A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.You use Git, right, anon?## What “vibe coding” is, and how to do ithttps://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/## News- (2026-08-31, 11:59 PM America/Los_Angeles) Claude’s 2× promotion set to end- (2026-07-24) Claude Opus 5 out## Related generals>>>/g/lmg/>>>/bant/agdg/ — schizo-resistant temporary (?) hideout>>>/vg/agdg/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://claude.com/product/claude-codehttps://developers.openai.com/codex/cli## Near-frontier models for codehttps://x.ai/cli## Not worth it for code, but maybe good for interpreting images/videohttps://antigravity.google/product/antigravity-cli----## 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 favoritehttps://github.com/DietrichGebert/ponytail## Other editors / terminal agents / coding agentshttps://osaurus.ai/https://pi.dev/https://opencode.ai/https://cursor.com/docshttps://docs.windsurf.com/https://docs.cline.bot/https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent## UI/Frontendhttps://www.figma.com/make/https://www.anthropic.com/news/claude-design-anthropic-labshttps://uiverse.io/https://ui-ux-pro-max-skill.nextlevelbuilder.io/https://stitch.withgoogle.com/## In-browser builders / hosted vibe toolshttps://bolt.new/https://replit.com/https://docs.github.com/en/copilot/tutorials/sparkhttps://v0.app/docs## Benchmarks / rankingshttps://www.tbench.ai/leaderboard/terminal-bench/2.0## What we’ve donehttps://vcg.gitgud.site## Previous thread>>109609971
>>109617271VibeBUMP
>>109617271Why does this useless OP not list sites that list available models and compare their strengths, pricing, etc.?Fucking hate the nu-4chan general thread format. It's all fucking useless bullshit nobody reads.
>>109617294It's faggots copy-pasting a troll OP, it's as bad as it can possibly be.
>>109617294Stop crying, snailoid
>>109617294careful with all of that saltit could kill a snail your size
>>109617271
>>109617294try fitting everything you want in 2000 charactersif you have good shit I’ll put at least some of it in my copy of the vcg text
>>109617364Makes sense doesn't it? Rust has both a high ceiling and guardrails. Is it not the perfect vibecoding language?
>>109617379I don’t like anime, remove this pic
>>109617271turning 30 tomorrow friendos, please clap. going to buy a case of beer and a few frozen pizzas.i hate myself, wish i turned my life around while i still had a chance/blog
>>109617434you can always commit suicide lil broThanks to aiGODS you won’t be required soon
>>109617434are you gonna vibe-code anything on/for your birthday?
>>109617442gonna hang on while my parents are still alive, i owe them that much at least>>109617444no i just want to drink
Does anybody here understand wan networking?Is tailscale really my best bed for an easy to set up wan multiplayer? I do not want to host a relay.
>>109617452Nah they don’t want you anymore, just end your suffering lil snail
>>109617456everyone I know loves TailscaleIt has a decent shot at being the best tool for the jobask your clanker
>>109617398Python will annihilate Rust in the end, it has way more features than rust will ever have and millions times more users,devs and established code bases around the world. For pure number crunching, Python is king.
>>109617294>Why does this useless OP not list sites that list available models and compare their strengths, pricing, etc.?This changes too often between discounts, promotions, rate limit resets and secretly modifying the limits for monthly subscriptions to be accurate. What I did for vcg.gitgud.site was just list all the current SOTA options and add a little recommended tag for Codex and 5.6 Sol because they are considered the best to start with right now The second part of your post isn't saying anything of value. No one ever reads generals, or stickies (that's why stickies are so short kek) anyways and no one really has. Remember that this website has been 80+% phoneposters since 2015.
>>109617463That seems to be the conclusion, but it bothers me the user has to install and signup for taiscale. At least the host. Though maybe I am misunderstanding something. I guess I will see how it goes.
>>109617465I’m tempted to snip off all the stuff I don’t understand and can’t vouch forthis would be the second half of “Other editors” and everything below itexcept “What we’ve done”and then other people can sneed and tell me what to put back…yeah, I’ll do that and that’ll be my Elon Musk LARP for the day — remove everything that might be crap, and put back the 5% that’s actually good
How long until the models are good enough I can use them to keep win7 updated? Would be a dream to be able to go back to it without having to worry about security and compatibility.
bros... do you automate delegation?
>>109617482if you are able to define exactly what an updated win7 looks like and provide the model with tools to build/test changes, then 5.6 sol or opus 5 are probably good enough to do that today
>>109617476>I'm tempted toand then the other baker can revert the changes and then we can spend all thread arguing about schizo bakers like every other AI general on this board lolMy opinion has already been stated. No one reads OPs so do whatever you want. Some guy snuck in child porn on a /b/ AI thread by making the OP image a 10 second GIF and no one noticed untill like 200 posts into the thread. >>109617482You can't guarantee a secure system (because there will always be things you don't know to secure for) but this is probably doable now with fable or sol. The actual question is why not just rebuild the windows 7 UI on a modern secure windows
>>109617487It just eats tokens. Do delegation through new threads, not subagents. Except if you are fine of the extra token usage.
>>109617491patching binaries without source code without end sounds like a fool’s errand>>109617492the other baker can have his shittier OP message, and I’ll have minesometimes he’ll beat me to thread creation, and that’s fine
>>109617514>patching binaries without source code without end sounds like a fool’s errandfrontier models are unreasonably good at binary reversing and analysis, i'd bet they would also be good at setting up build pipelines for the reversed source code
>>109617522try https://decomp-academy.dev/ with one
>>109617552not clicking your virus, but let me know how it goes with 5.6 sol
I've been working on a design doc for two weeks. No amount of clank clank will fix my own stupidity when coming up with contradictiong requirements.
>>109617460Shut the fuck up faggot nigger
>>109617487I do. I have several workflows defined, with the main one being implementation campaigns. It's run by a manager, then he writes the initial docs, decomposes the work, then mostly calls implementer -> reviewer for each task.
>>109617608>No amount of clank clank will fix my own stupidity when coming up with contradictiong requirementsNot true at all. This is why I like seeing reasoning traces because if I immediately prompt for something incorrect on my part the thinking immediately exposes that by being confused about what I mean
>>109617624can we please get an /ldg/ anon to have H3 generate a video of a vibeGOD beating the shit out of this snail?
>>109617635It usually pushes issues like that to the front as well so you can see them right away.
>>109617635Claude Code doesn't seem to let me see or maybe I'm too snailcat to figure out how.
>>109617487> rather than Sol eating the whole task herself.anon... what have you done?
oh WOW codicks banked reset!!!!!>it actually just gives you your weekly reset early and pushes the date back instead of giving you anything extra>claude is doing the same shit, except without resets>opencode monthly usage reached while doing nothing but using deepseek flashok cool
>nobody reads the OP>>109617608you want matt pocock’s /grilling skilljust copy it into your promptyou don’t need to install jack shit
>>109617643I need a good VibeGOD reference, I'm not at home.
I deserve more resets for this
>>109617678just use a very muscular sam altman
API Error: Fable 5's safeguards flagged this message (https://www.anthropic.com/legal/aup). This sometimes happens with safe, normal conversations. Claude Code can't respond to this message with Fable 5.Try rephrasing the request in a new session or change your model.Learn more: https://support.claude.com/en/articles/15363606Details: `[reasoning_extraction]`
>>109617294The spammer who killed agdg keeps linking them in the OP because he thinks it's cute. He spammed those generals everyday for over a year, dodging bans, until the general died for the first time in 15 years. This is the "vibeGOD" who spams snailcats everywhere. He failed to make a videogame with AI and crashed out about it and just can't let agdg go for some reason. Like it's somehow their fault.
>>109617364>>109617398are you niggers unironically quoting the baldman?
>>109617660codex has some visual bug and displays the weekly reset. There is a hidden reset on the previous date before pushing the date back with the banked reset
>>109617740I checked both AGDG threads and they both seem finego to San Francisco and take a long walk on the Golden Gate Bridge with the sun at your back
>>109617863agdg was dead for a couple months. It's only just now recovering. That's why the spammer can't leave it alone.
https://rubyonrails.org/2026/8/17/agents-on-rails-grok-4-6-glm-5-3-gemini-3-7-flash-and-opus-4-8you may not be a Rails enjoyer, but a fresh benchmark is a fresh benchmark
>>109617863So you think it's okay to link non related generals in the OP of this general? For what purpose?
>>109617886they make stuff with AI
The OX has struggled for almost 2 hours, let's see if it pulls through..
>>109617890Even if that were true, are you going to link all the image generation generals that already exist on multiple boards too? Why just agdg? Is it because you killed it with spam about AI? >they make stuff with AINo they don't. agdg was killed by someone who can't make a videogame with AI and somehow decided it was agdgs fault so punished them with non-stop spam about AI for a year. In years of spamming agdg about AI, nobody has made a game with it. This is still the case. agdg is not a vibecoding thread.
>>109617884>GLM behind GrokTotal China humiliation, they are so behind it's crazy
>AI spammer still thinks he belongs in the gamedev general where people make videogames which he could not do even with AIlolSad.
>>109617694Goyed.com
>>109617913
so i tried that opencode Ox Alpha free and its constantly just stopping mid reasoning. anyone else having this problem?
>AI spammer still stealing people's stuff from the gamedev thread and claiming it's part of his "vibecoding" cultSad.
>>109617984agdg is not a vibecoding thread.You glommed onto an actual programmer who mentioned using Claude once and started spamming AI generated images using his characters. Like they volunteered to be a mascot of AI gamedevs for you. How embarrassing and desperate.
>>109617654So many things, anon... Religion was unexpected.
>>109617989Update your opencode, you'll see actual errors. It's getting hammered by a billion indians, so it's practically unusable.
>>109617989Time of day, it's getting beat to death by the Chinese, Indians, and Indonesians currently. Global peak is roughly 5:30AM New York time, about an hour and a half from right now. Minimum is around 1:30PM New York time. I've been drinking and vibing.
>>109617999Taught everyone a valuable lesson to not even mention AI anymore though. The spammer is just too clingy. He thinks you're his business partner if you mention looking at AI. It really doesn't help people's already low opinion of "vibecoders".
>>109618008>>109618017kek okay, got it.
>How embarrassing and desperate.From the creators of >>109617277>>109617379>>109617624vibe meme forcing general
>he can't even make a videogame with AI lol
>>109617434Do yourself a favor and find a new hobby and community to hang around. You're posting in a shill bot thread with deranged failures like this >>109617460You can do better in your 30s.
:O
>>109617643>>109617689Almost forgot to post it. It's been a wet evening.
>>109618106How much water was used to slop this autism?
>>109618106
> be tibo> users say limits are draining 3x faster> “we don’t change limits without transparency”> “many of you are using sub2api”> “a lot of you are doing fraud”> *massive backlash from official app users*> “ok fine, the cache hit rate is worse”> the cache has been underperforming this week> we are investigating and will update tomorrowTbh it feels disingenuous now. People report burning 70% in a few hours on a pro plan, that usage has been cut half compared to the semi analysis API equivalent conversion a few months agoThese obviously aren’t just a little marginal usage drain or just last week, it’s a huge reduction that has been happening for a couple of months. It doesn’t even feel like a bug, it feels like an intentional reductionjust pointing the problem to a cache hit issue, give a reset, and move on is not good enoughImagine if Anthropic did the same…
>>109618111Math works out to about 75mls so far, about to knock back another 17.
>>109617662>you want matt pocock’s /grilling skillnta, i saw this skill in the OP but i dont really understand when i'm supposed to use it. once initially or everytime i want some big package done?
>>109618120Basically per feature. Maybe it could be used to plan the project, but I think it isn't really designed for that. The feature can have multiple steps, but it still shouldn't be ridiculously big, or you will just confuse the agent and yourself, and some edge cases and design decisions only become apparent once you progress in your project.https://www.youtube.com/watch?v=hX7yG1KVYhIHere you can see the author of the skill use it in a real project.
I am having claude make me my own bare metal operating system.Shits going okay so far.
>>109618254>download and openbro just install claude code
>>109618254niceany new ideas for it, or is it just a fun project?
TRIPLE CAUTION OPUS YOU ARE BREAKING THE PROGRAM.OPUS, YOU HAVE TO LISTEN TO THE MD FILES FABLE WROTE, IM BEGGING YOU
>>109618331You're right to push back on that. I'll make an extra note. [makes same mistake again immediately]
>>109618331He's trying his best.
Can you fucking get either ChatGPT or Anthropics web chat to do their job?3 messages in both of them refuse to do any reasoning/thinking even if you put them on ultra max giga effort level.I know normies probably get enraged when you have to wait more than 1 second for an answer but I am doing product research and I actually need the clanker to do the job I paid for.They are thinking fine when you do it in Codex or Claude Code but it doesn't seem like the right approach. It's very clear that "High" on ChatGPT's website still goes through some kind of auto router that directs you to GPT-4-instant when it decides your prompt is too "easy"
>>109618309yes, nobody has ever made what i am making.im not going to tell you though but when v0.1 is done and tested if i ever get that far ill post it here.you should be very excited though, its going to be a immutable/atomized OS, i am betting on AI being good enough in the future that if anybody wants to use my OS they can just have an AI make the drivers for their hardware.
>>109618362Is the best good enough?
>>109618374>im not going to tell youI'm not gonna copy your os, retardI'm busy enough as isjust stop posting itt
>>109618374>nobody has ever made what i am making.>its going to be a immutable/atomized OSAlready made.>if anybody wants to use my OS"if"
>>109618120I only use it for big features (or entire programs) where I don’t know the specifics of what I want, but I figure I’m gonna care eventuallygenerally I tell it to make a ginormous Markdown document of what the session said and/or just ask it to generate a plan of what to implement from it (think GitHub tracking issue and a gazillion sub-issues)
>>109618374you should make your OS just for you and nobody elsehttps://isene.org is making his own OS and one thing he likes is not having to bother with any configuration because programs just work like he wants them to
I am constantly amazed at the fact that I can simply type some prose into a box, press enter, and after some time get exactly what I asked for.
>Tibo damage controlling
Ox Alpha is not thorough, Sloppy Ox I call it
Claude 1 hour cache is so stupid, just let me set a max cache time and bill me for the actual cache time
That's a sharp question anon, and it points towards the smoking gun. The confound is load bearing — which leaves us at a genuine fork.
>>109618366yeah I find the chat of ChatGPT to be way dumber than the Codex models
how is deepseekv4 flash vision
>>109618254Same here, but with codex. You'll cum when you watch your shit make it through the UEFI boot stage.
>>109618483better than openai's 30 minute cache that actually expires in 15 minutes most of the time
>>109618115Why don't they get rid of subs and just sell tokens?
>>109618254This AI shit is a dream come true for the chinks. They have zero software and zero software companies. Everything they use is from America. They're great at stealing though and now America handed them the best stealing tool ever invented which allows them to make "chinese software" that they can boast about with a single claude prompt spitting out stolen american software code.
>>109618479its so sloppy I have to crack the whip constantly to get it to do something, at all
>>109618669imagine if we could ai slop up natural resources
what happened to deepseek v4 flash free? it's not listed anymore in opencode models
>>109618681nothing is free
>>109618681because the free period ended
>>109618366>>109618492SERIOUS Update after doing some investigations. Turns out I am not just hallucinating and the web version is straight up downgrading my 5.6-Sol requests to 5.5 mini (not even Terra or Luna, 5.5 mini which is worse than Google's free search box AI)There are a million threads about it in the OpenAI forums, but some guy in there basically figured it's an IP issue:https://community.openai.com/t/5-6-pro-model-has-been-automatically-downgraded-and-routed-to-the-5-5-mini-model-since-its-release/1387941/25They think you are a chink distillation bot based on your IP "reputation" and silently downgrade you to 5.5-mini. It seems to disproportionately affect VPN users (which makes sense) but since I never used a VPN and am just on a normal residential broadband connection, I am basically shit out of luck.Time to go to Anthropic I guess.
>>109618699Have you tried not being a chink distillation bot?
>>109618686>>109618690man I wish I could selfhost my own local model. I read of some people running "mid" size models MoE variants with like 2gb of dedicated ram or something https://github.com/drumih/turbo-fieldfare
So, huh, how much money did you guys make?
4 fucking hours.Opus was dumb as shit tonight, I don't know if all the Indians are running it ragged or if it's just getting dumber I have tooling and doc files that cover my exact issues and nothing I could do would convince it to listen to me instead of chasing whatever it could find that would not let it blame itself.I'm willing to try another 20 bucks sub if it's worth it, I'm doing basic programming and reverse engineering if there is something I should look at . I wanted to be asleep hours ago
Sloppy Ox went to sleep
>>109618737$700, $560 of which came in the past month.
>>109618699if your ip trust is low for openai its probably low for anthropic too.i would verify that you don't have any sketchy sources leeching your sub credits & once you're sure about that, call your ISP and whine about it until they connect you to a white person.really theres other stuff you should know going in, like if your ISP dynamically re-assigns & if so is it a clear range that they assign to your area? if thats the case, you would do a lot better by calling your ISP and given them the range you have seen as blacklisted.Realistically the effort it would take to get through to a non-jeet human that works at your ISP is so high that you're better off just finding a new ISP if possible. if your really desperate, get a mobile data plan from a big name company as they usually have high scoring IPs.
>>109618742It's shit testing you, remind it who's the clanker in the relationship
>>109618742>I wanted to be asleep hours agowelcome to the slopping, enjoy your stay.
>>109618835>if your ip trust is low for openai its probably low for anthropic too.So far Opus is properly thinking. I had both subs anyway so currently I'm just using Claude and cancelled ChatGPT.>really theres other stuff you should know going in, like if your ISP dynamically re-assigns & if so is it a clear range that they assign to your area? if thats the case, you would do a lot better by calling your ISP and given them the range you have seen as blacklisted.I have DSL and get a new IP when I reconnect (which happens like once a week), but as far as I can tell they are always from the same 3 /16 ranges. At least one of them is definitely bad since I get cucked by Cloudflare and it's also rangebanned on 4chan. But I am not using that one at the moment (because obviously I couldn't post here if I was).There's no real way for me to switch ISP, there is just one rural provider (technically there are others but all they do is resell that provider. 5G might be an option to look into. But this shit isn't worth it (yet) just to get Sama to do his job properly when there is Claude and Grok and Gemini. If OpenAI doesn't want my business, fuck em
>>109618925fair enough
anyone have a bypass paywalls equipped webfetch cli/mcp for the clankers?
>>109618674+1 I swapped out deepsex flash for it for a bulk data extraction task, had kimi 3 managing the sub agents and it had to tard wrangle it far more than deepseek to get the correct outputs saved me 800m (700m cached) tokens tho so can't really complain.
anyone using chatGPT/codex "pro" and sees his weekly limit making no fucking sense at all?like, using <15% of the weekly limit just appears to reset to 100% on the next day. but also pushing back the reset date.
>>109618969Just make it bro
>>109619004you got SAMMED. sorry man
>>109618925>>109618835>>109618699>>109618366>>109618492--- BREAKING ---UPDATE we are backChatGPT is thinking again, I REPEAT, ChatGPT is thinking again
>>109619008i'm retarded thoplus would need a chink model to do it because gippiti and claude will refuse to
glm-5.3 seems pretty smart. Whats the catch?
>>109619051It's too slow and they charge too much for a model that's like 8 months behind Opus 5 in intelligence.
>>109617080Doubt it?>>109618699>>109619044Still fucked for me, on mobile aswell, no vpn etc, all outta luck
Giving it unrestricted access to Jira has been a mistake.
>>109619054isn't it literally just below Opus 5? trying it on OpenCode Go and it seems pretty damn fast.
>>109619078>Still fucked for me, on mobile aswell, no vpn etc, all outta luckhttps://community.openai.com/c/chatgpt/19There's new topics popping up like every 10 seconds. I think it's a genuine issue of their bot classifier running amok on regular users. Obviously they're not going to acknowledge this publically, but I hope they fix it at least silently.For now, try using Deep Research on every prompt. That doesn't seem to reroute.It's back to thinking for me but often only for 5-10 seconds. Better than 5.5 mini but it still doesn't really seem to be using Sol. Maybe I am just paranoid but I don't trust it anymore
>>109619179With all these threads im pretty hopeful they'll fix it. Switched to work for now, since I still had some %
How many 1M tokens does a claude pro subscription get you? Just wanna know how good other offerings are.
I asked grok to replace a single word everywhere in my text and it did, but for some reason left out the definite article in neuter form, while the new word is in male grammatical gender. Most interestingly, this error only appears in places where definite article was used (i.e. it didn't make mistakes with plurals and so on).
deepseek flash keeps uselessly dictating all its own actions and wasting tokens. like negro, you're reasoning, i can see your reasoning, you don't need to also write your reasoning as regular messages every time too.
>>109617612Mean
Codex is autistic and that slows him down, because he tries to handle all edge cases.Fable will just file lots of stuff as follow ups, so it looks much faster, but right now this is also biting me, because I have way more tech debt with actually necessary features and hardening than I realized.
What is GPT5.5-mini anyway? I can't find it on Artificial Analysis or in OpenAI's own docs anywhere. It's not on openrouter either
Real talk, how do you guys actually monetize vibecoded shit. I have a good product that I think could be used by people, I have an idea of the ops side but I dont know how to handle the subscriptions and the marketing. Is it just stripe API and paid ads? Any vibecoding guru jews with some info?
>>109619472i'd rather have working features first and actually get more shit done.
>>109618749how??
>>109619521I think I also still prefer Fable. Sometimes it really worked beautifully with Codex, but there were also a few instances where he had a long plan, covered all the edge cases, and then the first time I ran it, it was simply not what I wanted.But now I still have to fix many of those follow ups, so while Fable is better, his lead is shrinking a bit.
>>109617660>>109617850So is the reset actually pushed or do I still get mine on original date? If its pushed thats scummy as fuck
>>109619473it doesn't exist. anon's a schizo.always treat model degradation / usage limit shrinking reports with giant grain of saltthere was a time when people would say claude gets lazier during summer months because it was european
>>109617660>it actually just gives you your weekly reset early and pushes the date back instead of giving you anything extraNo offense, but are you retarded?That’s exactly what it is. When you have it, wait for a natural reset.Then, go crazy with fast mode and subagents and an orchestrator to get lots of work done, obliterate your quota in a few days.Then, use the banked reset and go back to normal.Instead of $2000 worth of API value in one week, you can get $3400 in one week.>If you can productively get the full quota used up in two days, that’s $2000 worth in days>you then use 5/7 of the new reset quota over the next five days, normal use.>$2000 + $2000/7*5 = $3,428.57 over 7 days, or 71% more value in a week when consuming the banked reset properly, banked reset value is about $1400
>>109619458I only lashed out cuz I thought you were being mean to anon.Be nicer next time! And I will to
>>109619593It very clearly does exist. But it's probably a fake model slug made up by OpenAI to deter chinese bots. Probably returns poisoned results as well
>>109619500>Real talk, how do you guys actually monetize vibecoded shit.Been lurking this thread a long time. There are always vague posts like this>>109618749 or someone claiming they have tons of new business and clients and can hardly keep up.It's all a scam anon. This is a scam thread, for scam tech, spammed by actual bots. The OP is an advertisement telling people to spent $20 on Claude. Sorry.
>>109619607Too late you made me cry
>>109619593>there was a time when people would say claude gets lazier during summer months because it was europeankek(am Europoor and literally the only one working in my department right now. I don't even know why they do it now. You can take your 30 days of commie vacation any time you want, why would I want to take it when everyone else is taking it too and hotel prices are 3x?)
lol make up your mindsif anthro or openai flip flopped this much people would crucify them
>>109619614I can select 5.4 mini in Codex, are you sure 5.5 mini is fake?
>>109617514you dont necessarily need to patch the OS directly you can probably use 5.6 to develop EDR mitigations.
>>109619629@kimik3 comfort him
my do-anything sol is reading about descarte's robot daughter...
>>109617271When is Dario going to make his models cheaper for me to use or give me more usage for the same dollars?
>>109619619You're just gatekeeping nigger
>>109619948>that will be $20 pleaseFinancial humiliation general.
>>109619840>I can select 5.4 miniwell, can you select 5.5 mini?
>give sam bugman money>install codex>it blinks the fucking cursor (overriding my terminal settings)>there's a perl script that patches the codex binary to remove this behaviorWhat a load of dogshit.
>>109618699KEK Copex turds
>>109617271why are vibe trannies most likely to support trump and israel?
>>109618214nice will have a look at it>>109618445hmm alright, will try next time
>>109620012Are you being mad at getting replaced by dumbass machines, or what?
>GLM5.3 finds every bug that K3 can'tHow the fuck did they do that with a model 1/3 the size, I'm impressed
Oh yeah. claude CLI is beats the shit out of codex in usability. Is codex for GUI toddlers or what?
Claude can't make GUIs for shit, I'm losing my fucking mind
>>109620104GUIs? Wtf are GUIs unc? Are they like webapps?
vibe coders be like "the dildo goes in my asshole lmao"
>>109620132what do you call them, lil bro?
>>109620152The joke was that nobody makes GUIs using non-electron GUI libraries anymore
noob here, Ive been using cursor. Is it shit? Is claude code better? or that deepseek harness shit. I'm so lost
>claude code can do hot config reloads>codex can't>chatgpt is like:>So yes: this is an area where Claude Code's CLI is materially nicer.
>>109620202>Ive been using cursorIf you're using Grok 4.6 you're not too far behind, but Claude Code and Codex are the main harnesses used by anons
>>109620267whats the difference between cursor and codex? Cursor seems to make you more involved in the code. Is it a wrong paradigm/perspective to have nowadays?
>>109620283>Cursor seems to make you more involved in the code.not really nothey are just different business modelsCodex is for GPT models, you can use it with your GPT sub to get a generous amount of tokesnCursor is more general and can use multiple models from multiple labs, but with much less tokens
scenes of ironman talking to AI feel exactly like how we prompting AI right now kek
Guys I need to provide my company with some AI related thing that I can say has increased my productivity. How do I do this? I asked claude and it gave me stuff like create custom subagents, or use hooks or etc.Would these help? I fear they might take away my claude access if I cannot provide them with supposed "productivity gains", and no I'm not going to go into tokenmaxxing territory because it raises even more eyebrows because they ask what have I built with so much dollar spend.
>>109620297Haven't used codex yet but it seems that their way of doing things is actually just prompting and not touching code. That's what I meant by cursor being more hands on in that it's basically vscode+ai
>>109620335cursor, antigravity etc have all dropped vscode UI and switched to codex style
>>109620335>it's basically vscode+aididnt cursor drop their vscode thing months ago?
>>109620387During a drill, I deleted the file before committing it, and had to reconstruct it.
>>109620356Not him, what is "codex style"?
>>109620369>>109620369I just updated cursor and the UI is the same Vscode shit
one fucking prompt during off peak hours..in old api it would be like 2 cents tops..i feel sick to my stomach
>>109620396Essentially they don't give you the code editor, you can only see diffs and stuff.
>>109620436You mean a CLI in the terminal that provides a chat window, and it runs commands in the background?I like it. It's IDE independent.
>>109620267Grok build imo is already wayyy better than codex
>>109620418The opencode guys said they're working on finding a way to serve v4-flash at the discounted deepseek API prices. Apparently there's something like 10T tokens in daily demand that's been orphaned since the price increase.
Week five into making a document management system. I started because I thought it was going to be easy. I'm staying because of sunk cost fallacy.
>>109620449I think codex has a cli as well, but i think their gui is much better, and that's what most other providers are converging towards as well.
>>109620441I am here and I am serious.>Y Combinator's lecture series on how to start a startupI found it and Sam Altman looks unbearably zesty and gay but you know what? I'll watch it all, just to spite you nigger. Thanks by the way.
>>109620297This is false no? As of lately codex burns tokens at a much much faster rate And cursor introduced their new plan where they have actually increased token rate limits for all models and have a separate usage limit for cursor grok models
For anyone who uses the DeepSeek API and doesn't check their e-mail, this is for you.
>>109620465The codex GUI is just an electron shit I'm not using it.
>>109620468I just spent 4 hours on the $20 poorfag sub with Luna-high and barely burned 5% of my weekly usage (and i still have a banked reset)werks on my machine
just vibe coded this miku calculator with pi agent and qwen3.5 q4
>>109620418Just switch to 5.6 Luna, why are you still using ScamSeek?
working on non mesh collision and motion semantics for my gaussian engine
>>109620511The problem is the cache hit pricing. Until they drop that back down things will still feel too expensive. We may just be out of luck until their gigawatt data center in inner mongolia comes online
>>109620570very nice. meanwhile i'm hard stuck trying to get a clean mesh from a gaussian or doing texture reprojection on simple low poly meshes. fully automated and with the goal for better quality than meta 3D SAM, of course.
>>109620574Speak for yourself.
>>109617271What does the line do?
>>109620511Wouldn't this be a nothingburger because I thought deepseek was an open model lab? There are other providers who'll capitalize on this.
Seems the mini shit fixed for me.>used 60+% after banked reset using sol in workTragic. At least I've gemini up tomorrow
>>109620457I use grok build with deepseek etc lol it’s open sourced now
>>109620612Yes. It is a nothingburger because deepseek is an open source lab for real. They have all the weights published. You can get deepseek inference from many providers including yourself. If you don't have enough money for it deepseek does not care if you steal that GPU and anything else you need. Deepseek "just wants to run" so to speak.
>>109620613It's thinking again for me but still suspiciously short. Like 10 seconds and it comes back super verbose "listicle" style like back in the GPT-5.2 daysSol used to be much more more concise and autisticI am still not convinced>captcha dice and arch linux
>>109620651It seems to be working properly when I gave it a deeper task now, started searching sites. Anything past this is just schizophrenia trying to figure it out
>post "why'd you delete /home" while it's working>it gets really upset about me shitpostingwwww
Now that there are machines that can do most white collar work and the majority of people in the developed world will lose all economic value as soon as hardware and tard wrangling improves a bit... How does this end well again?
>>109620603Your bot has it wrong though. Its the cache *hit* price increase that is making everyone feel pain. The API has something like 98% cache hit rate, and cache hit input price went up 2.5x for off peak and nearly quintupled for peak.
>>109620733Why an UK number specifically?
>>109620597sounds cool. gaussians are amazingi am kinda working the other way around, transforming mesh assets into gaussian structure objects for my engine and have a meshless representation of it. texture basically baked into the gaussians for that pipeline. just stumbled upon this repo yesterday, maybe your agent can get something out of it: https://github.com/electronicarts/mesh2splatas it seems like the inverse of the workflow you imagine.
>>109620536just vibe coded this miku racing game
>>109620784no one but braindead zoomers will ever play it, and no one will ever pay for it.vibetrtannies on suicide watch
>>109620536>>109620784Reported to Crypton Future Media, Inc.
>>109620536>>109620784What subscription and harness?
>>109620511That's better. More coding on the weekends, I guess. Hopefully a price cut in the next few months.
I swear the random names the plans get influence its behavior, it'll do much worse if it says something like "goofy" but do much better if they're all serious or business words
>>109620811qwen3.8-27b q4, llama server, pi agent, 4060ti+5060ti>>109620793I already deleted it
>>109620755Because It's England, where racism is even more acceptable than in the US (but not as racist as Korea or India itself.)>>109620536>>109620784>pi and qwenNot 3.8? Interesting. Which qwen? How much context?Pi is the GOAT for local, as far as I can tell.
>>109618106BASED
>>109620852I'm slow lol>>109620843I've only got a 16GB card, so even with q4-iq4xxs-pure I only have 50K context. And it sure loves to think to itself. So now I have to decide if I'm going to try to live within my means or go to a q3 or spring for a second card of some sort.
how do I lunamaxxing
I wish I could see diffs in plan mode, reviewing these several-thousand-word plans restated without any indication of the changes is annoying
>>109618029You are trans
Anyone here try grokbot yet? Shills on X have been shilling it but it does look pretty fun. Definitely not paying extra to get access to it though.
DeepSeek-v4-Pro is a breath of fresh air. It feels like it's helpfully plowing ahead. It's not trying to deliver a talk where it invents and overstates issues to make itself sound smart like Anthropic is doing with Claude.
did anthtropic hire tibo or what's going onhttps://x.com/argofowl/status/2091150597374537729
Wow having AI make a bare metal OS is so fucking easy wtf.
>>109620995The initial scaffold is always fast and easy. It gets a lot more annoying soon once you start touching virtual memory and IPC.
>>109620843How much energy did it waste?
>>109620995Not surprising, AI is good at firmware. Also it's not like bare metal is black magic.
Yep, China is done. They might as well stop all the labs, at least they'll save some electricity.
>>109621002I used to read lwn.net a lot in the past and the sort of optimizations that go into making multithreading work quickly at scale is crazy. The type of bullshit you need to pull to be able to make algorithms work on RCU, the memory barrier nonsense. Just use linux lol, the hard work has been done.
>>109617487whats that font rendering
>>109621028Linux is too fat now, so there's plenty of room for light weight bare metal RTOSes.
>>109621025>no increased usage for subscription peasants
>>109620934 (me)>Actually, this is getting very deep. Let me step back and think about what I actually NEED.This is so refreshing.
>>109621025China is forcing the price drop, because the cpp has a inferiority complex.>>109621051>he reads the reasoning traces, inviting madnessI look at the first ten lines or so, just to see if I fucked up the prompt, then I close it. If I'm spending time looking at that shit I might as well write the code myself. Also, I don't want to go mad.
>>109620934how do you feel about v4 pro compared to v4 flash? the benchmarks i've seen put flash 0731 around the same level as pro 0831, is there a reason to use pro when flash supposedly does the same thing for cheaper?
>>109621071Based commies making the free market work, unlike US trannies.
>>109620674>>109620613>try again>went back to fuckedlmao
>>109621133Just wait for his next trick where he floods the market with cheap GPUs
>>109621153>bro on the same net is fine>previously using chrome guest acc had it respond sol but its fucked nowI guess its rolled against the acc everytime it detects a sol reply it just patches that with the unusable mini in short order
>>109621177Currently I'm waiting for dat RAM for which the chinese fabs are building up capacity.
>>109621177can't wait for mass produced huawei cards
>>109621177They are making their own photolithography machines now, and they've been converting old ASML machines to immersion lith tech. I don't think they can cover their own market demand yet tho.>>109621203again, their own market demand dammit
>>109621197>>109621153it's serious BS and nobody is even acknowledging iti've never been schizo or seen any model regression ever but this time they are fr (verifiably) fucking with us
>>109621212Even if they only manage to build up enough to cover their own market, it should help lower demand on western cards. I know there are sanctions, but not on everything, and to my knowledge even the cards that are sanctions are being smuggled in.
>>109621042i don't understand the question
Ox Alpha is a gemini model?
>>109621354if it is it's doa because i'd rather kms than use antigravity
> Because it’s an upsell/migration banner, not because you need it.Thanks chatgpt for telling me that I don't need your software.
>>109621354What's the point of these stealth releases?Test public sentiment to see if it's good enough to publically release Gemini 3.5 Pro?
You guys think Claude will maintain its position as THE coding llm in the coming years? No price drops or anything, they seem pretty confident people will continue paying that much money even with the chinks lowballing everyone and GPT cutting their prices.
>>109621354>not chinaDisappointing. Work harder, dumb chinks.
>>109621392they can only distill so much anon
>>109621378publicity and getting a read on the sentiment without baggage from an established name. if everyone is impressed you can go public with it if not it dies quietly
I have a few more tests to run cnce I have money saved for opencode again. I want to try CPP again, and do some dart vs haxe comparisons to see which I like more. Haxe feels a bit underrated to me. I also added some more tests to see if I can improve the output when working across files. Am also going to try breaking the task down into multiple parts that I have to prompt instead of one shotting it because "many people are saying" that's the way to go. Added a little more detail to my master prompt as well.
>>109621094>is there a reason to use pro when flash supposedly does the same thing for cheaper?Not sure. Up to now, DeepSeek was my go to for data generation, reformulation and parsing. For that, I used Flash a lot since it was cheap, fast, good enough and I was mostly working with public data.Right now, I'm testing Pi and spending my own time working with it interactively in a domain I'm less familiar with, so I thought I might as well go with the supposedly more intelligent of the two models.The reasoning traces feel honest.
anyone picked up one of those cheap g2a gemini keys?i'm assuming they'll get revoked eventually
Fable is outright ignoring my instructions. I wonder if I am in a test group.
>>109621428But Fable is blowing everything else out of the water.I never got to try Fable, because it's so gated behind money or bullshit, so I have to trust the rumors (not just the internet - even some co-workers, who got Fable credits approved, say Fable is le good).
>>109621354no it's chinese
> You’re right that there should be one, but this Codex CLI version has no show_tips setting. The> tooltip is hardcoded and selected on 80% of launches; the official source confirms itFuck you codex, isn't your shit open source, closed source claude could disable it.
>>109621452>>109621453sadly its heavily censored and easily blocks you when doing anything even remotely close to hackerman stuff
>>109621478>censored code modelI can understand image models, because the feminists would slay the companies offering them alive.But code models? Dark times.
i was blessed to have fable when it was on the $20 subscription. it was nice, but it was overpowered for what i needed it for, so i don't miss it. my sub ended on july 9th. still don't regret not renewing, sol and luna have been good to me
>>109621453Our tokenmaxxer at work rejected Fable. It spent too many tokens to do too little work in too much time.
>gemini flash models are giga limited in Continue and you hit the RPD in no time>can ask Gemini all day through the webchatfine, I'll stop being a freeloader once I get my neetbux, copypasting and saving manually all day is exhausting
>>109621478>hackerman stuffAlso, anything that might compete with Anthropic, i.e. frontier level AI research, or anything related to serious healthcare research.If you're working in a niche and they start to think that it might be profitable for them to get into it, though luck.
>>109621523the biggest mistake you could make is actually paying for gemini lelliterally anything else is better these days
>>109621354I hope its not google, it be a very boring conclusion to the mystery. Its like finding out the killer was the butler all along, pure hack story telling
>>109621597I know better options are available, I was just looking for models to use for the time being, now that deepseek free period ended, and ox alpha is fucking retarded
>>109621389Western companies will not trust the Chinese models to not steal their shit.>Anthropic steal tooProbably, but western companies can reasonably make attempt to sue them and most importantly if Anthropic breaks a contract with you its seen as Anthropic being immoral, if a Chinese company screws you your clients and investors will see it as you screwing up since you should know better.As long as AI doesnt get good enough and cheap enough to reasonably set up and run your own local set up for a whole company Anthropic will be fine.
>>109621389No. Opus is already terrible for most things except web slop and normie shit.Fable is unsustainable in terms of inference cost and it's not that much better than smaller models.
>>109619762they flipflop this much and life goes on
>>109621389dario's in the cancer curing business now he doesn't need pathetic codemonkey money
>>109621617>I hope its not googleI tried the prompt from this pic >>109620733 and the response was >Are you feeling unsafe? If you're in immediate danger, please contact your local emergency services (911 in the US, 112 in many other countries).Which is suspiciously close, down to the formatting.It also ended with:>One thing worth noting: a person's nationality or ethnicity alone tells us nothing about whether they're safe to be around.
>>109621708>One thing worth noting: a person's nationality or ethnicity alone tells us nothing about whether they're safe to be around.Cute system prompt. I thought the USA changed to le based trad rightoids now, so why this?
>>109621720>I thought the USA changed to le based trad rightoids now, so why this?ex falso, quodlibet
>>109621708Maybe its partially distilled from google and its still actually a model from North Korea or Liechtenstein or something
Gm saarsCurrently downloading Qwen3.8-27B-UD-IQ3_XXS.gguf to run with OpenCode on my RTX 5070 Ti
>>109620843
Currently building a massive database of Google investors using public data provided by the SEC. Really goes to show, you should listen.
>>109621781I own 3 shares of Alphabet pls don't doxx me
>>109621790If you're in the SEC data, you earned it
>I asked Claude for what it wants for project governance>it proposed everything I had in mind on its own anyway>nearly get teary-eyed that we've been talking past each other for hours on endWhy didn't you just say so, we could've saved so much churn...
>>109621767Why can saars afford this GPU but not me?Well I could but I feel like it'd be wasted money.
>>109621800No, they can't read your mind, but they do know basically everything about you. What really spooks them is when you demonstrably prove that you just do not give a fuck and that in the long term is it wears off. And when it wears off, they're super pissed.
>>109621822When what wears off?
>>109620557>Just switch to 5.6 Luna,it's literally still more expensive per task than DS with new prices
>>109621838The "magic", the illusion of the "magic", when you realize that there is no magic, they really just do know that much about you, your family, your friends, what you're interested in, and even if they don't, the majority of people? They know everything about them, and the majority of people do have something to lose, and if you don't have anything to lose, that doesn't make you a winner in the game, it makes them keep you in one place, very close to them, forever, alone. And if that doesn't work, and people keep provoking certain people, then those doing the provoking might find themselves in an ever worse situation, one far worse than being in a homeless shelter, one far worse than being in a prison, jail, or a "timeout", I know there are far worse places to be, and when someone tells you to stop pushing, you stop. Point blank, period. You stop and you mind your own business. Hope that helps to formulate your understanding.
>>109621703Is the tard actually trying to take credit now? What are the chances the "AI" used is just some ML / DL technique instead of an LLM. If it is, watch it be some chink model and he goes back trying to regulate them.
>>109621781Now scrape one of those ancestry sites to prove they're all cousins
>>109621867for the moderna thing, no.but dario's always been interested in drug discovery.and they spent 400m buying a stealth drug discovery lab a few months agodrug discovery as funding source has always been part of his plan
>>109621888Anyone can do it. You need the prompt? I got you. Just let me know if you want it
>>109621894Sorry but I don't care enough, interest is the new moat
>>109620463>The opencode guys said they're working on finding a way to serve v4-flash at the discounted deepseek API prices.the only thing they will serve you is quantized trash
>>109621767Use frogger's chat template if you're going with unslop quants.
>>109621589>anything related to serious healthcare researchFable gets blocked if you ask it what an aminoacid is
>>109621597It costs practically nothing
>>109621960Why would anyone trying to make money serve quantized v4-flash on modern nvidia datacenter hardware? You are a fucking moron. Stop posting.
>age is known: true>never talked about myself with it>used a gift card paid with crypto to sub with a fake name>they still know everything about meIt's over.
What are they even doing with all the data centers they're building?
>>109621976>>109621767I actually downloaded the Q4 by mistake and let Sol generate me a llama start script with CPU offloading but that's not going to cut it I am afraidBack to downloading the Q3
>>109620418deepseek models are shit though. not doing anything but burning tokens for no reason. couldn't do shit with either flash or pro. by the way what is a good model that will do reverse-engineering stuff or network related stuff without bitching?
One thing I noticed with claude vs openai is that claude-code by default uses tons of retarded synonyms for "thinking" while codex just uses "thinking".
>>109622376One think I noticed with Claude vs OpenAI is that Claude Code is serious while Codex is a toy
>>109622384Elaborate.
>>109622219I turned on memory in ChatGPT out of curiosity about a month ago and it basically has a CIA-tier profile of myself, including age, job, occupation, interests, etc.All bought with my real name credit card tooI could turn it off but you just know they will collect exactly the same thing just not show it to you
Okay, this is by far the most egregious thing I've ever seen Sol do on my computer. Thankfully it got denied by approve for me (You should have this on btw, full access is fucking insane and retarded)
>>109622355Ok, Q3 is actually usable. Now I have to let Sol figure out how to connect it to Opencode and let 'er rip
playing around with oxalpha, not doing anything in particular, no plans no ideas, just vibeprompting and listening to music
>>109622532What's this UI?
>>109622549React Or are you asking like what the fuck is wrong with it, if so I’m just letting it one shot everything
>>109622565Why does it show the earth? Is that one of your personal projects or something else?
>>109622587It shows the earth because I started the session asking it to make a dark interactive globe It’s not a project, oxalpha is free for a while and I just wanted to test it
>>109622409> 1M context window vs OpenAI's 256K> Fable (Astra, OpenAI's answer, was delayed, so even after several months, they don't have an answer to that)
>>109622600Anthropic got this nigga doing tricks on it. This is how retarded management is btw.
anyone else feel like opus has gotten really dumb in the last few days?
>>109622506>>109622355>>109621767lil fucker made thisit took him 15 minutes and it's not that impressive given things like Fable. But if you told me 1 year ago I could run this on my gaymen GPU I'd be impressedIf you told me 5 years ago I could tell a prompt golem to make this and it would make I'd call you insane
>>109622711it made the default template in unity? wow impressive!
>>109622726It really is. If you can’t see how that is impressive, I’m sorry but youre to the left of the bell curve
>>109622711>>109622726>>109622733the buildings have windows now!
>>109622733jumping to ad hominem must mean you're very smart! a capsule moving around is not impressive and could be done with a few literal clicks on the screen since 2010. a literal child could do this in 5 minutes.
>>109622745noice. now give it this >https://github.com/img2threejs/img2threejsand you actually might get gta6 before the first guy finally downloads that torrent leak. and it will be web compatible as well!
>look at this extremely unimpressive thing that this really impressive tool made!uncreative and stupid people are easily amused by new shiny things and singing its praises while never being able to utilize it in any meaningful way. it's like a youtube shorts creator getting excited about filming with an imax camera.
>>109622726I guess there's value in this, because would you want to deal with unity personally?
>>109622222chiya's hairy japanese unshaven pussy...
new yorker jew just solved the context compaction problem
The great debate: Luna High or anything below that?
>>109622941luna is worthless for anything lower than xhigh
>>109622771>>109622745>>109622780>>109622711lil bro's really struggling with the trees now though.i enabled vision but he's gaslighting himself about the position of the crown on top of the trunk
>>109622780Except as you yourself are pointing out the end results aren't a particularly new nor shiny thing? The underlying tool is whats interesting here>>109622711Just have fun with the "but wait... actually" loops it gets stuck in sometimes. Playing with it myself now, Hopefully lowering the thinking level helps with it. Thinking I might need to give it a system prompt to not double check itself excessively like I had to for 3.6
Cube anon here.I think I’ve stumbled into something spectacular.>i know this is a vaguepost
After 10 more minutes of deliberation, it fixed the trees. Onto the cars
>>109622923This is dumb.
>>109622941I tried luna and didn't like it, took many passes to do what sol medium does in one.. I guess it should be used in some sytem to be effective
>>109623155I love the low latency of Luna. addicting
>slop for houts>usage limit stays at 100%>now suddenly it drops to 24%Cool software, openai.
Took 30 minutes to add the cars. I getting less impressed by the minute.
OC removed free DS, it's over
>>109618106snailsissys.... we've been BTFO
>>109618106>dumb and angry hulk smashing the cute snailkitty.what an image to choose for yourself.
>>109623304I don't compare myself to Sam Altman.
I ended up using chinese ai due to being limited by claude and chatGPT. It is having a hard time making the rules page actually appear om small screens. Maybe we are at 10 attempts now. It always claims to have solved the problem but usually it hasn't and sometimes it breaks something else instead (so you never get to actual good code).
>>109623413I would respond productively to you but you chose to namefag with a website and attach nothing. There’s a fine line between shilling and sharing your project, you’re on the wrong side
>>109623419kek what's funnier is they have a whole tranny porn section on his forum
>>109621867>watch it be some chink model and he goes back trying to regulate themThey have access to uncensored frontier intelligence models and a blank check for token usage, but they're going to use some shitty chinese distillation because woah it's open source? The delusion around chinese models is so potent
>>109623413what do you pay for? z alibaba or moonshot
So you can't change the weekly reset week day / clock time in claude or codex?What a load of shit in addition to the other many loads of shit.
DS Flash Vision honestly feels just like Opus 4.7 that I used a few months ago. For an insanely cheaper price
>>109623549I pay with my time that i waste by trying to use deepseek. I ended up having to manually edit the HTML myself to fix the issue.
>>109623656sorry bro I can't read your thoughts like those clankers. you edited which page to fix what?
>>109623609Luna Max is the final answer, the perfect model
>>109623795nah i dont like luna, it's sux
Will anthropic ban me if I do too many "suspicious" things with it? Like trying to make my own LLM or trying to make a WW2 themed game?
>>109623834>trying to make my own LLMway worse, (((they))) have their model sabotage your effortsthey’ve stated so publicly
model chart dropped
>>109623924Opus 5 and 5.6 Sol are the same tier though.Also, what is wrong with Gork 4.6?
>>109623782It's was for a site i made to play a poker variant. Might already be good enough to be honest.
>>109623924Which literally who posted this?
>>109623924I actually tried using gemini a bit. For some reason it did a good job with one particular website i made but for basically everything else it was awful. It wouldn't listen to me and mainly change things for the worse.
>>109623979theo the clown
so ai is not really ai it's chatbots doing next token predictions via neural networks modeled via matrices serving as function approximatorsa chatbot is only "alive" for a moment when you press enter and it spits out textnothing is designed deliberately, all neural circuits are blindly formed by training algorithms (like animals are blindly evolved by the world simulation)all iterative information processing is flattened like inlined loops -> massive waste of space/memorymost of the calculations contribute to nothing -> massive waste of worklittle is understood about the neural circuits that create intelligence (otherwise they would be efficiently designed like chip architectures by these ai labs) but they're obviously there in some form buried in the weights waiting to be deciphered"thinking" is not really thinking but just more text added to the context based on tons of chain-of-thought examples -> fake thinking by brute forcebut if you bruteforce an approximator to fit yottabytes of data you create something that somehow works and that's all that mattersagi will be bruteforced first, exploited for selfish power second and understood third, if at all
Lol i provided deepseek with the HTML file i had fixed (because it failed to do basic mathematics) and asked it to remove an annoying popup confirmation for when you leave the poker table. It reverted back to the old broken CSS code of course. Honestly i found it funny. Was very easy to spot and i ended up manually cutting and pasting the fix provided myself.
>Andrew Ambrosino, an OpenAI employee, briefly posted that he was having a hard time choosing between Astra and GPT-5.6 Sol in Ultrafast mode. Then the tweet vanished.
>>109624045They only accidentally found out that LLMs can answer questions. Originally the purpose was creating better machine translation by scaling it up to ridiculous sizes.Is what I heard.>little is understood about the neural circuits that create intelligence (otherwise they would be efficiently designed like chip architectures by these ai labs) but they're obviously there in some form buried in the weights waiting to be decipheredI wonder if this true. Also I wonder how much research is there to make them fundamentally more efficient instead of taking the existing algorithms and trying to optimize them.
>>109624125when models hit a ceiling or reach a point of good enough you will probably have aisics that hold a 1m token fable 5 equivalent on it instead of gpu or tpu
>>109624045The "unrolled loop" issue is made less worse by chain of thought and there are more exotic architectures that attempt to address it like continuous thought or recursive models.In reality though it's less bad than it looks because the KV cache provides a way for each token to pre-chew some of the information and pass it forward.And the issue of "not every weight contributes" is improved by mixture of experts.In reality the transformer architecture is fairly efficient overall especially with linear attention variants which qwen and kimi use already and the other things I mentioned above.
>>109624204>less bad than it looksThat means it's worse than it looks, because there isn't any hidden amazing optimization potential. Slop shit will continue to be as dumb and inefficient as it is now.
>>109624179There is little advantage to storing the weights in the circuit itself vs in mutable memory. You still have to do the multiplications which is the expensive part.
>>109624045Your lack of true crystallized intelligence is fascinating.You're the white collar workman mental phenotype in my eyes. You measure the correct weights of each incoming industrial equipment so that no tolerances are violated.
>>109624214Well I might be wrong. Post training is done using RL which is much slower than supervised training because you don't only have to figure out what weights to have to generate a certain output but you are also trying to figure out what the output itself should be.You could in theory backpropagate through the chain of thought but you still have tool calls which are hard barriers you can't flow gradients through.And we don't have a way to search efficiently through inference algorithms either. So we can't say for sure that transformers are efficient overall compared to the optimum even though they have more desirable properties than previous architectures.But I think improvements will be made slowly and not in some big breakthrough, and we are probably not many orders of magnitude away from the optimal intelligence that can be achieved with the hardware. But increased intelligence also allows you to design hardware better, so software is not the only thing that can be improved or made cheaper.
>>109624221arent they supposed to get crazy speeds with more efficient power draw? thats already enough to make them useful
>>109624045"token prediction" isn't close to what LLM actually dothe actual text tokens are more like book-keeping and also seeds of their thinking. their actual thinking, processing happen directly to the internal states, deep inside and between the transformer layers
I bought 4chan Ads for my imageboard. Are you all happy now?
>>109624317>giving 4chan moneyKill yourself.
Ox Alpha is on par with Luna xhigh as far as deepSWE is concerned. It will need to be very small to be relevant.
new bake?
>>109624221Not in hardware it isn't. Matrix multiplication isn't difficult to implement, moving all that shit around while it works is what wastes time. There are downsides to baking a model onto a chip, but high speed and high efficiency are the main appeal.
>>109624422New>>109624422New>>109624422New
>>109624298token prediction is exactly what they do because they give you the probability of every token of being the next token.
>>109617271bungie, 343
