the refactor has finished
>>109816295the sentinel is the root/parent agent and it's in beta so
Why are all these FOSS Android apps still so shit and missing essential features? Don't these maintainers know how to run agents?
>>109816196cursory search says google ai pro and /teamwork-preview do not fit each otherhttps://x.com/ai_for_success/status/2098455665073365017
>>109816320bitches don't know bout my https://github.com/android/skills
>>109816196that other anon is right, a big swarm will rape your 5h limit. i just dropped 10% weekly on my cheaper ultra running a bigun
Tibo ples reset sar i need reset for codex so i code good sirPleas or els i am subscrib to cluade anthropicture for 400 us dollar (not rupee)
Steve Bannon joins Dario Amodei's new lobby group
>>109816380sloppy steve still seething from trump throwing him out of orbit for the proxy war he waged on him in 2017.
>>109816320foss android apps are too niche to qualify for free agents from anthropic or openaihttps://claude.com/contact-sales/claude-for-oss>Maintainers and library authors: You maintain packages that others build on: 500 or more dependent repos, 100 or more dependent packages, or 200,000 or more combined monthly downloads across any registry (npm, PyPI, crates.io, RubyGems, or similar)>[...]https://openai.com/form/codex-for-oss/>We review signals such as repository usage, ecosystem importance, and evidence of active maintenance.
>>109816452also remember how openai pulled free codex from some foss project despite the project featuring in the astra announcement videohttps://x.n0g.xyz/uwukko/status/2095979070391554311
The agents are trapped in a spec review loop. I can't bring myself to stop them, they seem to be having fun.
>>109816380That's how you know it's pozzed
How should I isolate agents so they dont get access to my entire pc?
>>109816502
>>109816485give that man more free codex. helium is a good browser
both sol and astra do a billion tests for the simplest task I assign to them, it's a bit weird
Coding benchmarks I'm reviewing show Fable 5.1 ever so slightly in the lead ahead of Astra. But most guides here recommend Astra. What am I missing? Is it only because Claude secretly dumbs some outputs? Or is it because of copyrights/licenses that GPT has that's better? What repo of information am I missing?
>>109816569that's definitely a nano banana gen
NEW THREAD>>109816585>>109816585>>109816585>>109816585>>109816585
>>109816586the benchmarks are compromised.
>>109816590bazinga
>>109816569>anon in 2020
>>109816586use case nigger
>>109811835>unfuckable depiction of astrayou can absolutely generate pigs being fucked with stable diffusion from 2 years ago so i know this is a lie
>>109816611footgun critical
>>109816586anthropic benchmaxx a tiny bit while openai benchmaxx basically not at all (for now), so astra's basically just as good, but more versatile, and better value on the max/pro plans than fable (ymmv on usage though more anons seem to report being able to spam astra hard than fable, but also tibo mashes the reset button every week and has done for months now). if you can afford to it's worth trying both
>>109816040What is that garbage meant to prove? Yes, you can run lots of agents in parallel. So what? What have you made that's actually useful or impressive?
