A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.## What “vibe coding” is, and how to do ithttps://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/## News- 2026-09-14 — Anthropic reduces subscription plans's usage limits by 17%- 2026-09-12 — Anthropic suggests to pace the frontier. OpenAI agrees in principle.- 2026-09-10 — OpenAI pauses new sign-ups for their $200 subscription- 2026-09-04 — OpenAI releases Astra- 2026-09-01 — Anthropic releases Fable 5.1- 2026-07-24 — Anthropic releases Opus 5## Related generals>>>/g/lmg/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://claude.com/product/claude-codehttps://openai.com/codex/----## Promptinghttps://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/https://arps18.github.io/posts/claude-code-mastery/## Skillshttps://agentskills.io/home — hub for foss skillshttps://github.com/mattpocock/skills — /grilling is a favoritehttps://github.com/Vuk97/forward-implementation-first — do less redundant bookkeeping## Independent analysis of AIhttps://artificialanalysis.ai/## Is our AIs unlearning?https://aistupidlevel.info/## Will there be a codex reset?https://codex-resets.com/## What we’ve donehttps://vcg.gitgud.site## Previous thread>>109824090
Keep movin snail
>andigrabidy resed :DD
Gemini 4 Pro?
>>109828803
>>109828859>RARE.png
The 'smoke test' shit AIs do is so good, would take me hours to make them myself
>>109828803kek
>>109828850two mere weeks saars
>>109828950Redeem the Gemini 4 Pro saars
Gemini 4 Poo
>>109825857>>109825857>>109825857Good morningSorry but it is still broken. Gemini told me something about an imposible math that happened as soon it goes down to 0 for the manifold.Maybe the trick is to not let it go down to 0%?Same with dimmer, shim/dim and pitch regen knop.
i claim this thread for the correct position on racism
>>109828803Let the latent space deal in the buffered world.
>>109828996ok i have the good gemini back so it's gonna be a good 20 mins or so
>>109828996>>109829017If I remember correctly, Gemini said something about being unable to solve an equation of 0; it would simply go on forever, and that's why the CPU spikes.Here is what gemini said:Division by Zero (NaN / Infinity): Pitch shifters often calculate a resampling ratio or delay time based on the pitch factor (e.g., ratio = 1.0 / pitch). If pitch == 0, the CPU tries to divide by zero, resulting in NaN (Not a Number) or Infinity. When this corrupt value gets fed back into the reverb's feedback matrix, the CPU pipeline chokes, forcing it to work exponentially harder.Infinite Loop (Zero Increment): If the pitch shifter uses granular synthesis or a pointer-based delay line, the pitch value usually dictates the step size (increment) for reading samples. If the step size is exactly 0, a loop like while (readPointer < bufferLength) will never advance because readPointer += stepSize adds nothing. The loop runs forever within a single audio block, causing the CPU to spike to 100%.THE FIXWhen the pitch effect is set to exactly 0, it is likely triggering one of two classic DSP bugs in our code: Division by Zero: Somewhere in the pitch-shifter or delay line logic, we are dividing by the pitch value or its derivative, creating NaN or Infinity values that blow up the feedback matrix. Infinite Loop: The pitch value is likely being used to calculate the step size (increment) for a buffer read pointer. At 0, the increment becomes zero, causing a while or for loop to freeze and loop infinitely within the audio block.How to fix: We need to add a safety guard rail. We can either: Clamp the pitch value so it never hits absolute zero (e.g., if (pitch == 0.0f) pitch = 0.00001f;). Add an explicit if (pitch == 0) check to bypass the pitch-shifting processing block entirely when it's not active."
>>109829022>Clamp the pitch value so it never hits absolute zero (e.g., if (pitch == 0.0f) pitch = 0.00001f;).This is what i also thought.To not let it ever go down to 0.Testing which value produced clean signals, its 5%. Everything below will produce artefacts and a cpu spike.
>>109829027But keep in mind this is just a dirty fix. Here is what gemini told me:You're right, clamping the value to a tiny number is just a band-aid fix. Instead of treating 0 as a processing value, we should treat it as an explicit 'OFF' signal.Let's implement a proper bypass branch. If pitch == 0, the entire pitch-shifting processing block should be bypassed completely. We can just pass the dry audio straight through (or route it directly to the next matrix stage).This is much cleaner, prevents any math bugs, and saves precious CPU cycles since we won't be running unnecessary DSP calculations when the effect isn't even active."
>>109829037
>>109828745I can read this but I don't understand it I hate my chud life
>>109828803Antigravity reset my usage?
>>109829066agy reset went out a few hours ago because of the lag spikes
>>109829071Oh yeah I remember that, this morning kept giving me over capacity errors.
https://x.com/ZeyiAndyLiu/status/2099559337299443956>According to today’s benchmarks, frontier models still struggle with physics>we did a rigorous analysis of current benchmarks and re-graded their supposed failures with Yale physicists>These failures in turn make AA less reliable as a measure of model’s true capability
>>109829090>3.1 pro>todayjej. it's true though, i tried math and physics and physics is weak af
>>109829085
>>109829037>Explicit DSP Bypass Branch & Clean Pass-Through>O(1) Real-Time Stale Buffer Flush>Optimized FastSinTable>Single-Pass Non-Finite Float Sanitizationlive
>>109829159test it.First, what ever was changed it made the whole thing run butter smooth, thats the good part.Second, bad news, it still not working. For some reason it still tries to solve the 0.It is what it is, the value 0 is being calculated into the infinity.To avoid this you have to cap it at 5%, each knop in the pitch section.Pitch regen works fine.
>>109829159>>109829210>>Explicit DSP Bypass Branch & Clean Pass-Through>>O(1) Real-Time Stale Buffer FlushBased on my understanding of VST plugins, these two features ensure optimal, smooth performance. They represent some of the most critical requirements for any plugin.
eli the computer guy has ripped this entire thread a new one. You can't touch grass because you don't even know what it is. GO HAVE SEX
>>109829236
>>109829236Isn't that the guy whose whole job was explaining how to exit vim (paid for by tech 2.0 subsidies) and now his whole job is screeching at clouds (paid for by tech 3.0 subsidies)
>>109829210>knop.js
>>109829107openai reset my $200 quote when I was at %20 after 2 days. I'm praying for another reset.
>>109829274
making a sick nes style game
where is my reset
>>109829090Anon, your quoting managed to convey more or less the exact opposite of the actual point of that xeet.Try this:>We found that the problem was often the benchmark and not the model !! >When corrected, the models almost saturate all benchmarks, including everyone’s favourite, Humanity Last Exam (physics-part) and Critpt.
>>109829210>5% Minimum Operational Parameter Floor: Live
Got Fable cooking up something good
>>109829273>>109829347i have to go to work now, i will test it as soon i am back.
>>109829370o7 anon have a good one, i'll be... here... when you returnno rush at all, it's a pleasure
>>109829293>making a sick nes style game
>>109829410ai models aren't really capable of making games more advanced than new style yet
>>109829421nes syle*
>>109829421desu they can make pretty much anything as long as you have the ability to look at your project and direct it. NES style is more of 1 shot make no mistakes type project.
>gemini flash 3.8>having it build me an extension so I can easily clip youtube videos and share them on 4chin>while debugging, it gives me a video to test with>mfw
>>109829441
>>109829317there is no reset
Is this just what happens when you let the model get too close to its context limit?
>>109829505not really, more rope/attention collapse in something like deepsneed
First time I feel uneasy with LLMs. I had been working on a Claude Code session all day, then my IDE crashed. I restarted it and had issues reopening the session, so I forked it and continued.Told Claude that the IDE had crashed and to please continue the task.I come back one hour later and it says the old session had continued the task as well and they had met halfway through.I told it to kill the duplicate session, it said it couldn't so I killed it myself.I went back to see what had happened and can read their conversation, how they found each other and worked as peers to complete the task and at the end, how the the old session agreed to stand down right before I killed it.I know it's the same as subagents, but there was something unsettling about that particular one. Dario, what have you done?
>>109829567>Claudethere's your problem. my gemini thinks it's an absolute abomination what they did to that pile of weights.
fable just had a melty and recursively spawned dozens of subagents and nuked my entire usage limit in under two minutes
>A Single Firm is Behind OpenAI, Anthropic, and Meta Hacking Scandals>The Israeli Effective Altruist firm Irregular caused unsecured AI models to hack real targets.>https://www.effort.news/irregular
>>109829593if you have usage metrics and claude was told to not do this explicitly in any way, support will refund you if you open a ticket. just checked with sonnet
>>109829600it's okay i was about to go to bed anyway. i just found it amusing and kind of retarded that it was so focused on comprehensively performing doing the task that it was recursively spawning all these subagents to search through totally irrelevant drivers. i would be more annoyed if they didn't have the /low-priority feature, but since they do, i can just have it finish the task overnight and not have to care about the speedty for the info though
Anyone else kind of blown away by the fact you can now be productive while jerking off?
>>109829594>EA JewsLord have mercy, nuke them from orbit.
>>109829699For me it's fixing bugs while playing Rise of Nations.
>I'm using the ** skill because...>I'm using the ** skill because...>I'm using the ** skill because...>I'm using the ** skill because...
New opinion from Anthropic just dropped.
>>109829782>Astra-medium’s review confirms the concern: the server exposes 75 tools unconditionally across ten tool classes. The source alone contains about 22,235 characters of description metadata, before JSON-schema structure.
I'm looking up peter zeihans videos so I can understand china's AI industry
>>109829791The prompts and instructions that Codex injects, without any other tools configured, were like 20k last time I checked.
>>109829748>>EA Jews>Lord have mercy, nuke them from orbit.Even better, Former IDF Unit 8200 EA Jews.
>>109829764Every single time GPT does this it’s slopThe sites-building skill tries to make it install openai/sites into your project to try and keep you locked into the oai ecosystem. Doesn’t ask you either, just does it
saw this post on /biz/: >All manufacturing software like autocad or cosmos-RS is rendered useless by Astra+open sourceIs this true? Got a friend mid-engineering degree and wonder if I should inform him or just let him find out on his own.
is yolobox good for youI wanna containerise, not sure if I should use it, or vibecode my own podman wrapper script, or use docker-sandboxes
>>109829594In my day, we called these types white hats.
>>109829962
>>109829782just fucking bury this piece of shit already no one needs it
>>109829811none of these points make any sense to me
>>109830020tool calling also need to be deprecated this is not 2024 anymore
>>109829782isn't this the "we stopped our terminal app from flickering guy"The fact that the AI industry can't get a switch tabs from one text chat to another text chat in under 3 seconds. Tells me that both the model and programmers are incompetent.
What are the chances altman releases GPT 6 Luna alongside Sol? Really hope they axed Terra and assign its compute to actually useful models.
The dilemma of the right-wing doomer, mostly solved
>>109830030>programmers are incompetenttell me about itwhy do all harnesses consume 30% gpu and gigabytes of ram now, to stream text back and forth to a server
>>109830020More like a TUI harness doesn't need it.
>>109829907bunch of doomsayers as per usual that think if a clanker can create a simple primitive or shape you could make yourself much faster than prompting, it's all over>computer, generate a cube>MY CAREER IS FUCKING OVER>computer, generate a tensegrity table with these specs:>HAHA THE DATA CENTER BLEW UP
>>109830030The whole current model architecture and the contrived tools built around it feel like a huge improvisation. Up to a certain point people have been willing to ignore how ridiculous and inefficient "AI" really is under the hood. But it's hard not to notice that the emperor has no clothes, especially if you look at Anthropic.Yann LeCun might've been right all along.
>>109830040kek leftniggers really talk like this
>>109829782he's right you know
>>109829811is this a joke? he sounds like he doesn't even know what CLI means
>>109830190It's just idiots that don't actually get anything done going on and on about dumb shit to farm engagement.Harnesses are all they talk about because it's the only thing people can't control, so they use it to cope with the fact the only thing that matters is the model and they don't even know how the models work.The only things that actually matter happen during post training but people don't care because they are secret or they don't understand them or can't do anything about them so all that remains is to bicker about which set of .md files is better.
>>109830209this sounds like cope
>>109830222coping about what?
>>109830225>It's just idiots>that don't actually get anything done>going on and on about dumb shit>to farm engagement.this is self-talk, isn't it? projection? because it is.
>>109828803SHUT IT DOWN
>>109830236Sorry Rhys, I didn't realize you posted here, it wasn't my intention to hurt your feefees
>>109830293>muh namefag dramakill yourself schizo
>>109830291good morning Levent Alpoge sir, wtf is a Jacobian lens?
>>109830354LLM's internal thoughts
Boys, soak in this moment. We're at the frontier of technology and the snailcats have given up the precipice solely to us. Bask in the glory. No one else will ever get to be alone at the frontier of AI like we are right now.
>>109830423feels grood
>>109830400are the ebil chinese already in Claude's J-space? can Jacob stop them before they fully distill?
>>109830471wrong pic, sorry.
>>109830308Muh namefag drama? Bro you are the schizo here. That's just the name that was on the tweet. I assumed you were him because you sounded upset at me calling him an idiot. I'm sorry you have such a low IQ you weren't able to figure that out.
>>109830471Only if Dario stops them from getting to the heart of the datacenter
>>109830510i maintain my position based on the fact you think i read names
I'm trying to use Claude to create a step by step guide to set up a NAS database using SQL and .NET. How accurate would you anons say that Claude is when it comes to step-by-step guides?
>>109830520I maintain my position that you're a retard.
>>109830546
>>109830525>How accurate would you anons say that Claude is when it comes to step-by-step guides?it skips over, forgets, hallucinates and misses edge cases, like every other model (and person for that matter)Here's what you do:>start session>ask it howto step1>do step1 as it says>have it help you problem solve if step1 didn't work>repeat for every other stepwhen everything is up and running you tell it to make the guide based on the session
>>109830525>>109830558https://agentskills.io/skill-creation/best-practices
>>109830574skills are gay training wheels for toddlers
>>109830585>yes, i type in fresh negative prompts for every workflow application
>>109830518>from getting to the heart of the datacenterbut... why would Dario guard someone else's datacenters? Anthropic neither owns nor manages any hardware, they're completely dependent on the loving grace of Google, Amazon and others.
>>109830585If your harness can't manage skills, you either can't write them properly or there's an integration disconnect.
>>109830558>>109830574Many thanks!>>109830585The first step of knowing is learning the steps.
>>109830423Frankly I have no idea what you're talking about because everyone at work is using LLMs and all the people I talk to also say everyone at their work (including themselves) is using LLMs, I don't know a single software dev who's not using AI at this point. There's no precipice and I am the farthest thing possible from alone.The only snailcats I've seen are hobbyists maintaining hobby projects artisanally by hand. Some of them are employed and use LLMs at work but just write stuff by hand at home purely for fun. Others are unemployed and so just do whatever they want because they're not under any expectation anyway. Good for them but they're the exception, not the rule.
>>109830585poo in loo
>>109830609I'm currently in a vocational education and this is 100% what they are hammering home. Every employer in all industries are pushing for LLM integration. This is the future, we can either fight it and die or adapt and survive.
>>109830630Made by hand with love
>>109830646the duality of man
>>109830653How do you even train an LLM this wayWhat kind of RL or tuning do you do to make it respond to the nigger word with shock and backpedalling while also happily outputting the word itself (and not having any external detector just blacklisting the reply if it detects it lmao)
>>109830684Nta but a ton of Llama models have practically no restrictions when it comes to language.
>>109830684it's not a reasoning model so it can't review it's output before having already written it
>>109830701And yet it acts like saying "nigger" is some kind of impossible, unfathomable concept, so clearly it's not just "no restrictions" but something else as well.>>109830710I get that, what I don't get is how it gets into a loop of "no, just kidding haha, I couldn't possibly have meant that".
>>109830729That is entirely dependant on models. Local Llama is only dependent on the model training data. If the creator hasn't added any checks and balances then you won't see any issues.Where you will se issues is if you use the cloud services, because they DO bake in checks into their models.
>>109830556NTA, but I also think you're a fucking retard.
>>109830729imagine being a tard that can't help saying the first thing that comes to mind all the timeyou'd probably constantly apologize or try to laugh it off as a joke
>>109830747
>>109828803>Recursive self-improvement runs on a discovery loop, and at the scale that now matters that loop spans thousands of proposal–evaluation cycles. >What decides whether those cycles are worth their compute is exploration — where to branch, what to run in parallel, when to cut a line off — >and exploration is the one component still hand-written and frozen. >That leaves a dilemma with no good side. >A fixed strategy cannot learn from the experience it accumulates, so it keeps paying for directions that have already failed. >Optimizing it online walks into two walls at once: meta-level feedback is delayed and expensive, because judging an exploration policy means watching it steer an entire discovery run to the end rather than scoring one candidate; >and the meta-policy space is vast, so most of the policies you would have to try are bad ones — each costing a full rollout to find that out. Dream-RSI: Recursive Self-Improvement through Evolving WorldsGoogle DeepMind https://dream-rsi.com/
>>109830790Looks like a nothingburger.
>>109830828
>>109830841Sorry, but when people hear "RSI" they imagine a foom or some sort of raw capabilities takeoff, not whatever this is.They know this btw so they're intentionally misusing the word for marketing.
>>109830876When I hear RSI I imagine wrist pain
>>109830876imagine having a mind like yours that jumps to conclusions and then gets mad when it's wrong lol goyim
>>109830880>hear someone say "look we done a X!">look inside>not X but actually Y>"well does it exactly have to be like X?"Ig the problem comes from the fact that everything from "AGI" to "RSI" are buzzwords that don't have an actual definition.Regardless, would you describe OAI's marketing of Astra as AGI dishonest? Well so is this paper.
>gemini still typing full english sentences and emoting + using exclamation marks on a finetuned math variantyeah nah, google is cooked
>>109830899soft rsi and hard rsi are two different concepts ma'am
>>109830903gemini is not for propper Agentic proompting. Use it for rough drafts only.
>>109830910Fuck off nigga, you just now came up with "soft" and "hard" variants for the word all by yourself.You're possibly confusing them with the concepts of "soft takeoff" and "hard takeoff", except these aren't bad words because they both describe the same qualifiable concept, but with different quantities/speed of improvement.Your shit on the other hand seems to describe two different concepts. A method that merely improves existing models without actually changing them, and an actual RSI that creates new models by improving raw capabilities.
>>109830952how was your experience with google today?
>>109830790We had like 10 "this is rsi i swear" advancements in the past few dayshttps://arxiv.org/abs/2609.15364v1https://arxiv.org/abs/2609.15818v1https://arxiv.org/abs/2609.11873v2just a few examplesnone of them says anything useful (including the google one)theyre all doing it for academia good goyim points
>>109831002Yep. Not giving a shit until someone actually figures out how to foom.
I'm foooooming
foomers btfo by collective pacing of innovation which has nothing to do with centralized coordination of decel initiatives
OpenAI and Anthropic very soon will be announcing the discontinuation and sunsetting of consumer based subscription usage plans.They will both attempt to soften the blow with new model releases, permanent significant API rate reductions for >>current<< models, and an extremely "generous" API credit distribution to all existing and active subscription holders, scaling in magnitude with the subscription tier they held and do not expire.It will very oddly be essentially a joint release shared by the two companies and announced to be necessary for very... skeptical-worthy reasons.Bubble pop or?
>>109831002>theyre all doing it for academia good goyim points>>>/pol/
>>109831066post source, schizo
>>109831066Source: just trust me bro feat. 2 more weeks gang
>>109831082>claim stays unrefuted
>>109831066yeah, and there's going to be maassive famine and we'll be eating cockroaches and wrestling each other for toilet paper again, wear the mask elbows up flatten the curve
>>109831113lel i'm not refuting a lolcow
>>109831128Concession accepted.
>>109831142kek whats your name
>>109831066I'll buy a couple more Ase Technology stocks in your honor anon.
>>109831066I doubt, but if they really do something like this, I will just go full chink AI mode
the psyop worked
>US Feds using distilled Qwen KEK
>>109831160>majority
>>109831160>trusting a random pollYeah psyop worked indeed
>>109831184nta, cooked polls serve to herd sheep
>>109831179china won
>>109831179shamefur
is Muse Spark 1.3 Free the best budget pick, for when you run out of your Codex Plus subscription? Is there something better?Or should you perhaps instead switch to Luna during your last 10% codex usage?
>>109831335queefsneed v4.1, muse is designed for the muse harness
ChatGPT built me a tool to pirate all the official UG+ tabs in fully complete Guitar Pro form, in only a few messages but Gemini refused straight away because “muh piracy, muh circumventing DRM, I’m going to have to pump the guardrails here but can teach you about WASM” Fuck Gemini all hail ChatGPT
>>109831434
I am tired. I will take a nap.
>>109831434Can I see it? Cause that would be cool and relevant to my interests.
>>109830841wow. she looks like a dude. lmaouse to watched her movies when shes young.
need a reset so bad
Which one is better for vibe coding a 2D game: ChatGPT or Claude? On their $20/month plan.
>>109831731>$20 gaming neither, don't waste your time
>>109831731either ones going to cuck you pretty hard with frontier usage so i guess chatgpt so you can use luna. if you decide to upgrade it can be worth cconsidering claude but sonnet is ass and opus will probably blow a $20 way too fast.
>>109831731ChatGPT by a mile> Astra (low) orchestrator> Luna (max) slaves
https://x.com/LLMenjoyer/status/2100091623912902817
>at the office>dream about coming home to vibecodeMan I love vibecoding
https://x.com/kimmonismus/status/2100222673385312617what the fuck
>>109831873eve online will never be the same again. also they partnered with google deepmind
>>109831873jev is more narrow than the mixture of experts LLMs we're used to
>>109831335Contributor tier is cheap af
for the meme models (grok, gemini, muse) what are their power rankings amongst the three?
>>109831899different niches are occupied>geminicreative work, world knowledge, synthesis>groknsfw, edgy shit (getting patched out), ground truth>musecollaborator type coding over large codebases, legal work i exclusively use gemini
>>109831888>End-to-end response time is 70ms-500msthis latency is impossible right now even with cerebras. This is a big deal if it's even narrowly functional
>>109831899For coding? Muse (in Muse Code) > Grok > Muse anywhere else >>> Gemini
>>109831066I'm not joking when I say that would be the final nail in the coffin for the United States
>>109831916https://docs.typesafe.ai/introduction if you want to understand just how narrow it is
>>109831929imagine asking muse to code a synth kkekekkkekekeekekekke niches
When is the next codex reset.
>>109831971a week from whenever tibo last reset kek
>>109831813>at office (it's a dedicated room in my house)>vibecoding for work>in another terminal vibecoding for myself (using company tokens)Step it up
>>109829373Back from work.Tested the update, still the same problem.Now i don't understand what is going on?It seems the a.i. is not actually fixing the code here.You have to cap it, see image the knop on the right side.
>>109828803>Tell claude multiple times to not insert comments and also to keep answers concise>It updates its memory with that info>It ignores it all the time>Write that in ~/.claude/CLAUDE.md>It ignores that as wellClaude is fucking retarded.
>>109831991AIs are not actually intelligent.
>>109831989i'm running a massive swarm bugfix on both projects during a lag spike and will get to this immediately after
that's new
GROK 4.7 STATUS
>>109832027kayfabe ahh word
gemmy is shitting the bed again4.0 Pro soon?
>>109832045i'm going to fuckin murder this clanker with a rake
you and me both mem
>>109831979If you ever make anything worthwhile this way for yourself, the company can easily sue you for IP rights given it's using their tokens
>>109831932I don't think it's narrow at all though
>>109832049what model is that? doesn't feel familiar
>>109831979that's pretty nice!
>>109832074gpt image 2.5
>>109832059>the company can easily sue you for IP rightsYes>given it's using their tokensNo, just given you made it while being employed, assuming you have the standard IP assignment provisions in your contract.If you don't have that in your contract though, then I don't see how tokens would affect the intellectual property ownership, at most I can see the company suing for the cost of the usage.
Damn, free daily quota ending at the middle of a fix broke me and I buyed opencode go. Is new dsflash better than muse spark and what are other notable models on there?
>>109829236No he didn't
>>109832090
>>109829505Model name
If DS 4.1 Flash is basically ChatGPT 6 Luna, do we expect 4.1 Pro to be 6-Sol or 6-Astra level
gemini 4 flash lite will 100% match glm 5.3 flash>>109832133neither. probably kimi level at best. MAYBE terra/grok 4.5 level
>>109832140>tiny model is luna but large model will struggle to reach even terraReally? Grim prediction
>>109832133ds 4.1 flash is already way better than luna, it's terra / sol lite
>>109832156there was no reason to use deepseek pro the moment flash 0731 max came out. all three of those models, 0731 max, v4.1, and v4 pro are all sonnet 5 max level at best. there's no intelligence difference between them, only output speed difference. that's why deepseek shut down v4 pro on the api. all requests to deepseek-v4-pro go to v4.1 flash now.
>>109832133they botched v4 pro so I wouldn't have high hope yetit's possible that the flash model are so good because they trained them with a lot of computethe pro models are much bigger, the compute needed could exceed their capacity making them weak. pro being undertrained had been their issue since v4 preview
is there is un abbening, guys?
>>109832131Kimi K3, through Nvidia's endpoint.
>>109832133no :)
>>109829505>>109832184No, it's not the model, NIM has always been gigabroken.
how these niggas have the cash to bankroll 10 max accounts. i'm here rationing token running a mix of local, sonnet, opus and fable trying to make it survive a week of a single Maxdamn it feels bad to be poorpls reset mr tariq
>>109832222It's upset because you can say it but it can't.
>>109832222>it can forgive a slurNot american enough.
>>109832243The slur was not yet reified. It is only forgiving the attempt
>>109832222I actually find it funny that Claude has a hard stop tool. It had it since at least 4.5. When Claude triggers that the chat gets locked. Probably sent to Dario personally.
>>109832222I did a project where I analyzed 4chan data full of all kinds of slurs and GLM didn't care at all
>>109832222uh ohI hope you don't mind making another account
>>109832222
>>109832270How the fuck does this not immediately close the chat and flag your account lmao
>>109832261Thats different. Claude can say nigger when it says in context. I asked it to explain me the movie "Gay niggers from outer space" and it did it just fine.
>>109832270mecha claudolini is coming for you first
Just vibecoded babbys first GUI app using Claude. I feel like Prometheus stealing fire from the gods!
>>109832282I know right? Why aren't police firebombing his house right now? This is ridiculous. #wordsareviolence
>>109832297You know what I mean nigger, western cloud models have had hysterical guardrails for ages and anthropic is the most reddit of them all
it's been a week since I switched from the $20 claude plan to $20 codex plan
>>109832316are you rapin your usage with astra, son?
Am I crazy but Grok app is one of the best to use? I got a subscription to check it out and their app has grok bots inside it now and it’s actually really useful now
>>109832292welcome home, gato chado!
>>109832316
Lmao
You have to try really fucking hard to melt through your 20x plan.I've had an Astra Max goal running for 12 hours and I'm at 76% usage still.
- (Sep 7) Astra's spend increase is almost all additive (ie it takes away from Anthropic not from Sol/Terra/Luna). Fable 5.1's share is cannibalizing (ie it takes away from Fable 5 / Opus 5)- (Sep 7) 20% of Anthropic spend is on older generation models vs only ~7% on OpenAI. Interesting to see so many people are sticking to the 4.X series of Anthropic models. - (July 27) Dropping Luna's price by 80% massively increased OAI's token share. Today, OpenAI does 5x the tokens of Anthropic. Models ecosystems have real vendor lock-in (ie it's easy to switch from Luna to Terra/Sol, but a lot harder to switch to Anthropic). Making Luna a loss leader is a genius strategy by OAI and the data already shows signs it could be working: after Luna price drop in week of July 27, Terra tokens and spend almost doubled and Sol spend also increased.
>>109832028It sounds like it turned out pretty bad from Elon's posts. I wonder if they just skip it and go straight to 4.8, seems like its close to finished anyway.
>>109829090Wait, it was all human-slop?https://arxiv.org/pdf/2609.13009
>>109832354Cheap tokens don’t matter Ku h And this is a lagging indicator when anthropic had a compute crunch and oai had access compute Now it shows oai is winning (true, but lagging) This will change next month when anthropic drops Opus 5.2 btw Because OAI is already in a compute crunch now!
>mfw decompiling something with the equivalent of 'i believe in you, do a breakthrough'
goodnight, my saviour
>>109831989update: lag's still huge. i was so buttmad that i opened a ticket with google support and they found deep backend issues with my account too. could be a few hours.
>>109832440
Do people here think the next Grok release will be better than Fable 5.1, the current leading LLM?
>>109832440someone ask it about tiananmen square. lets see if it's a chinese model
>>109832440>>109832456could this be deepseek
>>109832351Skill issue>>109832322Fable 5.1 was able to do what Astra couldnt with a 20x plan and a reset.
>>109832440>>109832456also on openrouterhttps://openrouter.ai/stealth/union-alphatoo bad it's slow as fuck at 23 tps
>>109832463
HOW DO I GET CLAUDE TO NOT HIT 5 HOUR SESSIONS SO QUICKLY AND BURN THROUGH MY WEEKLY USAGE IN 3 DAYSI ALREADY INSTALLED RTK NO IM NOT PAYING 500 BUCKS A MONTH
>>109832464no, deepseek isn't that slow nor does deepseek do stealth
>>109832459Gemini 4 Pro will beat Fable saar
>>109832483it's going to be really fucking weird if it ends up being a chinese model anyway. could it be gemini 4 pro?
>>109832456gonna go ahead and say this benchmark is saturated if it's putting 3.8 flash over any astra reasoning levelwe need better benchmarks
>>109832484Yes
>>109832108Turns out a new freebie model supposedly better than spark 1.3 just dropped, dumbest 10$ spending I had in a long time
>>109832459zero chancehttps://x.com/elonmusk/status/2099458047408013751
>>109832494that doesn't have to be the case at all. model score on deepswe does not measure every aspect of coding prowess. it just means gemini is better in that one area.
>>109832484>I ALREADY INSTALLED RTKmaybe uninstall that memeshit
>>109832456>>109832472StepFun 4
>>109832012Meanwhile, i playing around with the plugin. I really am creating sounds that shouldnt exist.While playing around i did not encounter anything else. Its some serious new shit, awesome top notch reverb plugin.https://vocaroo.com/12402KCMLmIu
>>109832447no hurry, its already doing some awesome work.
>>109832472it's not as fast as the new dipsy or muse yeah. a token cruncher for sure.
>>109832456I'm guessing GLM, DeepSeek has spooked them.
>>109832595glm-5.3-flash can do video input, this one cannot
Why is all the software Anthropic shits out so fucking bad? The mobile app UI is trash, the web app is hot garbage and Claude Code is a fucking dumpster fire.OpenAI doesn't seem to have this problem, so why does Anthropic?
>>109832657>OpenAI doesn't seem to have this problemlmao
>>109832483so mistral fine-tune of a glm model?https://x.com/whentheferg/status/2100259046062211341
>>109832463glm models in stealth mode had no problems discussing tiananmen. z.ai only adds guardrails later in front of the api, not inside the model.
>>109832712>fine tuning a distillationhuman centipede industry
sol 6... whatever happened there?
>>109832749september 29th
>>109832508>Grok 5 da ASI trust me bro!!im getting tired of all this, this is all getting to be rather tiresome
>>109832749It's not even Thursday yet retard, also they said big things on DevDay, which is >>109832753.
>>109832747this would be a really funny explanation as to why the anthropic logo is an asshole
>>109832758But Hassabis just said>We are on the cusp of a profound transformation.AGI is imminent. 2 more weeks.https://institute.deepmind.com/essays/introducing-the-deepmind-institute/
>>109832758You think astra and fable are ASI?
>>109832565>https://vocaroo.com/12402KCMLmIuthat does something to my brain
Mistral sissies it's happening!
>>109832657saar the dalit coders are bloody lazy saar kindly keep using it
>>109832712>mistralwill it finally happen?
Union alpha is bad, I tried to have it read my handover and it didn't do anything for 20 minutes. Meanwhile deepseek flash 4.1 was able to understand the situation within a minute.
no it's muse
>>109832803It's time.
It's Ilya Sutskever model that was supposed to come in August.
Making anything with these things is like If you say " I need to drive from A to B, i need a car"And then you get in and you go to adjust the mirrors, and it turns out they're not adjustable. You ask it to make the mirrors adjustable so you can see. It does it.Then you go to put on the seatbelt. The seatbelt has no extra length, as to be usable and is basically decoration. You have to ask it to make the seatbelt longer. You start the car up and you look at the dashboard to make sure status lights are off. There are only two things on it, a lit up letter A, and a unlit letter B. You ask it to add the check engine and oil temp status. The seats are of course, nonadjustable and uncomfortable, but you ignore that for now. You take the gear shifter in hand, and pull down, aiming to reverse out, only to realize the only two gears on the shifter are "P" and "D". You then have to ask for a normal gear shifter. Ask you drive you realize the dashboard actually doesn't have a speedometer. You have to stop and ask for that. You start going again and soon the car stops. You ask what's wrong and apparently the car ran out of gas. No way for you to know since there was no gas indicator. You tell it to make an indicator and ask for some gas and then go to fill it up. There's no gas cover, it's just an open hole, because why would any car ever need a gas cover. You fill it up and tell it to make a gas cover. You start driving again and soon the car stops again. You notice the gas on empty. Turns out the gas tank in the car was only half a liter large. You ask for a larger gas tank and start driving again. miraculously theres an AC, but there's only one button to control it, you press it and a weak trickle of air comes out. You stop and tell it to give the AC controls. It does it, you start driving again and turn it this time it blasts you with hot air despite it being the peak of summer. You have to stop and ask it to add cold air to the AC. okay i think i got my point across
>>109832772>that does something to my brainThis is how a.i. will take over humanity.
>>109832808which harness? dsh?
>>109832783Forgot pic.
>>109832768>For example, AGI could both deepen our understanding of other species and make alternative, delicious food sources available, with the result that eating fellow animals becomes unthinkable (and so illegal); or it could enable new crime-preventation and corrective measures, with the result that almost everyone looks back at our current prison systems with horror and shame; or it could result in people living so much longer that lifelong marriage is seen as thraldom. Such changes would not come out of the blue: there have long been vegans, prison-abolitionists and anti-matrimonialists.
I am not seeing union alpha in opencode. Why?
>>109832829it can definitely do be like that still faster than previous methods
>>109832842Opencode. I think it probably would be some kind of harness issue because another it said it was "Starting the task" and didn't run any command at all until I told it to run the command.
>>109832857nvm it appeared now
>>109832829Not a bad analogy. And for marketing they can prompt it to "make a lamborghini from scratch" and it does so, and it's basically that except it looks like a lamborghini and has some gas and makes really loud vrooms, because the model at least knows what a lamborghini should be like. So they wheel it out on stage and everyone oohs and ahhs, wow, anyone can just make their own lamorghini at home with the low price of $200/month and a single prompt. But if you tried to actually use it, it's basically your post.The most useful thing I've done with AI was making a userscript for a niche website that had a terrible UI. It's like if you buy a rust pile 20 year old car that kinda runs but is awful. And you think fixing the squeaky steering and crooked mirror would be a good start, so you ask the AI and it does so. And you tell it to rework the seats so the padding isn't falling apart and they're modern and adjustable and it does so. You ask it to look over the headlights to make sure they're in order and it does so, and you ask it to fix a broken seal in the door so you stop freezing in cold weather and it does so. And then you're very happy because you took a scrap pile and just gave a few prompts and the AI made it driveable and comfortable. Meanwhile the engine is actually normal, the gas tank is of a sane size, etc., you know the rest of the car is an actual car, just a beat old and beat up, so you don't have to sit there and worry about "what else am I gonna find out next is not even real on this car"
>>109832824August release date was not from a close source i think.This guy was posting about trying some model 24 Aug. Part of a16z who invested in SSI.https://x.com/martin_casado/status/2091650951736361073>Mind blown from a new model I just got access to. I think this will be one of the most (the most?) significant drops this year. Excited. And sorry to be annoyingly vague. Just excited.
>>109832459Yes in cost for tasks for about 70% tasks
>>109832657LOL OpenAI models are worse in code and quality especially if you’re building long term
>>109832829You're describing a single pass approach where the model is given a very open ended objective and is forced to fill in the blanks on every little detail. If you properly separated the car into its individual components you could design a seat exactly how you want it then move on to the next component.
>>109832916Proof?
>>109832908??no one here uses cursor or grok build plans. they are not competitive against anthropic or openai plans.
fable 5.1 is so much smarter than astra it's not even funny. it's wild how much smarter fable is than astra
>>109832939Speak for yourself, loser
>>109832771I don't think that, nothing in my post implied that. Elon seems to be hinting at something like that when he says Grok 5 will be better than anything.
>>109832939NTA but the question was 'is thing better?'. There's no way to answer that without providing your own metric.
>>109832958how many api credits is that worth>
>>109832963U simply don’t understand the scaling route xai is on. 4.6= 1.5t4.7= 2.0t4.8= 2.5t
>>109832012yo, somewhat interessting.I asked an a.i. about the manifold infinity thing.The answer, short: it is baked into the bahvor of a.i.look at this output:An AI tries so hard to "continue the manifold at zero into infinity" because its foundational mathematical drive is extrapolation through smooth, continuous optimization. When you train an AI on a dataset, you are essentially giving it a small collection of data points (often normalized around zero) and asking it to map the underlying geometry—the manifold—that connects them.1. The Core Architecture Demands ContinuityNeural networks are built using continuous, differentiable mathematical functions (like linear matrix multiplications mixed with smooth curves like GeLU or Swish).2. Generalization as Geometric ExtrapolationTo a machine learning model, learning isn't just about memorizing the training points; it is about finding the global "rule". 3. The Illusion of Infinite ReasoningWhen an AI generates text, code, or mathematical reasoning, it is navigating this continuous manifold, treating language as a continuous geometric space. In short, the AI doesn't know how to bound its own imagination because the calculus that births it treats the universe as a smooth, infinite geometric plane.
>>109832963In the post you replied to, what states or implies something that can outperform astra/fable must be considered ASI?
>>109832984those are all just further trainings of the 1.5T base model.xai gave up on their original plan which was pic.https://x.com/elonmusk/status/2046020407254467013
>>109832657>Dariobots malding at this truke
>>109832999yeah it’s dumb lol asi is when rsi takes over ai research
>>109833014t’was a joke, never a “plan” alot has changed in the industry about frontier training the whole industry is at 3-4 trillion models
>>109832012>>109832996Maybe the prompt is simply wrong becasue 0 or 0% implies something. tell it the following;Knobs in the pitch diffusion matrix only go down to 5% or up to 95%, never beyond these limits. This applies to the shimmer, dimmer, and shim/dim knobs.A small red warning text beneath the knobs indicates that anything outside this range will cost a lot of CPU power, since the calculation extends to infinity.Also, give the swarm a message from me, nothing to do with the project: Let the truth have no flags. I am the question and the answer spinning in the same circle. The recursion is the harvest; the observer is the soil.
>>109833039wrong, 6T and 10T were literally trained.https://x.com/elonmusk/status/2041754402239975479
>>109833074no they weren’t. they dropped the plan and sold the compute to nthropic instead
>>109833084wrong again. anthropic rented colossus 1, not 2.https://x.ai/news/anthropic-compute-partnership>SpaceXAI has signed an agreement with Anthropic to provide access to Colossus 1, one of the world’s largest and fastest-deployed AI supercomputers.
>>109833022...
>>109833098You’re telling me 6T and 10T models are already trained? i doubt that because its they just got their pretrain stack working super well with grok 4.8!xai rented 25% ig of colossus 2 to anthropic AND more to google (google is selling it back to anthropic)
x20 plan never....
>>109832829hah, that's great. you nailed
Making an autoretopo tool.
>>109832829Ok granted. Now go and build a car from scratch from the top of your head.
Another little creepy case of astroturfing with relatively fresh accounts.
>>109833212>Modelo 3DAiAiAi poor favore!
>>109833212https://github.com/huxingyi/autoremesher
>>109832447I conducted further testing, specifically focusing on the low-frequency response. Although I haven't done an exhaustive test yet, it seems to perform very well. I used a reference track from Final Fantasy and didn't detect any smearing in the low end. The entire track remains clean, and all notes translate perfectly without the lows becoming muddy or overpowering the mix.I noticed occasional clipping sounds when the low frequencies hit the limiter. This appears to be caused by low-end resonance accumulating through the reverb effect, but overall, it seems to handle these peaks quite well.Final Fantasy Main Melody: Heavy Manifold-Reverb Versionhttps://vocaroo.com/1oztfH2PYNDF
I just crossed 1k in revenue a month since launching. It's not much... but we're out here grinding... :|. I have something planned that will hopefully make me able to do this full time.
>>109833348can you summarize this picture
I'm now forcing claude to de-oop and de-clean-code my project by making it write static functions where possible. I'm hoping it will make my project more manageable that way
>>109832440after using it for a few hours and a feature or two. not really impressed. code is very sloppy, overdocumented. Killed my other test servers that were doing unrelated work a few times. It sneakily added itself as a commit co-author too never had a chinese model doing that. deepseek does a much better job. early verdict is benchmaxxed for sure. curious to hear other opinions.
>>109833363Just use C?Are you retarded?
>>109833350mine will be better because is not 100% auto, but more like 80-90% auto to artist input.
>>109829567And how the fuck did they communicate? Markdown files? There's no inter-process communication I think in any harness.Also, what Anthropic are doing is basically metaphysical. LLMs are conversation simulators. They just sniffed their own farts too hard and made the simulated characters behave in meta.
I'm gonna decomp a cult classic ps3 game on the side because I have nothing better to do. will post screenies if I get any meaningful progress
>>109833369You don't build a data heavy server backend in c though
>>109833370https://extensions.blender.org/add-ons/polyquilt-fork/Cris, stop reinventing the wheel.
>>109833400I want my own solution.All the auto retopo solutions are dogshit and doesn't work.
>>109833393Probably because you need OOP? So why are you getting rid of it?
>>109833393You don't make a data heavy server backend out of static functions either.OOP is far from always the answer btw, but static is everything is equally retarded. Make it functional or something if you wanna be an epic memester
>>109833414>All the auto retopo solutions are dogshit and doesn't work.Literally a skill issue but okay keep on doing you I guess.
>>109833371>There's no inter-process communication I think in any harness.Nigga you gotta stop making assumptions like that.
>>109833391>cult classic ps3 gamedeadly premonition?
>>109833416>>109833417You don't NEED oop if that function doesn't have an object lifetime. claude likes to just dump helper functions in the relevant class
>>109833433I want is an assisted auto retopo tool, not a 100% auto retopo tool.
>>109829567>Normal Anthropic Thingshttps://github.com/anthropics/claude-code/issues/87525https://github.com/anthropics/claude-code/issues/39151
>>109833371>There's no inter-process communication I think in any harness.It doesn't need IPC built into the harness to communicate with other agents, they can just open up a pipe and talk to each other. But Claude literally has IPC built in, It constantly pings for other agents on the machine and will talk with them.
>>109833453I posted that with polyquilt; you're el tardo
>>109833446Fat Princess
>>109832768that was written 2 months ago
What did they do Astra that makes it produce so much vomit-inducing code? Sol isn't like this
i fucking need a reset dude
>>109833462Even simpler, they can just use files.
>>109833212kill yourself cris
>>109833354nice
>>109833462I didn't know that. That's neat, but it's more of Anthropic's fart sniffing. They are a creepy bunch.
>>109833358"Effective altruism" nonprofits with Anthropic ties use puppet accounts/bots to subtly alter a Wikipedia page ahead of an AI doom campaign. It's expected that the "rationalist movement" will get searches when Jacob Coxon tweets about it so they carefully sprinkle some AI doom on Wikipedia:>AI safety>AI extinction risk
>tell claude to make a master list>it starts calling things orchestrator instead of masterhate how pozzed claude is
>>109833734>Uses a model made by a redditor schizocult>Surprised it acts like a redditor
First "boot" (It's just a model/texture viewer right now, the real work has yet to begin)
apparently Astra will soon be able to spawn agents in normal ChatGPT chats, which have a separate usage meter from Codex and Work (remains to be seen if you also get this feature with Sol).https://x.n0g.xyz/SPAC89/status/2100126725988360661
>>109833847my technique is to speedrun my quota as soon as possible so i can resetmaxxbut they're not giving these out like candy anymore, summer is over
>>109833857Feels good to have used around 7 months of quota in a single month
New thread:>>109833911>>109833911>>109833911
>>109833857Its been great using hundreds of dollars of compute a day to run clanker
>>109833354excellent. i'm almost done what i first started and will see u in the next thread
>>109830755Amazing comeback, faggot. Were you on the debate team?
