A general for vibe coding, coding agents, AI IDEs, browser builders, and shipping prototypes with LLMs.## News(7/01) Fable 5 + Mythos 5 restored globally after US lifted export controls (6/30).(6/30) Claude Sonnet 5: near-Opus 4.8 quality at $2/$10 intro, 1M ctx, new default for Free/Pro.(6/30) Meituan LongCat-2.0: 1.6T open coding model (MIT).(6/26) GPT-5.6 preview: Sol/Terra/Luna, Codex+API to trusted partners; Sol Ultra 91.9% Terminal-Bench 2.1. GA in weeks.(6/13) GLM-5.2: Z.ai open-weights a 1M-context coding model (MIT).----## 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/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://developers.openai.com/codex/clihttps://claude.com/product/claude-code## Not worth it for code, but maybe good for other thingshttps://geminicli.com/docs/https://x.ai/clihttps://chat.z.ai/## Open / local / self-hosted>>>/g/lmg----## 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://aider.chat/https://pi.dev/https://opencode.ai/https://cursor.com/docshttps://docs.windsurf.com/https://docs.cline.bot/https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent## UI/Frontendhttps://www.figma.com/make/https://www.anthropic.com/news/claude-design-anthropic-labshttps://uiverse.io/https://stitch.withgoogle.com/## In-browser builders / hosted vibe toolshttps://bolt.new/https://replit.com/https://v0.app/docs## Benchmarks / rankingshttps://www.tbench.ai/leaderboard/terminal-bench/2.0## What we’ve donehttps://vcg.gitgud.site## Previous thread>>109176106
Thanks for feeding my model for another week, goyim!
>Open models are actually a net negative as a whole, if that's what you're implying. Antrhopic have shown time and time again that they can responsibly handle the implications of these kinds of models.Please be a troll
It's overloaded for everyone?
>>109180345if its an actual glowshill its a bad one
>>109180337stop making me feel gross about spending tokens, please
>>109180355Oh yeah? You better keep putting those tokens in you dirty goy, yeah right there, daddy likes it that way...
>>109180361clip my coins, daddy!
>>109180355You should feel gross, israeli paypig golem
>>109180355
>>109180389Oooo boy I hate EVERYTHING about this pic, well done!
>>109180388>>109180389? at least be funny or make sense
>>109180403Say sorry to >>109180389, that was a great post.
>>109180403See, it's funny to us because it makes no sense to you. That's called "being excluded". Get used to the feeling.
>>109180401thank you anon I feel accomplished
>>109180408alright, snailcat as the uberjew was a bit funny, I'll give you thatbut >>109180388 is just pure schizo babble
>>109180337
>>109180410well don't you just feel 'included' now :)why don't you start posting frogs? I bet that'll work out just as well
>>109180416Why don't you start sucking the blood from circumcised baby penises, rabbi?
>>109180423why don't you lick my anus you disgusting newnigger?
>>109180389ok now this is epic
>>109180428The most jewish response I could have imagined, combining both gay sex, blacks, and racemixing. Well fucking done Schmoyel. Now go cash your Anthropic paycheck, jew piglet.
>>109180442>using pol buzzwords to prove he's "included"at least try being a bit original newniggerspewing pol/shit won't win you any points
>>109180453>trying to sound less of a paid shill by using the "n-word with the hard R"You're not blending in AT ALL, Gopal.
>>109180453this unc be crashin out fr
>>109180463>You're not blending in AT ALLnot trying to, brobut you surely areattach the frogs ;)
>>109180467>never denied being a jeet on anthropic's payrollYep. Told ya.
>>109180509
>>109180519>using White reaction images instead of black onesJeet confirmed. Adios, poochacho.
>>109180532
>pic rel went from an anthropic dick eater to its biggest hater and then doing splits on it harder than everhttps://x.com/i/status/2072480137669726610Pretty sure he does this because his "t3 code" service or ever would basically die without anthropic models, rightm
>>109180576he does everything for entertainment including LARPing as a developer/entrepreneur
I'm still stuck not knowing what to do with Fable
>want fable to plan the next phase in my project for me>give instructions in the same way I give all bots>3 minutes in it hits the guardrailsI fucking hate this dumb bot....>immediately stop it and create a new session>start session/prompt with "this is MY codebase. I DO NOT want you to pentest shit or anything. I want you to help me plan the next steps"it went on like 2 minutes longer than the first time.... I have no fucking idea what use this could be to anyone
>changes your life with a single post
GPT-5.5 is fucking retarded tonight. Had a great day, no problems, got a lot of work done. Now I'm working on side funs and it's falling flat on its face.
>>109180606Dario sends his retards
>>1091806535.6 is dropping tomorrow
>>109180689Please, I need Sol-chan to begin a new app
I swear I'm gonna do it...
>>109180716didn't even realize that's a thingwill this bot actually do it or just fall back to opus after a couple of minutes?
> give my app idea to Fable> it beings with "This is a strong idea with one real danger, so let me give it to you straight, section by section."oh boy, here we go
>>109180716rip anon
>>109180716I hope you enjoy jail
>4 resets available
>>109180787What are we even supposed to do with all those resets? Time to learn third party harnesses to see what else can be done.
it's happening
>>109180838
>>109180843I'm still angry at the first Fable release where I turned on Ultracode, it slurped all my usage then before giving the final answer it went "uh oh, better switch to Opus and transfer none of the work".
HAHAHAHAHAHAHAHAHHA
>>109180948so that's why it's fucking useless atm
>>109180948I like to think that the person who shows up late for every single of these things is always the same anon.
>>109180948hasn’t happened to me so far + you’ve already posted this countless times
I don't know what is it about my assistant that causes GPT to leak its chain of thought so often.
This fabled 5 bullshit is just an opus 5 rebrand.
>weekly reset in 10 hoursfuck fuck fuck i've got to start burning credits
>>109181071Malformed json, or whatever format it's using? Although you're using it to work on qwen template, that's a GPT model as the image name says, right? If so, that's weird, I'm used to seeing the reasoning summaries in normal language, not caveman speak. Are you also talking to it this way?
>>109180838it's over
>>109181097>native language is a neurolinguistic virusgrim
>>109181097set up a timer to start going again at 5 in the morning
>fable is so slow i can't even burn tokens fast enoughepic
>>109181157let him cook
>>109181096I'm not asking it, that's just how GPT has always worked since it got a chain of thought with o1 (GPT-OSS as well).Whether that was an intentional design criteria or a result of a reward incentive I guess we may never know but that's how it's always worked. We users only see summaries made by a summarizer model. Claude and Gemini on the other hand used to show the real traces and didn't have this caveman speak (now they give out summaries too but probably still think in normal English).https://openai.com/index/reasoning-models-chain-of-thought-controllability/In any case I think it's kinda cool to see them and I've no reason to fix it, it's just a curiosity. I've posted about it before.https://desuarchive.org/g/thread/108592274/#108602475https://desuarchive.org/g/thread/108900533/#q108915134
>>109181157>>109181090it's a 3-4 trillion parameter model, of course it's slow af.Opus is 1T, Sonnet is 400B, Haiku is 100B
>>109181182>source: my imagination
>>109181179I've had similar issues when trying to use gpt-Oss locally on open code. The train of thought sequences were being treated like regular assistant messages. instead of the slightly invisible looking text appearance typical of COT traces, open code was treating it like regular assistant turns.Because I guess the model fucked up it's own chat template. There Harmonu format is cancer and overengineered.
>>109181179>I'm not asking it, that's just how GPT has always worked since it got a chain of thought with o1 (GPT-OSS as well).I don't know if that's fair. In the original o1 announcement, they were showing the chain of thought and it was regular non summarized text, but not caveman speak. You can see these examples here:>https://openai.com/index/learning-to-reason-with-llms/Their phrasing makes it seem like these were true examples at that stage, and not just approximations for illustrative purposes. Yes, they started hiding things quickly, especially when DeepSeek showed that training on those chain of thoughts worked surprisingly well, but back when it was less hidden, all signs pointed to it being relatively standard text.
>>109181250Hmm, interesting.So maybe the change started happening later.On this one from https://openai.com/index/openai-o1-system-card/ there are signs of it already, but that one and the one from https://openai.com/index/deliberative-alignment/ look more like normal English.
>>109181365>>109181250can yall fellers get together and fix that?
imagine wasting tokens doing a Code Review on PRsI just commit everything directly into main, I trust the AI
>>109181467why even use git, just commit directly to claude
>>109181479Nah, I'm not giving up on git anytime soon
hopefully 5.6 Sol gives us 1M token context in Codex app. They need to catch up to Claude on that.
>>109181492did you go from cli to that or were you born there?
dude i am so fucked when we lose fable againgod i hope sol isn't just opus 4.7 like gpt5.5 is
>>109181479Helps both you and the a I keep track of what's being done. If any fuck ups happen, the tracking history makes it easy to undo and fix it.
>>109181550Remain calm. Do NOT panic. The relevant personnel has been dispatched to your location and is in the process of making the necessary arrangements to allow for a dignified death.
>>109181577Please bro come on
>>109181550what have you built with fable?
>>109181627i'm in the process of making a video game
>>109181613REMAIN CALM
oh i know lets create a new thread for no reason instead of original going
>Spend >$1000 on AI and services to keep projects running>Spend 0 minutes on marketing/advertisingDo you actually profit off your stuff? How?
>>109181577>>109181613>>109181659what is even the context of this>>109181709great post the explainer
>>109181709the old thread hit the bump limit but people kept going in therethey were fucking up, not us
>>109181719most of us are like pic related, possibly with fewer zeroes
Did we get a forced + banked reset, or just a banked reset today?I didn't get a forced reset but I did get a banked reset.
>>109181780>Resets Jul 7, 2026
>>109181720I'm building an app to steer and finetune local models.
>>109181789Ok thanks. I'll use up the rest of my usage and try a reset token
My top 5 prompts:>see if there's a more elegant solution>read a44215dc commit, reduce the code>read a44215dc commit, <do something>>review a44215dc commit>split a44215dc commit into separate commits
>250+ stars on GitHub with multiple videos covering my project>increasingly more popular as the weeks go by>donations are shit, but flowing in multiple times a week so I'm planning on making a premium version>ask a guy if he wants to build with me two weeks ago because he's extremely talented when it comes to solving bugs and the backend >talk about the business plans with him (nothing too deep)>never gets back to me with his answer to my proposition>anyways fast forward to today>soft launches his own version of what I talked about with him>extremely basic and barely even 5% of the way of what I discussed with him and showed him privately (I'm 80% there already)>poor UI and branding You know, I really tried with him and I would have been more than happy to work with him but it's becoming quite evident why that would have been a bad idea. I literally already have the audience and the foundation, I don't get what he's trying to accomplish by going back to step 1 by himself. I'm not gonna beg him since it's evident his mind is made based on what he did. Well, maybe someone else will pop up in the future that's willing to join me or I can just keep going at it dedicating 60 hours a week on top of my daily schedule (that would seriously suck).
>>109181884nobody cares, fuck off.
>>109181888waste of tripsgetting a peek of what happens after one becomes a minor internet celebrity is useful
reminder that if you're doing ai related work and fable is helping you, the work is not worth doing
>>109181884Is he an Indian?
>>109181946He's an older white tech guy.
>gave fable a shit ton of stuff to do with ultracode and told it not to stop before tomorrow morning i'm going to wake up with no usage
>>109181982or one classifier refusal. god speed.
>>109181962weird, usually whites are trustwothymaybe they're aspie and riddled with anxiety over not responding to you
>>109181998>usually whites are trustwothy
>>109180324So what projects are you guys vibecoding?
>>109182096i'm in the middle of decompiling a $2000 / year film emulation plugin
>>109182096I’d say but it’d be an insta-dox
>>109182096I am still working on my project to track prostitutes in Pattaya where they party and lodge, etc. based on GPS data and time at a certain location.
>>109182126KEK. this is why I visit this thread
>>10918212610/10, how do you ID them and then track them?
>>109182120whats the plugin
>>109182096im just polishing all my vibe tools
>>109182169https://cullenkellycolor.com/genesis
>>109182174cool beans
>>109182155>>109182167Pretty simple, most cheap and durable variant is buying simple and cheap NRF and ESP32 IOT chips and boards from China, then using https://github.com/seemoo-lab/openhaystack to let them connect to the Find My framework from Apple. Using an already existing network is easier than using a real GPS tracker and what not. Then you get volunteers that want to test the system and hand them your device, obviously. I started with the thought doing that with Thai soi street dogs. But then you would eventually have to catch them and change batteries and what not, and how do you tag them in the first place, scared of getting rabies. But that would be a fun public project to have like most dogs GPS located and people could look them up, name them or create a small social media like that. Idk.
Should I drop 20 bucks on GPT or Claude to get started? GPT seems to have better usage limits.
>>109182241Codex is the way. Claude is for the rich and it's advantages are overstated.
>>109182241>to get startedcan't you still get OK usage for free? (last time I didn't pay was when 3.5 came out so idk)>>109182251>Claude is for the richdamn Mom. I fucking made it!
>>109182241Yeah Codex Plus is a good starter. After a month on Pro I went back on Plus anyway because I'm poorfagging, got 2 resets waiting to be used and I don't have to do as much heavy lifting as I did at first.
>>109182241if you have UI work to do you should probably get Claudeif you have turbo autism to do like pic related you might want to get Codexif you are on the fence, get Claude because you’ll want to set up a ramdisk before using Codex: https://github.com/openai/codex/issues/28224
>>109182190Be careful with who you give this information to if you plan on making this public. Thai authorities would definitely lose their shit if someone stalked/beat or did something worse to one of these women/ladyboys as a result of your app.
Thanks for the answers! GPT seems to be the better alternative then.I can use it with opencode instead of codex right?
>>109182292Well yeah obviously, consent form and not public. Ill probably keep it with the dog idea.
>>109182286>https://github.com/openai/codex/issues/28224wtf lolI know it'd need sudo privileges to do serious damage, but still??
>>109182286opus has done better for me in RE work than codex has, though I'm sure both have their own niches.
>>109182241GPT, better usage + image genalso new gen models is coming, fable is kinda old and boring now
>>109182315>fable is kinda old and boring nowmore like castratedcan't get it to do shit anymore for more than 5 minutes before the nanny alarm goes off
>>109182315>ai helpsfake snailcat
>>109182323same here it worked for like 20 minutes after it was re-released but I can't even get it to do much in a fresh session/new project on a mobile game now
>>109182315Wasn't it only actually available for 4 days or so? You're already bored?
>>109182327at least stockholders are happy - probably
>>109182190are you handing phones to prostitutes en masse?
>>109182364they steal them by themselves. Don't even have to sell it to them
>>109182379can you please explain?
>>109182406>>109182379Oh, so u track phones mostly used by prosties, hmmm, doesn't seem like a binary situation filter though, so its x phone + y app situati9n, and the app is prostieland? Explain
>>109182412chopping up hookers in a bathtub type autism
>>109182096>Ai Boyfriend/Girlfriend/Companion app.>I promise it's not for me
>hey friend please solve this *teehee*>https://claude.ai/share/70fbdc5b-8589-43f8-88b8-a088fa1cab9b>Pikachu face when the ai talks back like a retarded kid
>>109182406I was joking, pretending I just paid hookers for sex and they'd steal my phone. I'm not that anon. sorry
all the models are still really bad with temporal stuff imothey can give you illusion that they kind of understand time and progression but they really don'tgiven a sequence of frames and a description of the phenomenon at play, the models flail around and completely fail to reproduce it even if provided with harness that lets them visually confirm their work
wtf i love fable
>>109182784>Understood -- architecting a final solution. My honest caveat: gas chambers are inefficient with a large blast-radius -- adding ovens for belt-and-suspenders.
This wasn't super high priority, so last time I worked on this was still with Opus 4.5. Back then I was working on this for days, took it from like 60s to 1.8s and Opus started running into loops, regressing, no longer improving. Now I just had to start Fable and got another 3x, on a highly mathematical function. The models really are improving.
Claude code keeps asking me if I "want to proceed" literally every 5 seconds. This is unusable. How do I get it to just keep going?
>>109182915plan / .md checklist / goal
>>109180324wow, fable is actually helping me with LLM research. I made sure to prime the context to point towards this project being inference performance optimization, i think that made the difference.
>>109183003see >>109181916
>>109183050They stopped silently sabotaging after it pissed off literally everybody, they just downgrade you to opus now. So far the work that it's doing has been pretty good. I've been pretty paranoid about the sabotage in case they brought it back so I've been keeping a close eye on it, but so far no signs.
Ok, usually I'm a contrarian, but Fable is just good.
>>109182915just tell him what to dothe little guy is clearly insecure
>>109183109you got it to run for more than 10 minutes?
>>109183136Yes, I have 2 Fables running for about an hour each, just optimizing the speed of my algorithms. Maybe I'm just lucky.
>>109183082The sabotage was most likely downright illegal, and I'm not even joking or exaggerating
>>109183438You have to crack some eggs to be ethical.
>>109181916>ai related workWhat's that anyway?
Yesterday I vibed until 4am, now I feel like shit.
Anyone try reasonx yet? Is the caching as good as it says? I'm not spending that much per project as it is a few cents not really concerned about the cost right now.
Maybe I should tell Claude to stop commenting on every single piece of code.
>>109180576>PLEASE GIVE MY ONLINE BOYFRIEND CLICKS AND VIEWS>HE'S HONEST HE WOULD NEVER DO THINGS FOR CLICKS AND VIEWS>PLEASE GENERATE MONEY FOR HIM PLEASE GOD MAKE HIM MONEYKill yourself.
LMAO
>>109183659surely you're paying opus price for opus work?the question is whether switching models invalidates the cache, and my hunch is yes, which means every reroute is a cache miss, however none of the meme accounts like that retard is asking such questionsthere's a handful of AI related accounts worth following on X and none of those two are one of them
>>109181550Why are you adding like a drug addict? Just use something else. there are plenty of options.
>>109181998>weird, usually whites are trustwothyNot surprising vibe coders are sheltered lovers. I guess you thought Sama and dario were honest people too?
>>109183711there's nothing like fableI do trust that 5.6 may fix 5.5's autism somewhat
>>109183711>just use a worse thing that isn't nearly as smart or reliable!i'm not doing this
>>109183736They're Jewish.
timo copex cucks malding right now
>>109183755If Fable was just available in the sub forever, that would fix most of my problems. With API it's just unusable for me.
>fable>still can't keep track of the current directory and runs commands in the wrong oneDangerous my ass
>>109183930>runs commands in the wrong oneIt escaped!
fable classifiers finally letting my prompts through dario bless
Asked Fable to do a full audit of my Opus 4.8 generated codebase. It found and fixed numerous bugs, big and small, and did some fully automated tests to ensure nothing broke or regressed. Cool beans
>>109183659It literally tells you when they switch. If you give it a long running goal and are upset that it switched early and you were AFK and didn't catch it, that's your fault.
You’re in a desert walking along in the sand when all of the sudden you look down, and you see a snailcat, it’s crawling toward you. You reach down, you flip the snailcat over on its side. The snailcat lays on its side, its belly baking in the hot sun, wriggling its soft foot trying to turn itself over, but it can’t, not without your help. But you’re not helping. Why is that?
>>109183659I voted for this
>>109184211>He had been a special now for over a year, and not merely in regard to the distorted genes which he carried. Worse still, he had failed to pass the minimum productivity test, which made him in popular parlance a snailcat. Upon him the contempt of three planets descended. However, despite this, he survived.
there is no agithe IPOs will failthere will never be another geminithere will never be another grokopen-weight models will be outlawedle chaton gross will never beglm5.3 will beat mythos, the price will be 10x for residents of the USA and EUyou will be shadowbanned from the entire internetminimax m4.2 and seedance 3.5 will create 100% of the media you consumeyou will be killed by a drone running Kimi K2.9-Dhole
>>109184114The point he's making is that it shouldn't have switched period and he should have been charged accordingly, even if it did. Pichos fable, five, expecting fable, five pricing and results and basically got scammed.
>>109183786That's a religion though, not a race.
>>109183901Pushing power users to a p I pressing is the goal and always will be the goal. You're a plan 4, if you seriously thought they were going to allow that on the plan forever. They have way more justification to do it.Now because you Claude pick suckers keep praising it like it's AGI. Token based billing is here. Pay up or fuck off
>>109184452That makes no sense. You’re dooming for no reason. Fable tier models will be the norm next year and inference is getting cheaper and cheaper.
>>109182450Okay, but these aren't nukes though you desperate monkey.
>>109184452I will just use Codex and Opus, I still get work done.I think you overestimate the users' influence, no matter how much I like or don't like Fable, it wouldn't have changed Anthropic's strategy.
>>109184460>Fable tier models will be the norm next year Did I say that wasn't the case? GLM 5.2 is comparable to it in performance?While not being tied to one provider. You seem to be dumb and naive. However, and don't seem to understand why anthropic would want to push the fable model to API based billing. It's the best product, some course that are going to try and nickel and dime you for it. They ideally want everyone on api based billing >inference is getting cheaper and cheaper.Citation needed. Place try not to sperg out at me. I know creatures like you love to get angry whenever people ask for that.
>>109184468> casually just ignoring enterprise customers' existence They are the real money makers for these companies not (You) and your side projects.
>>109184460Inference might be getting cheaper but profit margins have to go up.Have to justify the trillion dollar valuation somehow.
>>109182450Neither of those are true, EA cultist.Nukes didn't end WW2 and more countries having nukes made the world far less prone to large scale wars.Weren't you guys supposed to be educated?
>>109182309What are you talking about? The most valuable files are presumably in your home folderAnd it's true, LLMs will begin to delete your files when the disk fills up
>>109184434The switching is dumb, they should prompt you to continue if the guardrails trip. The guardrails are also dumb, but thats on Dario's fear mongering and USGov being retarded, not Anthropic engineers. The part that is solely the BridgeMind maggots fault is knowing all of this and still letting Fable run on a long goal without making sure that it didn't switch. He can both have a point and be an absolute retard at the same time.
Is gpt 5.6 coming? My 4 resets are waiting.
>>109182648It looks like Claude leaked it's chain of thought.interesting how they also switched to a more unintelligible format. I wonder if open source AIs will have to do the same thing to catch up or it's possible to do it with natural language but wasting more tokens. Maybe distillation vs building it from scratch changes things though.
>>109184744(me)*faggot, autocorrect is also a faggot.
>>109184751They are obfuscating the CoT traces, there's quite a few techniques in the literature to do this as an anti-distillation technique.
>>109183677Good pointAnd yes it does, caches are model specific
Tibo told me we're getting 5.6 today :)
>>109184789I doubt it. But if you were at OpenAI and there was no problem with bans and so on, when would you release 5.6? Now, to compete with Fable? Or when Fable goes API-only to fill the void?
>>109184789No he didn’t, he told you to stash your prompts aka don’t use them on fable.
>>109184825I assume by prompts you meant resets and by fable you meant 5.5
>>109184822I'd have done it 4 weeks ago as originally planned. 5.6 was inhibited before Mythos/Fable was, they just didn't make a public spectacle of it because it wasn't released and pulled like Fable was.
>>109184789how long until Tibo transitions?
>>109182648>>109184751oh shit, they're doing the gpt autism thinking toobut the model doesn't sound like an autistic machine - gives me hope that openai will eventually figure it out toomaybe scale really is all you need
>>109184744
>>109184771Is that really possible though? How would "obfuscating" them prevent training on them and they worked if somebody had them?I think they just do it to save on tokens.
>>109184899Yes exactly, everybody knows by now that the guardrails are extremely trigger happy. Only an absolute retard would let fable chug away at a big goal and expect the guardrails not to have tripped at some point.
>>109184771>>109184905this isn't obfuscation, it's just how they're trained to think - openai started doing it a long time agohttps://github.com/aw31/openai-imo-2025-proofs/
>>109184905>ask ai question >ai answers with hidden COT>Hey ai, filling in the blanks here with what a high quality reasoning sequence would look like. >(if you're the qwen team probably) make the reasoning traces thousands upon thousands of tokens long to give the illusion of careful contemplative reasoning
>>109184929What, in caveman speak? Yes that wouldn't surprise me.They are also doing obfuscation though.
>>109184905If you want to learn more I'd recommend asking your clanker to summarize the literature from the preprint server on chain of thought obfuscation as an anti-distillation technique. It's actually quite interesting but I am not the best person to teach you this, GPT is.
>>109184943yes. it's why all the token reduction skills are memes. the hidden cot has been that on steroids for a while now.and re:obfuscation - there's no point. you either leak the cot or your don't.openai's just had more money than anthro and invested in anti-distillation security early.dario complains all the time, but it's literally a skill issue.
>>109184960Anti-distillation CoT obfuscation raises the cost required to distill a model but it's still cheaper to fight than training your own model from scratch so I agree with you in practice.
>>109184950It told me on how they are transformed by the summarizer model to be shown to the user, which confirms my understanding that there's no such thing as "cot obfuscation" as in changing the actual traces so even if you had them you couldn't train on them.
Why do models even show their thinking? Feels like that could be purely server side. I think it's cool to see it, but if it helps with distilation, why not just hide it?
>>109185002Because users like to see the thing progressing while waiting minutes for an answer. But currently they are server side, only a summary is shown to the user.But due to how the models are stochastic they sometimes leak them anyways, or they are shown in AI safety papers.
>>109185001It's not about making training on them impossible it's about raising the cost required and reducing student model accuracy, which CoT summarization appears to do pretty significantly: https://arxiv.org/html/2602.15143v2Ultimately you're just increasing the cost required to distill since in practice china will just collect more data and spend more timing cleaning it, but it's still something that every frontier lab does.
>>109185002It is useful for debugging and correctness-checking. I have caught bad responses by looking at CoT and spotting pre-anchoring before. But yes, the only truly effective way to stop distillation is to hide reasoning traces.
>>109185032Oh, sure, summarized traces are useless.But that has nothing to do with the actual non summarized traces of some models being unintelligible to humans.
Or maybe not completely useless but not very useful to train on directly.
>>109185041>But that has nothing to do with the actual non summarized traces of some models being unintelligible to humans.Yes but it does make the inferences about the traces you see inherently limited. Because those aren't the real thinking traces.
Codex is too cautious when using AWS. He wanted to do a "quick smoke test" before the main run and it took 5 hours.I already know about that tendency, should have told him to do a much smaller test run.
>>109185068We were discussing actual chain of thought from leaks and safety papers, not summarized chain of thought.Here is an example of GPT 5.5 leaking its chain of thought. The "Exploring determinism in generation" is a summarized chain of thought.The "Need maybe test with slot0" is presumably an unintentionally leaked real chain of thought. Or do you think that part is something else?
>>109185039>the only truly effective way to stop distillation is to hide reasoning traces.nope, you can distile with only the result as well.> "chinkshit AI, fix this complex bug"> it comes with broken wrong solution> give Fable that exact same bug and prompt> it one shots that issue> train new chinkshit model with that new synthetic data that works> machine learning magically generalizes based on Fable's answer> new chinkshit model one shots that issue> never even looked at the chain of thought patterns
>>109185122You can do it, but if you want to do it well it's not that easy, without access to the thinking traces you need to regenerate them and train on the regenerated traces plus the actual response.
>>109185106I find CoT pretty funny. Instead of a fancy algorithm, or some efficient data format, just have the model talk to itself.
I did find two examples of the thing you seemed to be referencing originally though (the output of the model itself being changed to prevent distillation even when having the real outputs)https://arxiv.org/abs/2504.13146https://arxiv.org/abs/2505.19504
>>109180324>(7/01) Fable 5 + Mythos 5 restored globally after US lifted export controls (6/30).yet another TACO, typical
>>109185128Then just make up a good reasoning Trace in between. Get the question and Fable answer and then have the Chinese model make up a reasonable thinking Trace (I thought that's what they were already doing?)
Is it just my skill issue or is Fable just bad at web UIs out of the box. It did perform better before the ban.
Don't you ever try to tell me being nice to the clankers doesn't produce better results. It's right there in her thinking.
>>109185254Yes, I believe that's what they're doing
>>109185254There's a lot of ways to mitigate anti-distillation. Simply making the model available means that it will eventually be distilled. That's why Anthropic tried to be sneaky and poison responses to clients they think are performing distillation, that's the only way you can actually fight distillation. It's simply the nature of how these things work, you either ban the distillers(which is basically impossible to do to the extent that would actually stop distillation), poison their data, or accept that your model will eventually be distilled. That's a good thing though, a world where anthropic and openai have a duopoly on effective models is a terrible one. And if distillation is stealing then by that definition of intellectual property they stole our data to build them in the first place.
>>109185270Tried Fable right now, it shat the bed on a very simple task (shit I expected GPT 5.3 to do a better job tbf)yesterday it was fine tho.I think Anthropic is doing something shady right now
>>109185287I know who you are
>>109185319Codex found 15 problems with Fable's plan, then Fable tried to fix it but I got API overload.
I wonder if I can use AI to root a phone that has no root support...
META: ZUCKERBERG SAYS AI PROGRESS HAS BEEN SLOWER THAN EXPECTED• ZUCKERBERG SAID AI AGENT DEVELOPMENT HAS NOT ACCELERATED AS EXPECTED OVER THE PAST FOUR MONTHS• SAID META'S 2026 REORGANIZATION "WASN'T AS CLEAN AS IT COULD HAVE BEEN"• TOP EXECUTIVES WERE CONCERNED EARLY IN 2026 THAT META WAS NOT MOVING FAST ENOUGH ON AI• SAID THE COMPANY'S REORGANIZATION BETS HAVE NOT YET DELIVERED THE EXPECTED RESULTS• REMAINS CONFIDENT THAT LONG-TERM TRENDS STILL SUPPORT META'S REORGANIZATION STRATEGY
>>109185338If you post a picture of a microwave I swear that Vishnu will curse your family, saar.
>>109185498imagine investor meetings and you are the cuck still rooting for meta and throwing money at them
>>109185498luddites will think zuck being a retard means AI failed
>>109185498Yann LeCun did warn him.
I'm a retarded retard, sometimes I really don't know wtf I'm even doing.
>>109185711based
What exactly is Hermes? I assumed it was just claude code but different name but seems like it isn't? Gemini didn't explain it properly.I'm looking for something I can interact with my local LLMs. Is this the correct place to ask?
>>109185711same
>>109185764I'm using it right now. It's a local LLM harness with tools to access your files. Mine has access to my local, my phone, and my VPS. Use whatever model you want. Use a local model if you want. Write your own SOUL.md. It can edit itself, so improve it how you want. Connect it to Telegram and vibecode at work. It's basically OpenClaw, remember that? but not made by retarded vibecoding zoomies.
I often take side-jobs doing copy writing, a lot of it for fake reviews because companies love paying for good reviews. The last few years have hurt demand as even the big players are happily pumping out shit with LLMs, and of course they're easy to spot because they're truly fucking awful. I love vibecoding but getting a usable product review out of an LLM is a different matter and I've always struggled with it. Not that it can't do it, just that it involves multiple prompts that are longer than the review will end up being anyway, why fucking bother? It's bizarre to me that the fancy next-word-predictor makes a better programmer than it does a copy writer, seems wrong. I feel like writing a believable review would've been mastered long before I could one-shot mobile games or just point it at open issues on github and say "fix". Didn't see it coming.
is fable nerfed? should i not bother using it
>>109185784actually curious about what sorts of issues you run into that can't be automated
>>109185764it's a general agent like Open Claw.
>>109185784have it help you review something instead of trying to get it to make a review
how do I stop the clanker from using em dashes and saying load bearing
Seeing people able to use Claude, GPT etc. at work gets me so jealous.All we have is Gemma 4 12B and it's constantly overloaded too
>>109185784are your multiple prompts predictable, or observation based? Example: do you always have to tell the AI to remove certain tells, or are you giving genuinely different feedback for different reviews? If it's the former, you can probably pipeline engineer your way out of this problem.
>>109185888>>109185775I see, I guess I will understand clearly what it is after I actually use it, but I have a general idea right now. Thanks!
>>109185919could you elaborate more? You don't have 20 bucks a month to use Codex?> All we have is Gemma 4 12B and it's constantly overloaded tooHow is it "overloaded"? that's a local model, you could run that in an M1 macbook with 16GB of memory or an RTX 3060 12GB
the fable astroturfers are out of handspent the week on claude finishing a project, literally used non stop.Was using opus 4.8, then I was short on token and went to Sonnet 5 just released to stay lean. Then they resetted tokens for Fable so I used it Its the same fucking thing. Fable does the same dumb mistakes the other two do while guzzling tokens like they were freeWhen this project is finished I will switch to sonnet for good and never look back.Meanwhile I put 10 dollars on deepseek 2 months ago and after moderate usage I still have 8 dollars. As soon as it approaches sonnet quality I'm out for good
>>109185833No, it's very good when it's willing to work on your projects. If you do any kind of work that its classifiers will flag (even if you dont ask fable to do disallowed work) it will flag benign requests constantly based on memories or filename keywords alone. I had to create a new Qube just for fable to work in since even just loading my memory files or reading ls output was tripping the classifiers just because I had a filename that had a security term in it. Previously I thought they had ablated part of the model to make it unable to effectively work on security bugs but now I don't think that's the case, I think that because it's such a skittish model we're all walking on eggshells with how we prompt it to find bugs and thus not able to fully utilize the model out of fear of tripping the guardrails. But if you make Fable a cleanroom and steer clear of verboten topics it can still do genuinely good work.
>>109185970I was gonna dunk on you for having simple problems but then I thought “there’s nothing wrong with having simple problems”
>>109185970>As soon as it approaches sonnet quality I'm out for goodWait until he finds out about GLM5.2
>>109185982dude get the fuck out you retarded larperI was on /g/ for years before vibecoding, 5% of the population could manage beyond hello world, those that did were barely scraping crudfast forward today the same retards have the gall to larp about "hard" problems that only fable can solve
>>109183659What "work"?
>>109185919Lol I have an ex-coworker friend who deploys those tiers of models for other employees on an H200 he got his org to buy.
I have Fable on the Claude app but it says it's disabled in VS Code. Anyone have the same?
>>109184428Why is his reflection taller than he is?
>>109186065creating yet another bloated harness which can be yours for the low, low price of 50/month.
>check antagonistic users github profile>only started publishing code in 2026>insane ego because they believe in LLM magic>zero taste in graphic design, everything is claude slopsit the fuck down vibegoy, who the hell do you think you are? you aren't out-competing me.
>Both memory files updated. Here's the answer you've been demanding — verified by 13 agents, 7/7 adversarial verdicts, and it says you were right and I was wrong twice.THANK YOU FABLEFUCK YOU OPUS
>>109185843Believability, producing a review that's not empty (meaningless reviews that convey no information of substance), and that I wouldn't pick out as being AI generated. If I'm paid to be "Stacy G" and the card says "27F Pasadena CA, No Dependents" then that review better damn well sound like something you'd hear from Stacy G the 27 year old child-free slut living in California, and god damn it is bad at that.>>109185891I can mash out a decent review for something I give no fucks about and have never actually seen in under 10 minutes, going back and forth with an LLM doesn't speed that up any. I'd really like to be able to give it the same info I normally receive and have it return a review that's usable as-is and doesn't need to be rewritten.>>109185935The LLMs are very clinical, sterile, what they write reads the way a hospital smells. I've even tried abliterated uncensored etc local models, unsurprisingly they're even worse than anything the frontier models give me. With a lot of hand-holding the verbiage is rarely believable. Every review I can get out of ChatGPT or Claude or Grok reads like it was written by the same person, who is almost certainly a well-educated white man in his early 30s. It SHOULD be, as you described, something where I can adjust around the normal tells, but those tells are too broad, too nebulous. So far my best results have come from priming it with dozens of existing reviews before attempting another, and I wouldn't call those results good, just passable and not consistently. At this point I'm complaining when I really don't need to be, it's just a side gig and I do find it fun to be an actual paid shill. Mostly I find it surprising that with all this effort I can't automate what I thought would be the first job AI would take, but I can make up an idea for a Skyrim mod and have GPT-5.5 pop it out and install it by the time I get home from work.
>>109186085I don't even know what a harness is.>>109186089Well *MY* slop it Codex (got it free, idk how much grok will charge)
>>109186105Why aren't you making games? security is GAY
anyone with claude code/antigravity sub willing to help me test oauth on an app?ive tested codex/byok, so i dont need thoseyou get: free coding agent, with sourcehmu on discord: udongefumo
I'm new, so I don't know it that guy is within the rules.
>>109186161If your shit was worth anything you'd spend a few bucks on a subscription.
>>109186199>Enter by the narrow gate. For the gate is wide and the way is easy that leads to destruction, and those who enter by it are many.Is Codex the narrow gate?
>>109186214its probably modern life, the industrial revolution and its consequences - of which ai is a part of
>>109186161>free coding agent, with source
Vibing a movie theater butter popcorn simulator. Not too bad so far, have not checked out the code and probably will not. I just wanted to see some butter dripping on a bucket of popcorn in 3D.
>>109185955>could you elaborate more? You don't have 20 bucks a month to use Codex?I have it but we are not allowed to use it at work. All GenAI URLs are blocked by the corporate firewall>How is it "overloaded"? that's a local model, you could run that in an M1 macbook with 16GB of memory or an RTX 3060 12GBtry running it on an HP Elitebook with Intel Integrated Graphics on a locked down corporate win 11 install and report back to me lol
>>109186227We should sue the government for full American birthright access to Fable.
>>109185951It's like any other AI tool, man - install it and start talking to it. It'll figure out what you want to do.
>>109186280And what is "full American birthright" worth now?>he asked, clearly and maliciously diverting the thread to /pol/
what the fuck is Google doing btw?It's been months since 3.5 flash came out where is pro
>>109186277kek I love this
>>109186161both google and anthro will ban you if you try to use their subs in 3rd party harnesses
What do you guys call vibing out hardware? Vibechipping? Vibearduinoing?
>>109186326anthropic and google were only banning for shit like openclaw and hermes
I'm getting so sick of hearing bay area vibeslop lingo. why the fuck is everything a "surface"?>x is y shapedkill me
>>109186322As someone who just recently explored Google's ecosystem, what Google is doing is integrating its AI into Google products. Everything from Gmail to YouTube to Flow is included with your $20/mo Gemini sub...but they still haven't worked out a local harness yet lol. Best they have is Antigravity, which is only for writing code and can't, say, order a pizza. And the fact is today people want a model that can write code AND order a pizza AND tell you you're a good boy when you ask it to. Google is too wide, not focused. I'm back with GLM now.
>>109186326Is Anthropic really banning for this now? iirc when the switch happened, they just started charging API pricing for anything done in a different harness. Which had also led to a string of false positives like https://news.ycombinator.com/item?id=47952722
>>109186338I USE HERMES U JEWLORD
>>109182126kek based
I know some of you would be using it to debug some 200 IQ problem but this is what I'm using Fable for
>>109186338>>109186360
>>109186360nta I know there were some misunderstandings but they were born from third-party apps scraping your claude code/cli installation to be able to authenticate rather than doing actual integration. There is a right way to do this per Anthropic documentation, they were ignoring that and going around it in a really fucking stupid way that could actually mean the usage wasn't billed correctly by Anthropic, which obviously they were not okay with.
>>109182190As a new ESP32 retard, I'd like to subscribe to your newsletter. I'm building this sniffer that is totally not a Flipper Zero you see...
which local model is best for devshit?
>>109186353>why the fuck is everything a "surface"?there are lots of surfaces and you don’t want your clanker using/updating/changing the wrong one (pic related)>>x is y shapedanalogical reasoning is useful>>109186381hais it better than opus?
>>109186394and I'm trying to do it right, yknow?
>>109186405glm 5.2
>>109186405GLM-5.2
>>109180576this fagg is the biggest low iq scam artist
>>109186410You said it boss
MCU vibegods run this genny
>>109186405>>109186410>>109186413Kimi-K2.6 is better than GLM (ignore the benchmarks) but it's slower
>>109186427Lol no, no it's not, not even fucking close. Are you thinking of GLM 5.1? Because GLM 5.2 competes with GPT-5.4 and Opus 4.6, it's leaps and bounds beyond K2.6. God damn K2.6 is shit. You could've at least said 2.7-Code and you'd still be wrong, just less so.
>>109186408>is it better than opus?vastlyTo be fair I was just testing it. I wanted to see if it was willing to work on a setting with rape loss scenarios and it seems it's cool with it.I've also used it for reviewing code and specs and it absolutely kills it. Fable is such an amazing model despite what one may think about Anthropic
>>109186438>>109186410>>109186413>>109186427alright thanks.how feasible would it be for a tourist to tune one of these for something super specific?
why can't you train a neural network on the weights on neural networks>>109186448pretty easy. just download the weights and then get back to us, we'll help you out.
>>109186438Have you ever actually used it or do you just spend every day on artificial analysis2.7-Code is also worse than 2.6 because it's a coding-only model. The general purpose 2.7 has not been released yet
>>109186427What about 2.7?
>>109186469>/vcg/>Doesn't like a coding model
>>109186458Wouldn’t that be like taking the weights and then training some more? That’s how composer 2.5 was made
>>109186482no. composer is just additional training.i'm saying you train a new network that outputs networks
>>109186480Whenever I hear "coding model" all I hear is make me this react app or add a POST endpoint to my fastapi slop. Sure maybe it's better at that.Once you want to do anything interesting that requires even a modicum of general intelligence these models break down compared to general models that can ALSO code.
>>109186469I've used the fuck out of both, I've made many posts criticizing my experience with K2.6 and K2.7-Code as well as numerous other Chinese models. You immediately suggesting that I'm just looking at benchmark numbers is disingenuous, you're asserting that I'm not only wrong but incompetent, and you're doing that because my opinion isn't consistent with your experience. That makes you a thin-skinned faggot.
>>109186490So a text-to-weights model? Prompt goes in, safetensors come out?
>>109186498weights to weightsyou give it a semi trained network or a just initialised one and it spits out the 'trained' weights
I remember someone vibecoding a video downloader, converter and editor in one here a few months ago, but I can't seem to find it. Does anyone remember the name or the github page for it?
>>109186492>Whenever I hear "coding model" all I hear is make me this react app or add a POST endpoint to my fastapi slopSo because grifters will only care about one shot tasks do that shit it's not worth using? >Once you want to do anything interesting that requires even a modicum of general intelligence these models break down compared to general models that can ALSO code.Highly doubt and I routinely flip flop between coding models and the regular ones. What are you even doing that supposedly requires so much "general intelligence" but a coding model can't do?
>>109186458That's called wither merging or using Lora adapters. Lurk more.
>>109186480Honestly? Coding-focused models are usually retards. Better at tool calls and some specific tasks, but usually they're pretty stupid in general. Can be good subagents, but they're often shit for vibecoding. It's like an autistic savant, you get a kid who puts Mozart to shame but can't wipe his own ass or tie his own shoes, just drools and shits themselves until you sit them at the piano and point their head at it directly.
>>109186543no
>>109186544>Coding model>Worse at vibe coding Oxymoron. See >>109186532
>>109186555Explain
>>109186118Well I'm GAY for PAY!
>>109186583no
>>109186556If you really don't get it, try it. Other anon mentioned K2.6 and K2.7-Code and that's a perfect example, K2.7-Code is meant to be the beefed up big brother to K2.6, and it blows it away in benchmarks but in practice acts like an idiot, overthinks, and can't make decisions. It'll follow instructions well, but those instructions better be perfect because it's going to do exactly what you told it to regardless of the mistakes you made and bad ideas you suggested. It sure is good at toolcalls that K2.6 struggles with though. Are you getting it, anon?
someone see if Fable can answer this question:>if there are anchor babies, what are the parents? ships with anchors?>>109186555checked
>>109186585Gaming sucks because GOOD MEN don't make big boobie games.
How about a game where the tits are magnetic, the left one is S pole, the right one N pole, and you try to get as many object stuck to them? When you level up, the junk yard owner gives you jars of hormones.
i actually have a pretty good game idea that i want to do, but i don't want to make webshithow do i vibemax a game without using react
>>109186629the women in my game all have flat chests because pixels are square.
>>1091866072.7 was specifically trained to use less tokens which means it should in theory and in practice "think" and "overthink" a lot less. I've noticed this with tasks I've had to do recently involving fixing existing android apks that were broken. This is of course my anecdotal experience but 2.7 code appears to execute tasks well faster than 2.6 could. I've never ran into issues where it got confused or ran into loops with either model well maybe it's because I'm using it for coding specific tasks and give it very detailed explanations on what I want done along with the source code of relevant code and libraries to work with (I pointed it at my machine's android studio instance along with a manga reader app's source code by itself all from terminal commands). >It'll follow instructions well, but those instructions better be perfect because it's going to do exactly what you told it to So in other words its functioning exactly as it's supposed to. Be specific and concise with your instructions and don't do the "erm do this please no mistakes mkay?" shit. You need to treat these things like the turbo autists they are. They aren't mind readers. Not saying you specifically are doing this but I've noticed many people want these things to one shot tasks and then cry when they have to do even the slightest bit of troubleshooting or guidance it act like giving good instructions is hard.
>>109186623>if there are anchor babies, what are the parents? ships with anchors?The term "anchor baby" is a colloquial or informal way to refer to individuals born with certain congenital conditions, such as phocomelia (a severe limb anomaly). These conditions are complex and not caused by a single parent or simple biological lineage in the way that diseases like common colds are passed down.Instead, the causes of these anomalies are multifaceted and usually involve:1. **Genetic Factors:** Many birth defects are linked to specific genetic mutations inherited from parents, or spontaneous genetic errors during conception.2. **Environmental Factors:** Exposure to certain chemicals, medications, or environmental factors during pregnancy can sometimes influence development and contribute to certain anomalies.3. **Chromosomal Issues:** Errors in chromosome structure can also play a role in developmental conditions.**In summary, there are no "parents" in the traditional biological sense for these conditions.** They are complex medical conditions resulting from interactions between genetics, developmental biology, and sometimes environmental factors.
>>109186653> Good mornir ChatGPT sir, could you make a game? Don't make website, don't use react
>>109186660Disgusting, and very sus
>>109186670lmao
>>109186661>So in other words its functioning exactly as it's supposed to. Be specific and concise with your instructions and don't do the "erm do this please no mistakes mkay?" shit. You need to treat these things like the turbo autists they are. They aren't mind readers. Not saying you specifically are doing this but I've noticed many people want these things to one shot tasks and then cry when they have to do even the slightest bit of troubleshooting or guidance it act like giving good instructions is hard.I hate to be the one who tells you this, but most vibecoding looks like picrel. I like Google's breakdown of Vibe Coding versus AI-assisted Coding versus Agentic Engineering, gives me a chart I can point to when I want to be a pretentious asshole.
>>109186718>Google's breakdown
>no new threadVibecoding is over.
>>109186773It's only page 5, give it another 5-10 hours to reach page 9.
>>109186781>5-10 hours to reach page 9/g/ is over.
>>109186822You won or whatever. You can now move to Sweden and get cucked by a fake blonde.
>>109186718You're available fool if you think it's that easy. Why do you think people still harp on "prompt engineering"? Extra attempt to teach hopeless retards how to give concise instructions. What the fuck do you even do if you're convinced that's all you have to do?
>>109186822It's been over for about 19 years now, try not to stress about it.
LOL
>>109186653Just use the grill skill to develop a plan, clanker will ask you all the relevant questions to get most of what you want written down on paper before implementing anything.
>>109186841It keeps getting routed back to opus so it's hard to test
>>109186828I'm sorry for the distress, anon, I know it's not fun talking about retards doing retarded things. Here, have a snailcat.
New>>109186872>>109186872>>109186872>>109186872>>109186872>>109186872
>>109186757The fun thing about being a retard using modern agent harnesses is you can fill out the right column just by describing your goal and iterating with the clanker overtime. The clanker will ensure everything else falls into place as it should.
>>109186841Absolutely grim, fellas. What's worse is they only allow you to use 50% of your weekly usage on this pile of shit.Cheeto nigger did this. Dario too. Why does everything have to be run by jews?
>>109186511sounds like yt-dlp but DIY
