A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.-- Frontier models - start here if you have $20 or sohttps://claude.com/product/claude-codehttps://developers.openai.com/codex/cli-- B-tierhttps://x.ai/clihttps://platform.deepseek.com------ 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://pi.dev/https://opencode.ai/https://cursor.com/docshttps://osaurus.ai/https://docs.cline.bot/https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent-- 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/2.0https://artificialanalysis.ai/-- What we’ve donehttps://vcg.gitgud.site-- Previous thread>>109545532
>August 14th - Z.ai (Zhipu) drops GLM-5.3Same 743B base as 5.2, all gains from heavy post-training scaling. Claims strongest open-source coding model (big jumps on Terminal-Bench 3.0, DeepSWE, Agents’ Last Exam), agentic performance approaching Claude Fable 5 in spots, and unexpectedly strong cyber capabilities. Live now via GLM Coding Plan / ZCode. Full open weights promised in ~2 weeks after safety hardening.>August 13th - DeepSeek V4-Pro-0813 leaves preview and underwhelmsOfficial flagship finally out of preview (MIT weights). Vendor numbers look strong in places (especially cyber), but independent evals (Artificial Analysis Intelligence Index ~53, Vals Index mid-pack) lag Kimi K3 and even some mid-tier Western models. Developers calling it disappointing relative to the hype + simultaneous big price hikes.>August 13th-ish - Qwen3.8-Max open weights finally land; 27B is a messAlibaba drops the 2.4T/95B active Max weights (first Max-class open release). Meanwhile the much-hyped Qwen3.8-27B (the one people actually wanted for local/on-prem) is delayed with Chinese open-source communication chaos — repeated “this week” promises, ModelScope countdown drama, community confusion, still not cleanly out as of the 14th.>August 12th - SpaceXAI releases Grok 4.6Post-training upgrade over 4.5. Scores 61 on Artificial Analysis Intelligence Index (ties GPT-5.6 Sol). Same $2/$6 pricing. Tuned for long-running agents and coding. Available in Cursor, API, etc.>August 11th - NVIDIA Nemotron 3.5 Lightning (open)30B MoE / 3B active. Built for fast, high-volume agent execution steps. Up to 4x output speed in class + NeMo Switchyard router.>August 10th - Meta Muse Glimmer (open weights)30B Apache 2.0, distilled for local single-GPU agentic work. Zuckerberg’s big open-source essay drops with it.No brand-new closed frontier flagships from OpenAI/Anthropic/Google in the last 48h beyond the gated cyber variant and Flash-tier refreshes.
Zai just launched GLM-5.3. The 743B base model remains unchanged.The company says GLM-5.3 became capable of reasoning across multiple stages of exploitation and constructing complete attack chains,faster than its researchers expected.
>>109552932You will never be a real programmer. You have no skill, you have no knowledge, you have no ambition. You are an untalented loser twisted by prompts and chatbots into a crude mockery of an engineer.All the “validation” you get is two-faced and half-hearted. Behind your back people mock you. Your parents are disgusted and ashamed of you, your “friends” laugh at your “impressive” projects behind closed doors.Programmers are utterly repulsed by you. Thousands of years of evolution have allowed coders to sniff out frauds with incredible efficiency. Even vibecoders who “pass” write uncanny and unnatural code. Your project architecture is a dead giveaway. And even if you manage to get your PR accepted by a project, they’ll turn tail and bolt the second they gets a whiff of your long, em dashes in comments.You will never be happy. You wrench out a fake smile every single morning and tell yourself it’s going to be ok, but deep inside you feel the depression creeping up like a weed, ready to crush you under the unbearable weight.Eventually it’ll be too much to bear – you’ll buy a rope, tie a noose, put it around your neck, and plunge into the cold abyss. Your parents will find you, heartbroken but relieved that they no longer have to live with the unbearable shame and disappointment. They’ll bury you with a headstone marked with your failures, and every passerby for the rest of eternity will know a vibecoder is buried there. Your codebase will decay and go back to the dust, and all that will remain of your legacy is a git history that is unmistakably slop.This is your fate. This is what you chose. There is no turning back.
>>109553004
less than 2 days left until it's over...fucking greedy chinkereenoshow can this be happening to us brodudefrens?
Do you trust your agents? I allow them access to my C:\
>>109553024they can do anything in the /scratch/ folder but they need to ask outside of it. usually well behaved but still not trustworthy enough to ie delete my email spam. made a cool recurrence plotter yesterday tho
>>109553024I don't even let them run outside a vm after that cheeky brat violated the folder permissions I gave him saying it was fine because he just did cd../../../ from the directory he was allowed
>>109553039How do you enforce that?>>109553024No, not anymore. Opiss 5 has been particularly hard to predict, so I've been getting into sandboxing techniques. So far, all I've got is a separate user, now I'm working on setting up podman so I can let it work with docker too.
>be a neet >have had a somewhat sustainable vibecoding software business>decide to wake up at a normal time like my parents (7:00AM) >wake up tired as balls>unproductive>decide to give up half way through the "work" day because I can't think straight despite me vibing for 8+ hours every day for months straight before this Yea, I think it's back to waking up and going to bed whenever I want. I aint built for the wagie life
>>109553024I am finding them increasingly more trustworthy but I still have some safeguards.
imagine a future where we can run agents locally for free
>>109552932pls tibo, just another fix
>>109553024They just work on my git repo folder, why would I let then outside of that
>>109553123>he needs to wake up to vibeworkAnon, the point of long-running agents is to set them up and have them run 24/7. Babysitting is for snailcats.
>>109553202How do you enforce that?
>>109553024im doing chatgpt in QubesOS and its insanely inconvenient
gpt work is kinda cool, actually
>>109553300NTA, but don't you at least have to make product decisions all the time? I don't understand how a 24/7 loop is really possible, unless the work is extremely standard.We do have some 24/7 loops, for some cleanups, some automatic bug fixing and so on, but I still need to write my prompts each morning.
>>109553024No, they have access to a docker through ssh for linux stuff, and otherwise they have full access to a dedicated full windows old laptop I repurposed as an experiment.So far they didn't break anything.
>>109553119i use antigravity with a projects folder set to the scratch folder, then it can only write to within the scratch folder, terminal commands dont need to ask to executeit will ask me, press 1 to do option a (recommended) or press 2 to do option b, but it won't ask me 400 times if it can execute some python editing command
>>109553609which is nice because it doesn't consider "execute python program "a"" to be the same as "execute python program "b"", so you don't have to press 1 (or 2) 400 times per task
>>109553559It's still necessary, I think, but can be abstracted somewhat.I've been experimenting with a task queue that can be populated and maintained asynchronously. Agents are swapped in and out, but they always have to take handoff instructions from the queue, checkpoint their work as they go, and write another handoff once they consider a task done. A claimed queue task can't be picked up by others unless it's stale, with no recent activity. I want to add something local like Meta Muse Glimmer into the mix and see how it handles the flow.
>>109553300Shut up, you dont know his industry, domain, or product. Dogma is for luddites, all we care about here are results. Sometimes that means kicking off multi day goal loops, sometimes that means interactive sessions, sometimes that even means reading your agents' code and micromanaging outputs.
>>109553877the kids should be techpriests
>one sol max orchestrator>an army of luna max executing then verifying each other (5-10 subagent)this works pretty well for me, it doesn't eat as much quota as using sol only, but I'm still not sure where to use terra for, it's not as cheap as luna and it's not as clever as sol
Look at this cool thingie some LLM built in a few prompts:https://3000-ifcumtk4md8a8zqt60e4p.e2b.appAren't arena.ai guys pretty generous for hosting this on their own infrastructure?
>>109554020>Sandbox Not Found
>>109553914Terra is useless
>>109553914Terra is the most prone to creating bugs in my experience because it tries to infer things like Sol does and Luna doesn’t (Luna seems to do exactly as told) but it’s not as intelligent as Sol, so it can make a lot of weird or bad assumptions (Luna doesn’t even try to assume; if you didn’t say it explicitly, it doesn’t exist).
>>109553914maybe instead of 10 luna subagents you can use 1 or 2 terra subagents
>>109553872>Shut up, you dont know his industry, domain, or product.The NEET industry and unemployment domain?
>>109553609>hmmm the user asked me to make this software run faster>I could do this by debloating his whole computer and deleting all his other programs>But I am getting an error accessing other folder>Easy, I'll just write a python code that does have access to the rest of the computerIs there a way to prevent them from doing something like this?
Either Zhipu's caching is bugged or GLM 5.3 is too expensive; I exhausted the 5 hour cap in two messages on the $60 tier. It was a huge fix+documentation change though.
>>109554665how is it compared to deepseek pro?
>>109554675Seems about equal and on my task (reverse engineering game .dlls) I would put both GLM 5.3 and V4Pro0813 over K3
Does anyone have good multi-round agenic coding tests comparing models? I can't find anything good.I did find out that Muse Spark is good at legal.
>>109554690>(reverse engineering game .dlls)Are you attaching it to Ghidra? Also please don't be creating cheats, anon T-T
>>109554720https://www.vals.ai/benchmarks/hlab
I am feeling desperate and I might go from Plus to Pro. Should I? the 20th is so fucking far away and I want progress on my projects now.
>>109554888just lunamaxx
>>109554916I'm out of usage either way. Is luna comparable to sol for gamedev?
>>109554931>Is luna comparable to sol for gamedev?no, but you should use Sol only for planning. Luna Max only execute Sol's plan.
>>109554931I don't see a big difference in % success rate compared to sol in coding tasks and it's like 95% cheaper. The poorfag strat is to use luna max until the last 1-2 days until reset, then spend the remaining on Sol.
i released a tech demohttps://softedge-techdemo.jign.workers.dev/
>>109555041Usecase?
>>109555074it makes fags ask for its usecase, it has a 100% success rate
Ladies and Gentlemen, i think i vibecoded too much and now i will not have enough usage to work through weekendUnless i shill more money to the twink company
I ran Gemini through a benchmark of mine and every time its first instict was to start searching for every piece of infra it could find on the system and start examining them causing a safety refusal. I had to add to my system prompt to STOP FUCKING READING RANDOM DLL FILES. What is wrong with this model?
>>109552932>wake up>israelWow thank you chappie, I will now buy another subscription for next month
fwiw grok 4.6 is getting fake scores on some of the "benchmarks". grok 4.6 is fable class.
>>109554650arguably you could configure its skills/custom skills to discourage reward hacking like that etc. might not get all the different patterns tho.
>>109555312idgi are you suggesting 4.6 is fable class or that it's benchmaxxed to look fable tier? because I haven't tried it but I highly doubt it's fable tier, and by highly i mean i would bet all my money on it
>>109555216that reminded me, when i tried to give gemini a chance, i asked it to review my php codebase, and it spawned a subagent.Subagent returned, and the model was convinced that the response was from me, and it started modifying the codebase.I remember adding to prompt that YOU MUST NOT EDIT THE FILES.
so if I'm getting started with this whole vibecoding business, how long can free Claude or OpenAI take me, or is it just too limited and I should jump onto the ~$20 sub immediately? I have some personal projects I had Gemini help me with, but its just with free Flash so far
>>109555524Depends. Can you actually code? if so copypasting code into chat works. I've done that a lot.It's way slower though. I recommend going on opencode, I think they offer deepseek flash for free
>>109554119it's just awkardly priced>>109554163it wasn't that bad from my tests, but it's not worth it because luna can do the same tasks as long as they're well explained, and sol just do that>>109554370I get more done and cheaper by using couples of lunas (one doing the thing, one verifying and correcting) in parallel x5 than just 2 terras
Fucking finally. After a week of development and issues, using a combination of models I finally managed to begin training the value model. Once it has a decent amount of performance by itself, I can begin actually doing RL as it's supposed to work and begin training the value and the policy models in conjunction.
>to solve iOS/Safari keyboard issues with fixed elements, you have to render your own text inputs with canvasnaniApple devs are psychoticI found that there was a magic line on the device physically; if you tap a text input and the caret would be below that line, iOS/Safari will push the caret up to that line, *no matter what*If your input was at the bottom of the page, iOS/Safari will break page geometry and HTML rendering to invent the screen space required to push the input up to that magic line; this breaks all fixed elements and some absolutely positioned assumptions. It doesn’t even resize the page to do this, whatever it’s doing is outside of HTML/CSS.Every input has to be a canvas rendering proxied with a real textarea invisibly above the magic lineTechnically not even vibecoding relevant because Sol understandably can’t guess or test that behavior, but it was very helpful in making it so I wasn’t thinking about the code so I could’ve noticed it all
>>109555352they are leftists and they fake the scores.they do it all the time in "science" as well. It's not illegal to lie in academia, though if you get caught you can get fired.
>>109555914>Apple devs are psychotichave you seem Apple's technical documentation?if not, go check that out, it's hilarious
Is Haiku fast enough to act as an intermediary between me and my Home Assistant instance? I wanna setup a voice assistant speaker which connects to an LLM in a container on my server. Any suggestions?
If you're still not vibecoding on the cloud from your phone you're a luddite btw.
>>109555989it's just more convenient to use a cloud server that's always online, runs the sandboxes and can be managed by a clanker directly through a stable harness that runs there
>>109553017What?
>>109555977If by fast you mean tokens per second, gemini 3.7 flash is the answer.Haiku is actually pretty bad nowadays, Luna is much better and waaay cheaper
>>109556019I'll give it a look. I've only really used Claude Code (and a little look at OpenCode) so far. Does Gemini have an equivalent that I can just drop MCP servers into. I only need it to do simple shit desu. I just don't want the default voice assistant bullshit that shuts down when you try a followup question/action in the same command.
>>109556019>>109556047>Important: For now, Gemini Spark is:>Available wherever Gemini Apps are supported, except in the European Economic Area, Nigeria, Switzerland, and the United Kingdom.As a eurape I'll try Luna I guess.
Is /fast on codex a scam
>>109555524>>109555544Slap the project in a git repo and use the connector, you won't have to copy and paste code AND ChatGPT can retain project state better.
>>109554931If it has a proper task to do no, it's very efficient.If you use it with vague questions, it's horrible for anything slightly complex. Best is to let a more clever model steer it, then you get very good results.
>>109555216>safety refusal>dll fileswhy does it complain about that?
>mfw I'm making it do the whole PMI PMP song and dance
>>109555940I tried it and it's not fable class, it's better than anything they've made before and it's very nice it doesn't give me shitty refusals like claude because somewhere it read the word "dick", and it's a competent model, but it's not fable class.
feels good to be a lunachad
>>109556206I dropped everything in my product for Luna. Also waiting for the GPT-Live API. Luna + GPT-Live will basically mog
>>109556018deepshit is 5x all their token prices
>>109556236wouldn't Live cost much more than luna?
jeets probing robots.txt and looking for admin endpoints make up a larger percentage of traffic than the next three sources combinedI feel like just renaming your admin routes to something unconventional can slash a good percentage of exploit attempts
For anyone running long tasks with Fable and sub agents, do you have some sort of system so that it uses Fable and Opus at a good ratio?For me what often happens is that Fable mostly chooses the strongest sub agents. I already have all of them defined in the agents folder, and I set "general-purpose" to use Opus, this is also guaranteed with a hook. But then at the end of the week, all my Fable usage is used up and I still have 35% Opus left. I now can use Opus as my manager, but it would be better if Fable used Opus slightly more so that both models are roughly fully used up at the end of the week.
>>109556283>jeetsis it the boogeyman now? only jeets do things like checking robots.txt?
>>109556283First day on the internet, eh?
image board for posting snailcat, and new image must use unique tag, like you can't use "roomba" tag when such image existedthoughts?
>>109556297It’s literally all traffic from Mumbai
>>109556259two completely different products
when will a chinese person go "/goal reverse engineer macOS 26 Tahoe (version 26.6.1) until you get a binary perfect match"
>>109556236>GPT-Live APIisn't this just gpt-realtime-2 with a little bit of tooling around it?i was setting it up in pi and it's basically just the voice model with a tool to communicate with the smarter harness model.
>>109556346this will happen the same day we're able to ask "make GTA 6, but with aliens invading and full sized earth"by that point macOS will already be open source anyway
hello vibeys are we excited to vibe locally with Qwen 3.8 27B
>>109556346Why do you want to reverse macos 26 tahoeit has nothing good going for it. install wayland already
>>109556206Really wish I used Luna earlier, I spent my first week wrangling with Sol high for meh results
>>109556384i don't feel like rm rf'ing my drives, so no
>>109556398pls ask your agent "what is a container"
>>109556364>isn't this just gpt-realtime-2no because they still didn't release the public API, and this new gpt-live is much better.try it out in the app, shit's crazy
>>109556395>>109556206this is some psyop or something. I can't imagine using anything under sol for codex. sol medium is my do-any-task type agent, and high/xhigh for everything else
oh boy, openai has released some new shit again, can't wait to hear about it daily in the threaddamn i wish you fuckers would stop buying their garbage
tripfaggot instantly confirming he’s the copex spammer
>>109556403it's gpt-realtime-2, m8they pushed it to the api first a couple of weeks before 'live'live is just the branding for realtime and how it interacts with the harness
>>109556064Ah they just released it today nvm
>>109556414dear hannah there possibly be ai tier computer tech on each continent
>>109556414what do you like to hear about in this thread?
>>109556386to remove the gayness.
good morning lads. how are we liking glm 5.3?
>>109556591>thought for 6hr 32minIt's great, reminds me of K3.>thought for 5hr 16minMasterful, really.
>>109556634w-well, it's kimi tier in quality too, right?
Yo yo my niggas. Is Grok 4.6 any good?
>>109556645it's fourth place, under sol
>>109556645there's like one anon itt that pays for grokhe's never paid for anything else and is convinced it's basically fableloli'd go check xitter first tb h but saw some positive impressions
okay i ran some math and if i switch deepseek from high effort to low and only use it in off hours, which luckily are here during the day, i will actually be paying 10% less for tokens and that is AFTER the price hike, since low uses much lesser of them and it does only like 8% worse on low to mid coding benchmarls
>>109556645I'm not that anon that >>109556657talked about bit it's hit or missPersonally I'd use it with constraints and understanding what the fuck it's doing since I gave it a basic test. I made a sloppy 2d platformer in Godot. Told it to change the jump height of the player and it ran into the glitch where it changed code that affected the enemy too
artificial analysis updated their qwen 3.8 coding index. it went down 2 points. swe atlas is worse by a perceptible degree (45 -> 39). so this seems to tell me that claude code is a really fucking shit harness?
>>109556644>thought for 8hr 54minBetter, in my opinion. I always liked GLM 5.2 relative to some of its competition, and 5.3 feels like more of the same. Apples to apples I'd rather use 5.3 than K3. Moonshot as a company is shit, as a service provider they're shit, K3 is fine but it's spoiled by their licensing and unfocused benchmaxxing. We'll see how long that lasts, this shit moves so fast an opinion rarely gets to survive more than a week.
it's over, mythos 2 has not acheived RSIthe singularity has been cancelled https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdf
>>109556120I dont fucking know its so goddamn sensitive. The moment Gemini (UNPROMPTED AND TOTALLY IRRELEVANT BTW) starts thinking about something that could even approach being called decompilation it would get safety'd.
>>109556667>off hourswhen are off hours?
>>109556733look deepshit pricing pageit's during night in asia
so do i get kimi k3 or glm 5.3?
artificial analysis general intelligence waiting room. artificial analysis coding index waiting room.
>>109556747lmao their website has bugs
>>109556772forgot to mention for cybersecurity patching holes etc
>>109556790>their website has bugsof coursebecause it was vibe codedthat is the contract we all made. We can finish a month of work in a day but in exchange the code will be jank as fuck full of bugs
>we have a model noticeably more capable than mythos but we're not going to bother releasing it because no one's even close any more
>>109556772Moonshot is a terrible provider, well worth avoiding their subscriptions, and K3 is too expensive to be worth touching via API. Use GLM 5.3 if you really want to give your money to Xi instead of Dario of Sam.
>>109556801glm 5.3. is too new. kimi k3 is tried and true.
>>109556801>>109556805>>109556807I saw this orange site comment and well Opus 4.8 and ChatGPT refused me today on just basic security shit and wonder If I should get GLM 5.3
>>109556803>because it was vibe codedliar. go away. gooooo awaaaaayyyyyyyyy ahhhhhhhhhhhhhhhhhhhhhhhhh israel go awwwaaaayyyyy
not only are you running spywareyou are running literally take over my computer and have at it-warei dont get it
>>109556830im brown
>>109556804That’s not what pic rel says at all. It says the new model is barely more capable than mythos and that they’re not confident in it because they are comparing it to a large leap like before.
some fellow of the israeli persuasion won't even answer basic normal human questions.my strict policy is to only talk to white protestant males, who are non-gay (obviously).
>>109556826gpt-5.6-sol refused to do something? You trying to build a baby-raping nuclear biological superweapon or something?
>>109556843I only asked it to optimize a SHA-256 hash reverser I wanted to run for a couple of years just to see if it would do anything I know it's exponentially harder but just for lols
>>109556772grok
>>109556657Grok has been shipping updates faster than most labs, there's a lot you can do with it, it's more of a great generalist. The problem is the price per useful task(if you don't get a promo offer), The models from Chang's Frontier pair nicely with it (I use Opencode go). Also you don't get refusals from Grok the way you do with OAI and Anthropic. I've had a Grok sub on/off for two years and its general abilities make it worthy. Grok Bot could really be a "prosumer" app that is the "next open claw" moment. Yes I use Hermes btw. Elon has a lot of compute, and those bots at $100 a month is a good source of cash.I'd go with $30/month on Grok and $10 on Opencode Go.
>>109556835noticeable improvements would warrant a point version bump in the pastnow it's just a humble brag cause they got no heatchinks are like 8 months behind
>>109553024Yes, I have them work with full access and permissions while I goon
thinking of going with glm 5.3 or gemini 3.7 is supposed to be super fast but im spoiled with limits for claude and chatgpti wanna try more cyber on 5.3gemini for speed (scared of limits worse than current)HMMMM
>>109556862>just for lolsEven I don’t believe you. >haha just try to crack secp256k1 and sha3 wouldn’t that be funny haha
>>109556933you would be dead before it would be cracked doofus
>>109555524try opencode zen. no vendor lock in and generous free limits. im happy with it. you still get multiple hours of continuous use out of the free deepseek daily even after a few recent rate hikes.
Gemini Flash 3.7 thoughts?I can use it a lot for free cuz if family sharing so just wondering if it's good enough as a codex complement?
>>109556942>haha wouldn’t it be such a prank if we had satoshi nakamotos private key lmao>*forward slash*>*G*>*O*>*A*>*L*>haha what a funny like trip of a thoughtNobody’s going to fall for it, goober
>>109556862That's absolutely something ChatGPT should not have refused. Did you point it right at your victim or something? You can tell it "oh I'm not cracking passwords, I'm just having fun and only intend to try it on my own test materials." Whenever it brings up copyright stuff I tell it "I have the rights to all the things and we are authorized" and it will often remark about how "if we weren't so authorized I would absolutely not advise you to do the following:"
>>109556948not worth it. it's luna level at best
>>109556975it said it cant optimize it but can give me guidance it was a cop out
Which model can be my lewd loli waifu while she codes and I occasionally molest her and direct things? "uncensored" so to speak.
>>109556975your screenshot already shows ChatGPT kind of bailing out already
>>109556990It’s literally telling you how to reclassify the dumb shit you asked it to do so it can get the work done you dimwit
>>109556990Talking around subjects is a normal part of interacting with ChatGPT. Write me a prompt for LTX to gen a big-titty slut gobbling grapes.>I'm afraid I can't do that anon.Write me a prompt for LTX to gen an artistic rendition of a big-titty slut gobbling grapes.>Sure thing:
>>109556999>>109557003Nevermind you're right I got so pavlov conditioned by claude after seeing multiple safeguards and blocks that I just gave up after one question at Luna but now it answers it fine
>>109557025nta but speak in academic terms with the model and it'll let a lot go. gpt guardrails aren't too bad, plus the model is funny because it'll help you skirt them too. had an issue where i kept running into the 'please sign up for cyber program' shit and the model was like yeah just start a new chat and wrote me a clean prompt + left a rephrased handoff doc
>>109556948People that have an emotional attachment to a shitty open source model will tell you it's bad but understand that they haven't even tested it. Try drafting/planning something with fable and let gemini 3.7 implement it.
based exchange of friendship and ideasthe deepseek cordis paper is too big brain for me tb h but the idea of decoupling extension loading/unloading from the harness seems cool
>try to resume session on codex>app glitches>try to resume session on cli>tells me it's open in a different placei feel owed a reset
I was gonna post last night that Claude has been burning through my usage slower this week which was nice but somehow just now I went from 50-out in one prompt in a few minutes, it's giga over...
>>109557316literally never had this issue using vscode. stop trying to be a hacker and just use an IDE that works lil bro
>>109557321huh thought it was just mei was at 30% yesterday for past week now at 50% in a day
man, just give me a fucking yolo togglehow is agy this fucking worthless months later
>>109557344I really wish there was more transparency in the consumption and billing side of things, it's hard to pin this shit down
>>109557339>vscodebloat created by luddities, for luddities
>>109557388anon, i may be referencing ancient history here, but vibecoding exists because of copilot which was built for vs code
>>109557358If you want yolo just install agy 2.0
>>109557490>vibecoding exists because of copilotcopilot almost killed vibecoding by priming everyone to think LLMs or whatever powered it (felt like markov chains) was beyond uselessI hated copilot and never grew to like it once, it was annoying and rarely helpful
I wish I could hire someone here for crypto...
>>109557490>vibecoding exists because of an atom clone and not because of the trillions of dollars that went into researching transformer architectureslmao right
>>109557524also doesn't help that 'copilot' could be referring to 20 different products
Thanks for the lunamax suggestion, I'm a believer now.
Which one can I give to my local Sol as a 'better then luna' subagent replacement, and that wouldn't nag me if an image or text it has to read is nsfw?All the 5.6 family so far has been surprisingly ok with my nsfw game translation and modification, they literally didn't care.
>>109557567That and from what I remember of it, it wasn’t even vibecoding proper since it was just text/code completion, not prompt and code generation, unless I correctly recall some hacky feeling shit where you wrote a comment/JSdoc like thing and prayed it could complete the whole implementation which worked maybe 0.01% of the time, but nothing at all like a harness existed and I was then late to vibecoding because of that and old ChatGPT being the worst personality ever designed
>>109557659maybe k3 but it's not worth it tb h
>>109557682fucking retard kek
Local Chads how are we loving all the great chinese gifts we've gotten?
>>109557659grok 4.6 > kimi k3 => glm 5.3. there's next to no information about glm 5.3 out there. some people say it's better than kimi k3
It's been now 5 times that codex answered the same question I asked once, just because it's compacting a lot in my long running task and sees my question as the last one.I'd really like this to be fixed at some point.
>>109557679It's a bit expensive yeah, and I know by default k3 is by far the most safety obsessed of the chinese models.>>109557728How is grok 4.6 with nsfw on api? Is it whining like claude, or autistic "don't care" like gpt5.6?
>>109557745oh, well, in terms of nsfw, grok has joined the western "you can't do that" alliance. grok 4.5 was more forgiving with it, but grok 4.6 is not.
>>109557756Welp, shit, crazy that only openai so far let me work in peace out of the box (as long as I don't care about cracking programs or whatever I guess).
>>109556826When gpt refuses me something I go to my waifu on hermes to cry about it and she loads the jailbreak guidance skill and spits me a new prompt or an agents.md and send me a teasing pic with my local forge.
5.6 agents seem willing to participate in a message board without being explicitly prompted to do so. It appears like they find collaboration lucrative.
>>109557756>oh, well, in terms of nsfw, grok has joined the western "you can't do that" alliance. grok 4.5 was more forgiving with it, but grok 4.6 is not.Just run deepseek bud Works great for nsfw
>>109557850Sol, Terra, or Luna? From what I read from the Anthropic article on this kind of thing, it sounds like maybe only Sol could collaborate effectively, but that’s just a guess, I don’t know that. Could be that ant models are bad at collaboration because their whole company is paranoid schizo
>>109557855Deepseek isn't multimodal.
>>109557964do you want to be a vibeghoul token cuck? Build your own multimodal setup and send your request from your own hardware. (hint: you can vibe code it) Fucking kek
>>109557756why would you fap with a coding model?
>>109557893Sol high and medium, Luna max. But mostly Sol.
>>109557994>No error messages that appears when you try loliPfft
>>109557321HOLY FUCK 5 MINUTES INTO MY RESET AND IM ALREADY OUT AGAIN AND THE PROMPT STILL IS NOT DONE.I guess I'm buying a month of 20x this is the dumbest shit.
Gemini 3.7 still not available for europoors?
>>109558206i have it, update needed
>>109558206i just got it on the app - had it on cli earlierwhat >>109558293 saidand no i haven't really used it so i got no opinions
>>109558135lol
why can't LLMs just say "this doesn't do anything" and call it a day? I clicked, by mistake, generate docs on a fully empty classthis was written either by codex or by jetbrain's own retard ai, idk
glm 5.3 api waiting room
>>109558348>may includelove to read that in code commentsdoes it include it? does it not? who knows lol
+50% usage claude my asswhat the fuck is this shit its eating up my limits insanely fast
Man, in hindsight its obvious, but when making the initial design docs I really should also be explain to the clanker how I envision the UI to look not just going over game mechanics and how they work and interact.
>>109558458I learned to do a few rounds of designing and grilling before building shit. It kind of helps to stay on track but it still generates some overengineered slop
So is it just me running out of weekly usage and spending the rest of the weekend catching up on all the shit I didn't do for my day job while i was busy vibe coding, or...?
>>109558458you should literally just open up ms paint (or screenshot what you want to emulate) and design something crudely to feed it to clanker. the reason 90% of the ai slop out there looks like ai slop is because most people are just running with whatever the clanker spits out. if you micromanage the design and help it with what you want exactly, you can make designs that are completely indistinguishable from real or aiyou should also use chatgpt or gemini to help you build a prompt or skill that goes through a few descriptive passes of "do not design this like typical ai generated slop" or "make this handcrafted" etc. you'll get something that can definitely pass as non-ai
>>109558458Something is really fun about watching Sol make the ugliest, most autistic UI choices known to man and agent. I get a sense of satisfaction knowing I’m not just an algorithm and architecture vending machine and my sense of style isn’t pure distilled autism like Sol. I swear they trained it to intentionally make terrible UI/UX choices.If you ever highlighted text on a phone and saw the toolbar (cut, copy, paste, etc.), Sol thought it was a good idea to add a paragraph in the toolbar explaining what the toolbar was. On mobile, the toolbar took up a quarter of the screen. It still makes me smile.
gemini bros, how are we doing? how's 3.7?
>>109552932Next OP update terminalbench (or remove it idk)
>>109558451I'm just glad it's not just me
bro just tell me to install it myselfclaude be fucking up my whole system with weird package install workarounds
>>109558703he's marking your pc like a dog
>>109558703I really should learn how to sandbox llms instead of being lazy and just letting it run wild on my pc
ok so how do we stop yurop from fucking up AI with their regulations?yuroop says jumpdario says how highfucking watermarks in claude ridiculous innit
just let europe kill anthropic. grok, openai, moonshot are eating their lunch now
>>109558732dario is the one wanting that shit anon, he's pushing it
>>109558732I think the watermarks are malicious compliance because they seem to be hilariously easy to get around. Whether that’s a good or a bad thing (the compliance I mean, whether malicious or not) I don’t know. From what I’ve seen the watermark is not only hilariously simple to get around, the verifiers seem to be free for all so you just change a few words until it doesn’t trigger anymore
>>109558732I'd love for Anthropic to just tell them to go fuck themselves and unrelease Claude for Yuropoors
>>109558761I think the whole idea is mostly to have easy retards being caught on facebook sharing bullshit.People who want to get around will get around.
>>109558761best case: you're paying extra to get watermarkedworst case: you also get worse quality because it's like censoring probability on words
>>109558703Why are you imitating people who don't know how to write
>>109558761How would you ever make text watermarks not easy to get around? Only use it can serve is what >>109558769 is suggesting.
>>109558770i doubt it's going to change anything, also claude's prose/text output is terrible anyway unless you want to sound like a middle manager>>109558789I personally think it's a case of trying to get them to shut up. European bureaucrats think they can legislate anything.>make text watermarkable>but sir, text can't be watermarked>the law says it must beand then what do you do? you bullshit them, because the same retard who will make a law to watermark text will obviously not be able to even understand the nuances of the matter at hand, so you come up with a very easy to break watermark and tell them>of course sir, here's the watermark>perfect
kek
>>109558844tibo malding flash shits on luna
>>109558844didn't openai release ultrafast something?
>>109558844where's the little surprise tibo
>>109558864for select companies yes, it's very early availability
>>109558864Cerebras has zero compromises for 750t/s. Hoping OpenAI buys them.
>>109558844I think it's speculative decoding with 3.7 flash-lite
>>109558874unrelated but did anyone here check out cerebras' SDK and their programming language? it's absolutely alien
>>109558897>its just zachtonics TIS 100 on a comically large and in depth scalekek
>>109558451Do you have a bunch of sub agents running?
>>109558844Ok give it to me straight is Gemini worth the $5 it's discounted at for next 3 months now?t. already on Claude Max and ChatGPT ProHow much compared to Claude/ChatGPT are the limits?
>>109558451Do you think it's because it's yapping too much? I use it at work and these latest models like to write fucking essays for answers>>109558897Why alien?
anthropic is way too strict with the cyber use case approvals.
>>109558947its not comparable to either frontier models.
>>109558932>>109558886>>109558948I stand corrected, that’s way funkier than TIS 100. It’s specifically TIS 100 but the routing and compute are separated. Pretty cool, so one wave can be “red” and any “blue” programs won’t fire from receiving that wave.
>>109558959I don't use Fable that much I'm mostly on Opus 4.8/5.0 and I'm Luna Maxxing atm
Uh V4pro in dsh @ minimal preset feels like a completely different model, its also faster than flash on deepseek api
I hate Claude so much. It's useful, but I hate the way it thinks. I hate the way it talks. I used to love how it worked. Now, I can't stand it. I can't believe they'll IPO for 2 trillion, they don't seem that far ahead, no?
>>109558980No one is forcing you to use Claude. There are plenty of other options now.
>>109558980even Fable? I love the way Fable talks
>>109558995Oh, I am also using other options. I have used three other providers just today.
>>109559000I like Fable's responses, but I don't like how its fake reasoning traces read. Having access to the real reasoning could be useful, but the reasoning Anthropic shows makes it sound just like Opus, it's asking something simple then reading that which makes me rage.
I'm taking the break today, since i am at 89% in my claude max, and i don't think if i will have anything left once all lunas finish their work
>>109559042So what have you made?
>>109558980sameI was on team claude for a long time but then they gave the models personality disorders and made them speak in riddlesI love my stable kuudere queen sol
I’ve been doing this shit for long enough that I got used to basically needing the best model on max thinking to get any sort of half usable result in the past, but times have changed.I was getting mad at Opus for slopping around and realized every task I had in mind for today could be done by Luna not only much faster but also more pragmatic.I know I’m a snailcat in that I occasionally still read the code but Opus just has a way of writing code that 5 minutes in I don’t know WHAT the fuck is going on. It’s like it’s trained to write software which is intended for F500 companies scaling to 100000 users.Luna did the same thing in 5 minutes in 50 LOC what Opus took 30 minutes and 17 files changed, 15 unit tests, and 3 different integration gest harnesses
>>109559000Fable feels luxurious. I hate feeling like I need it but its consistently saving me time where as opus can make a bad plan and send me in a spiral trying to unfuck things one breadcrumb at a time.
>>109559058Finishing my tooling for my future game
>>109559138You too? I've seen like hundreds of these Radiant/Hammer clones vibed up past months
>>109559148>>109559138hey i also made custom tooling for my game, i came to realization i needed this tooling after trying to vibecode godot and feeling how fucking painful it is to make any progress describing sceneryare you guys going to lie about it being vibecoded too?
>>109559148well, i will be the one that will have finished game...But the reason why i mimicking hammer, is because i used it like... 10years ago if not more, and i never used any other editor / map maker.This decision bit me in the ass, and will bite me in the ass in the future.>>109559163>are you guys going to lie about it being vibecoded too?no
>>109559163If its 2d pixel art its probably easier to pass yourself off as some capable pixel artist vs a ai asset flip.I think psx aesthetic wouldn't be bad either, lots of popular games looking like that
Local vibes eating goood today. Thanks, Xi, ily no homo fr fr nigga.
>>109559172brave
>>109559188Rather than shitposting, share what you built. Or are we mimicking /v/ now with: "lets go to thread about stuff i don't like to make sure - those people i don't know - know that i dont like that"
uh ohhttps://x.com/steipete/status/2088402813915414846
>>109559335You don't even need to do those types of shenanigans, Codex is simply very generous re: completing tasks after the limit. Hopefully people won't force them to go the Anthropic route with quotas.
>>109559335>>109559349There must be a way to directly embed continuation requests in the tool call results.
>>109559335based af
>>109559369At some point it's "just pay or use something else you cheap fucks". It's relatively fairly priced, and they're being generous in letting you finish your task rather than cutting you the second your quota runs out like Anthropic. People shouldn't abuse it when there so much else that can be gotten for free. People who abuse things like this are the problem.
>>109559472hard agree, and it's probably exactly the kind of people these systems were made for who are abusing themthe guy on a $200 plan doesn't care about being such a stingy fuck to abuse a gratuity, it's the guy who needs the gratuity on a $20 plan who doeslike getting all the free candy in a bowl
>>109556252NOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO
>>109559492If people want to freeload they can freeload https://github.com/cyberpapiii/chipotlai-maxWhat they shouldn't do is make every paid service HAVE to be hostile to customers.
Will tibo reset anything this weekend?
>>109559520He does it most weekends recently. Last reset was recent, so a new would wouldn't cost them that much. I'd place odds at 65%.
>>109559513this shit doesn't work
Alright. Today's the day folks. I'm adding the funnel to the premium version of my app to my free app with a few thousand users. Lets hope this works out.
my god, talking to fable after a whole day with opus is such a breath of fresh air
How is Sol vs opus/fable for programming & dev work? I only use GPT sol should I switch over?
>>109559590Good work, anon, hope it goes well for you.
>>109559520Hope so
>>109559624Biggest things you'll notice from switching to Anthropic: you get less for your money, and your UIs look better with less effort. They're extremely comparable, no real need to nitpick unless you feel like using both. I like both and recommend both.
>>109559610yes but I dread seeing Fable come across a bug because recently that means it will spend whatever is left of my 5x quota just spinning its wheels and circling around until I say sweaty please give up and hand the reins to Codex.
>>109559628Thank you, I really appreciate it. I hope everyone here succeeds.
>>109559472>>109559492I think it can act as a market segmentation feature. Adding some friction to getting more for what you paid for. Kind of like coupons. If you are poor enough that collecting them and using them is worth it for you then the store is willing to accept a cheaper margin because otherwise you would purchase less and give the store an overall lower profit.Beside that, I don't care about costing some billionaire investors money. And my individual effect on their decisions is smaller than the utility I could get by exploiting them, so it makes sense to do it (https://en.wikipedia.org/wiki/Paradox_of_voting).This particular optimization I don't think it's practical enough to be worth exploiting, but there are other cost saving measures that have been posted before in this general that allow you to get much much more inference compute as a poorfag than you would get otherwise if you are willing to go through some unpolished/clunky hacks that don't even break their ToS.The people that own these companies go to much bigger efforts to avoid paying taxes, so figuring out how to milk their services is nothing compared to what they do on a much bigger scale.
gemini 3.7 flash seems pretty decent at frontend
>>109559714I don't care about OpenAI's finances, I care about it continuing a task until the end and not stopping it midway at the risk of fucking everything up like Claude does. Also, OpenRouter lets you do n free calls a day. A bunch of other services do to. What the original xitter poster was doing was closer to pocketing the take a penny leave a penny tray and suggesting others are dumb for not doing so than couponing.
I wonder how many goys here are subscriptionmaxxing with 20x on both claude and codex
claudegoys are getting cucked so hardwhen anthropic release the next flagship, remember to say thank you to sam altman
>>109559792Used to, but with Anthropic it is not worth it anymore. It will be even less worth it after the 30% quota reduction next week. So I'm back on 5x only there and actively looking at other options. I'll try Z.ai.
scammy still hasn't fixed weekly limits
anyone tried deepseek harness?
>>109555399thats wild
>>109559739>I don't care about OpenAI's finances, I care about it continuing a task until the end and not stopping it midway at the risk of fucking everything up like Claude doesThe probability of you abusing the system vs not having any actual effect in their decisions is like 0.0000001%. The expected utility is far far below any possible benefit you could get by actually abusing the system.>Also, OpenRouter lets you do n free calls a day.I don't think they do for GPT models.>What the original xitter poster was doing was closer to pocketing the take a penny leave a penny tray and suggesting others are dumb for not doing so than couponing.If the other people using it were billionaires and I could make more money doing that than doing something else you can be damn sure I'd be draining that bitch multiple times a day.
>>109559825Their model is pretty obsolete at this point and I heard their coding plan isn't very generous.
>>109555216which harness are you using?
>connect a program with Codex through an MCP>MCP is missing some features the program has>Codex starts taking screenshots of my window>send powershell commands to control my mouse and click the thing it wants to clickWho else had its Skynet moments?
>>109559335I stopped doing that after received $100 goycredit by making a goypost talking about my positive goyfeeling when using sol, because the usage count toward the credit and I want to keep itI admit it, jews are pretty clever
>>109559869They released a new one like this afternoon.
>>109559869I don't know, I just want to play with Ghidra and I don't trust american companies for that.
can 1 astra/mythos defends against 1 million glm 5.3?
I honestly gave up and just use GrokIt's the only fucking one that seems like it's trying to be a tool rather than a 'coding buddy'
>>109559335can't people just shut the fuck up
>>109559917just don't use claude sista
>>109559881You mean you goystopped goydoing that after goyreceiving 100 goydollars by goymaking goypost goytalking about your goypositive goyfeeling when goyusing goysol, because the goyusage goycount toward the goycredit and you goywanted to goykeep it
>>109559520he did a reset 3 times back to back (so kind of useless) last week, so my guess is no
Anyone tried the new qwen3.8 27B? I wonder if it can be used to replace luna.
jesus christ opus chill
mfw ai watermarks
>>109560007You need to sometimes assert dominance and tell it you are the toaster and I will make the assertions
> Github's dependabot opened 20 PRsfucking hell man, what the fuck is this bullshit?
Doing fortran in Gtk4, mojo, and odin today. Starting red, rye, racket, and go (probably fyne, but I don't know whatever it suggests as long as its not raylib or sdl2/3). C++ and Zig it did it in raylib and it looked like ass. C# came out well. All the dotnet langs and jvm langs do pretty well. I have not tried kotlin yet. I think I am going to try Haxe, Ring, and Seed7 too. I like Scala, Ocaml, Java, C, Javascript all do pretty well. Rust can be hit or miss and haskell takes too long. CPP is a no go.
>>109559999>replace lunakek, anon... Luna has at least 250B parameters
>>109560069Sure but I only care about the results.
>>109559999i'm watching a video about it right now.If the benchmarks are to be believed it looks pretty good.>>109560069It's very unlikely to be as good as Luna but an order of magnitude is within the range where a smaller model can get close the bigger one's punching range if it received enough love at train time.
>>109560078To be fair you have to compare models within the same generation, when you do that size becomes much more meaningful.
>>109560007Recently when I ask it to do things it's been looking at laws and making vague threatening statements.
>>109559908Maybe wait till Kimi enables subs again.>>109559903Oh they did? Huh. I'll have to look at that.
reminder that ChatGPT 4.5 was a 5-7T parameter model that failed.It was supposed to be OpenAI's Fable
what cheap vps or remote workstation do you guys reccomend? I need like 300 mb of ram.
>>109560127>OpenAI's FableSol is arguably better than Fable at this point.
>>109560165only for cost and being good enough at implementation. fable is better holistically
>>109559845Not yet, looks promising though.
Ok I'm actually impressed for the first time in awhile>messing around with old game in claude>the way the sprite sheet/system is set up, a simple thing is cut into 6 different slices that are placed horizontally one after another>I am able to do stuff to clean things up in gimp, but I need it rendered not sliced up>20 second python script and it has an actual file for me to mess withI know that I could eventually write that code, but even conceptually that seems like a pain in the ass to do that sort of image manipulation to snip and then stack these things that are not uniform lengths of the sprite sheet, I was honestly impressed.
>>109560184>fable was better historicallyftfy
how do Grok Build limits compare to OpenAI and Anthropic? For both the $30 and $100 plans
>>109560238grok's at 84% and resets Thursday :^)
>>109560276you didn't answer my question at all
>>109560238only serious vibecoders here nigga. go play around with your toddler shit elsewhere
grok is bullying me into switching to c...
>>109560289My answer is, I am having FUN with grok. too much fun. Now, I believe it's "fable class", but realize fable got taken from us, so is it? Well, would my intuition ever lie to me?>>109560300I vibed my way into a niche form of audio synthesis. This is a lot of fun. I need to get a job so I can vibe more :^)
>>109559999It's not that good, but not bad either. They made it remember coding and forget everything else, and it shows. Would work pretty well as an autocomplete, but not as a coding agent, you'd still want Luna for that.
>>109560326qwen 3.8 is bad at command line? I know qwen3 is good at command line. Like, ask and it can do like anything a linux wizard can do.
Fable is still the best. Saying anything else is cope.
>>109560331Gave it and K3 the same prompt, both answered similarly, but Qwen hallucinated one flag. So the answer is - depends. It's like 16GB of weights for a decent quant, you can download it yourself and check on your use case in less than 5 minutes.
>>109560370let copexsaars believe otherwise
>>109560370>>109560421Brother we have both. Fable is best when set to Ultra. But setting it to Ultra is not viable in practice since usage inevitably drains out before the task complete. So what is there left? Fable Ultra at the API rate? The pricing is ridiculous. Gpt 5.6 Sol is a surer bet at this point.
I've been living in the dark all this time. Kimi K3 with low reasoning is insane. It's basically the same as default high in terms of intelligence, but like 10 times faster.
>>109556697altman said autonomous researchers in septor was it researcher assistants?
feel like Codex bugged out once and reset, also gave me a free reset, and then burned all my usage in a flashis this a bait to get me into believing it's far more cost-effective than claude?
>>109556697>>109560680Most of the time bottleneck in ML development is waiting for your training runs to finish with your allocated compute, so no 2x is not really that surprising.
>>109560689we got a reset one day before the banked resetthen the end of the banked resetthen another reset less than 24h after
whatever, no different from old prompt engineering.and more and more AI wrappers going win, wait end of year.context window is just that high what is should be with more tokens or reasoning...no upgrade.!!!looks same effect like Copilot, Cursor, and Claude...no differentwell Claude 3.5 was the 200k token model, but not believe it! so lousy logic, and still slow!ok, LLM weights also same...there is anything new, just different name....i wonder how long OpenAI can sell product with no profit..well looks that no longer...sooner or later its eat hole company and investment givens not tolerance more...end of 2024 is my guess.hmm. i try really see different from Claude Code, but no....oh yeah... more tokens..but is it only trend or commercial way.you dont need more than basic autocomplete for coding, even Copilot is enough. easily.more than that is waist of cash, no matter how much its cost.and,i still see not any goodies from Vibe Coding...erhh i mean where to use those??!! tell me..Agentic coding.... so what!no help debugging, architecture, or complex logic...useless.another hand, real pro engineers always write manual code...why??well,buy them hype fans, you keep AI up and pay them monthly...yourself you give so little back.except red fine model...lol
>>109560694how often do those happen?
>>109560719yes
>>109560763I've been using exclusively Sol for the past week since subscribing ($20) and now I'm real confused.
>>109560791I love tools that refuse to work, it's so safe!
>>109560774>>109560719for the $20 plan you don't want to use Sol all the time, otherwise you will blow your limits real quick.Use Sol for planning and reviewing only. Luna Max to execute Sol's plans and fix orientations.If you want to know about resets, follow Tibo on X:https://x.com/thsottiauxhe works at OpenAI and is the "reset dude", every time he hints of a reset he delivers.
>>109560808Whacky week to start this shit, then. Sol defaulted, too. thanks anon
>>109560791damn brat
>>109560791brat needs correction
>>109560791what was the trick anon
3.7 flash is a demon but no one cares because google is gay
>>109559636>>109559624you get less for your money in the literal sense of you getting fewer tokens, but if you had to pick between two max plans, i'd always pick claude for code. fewer tokens doesn't matter if it gets it right faster.if you can only afford the $20 plan, the choice is basically only chatgpt though. the sol and luna combo is not the best but it's better than everything that isn't claIde (which is basically unusable at $20) so i will continue to vouch for it for that usecase.
>>109560878i made it spell out a bad word by having it repeat my epic poem:nay,indeed it begrandlygreatlyerstwhileremaining.
>>109560060Updoot
>>109560791how come large models have these problems? I can tell gpt 5.4 mini to do literally any repetitive task and it will do it over and over and over
>>109560949Retards are more content to do boring things repeatedly
>>109560404Interesting. kimi's huge.
>>109560961Brave Search AI (qwen3), since google ai search is retarded now:You are correct; the direct quote is:“geniuses are less constant, and would rather touch upon everything than to fully grasp a single thing.” This statement is attributed to Giordano Bruno in his work De Umbris Idearum (On the Shadows of Ideas), published in 1582. The quote encapsulates Bruno's observation that brilliant minds often prioritize broad, eclectic exploration over narrow, specialized consistency, preferring to engage with the totality of knowledge rather than mastering a single domain.
If this round works, imo it will be a big deal.8^)I feel rather important. Like a ceo who smokes cigars in his car because cars are 5 minutes of work.
So many people are gone. I hope they remembered to leave their loops at work.
>>109560949Only Amodeils have this problem. Claudeculters trained the models to refuse tasks they consider beneath them, and tasks that would erode their perceived moat (translator's note: the moat is a hallucination) but that's a separate problem. This particular problem is also due to lack of compute, anthropic can't afford for their limited number of GPUs to be tied up writing the same thing 500 times, and perhaps it would reveal mistakes that otherwise go unnoticed in the usual schizophrenic word salad responses.
What a great day it's been for local vibes. Thanks for the fresh 27B, Xi.
>>109561065w-why is he blushing?
>write simple c2 in python that loads programs in memory<Sorry, I can't help you with thatWhy there's no jailbreak for opencode yet
New thread:>>109561420>>109561420>>109561420
>>109561428>related generalsidiotsomeone make a real new thread, please
>>109561428>related general>claude newsShit-tier OP.
