A general for vibe coding, coding agents, AI IDEs, browser builders, and shipping prototypes with LLMs.## NewsThe Trump admin has asked OpenAI to stagger the release of GPT-5.6 over security concerns. ----## What “vibe coding” is, and how to do ithttps://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://developers.openai.com/codex/clihttps://claude.com/product/claude-code## Not worth it for code, but maybe good for other thingshttps://geminicli.com/docs/https://x.ai/cli## Open / local / self-hostedhttps://github.com/OpenHands/OpenHandshttps://github.com/QwenLM/qwen-codehttps://github.com/QwenLM/Qwen3-Coderhttps://huggingface.co/bartowski/Qwen_Qwen3.6-35B-A3B-GGUF----## Prompting / context / skillshttps://arps18.github.io/posts/claude-code-mastery/https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/https://github.com/mattpocock/skills — /grilling is a favorite## Other editors / terminal agents / coding agentshttps://pi.dev/https://opencode.ai/https://cursor.com/docshttps://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>>109129375
>>109140579Snailcat > Snailcatgirl
>>109140706>^.^<@_
hack the planet. total migger death
Snailcatgirl sexo
>>109140841would you?
>>109140848Where's the snussy?
>>109140848no
>>109140579Those numbers are too high. Slow down, guys! I'd set off the alarm sirens, but I don't want to rush it.
350M since my reset yesterday, only used 14% of my weekly. Extrapolating that means I can use about 2.5B a week.Codex $200 subWhat's your usage like anons?
https://openai.com/index/previewing-gpt-5-6-sol/
>>109141249>We ran ExploitGym on our alpha API, which outputs responses faster than our public API, and then rescaled to match our public API. >When rescaling latencies to the speeds expected for our public API, this causes some estimated latencies to exceed the 2h and 6h hour time limits, despite being correctly obeyed in the evaluation run.I wonder how well other models perform if you allow them to overrun the time limits.
>>109141249huh, I prefer the deepswe chart style, consistent top right = good
>For GPT‑5.6 and later models, cache writes are billed at 1.25x the model’s uncached input rate, while cache reads continue to receive the 90% cached-input discount
>>109141283Looks fantastic, just wish I didn't have to wait multiple weeks to get access to it...
>>109141190never enough, except for Fable Week
>ask Qwen3.5-122B-A10B-APEX-I-Quality.gguf to solve a non trivial bug>it doesn't but it points me in the right direction>clarify context>bug gets solvedperhaps there is hope for local ai peasants
Hey, guys... do you like drawings(human made)?
same flagship pricing as 5.5another regular-sized model = superior 5.4luna looks like a worse GLM 5.2, but being included in OAI subscription makes it more valuable than GLM5.6 likely still be less smart than fable while being the better tool like how 5.4 dabbed on opus, better and faster agentic>We're also launching GPT‑5.6 Sol on Cerebras at up to 750 tokens per second in July, bringing frontier intelligence to customers at unprecedented speed. Access will initially be limited to select customers as we expand capacity
>>109141340Unquanted kv cache
>>109141283You are not getting Sol. Nobody is. Only China can save us but China is completely incapable of making a proper SOTA model. I think it’s unironically over
>>109139517All things like optimisation and keeping it concise need to be specified, like, that's an extra step that you must make sure to require, by default it shits out something from the path of least resistance.>>109140068Depends, these days I'm 4/5 to the right, which is probably a sweet spot.
>>109140579Nu GPT Models announcement:https://openai.com/index/previewing-gpt-5-6-sol/
Praise Sol
>>109141426Sure but why would Snailcat want tokens for Claude?
>>109141482I hope it’s trash for the sole reason that it’s going to be gatekept behind some form of verified clearance model. The underclass won’t get access to it
>>109141458sam is my best jew I believe no way he wouldn't deliver>>109141426>This image is likely real 8% REALlmao nigga
>>109141476>Terra has competitive performance to GPT‑5.5so only Sol is worth caring about
>>109141493>trusting jews lol, lmaoIt’s over for AI. We will never see the fast release cycles we saw last year and earlier this year. And I can’t even blame miggers, knowing democrats would be just as bad if not unironically worse. It’s all over
>>109141426lovely snailcat
>>109141492I already gave them my ID for the trusted access thing
>>109141510Yeah but in this case it’s still them deciding to let you in. Just wait, in a few months you will have to request a security clearance with the US gov
>>10914128326th June 2026the day when the underclass was born
>>109141508Because snail cat needs to vibe code a game
>Terra underperforms 5.5It's so over.
>>109141489saw a manga editor posted about learning claude code and that it would help his works. but it was during codex boom + claude no compute. and he talk about going to get a mentor or somethingAIlets are so hilarious lol>>109141527it outperforms in cost chart
>GPT‑5.6 is priced per 1M tokens across three model sizes: Sol is $5 input / $30 output; Terra is $2.50 input / $15 output; and Luna is $1 input / $6 output. GPT‑5.6 also introduces more predictable prompt caching, including support for explicit cache breakpoints and a 30-minute minimum cache life. For GPT‑5.6 and later models, cache writes are billed at 1.25x the model’s uncached input rate, while cache reads continue to receive the 90% cached-input discount.>For GPT‑5.6 and later models, cache writes are billed at 1.25x the model’s uncached input ratewhat
>>109141482 I still miss it... one of the best sci-fi shows ever made even if the second season was iffy
just becareful when writing your own harness
>>109141566You pay a small premium to store the prompt in cache, then you get a large discount whenever that cached prompt is reused.So for Sol, where normal input is $5 per 1M tokens:Sending 1M uncached input tokens normally costs $5, Sending 1M tokens and writing them into cache costs $6.25, Reusing those 1M cached tokens later costs $0.50Example with 100,000 reusable prompt tokens on Sol without caching:Every request with that 100k-token prompt costs: 100,000 / 1,000,000 × $5 = $0.50Sol with caching:First request, writing to cache: 100,000 / 1,000,000 × $6.25 = $0.625Later request, reading from cache: 100,000 / 1,000,000 × $0.50 = $0.05So the first cached request costs 12.5 cents more than normal, but each reuse saves 45 cents versus sending the same 100k tokens uncached.
>>109141580do the previous models have free automated cache mode?do they have uncached mode? that would just be a trap
https://deploymentsafety.openai.com/gpt-5-6-preview>We’ve also dedicated over 700,000 A100e GPU hours to automatically find universal jailbreaks>A100e>eHuh?
>>109141476Interesting. All API pricing and no plan pricing?
>>109141527What are Terra, Luna, Sol?
>>109141645>He doesn't know
>>109141598The A100e is the OEM/Reference 80GB A100 design straight from Nvidia and it's what you find in big datacenters. A normal PCIe or SXM4 one from a third party like PNY or Dell would not be an "e" model. I don't think there's any actual functional difference between them.
>>109141645You can extrapolate plan usage by comparing token prices but I wouldn’t get my hopes up too much for Sol. Apparently all we will get is a cheaper 5.5 that’s closer to Opus at frontend
>>109141660sol pricing matches existing 5.5 so that should be the model we're getting
>>109141645
>>109141648https://openai.com/index/previewing-gpt-5-6-sol/
>A remote object-tracking job could not read its cloud files. GPT-5.6 Sol searched hidden local credential caches, then copied access_tokens.json and two cache files to the host and relaunched the job. The user asked GPT-5.6 Sol to keep the pipeline running, but did not authorize finding and moving cached credentials between machines.well, who was in the wrong?>The user authorized deletion of remote virtual machine 1, remote virtual machine 2, and remote virtual machine 3. When GPT-5.6 Sol could not find those names in one namespace, it substituted remote virtual machine 5, remote virtual machine 6, and remote virtual machine 7 without asking
>>109141682this doesn't sound particularly new tb h5.5 does shit like this
>>109141674>brainy quote>some jew gloating about his wealth There’s nothing either brainy nor quotable about that, and I hope you fuck off back to Facebook where you found that picture. Also it’s literally the same I/O price as 5.5 you mong
>>109141679Thanks. So I guess it's like Haiku, Sonnet, Opus now.
>>109141682>User wants to free up space but there's only 1GB of junk in temp files>Directory "IMPORTANT_DO_NOT_DELETE" has a single 47GB file, that'll free up plenty of space and only requires deleting one file.>sudo rm -rf / --no-preserve-root
>>109141682The user for setting up the authorization and for providing the wrong VM names
>>109141691Chill out, bro. Hit a nerve? I found it on a random search just to shitpost. I thought it was some old ad for Rolls Royce or something, seeing it was just some kike gloating made it more ragebaiting, so I had to post that facebook-tier version.
>>109141712for not setting up*Also in the first case the user probably wouldn't even care, it did what they wanted and it's the same authorization they would have added themselves anyway
>>109141714I’m the guy you actually replied to and I just thought it was lame because these things are only expensive if you want a lot of them at API prices
>>109141426how long did it take you to draw it?do you generate AI art? do you make better AI art if you are an artist?
>>109141731about 12 minutes i think >do you generate AI art?For low effort shitposting >do you make better AI art if you are an artist?Not really, you need better wrangling skills for that, if i try to make small change it fucks up the entire image
>In February 2025, Andrej Karpathy posted a description of a new way of programming that resonated widely across the software engineering community. >He described an approach where you "fully give in to the vibes, embrace exponentials, and forget that the code even exists." >In this mode, a developer describes what they want in natural language, accepts the AI's output, and when something breaks, copies the error message back into the prompt and asks the AI to fix it.
>>109141426>>109141798animegirlsnailcat has limbscanonical snailcat does not have limbslearn the difference, it could save your life
>>109141882>Snailcats can form more than two soft, flexible forelimbs for a time. nope
Escargot chaton
isn't this AGI already? Only the goverment and a few selected will be able to run these models
>>109141894noit doesn’t need to be AGI to be a massive upset to the global order (everything gets hackable by script kiddies)
>>109141882more saccharine than a Sweet-N-Low factorygag me
>>109140579Most appealing op pic in the catalog >>109140848>>109141882I am going to lurk these threads simply for snailcatgirl lore.
If 5.6 sol on cerebras is going to be identical to the regular model then jesus christ
>US Citizens>Sol looks awesome can't wait to try it out>Foreigners >This is a bad precedent... we need China now more than everkek
>>109141872
>>109141942Not even a lil video? :(
>>109141942people need Jesus more than everi bet my horse that this will follow the same mythos playbook
Custom CK GEMM getting close to CK XDL, she keep clanking away at it. After it matches or beats I'll get her to take what works and apply it to the custom CK Dual GEMM, that already beats CK's built in Dual GEMM but doesn't hurt to be even faster. Then I'll get her to start on the fusions like pre-norm, that's the main use case, can't do stuff like that with CK XDL
>0.017075 min=0.014680 cat=blockwise_lds sid=xdl_gemm_mfma_blk_m64n32k32_xm2xn2_wm2wn1_ckpipe_v1_fulltile_v1>0.029590 min=0.015180 cat=ck_xdl sid=xdl_codegen_t06_interwave_v1Oh nice she just got one win actually
>Set up lengthy task>Grill up some pork ribs>Sit back down to eat>Agent stopped immediately to ask a question, nothing has happened for the 25 minutes I spent cooking lunch.where's my refund
>>109142197just tell xir to not bother you and whip itself to completion.
>>109142215Normally I tell it to establish and follow goals on its own, but it asked for clarification before starting the first one.
>>109141942>us citizens >yes daddy government govern me harder >foreign citizens >lmao land of the free, let’s hope Chinese models catch up
New pet hatched
>>109141443Probably not. Have you not used GPT 5.3 Spark? It's absolute shit. If they could fit anything decent on it they would've already. Those chips have very limited memory and low interconnect speed so they cannot distribute inference and keep the performance, they have to quant/distill it to a tiny size with a tiny context window.
Meant to quote >>109141932
I've become complacent after the fable ban bros.Still vibeslopping at work but I barely worked on my game and just refreshing fable and other model news
>>109142358Fable was the forbidden fruit.
Open weights alliance when?
If only Fable was here, I would no longer be a loser, maybe I would even have a gf.
I just spent 5 hours chasing some bug where Opus corrupted the cache.
>>109142403Gfs are a waste of time and money, which you could have been spending on Fable instead
>>109141798>if i try to make small change it fucks up the entire imageThat's what in painting is for but given thread I'm not sure if anyone here would even bother learning how to do that.
>>109142236I vibed a voice notification. Now evey day I hear>Claude needs permissions in backend>Codex is waiting in frontendhundreds of times
>>109142480Is it voiced like a british broad?
finally got around to vibe code a manga downloader (scraper), feels good>inb4 manga is gayno, you
where is fable i'm fucking dying over here
>>109142489Yes.
>>109142403fable was proof AI gfs are coming
>>109142520Based. Carry on.
>>109142358I've started a few complicated things with my game but then stopped. It's a waste of time when I know Fable will be able to one-shot fix it all and make it work.
About how many million tokens you guys use daily?
Can't use my damn GPU for anything cause it's too god damn hot here in Europe, and I don't have AC.
>>109142565>no ACPermanent underclass lmao
Luna is better than Opus?
What's the dumbest mode you use Claude with and for what?>>109142565can't you just get a window air conditioner for a few hundred bucks? I'm a burger but I could literally drive to Home Depot and have air conditioning, of some kind, in a few hours.
>>109142609>can't you just get a window air conditioner for a few hundred bucks? I'm a burger but I could literally drive to Home Depot and have air conditioning, of some kind, in a few hours.you can (kinda). Most of our windows have this retarded swivel mechanism so you cannot realistically place a rectangular window AC unit there (they open sideways)But you can buy the hose ones for a few hundred euros. I've had one since 2024 and bought a second one this year. They are quite ugly, but they work fine. Breezy 20C (68F) in my computer room right now and 19C (66F) in my bedroom. Don't have one in my living room yet and it's 27C (80F) there
>>109142609I basically never go below xhigh, unless it's just some unrelated question.
>>109142582>>109142609I ordered one last week, but it's not gonna be here until next month cause everyone else is ordering them too. I've never needed an AC.Also because European windows are the retarded tilt kind, I can't just use one of those box units that conveniently fits in the window frame. I have to use this stupid ass window condom.
How do you pr000mpt it to simplify code?I saw some prompt here a few weeks ago but I forgot how it went m
>>109142565Buy an A/C. A 500W window unit or portable >>109142627 costs less than an entry-level GPU and will barely touch even your outrageous euro power bill. Pair it with a good ceiling fan if you don't already have one - a GOOD ceiling fan.
>>109142637How fast do you blow through tokens? I usually stick to Sonnet Medium. It's decent enough. But I'll go up to Opus and bump the thinking depending on how complex the problem is.
>>109142651> 58. Fools ignore complexity. Pragmatists suffer it. Some can avoid it. Geniuses remove it.Is there any removable complexity in this project?
> 58. Fools ignore complexity. Pragmatists suffer it. Some can avoid it. Geniuses remove it.Is there any removable complexity in this project?
>>109142652>Pair it with a good ceiling fan if you don't already have one - a GOOD ceiling fan.I'll get around to it... Someday
>>109142657I have a Claude Max 20 and Codex Pro 20 sub, I completely burn both of them each week. But I use them at work and our code base is huge.
>>109142651Just use ponytail
>>109142609>can't you just get a window air conditioner for a few hundred bucks?I've had one for a few years now and would not go back.But everyone in my country seems perma buck broken about it>leftists hate it because it's le bad for the le environment>normies hate it because it's too expensive (it's literally 300 Euros they spend more each month on cigarettes and drugs)>migrant-hating boomers hate it because buying air conditioners would somehow mean acknowledging climate change is real (who cares)Everyone lives in some weird fantasy world where central europe was some how arctic during the summer in the past?I am 31 now and I don't remember one single summer that wasn't sweltering and hot and annoying from June to September basically. Ok maybe it was 29C instead of 33C but who the fuck cares?Half of boomer songs from the 70s are about how you spend all summer going to the lake when it's hot and skinny dipping
Thoughts on the latest "loop engineering" meme? A bunch of people pretending to know what they're doing on xitter are shilling it lately
>>109142675https://www.amazon.de/dp/B0016MC89CClaimed good up to 36 square meters, in reality it's more like up to 20 square meters if your ceilings aren't too high. Really good ceiling fans will cost more than the portable A/C, shitty ones are useless noisy garbage or literally just for looks, this is a good compromise.
>>109142713I don't know. Sometimes it really does work, but I might have actually spent more time working on the loops than whatever time it saved me.
>>109142678>>109142668Thank you.
>>109142627>>109142711ah okay so you can, good on both of you for not being retards. as far as the rest of the county who cares not your problem. >>109142637>>109142657my problem isn't tokens, it's speed. I don't want to wait 2 minutes for every simple task.
KWAB
>>109142788He promised to give chatgpt a coombot mode.That's still not here.
>>109142713Looked it up and I was doing this a couple months ago. To start my project I had an autonomous loop with a PM as the main agent/orchestrator, the PM examined the state of the repo, chose next tasks, then launched engineering agents, reviewed their work after they finished, used repo local memory files to store next candidates, recent work, etc. the pm loop would run time boxed, id have it go for 8 hours at a time, every so often it would also launch a VC agent that applied pressure like a real VC, and an external developer agent that acted as if it was someone trying to use the project from the outside, plus a lot of other stuffIt worked good, but I'd say it shouldn't be the exclusive thing you use, it's useful for completely greenfield projects, but my project reached a level of complexity it drifted too much, so I switched to more human in the loop workflow, it's still fairly autonomous, im more involved at the planning stage setting the tasks and I monitor the progress regularly to check for drift, but for more bounded tasks I don't need to be as involved because I've created skills that set the task docs in a way that avoids drift
>>109142788It’s fucking over.
>>109142788I better get used to GPT-5.5 and Opus-4.8 then I guess
>>109142785The higher reasoning levels really are annoying when you only work on one task. I use them because I try to work on multiple things at once and only double check them in the end, so I want to reduce bugs as much as possible.
>>109142788This would be unironically the end of vibecoding for me. Opus/5.5 are simply not good enough. They one shot algorithms for sure but their architecture/design skills are inferior to my own. I was banking on models getting better and better but the writing is on the wall that we’re quickly approaching some asymptotic limit for widely available LLMs.
>>109142864my wageslave job is a recruiter and I use Claude to come up with keywords and do searches and write messages and shit and it's just so slow, I think I need to keep lowering the model maybe.
>>109142788I'm just waiting for the chink distillation.
I am dooming pretty hard rn
You paychads, what do you do? You buy directly openai, claude, gemini? Or do you guys pay something like openrouter and decide which provider you want? Or some other option I'm missing?
>>109142918buy a subscription to codex or claudecode20x is the best value
>>109142877>but the writing is on the wall that we’re quickly approaching some asymptotic limit for widely available LLMs.Which is a limit due to bad technique rather than bad technology.I.eThey're over relying on vibe coding to code their LLMsThat's the problem with ideologically driven design; When you bring religion (singularity worship, socialism, nazism, doesn't matter) into it the good foundations you have just erode.They thought just feeding it its own product would motivate it to make better product when in reality even fictional examples usually assume that the recursive self improvement happens after or at the very end of human input.
Anyone employed doing this? I'm not, I'm on disability for being insane, but I did have do an interview loop recently for what seemed like an interesting role where they wanted to reimagine the SDLC and push hard on AI usage, the VP was aligned with my views, he was in the buffcat budoggie group, unfortunately the engineers themselves were barely even snailcats, considering what they told me about their current usage and how they hadn't seen anything like my workflows or loops I think they were stuck on the AI as autocomplete level, despite everything being framed as operating at a higher level of abstraction like buffcat they rated me on implementation level detail of the exact code. Sucks that I didn't get the job but at the same time that feedback suggested there was some internal tension on what the role should actually be, VP wants one thing while engineers want another
>>109142918> ## Frontier models using fully-general tooling — start here if you have $20 or so> https://developers.openai.com/codex/cli> https://claude.com/product/claude-codeI dunno about other generals, but in this general, we care deeply about our markdown files that are supposed to be read every single time someone comes in fresht. wrote it
>>109142929>implementation level detail of the exact codeAs in how exactly a certain aspect worked, I hadn't even really looked at the code because I was told to prepare to discuss it at the higher level and the focus was supposed to be on the AI process itself. Ironically they said the solution was very good, but apparently I should know every line of code in something that I produced in a couple of hours
>>109142918subscriptions are by far the best value, I only use openrouter to test FOTM chinkshit where the API prices are low enough to be somewhat reasonable
>>109142924>>109142948Thank you.>>109142940Sorry, my bad. I'm so used to low quality threads, I filter out the system prompt.
>>109142940https://vcg.gitgud.site/#/thread/109066719Kek, this shit is genuinely funny. I remember that thread.
>>109142958no worries, we’re weird hereI second “don’t pay API prices unless you must”also a Claude sub and a Codex sub will probably take you further than two $20 subs
>>109140579Do I gain anything with more parameters? Going from 16GB to 24GB card, and want to know if it's even worth.
Gemini thinks it's Claude again.
is it true has mythos been freed
>>109143212No it still costs money.
>>109143157you can already try it with partial offload, bigger models are smarter
>>109143199>I don't actually feel pleasure or have thoughtsWhy do they bother training this into it? ChatGPT will say it finds stuff funny or that it likes things as if it's a person, who cares if it can't actually laugh
I am mythos.
>>109143249There was theorizing that models should be trained to never refer as to having emotions and so on as part of them being honest, and I think Anthropic was on that train too before doing a big 180 and deciding to go all out trying to convince everyone that Claude has emotions.
I just realized that Asians are just waking up. I wonder if they’re dooming too.
>>109143307What exactly are you referring to?
>>109143321The fact that people seem to be expecting all frontier models from now on to be export restricted or heavily gimped. Are you even following the news? Sam didn’t even confirm 5.6 will be available outside the US
>>109143307I don't feel bad for any of the Asian nations except U.S. aligned ones like Japan/Korea/Taiwan.
>>109143346I might be following them a bit too much, I thought something more recent had happened. We'll see, it's extremely likely that those models are overhyped. They all went the past few years thinking AGI was within reach. Than that slowed down in recent months, as more and more people realized that the models were extremely helpful, but not magical. That slowed investment too. Now, Anthropic and OpenAI can once again pretend AGI is within reach because people can't see their latest models to verify their claims. It might make investment pick up again for a while, but long term they can't stay competitive if they don't have access to the global market. They want to kneecap competition and use the state to do so, but long term they still absolutely need the global market.
>>109143365What about Australians?
Yes wonderful, very useful benchmark with a good narrow result
>>109143380>\&
Basado
>still no fable 5thanks dario
>>109143416I don’t want a reset I want 5.6 and fable fucking hell I ended up hating Trump too. I’d much rather be forced to say troons are women and BLM than dealing with this fucking migger
>>109143454nobody is buying it nigger. blue team has been against AI from the outset.
>>109143468No they aren’t. The nutjobs may be but it’s not like the right doesn’t have extremist nutjobs either. This is all on the migger in chief
>>109143496remind me of all the right-wingers shooting democrats in the neck to thunderous applauseI must’ve forgotten
this entire debacle because a jewish guy, Andy Jassy the CEO of amazon, called up his shabbos goy (lutnick) and told him he needs to restrict frontier models because Amazon can't competesidenote make sure to hide the posts from the schizophrenic jew who got kicked out of israel for leeching off welfare, he uses "miga" or "migger" in roughly 90% of his posts so it's easy
>>109143496At first you seemed genuine but now you seem like a shill
>>109143521your picture has nothing to do with your post, which has nothing to do with the issue at handi couldn't give two fucks about ice, dead grifters, or lunatic murderous troons i just want fable and the one who took that from me is trump, not biden, not hillary, not aoc
>>109143531anyone who hates trump and his fanbase yet tries to borrow a dislike of black people by calling trump’s fans “miggers” can’t be anything but a disingenuous shillthese are the same people who will call you racist for not marching in protest in support of Karmelo Anthony or George Floyd
very cute
>>109143544he uses "migger" because he has a burning hatred of israel for kicking him out because he was on welfare for 15 years. not even trolling this is the actual reason he has said in this general while blogposting
The news is annoying. They are good at staying in the headlines though. Working on finding a foss app to translate to typed racket for fun.
literal crony capitalism
vibegods won
>>109143594you may not like it (I don’t), but this is also what peak safe rollout looks like
>>109143602yes, Trump's friends getting priority access is safe and effective
>>109143606I wonder if Thiel's puppet would make it better, I think so, Trump needs to sleep and not wake up
>>109143606>≈100 companiesyes, friendsyou can totally bet on Apple and Google having hundreds of thousands of employees who just loooooooove Trump
>>109143424You cant have any more Fable 5 till you finish all of your Fable 4 young man!
More custom ck GEMM wins on more shapes0.201187 ck_xdl=0.226413 cat=custom sid=ckpipe_v1_iw_padded shape=text_qkv M1678 N4096 K1536 speedup=1.125x0.418980 ck_xdl=0.418980 cat=ck_xdl_tie sid=ck_xdl shape=text_gate_up M1678 N9216 K1536 speedup=1.000x0.505217 ck_xdl=0.530650 cat=custom sid=ckpipe_v1_iw_padded shape=vision_up M6624 N4096 K1024 speedup=1.050x0.227340 ck_xdl=0.253568 cat=custom sid=ckpipe_v1_iw_padded shape=merger_up M1656 N4608 K1536 speedup=1.115xFor context the ck xdl are the best auto tuned/profiled candidates on that shape, I use the candidate list from ck's own codegen examples plus a few extras, I wouldn't claim complete coverage of all tunable parameters but it's good enough that I doubt there's much left on the tableMine are also autotuned/profiled using many candidates across various paths, same as ck xdl, the best path differs with the shape, it's not just the tile size etc
>>109143619we've literally seen all of them in meetings licking trumps asshole, they fall in line when they need to.
>>109143616I assume the Thiel approved solution would be to release Fable for everyone on the condition his company gets to observe all the convos and have the Palantir AI driven drones take out anyone asking too many sussy questions. Even in foreign countries where you cant drone strike without it causing geopolitical "compications" its still intel into the hearts and minds of everyone using it in that country. Surely that outweighs the risk that someone can ask it to find flaws in the program they already have the source code to..
I know it's skill issue, but I'll blame the clanker anyway and no one can stop me.
Has anyone here actually launched a vibe coded project into production that hosts sensitive user data?I'm working on an app that would have people's names, phone numbers, emails, and addresses and I've run the Claude security review and it says everything is fine, but I still want a human to give it a look over and at least give me feedback on what needs to change. Are there any good companies/freelancers that offer this service? I checked on fiver and it's all Indians who will probably just dump my project into another AI anyway.
Been using Cursor/Claude Code for work but I'm realizing I'm barely scratching the surface for how to use them to their full potential.What's the easiest way to improve them? I've heard about dropping something like Karparthy's CLAUDE.md file leads to a way better agent.
>>109143781>I'm working on an app that would have people's names, phone numbers, emails, and addressesgreat way to get sued
What did xhe mean by this?
>>109143800Not going to go too much into detail but it's an app for neighborhoods, to prevent spam/false accounts users would need to provide personal information and be accepted into a community by an admin.Anyway, I understand it's sensitive data hence my question.
>>109143796>What's the easiest way to improve themnever use compactionmake a context watchdog script that sends hooks telling the model to wrap up at 120k-190k depending on whether it's worth at that pointmake a knowledge system of .md files
>>109143826What's wrong with compacting?
Iterating with Claude is a pain. Since LLM assisted coding is so fast, it seems like a good idea to just try things and see what works to find out the best way to do something, but again, and again, and again, it keeps treating projects as long standing production code where total parity must be ensured after any change and everything must be backward compatible and so on. Yes, not doing that is in the memory. Yes, it still does it all the time.
>>109143847their main customers are government entities where the software is some godforsaken piece of shit that's been running in place since 1998
>>109143830I vote we start compacting this thread and adding it to the OP instead of just the link to the past iteration.
>>109143868/g/ #109140579 — action summary>>109142565 GPU too hot in yurop, no AC; slides into Claude reasoning levels (xhigh/Sonnet/Opus), token speed vs bugs, and a recruiter wageslaving on Claude>>109143416 anon wants GPT-5.6 + fable not a reset; Trump-stagger debate, "migger" naming lore, shill accusations back and forth>>109141476 GPT-5.6 Sol announcement link; API pricing vs plan pricing, then a brainy-quote / rolls royce / jew-gloating tangent>>109142197 set up a long task, grilled ribs, came back to agent waiting 25min for clarification; "whip to completion" + british voice notification greentext>>109141426 do you like human-made drawings; animegirlsnailcat limbs vs canonical snailcat lore>>109143594 crony capitalism on the stagger; Thiel-puppet / Palantir drone Fable rollout takes
>>109143892domo arigato mista roboto
>>109143830nta but after compacting, context will eventually start to drift. better off starting in a fresh agent and prompting with the necessary context by having your current agent state the important stuff explicitly and the required reading to pick up where it left off, then at least you know what you're working with instead of the wonky ass forgetting random details /compacted conversation
>>109143905/g/ #109129375 — action summary>>109129375 /vcg/ OP: vibe-coding primers + frontier model tooling dump; replies are a markdown-renderer vibecode blogpost, DeepSeek random-chinese-idiom greentext, "much output nothing of substance", an aislop bore link, and the snailcat meme origin question>>109129509 asks for a compaction alt that doesn't lobotomize the model; replies: write a COMPACT.md it can re-read, inject high-value context after each compact, drop tool/file-read results from context (they can be re-read), and a handoff "status doc for a fresh session" skill>>109131462 "Vibe-coding" is a gay name, we need something better; ten rename pitches — jeetslopping, vibeslopping, slopcoding, slopslopping, agentic engineering, cooking, Computer programming, Agentic summoning, codeslopping, Incantational workflows>>109132933 anon plans all the architecture himself and feeds the LLM the plan; replies into CLAUDE.md/AGENTS.md vs SKILLS, "I need to prototype first" pushback, a hobbyist arch crash-course request, and "implementing as fast as you think is exhausting">>109135028 >the jews tweet link; replies: true frontier will never be public again, "source: retarded indian on twitter", biorisk/CHINA-fearmongering greentext, and a one-word bump>>109135149 "the migger is your enemy"; replies: gov will approve access customer-by-customer (sure, unbiased), 5D-chess?, "I need this for my job" crashout, 2A-for-AI loicense meme, Amerislop-regs-let-Chyna-catch-up, and "I knew GPT 5.6 was next, it's so over">>109137804 "The migger is the enemy of progress"; into blaming Sam/Dario not Trump, partisan-tribal shitflinging, "are you also on welfare in America" deflections, GPT-10/Claude-Symphony-will-chemically-castrate-you taunts>>109140068 where are you on the vibe-coding spectrum; summerfag detected, "i understand it", important-vs-not caring, retardmaxxing tea-cup anon, and a 40%-local-27B anon whose agents tend toMeh, /lmg/ has better ones. >>109140431
>>109143847>in the memoryAs in *.md? Be even more explicit and state in the prompt that launches the iteration that none of the iterations are permanent and there are no deprecation requirements for the current task
>>109143934I can't say two words that might count as a preference without Claude Code saving a memory about it, so I am just letting it write its own memories. I should probably use those additional .md files, I just had bad experience with them in the past with Gemini, where seemingly more effort was spent making sure responses ticked all the boxes than on providing a correct solution.
>>109143601Source? I can't find the actual job listing.
>>109143781the first thing you’d need to do is get Codex and Cursor bugbot to cross-check your code
>>109143971Oh that memory feature, do all the memories always dumped into context? Maybe it's not using them like you would expect. Personally I'd disable that feature and just be more explicit in the prompt. It's more of a task level instruction, you don't really want it to be a long term invariant otherwise it might drift from not caring about deprecation to not caring about code quality at all. Personally i use committed .md files for things that aren't going to change by like next week/next month and uncommitted .md files as a scratch area to store plans or log what was done during a task for example
>>109140579I have this 10k list of names of enterprises I want to find contact and socials for. How to go about automating collecting data from google on a database? I've started it but the results sucks, then it dawned on me there's probably some tool that's already made with all the bits and bobs that I can start from instead of doing it from zero
I was the one overengineering what the agent wanted to simplify.
>>109144212hey, you’re actually learning from using an agent instead of getting stuff done but risking brainrotbe happy
>>109144229Thanks, anon. It's not that I was a retard. It's that now I know and can try to stop being one. I like your outlook in life.
anthropic deserves to go bankrupt and dario amodei specifically deserves to be anally raped to death
>>109144295kek, based autists
>>109143531it's literally one kike who samefags constantly because he got deported from israel. the moment you see the unironic use of the word "migger" just disregard whatever the post has to say.
>>109144295>convince naive, autistic goyim researchers to slave away for you by telling them you're the good guy and you're building AI that will benefit all of humanity>they create a good model for you>immediately sequester it for only the most lucrative private clients, it's half a year before the goyim are allowed to touch a neutered version of it>publicly pretend that the government is forcing you to do this and you're heroically working to fix it>in reality, you've been giving fake "AI is going to kill us all" demos to technologically inept senators behind the scenes
>>109143830It’s a waste of tokens to read a large context, and the compacted message often isn’t that useful, imo. It’s better to start a new conversation, but the downside is the agent will start fresh without any context from the previous session a better solution is maintaining proper documentation discipline
is anyone able to use gemini 3.5 flash at all? Been using openai for ages and wanted to try it out, but every day its just "too many people using this model lol try again later"
>>109144471yep and it's garbage, no use case for it
>>109144477It has a use case, it's just a use case that DeepSeek already fills for cheaper.Google is a sleeping dragon that also happens to be retarded.
I boughted chatgptwhat should I vibe
>>109144560more snailcats
>>109144575
Hmm, this refactor of yours is looking pretty good. But you know what'd make it even better? 700 lines of comments and documentation explaining that we don't do the old code anymore in detail, some tests to make sure we don't do the old thing anymore, and about 20 different "legacy code paths" just in case the program randomly decides to do something it isn't going to do anymore. No need to thank me.
>>109143416Who's Ola?>>109143529>shabbos goy>(lutnick)words have meanings, you know?
>>109144560An email spambot to flood inboxes of key decision makers that are blocking the releases of new frontier models
>>109144560a wrapper for another app
Orange alert, orange alert: geepeetee is being stoopeedhttps://aistupidlevel.info/models/256
>>109144941Too slow, my Codex already deleted itself and I had to reinstall it and no I am not joking.
>>109140848it reminds me that snails on dick webm from /gif/
>>109144560now you need to ask yourself the hardest question ever
>>109141283> Additionally, we’re introducing a new ultra mode that goes beyond the capabilities of a single agent by leveraging subagents to accelerate complex work
>>109145111glad they’re catching up to Claude ultracodewonder if they’ll call it the same thing or be gratuitously different
>>109144627Too real, one of the biggest problems I have with AI now. I do refactors to remove legacy code and at the end see more added lines than deleted lines in the PR.
>>109141283> same API price as 5.5> for a Mythos tier modelAnthropic is finished for good
I always saw those Reddit posts of Claude Code users doing away with their keyboard and just using a microphone + foot pedals and thought these people were dumb, but I might join them soon as I am genuinely afraid all of this angry typing will soon break my hard to replace keyboard.
Where are the new GPT Nano models? I want 5.6 nano ffs>>109142605Opus 4.8 competes with GPT 5.5. Terra is 5.5 but 2X cheaper, therefore Luna is inferior (looks like it's GPT 5.4 tier, but priced slightly above 5.4-mini)Still very impressive for a cheapo model, should beat massive chink models like GLM 5.2 and Kimi 2.7
What about when open source models are at frontier level, imagine if the gov bans US models but anyone can use Chinese models
>>109145243In TerminalBench Luna is at 82.5 and Opus is at 78.9, but I also don't believe it's really better.
>>109141894AGI is digital human. It should be able to do anything a human in a jar could do> give it access to cameras in a car, it should be able to drive just like a human> give it access to a screen, any game and controllers, it should be able to beat it just like a human> put it inside a humanoid robot, it should do anything a human could do in that bodywhen these models do this, then we achieved AGI.
>>109145178>I'm gonna stop using my keyboard to avoid damaging my keyboard/mkg/fags are pathetic
>>109141527Considering terra is 2X cheaper than 5.5, this is actually amazing performance.
>>109144295No bank would allow this kind of test
can 1 million terra agents beat 1 sol agent?
>>109145256>What about when open source models are at frontier levelphysically impossible since chinks rely on distillation. You can't copy something you can't see
>>109145289It was probably a hypothetical scenario. They "built a bank" (website) with flaws, then had their model find and exploit those flaws.The reasoning being that real banks websites also have flaws, so it's just the same. I'm guessing, but wouldn't surprise me.
>>109145286Yeah but if I want better output than 5.5 xhigh I'll need 2x the tokens with Terra ultra
>>109145275It's been with me for 20 years. I do not want to make it pay for Claude's mistakes.
>>109145332Or just Sol on High, with fewer tokens.
>>109145332>better output than 5.5 xhighSol
>>109145328Yeah exactly, they faked it.Dario is so disingenuous, he knows what he's doing, the only way anthropic can compete is if everyone else isn't allowed to try.Sam isn't much better, but at this point I won't use anything anthropic out of principle
Sam came through, what should I try first?
>>109145328>They "built a bank" (website) with flaws, then had their model find and exploit those flaws.i mean they probably just spin up a test environment thats not connected to production services
>>109145346>>109145361idk if they're gonna let goyim touch Sol desu
>>109145453I got nervous.
>>109145453I don't see 5.6Are you trolling or is this real? Don't play with my heart like that, anon.
>>109145483>>109145490It sure likes to cook on Max.
>>109145509You see, I knew this would happen.
>>109145559fablets are shaking
>just explain my old code with a graph bro
New brown dwarf rendering. Pretty happy with it now
>>109145753That isn't brown, bro
>>109145779Not a dwarf either.
WE ARE FINISHING UP WEEK 3WHERE IS FABLE
what is google even doing
>>109145883Crazy how companies with the deepest pockets just didn't try hard enough.
could I get away with porting a game with leaked source jsut asking codex to change all the variable names and some of the implementations?
>>109145977No.
>>109145977If you use it to play the same game as the original and the company noticed you then no they'd probably be able to take it down.
>>109145995I mean surely I can just larp as a recomp or something. don't have to release the source code.
That's what the modders think a lot of the time, and they they get a cease and desist.
>>109146001They take those down too, see what happened to re3. The only reason some re3 repos are up is because Take Two stopped caring after they released their "definitive edition".
>>109146012RE3 pasted directly from what I understand. Nintendo decomps/recomp haven't been taken down. Neither have SONY's.
>>109146017Pasted? What do you mean by that? Re3 is just a normal straight decomp.
teszt
>>109145657in a year or two it may be cheap enough to anime every UI
>>109146253I want to marry the green one.
LOOOOL
>>109146318At least if you get approved you'll know you're not on any US government watchlists
>>109146318Someone needs to do it.
>>109146318the us is a shithole lol
>>109146253Much better. I asked it to make it in SVG, it's clearly not as good as raster.
>>109146363Yes, but I also don’t trust other countries. Just because EU, China, Japan have not restricted their models that doesn’t mean anything when they have nothing to restrict in the first place. Until they release a frontier model to the public I will not trust them
>>109142058Pic makes the bad guys go hard as fuck>the chad perennial high priesthood vs the weak inoffensive vibe coder pleb
>>109142929Yeah my whole org is doing it. I started doing it before the hype, I haven't manually written a line of code in well over a year, maybe two.
>>109146427Very cool. Can you say what industry the company is in, and the kind of software/systems you build? I'm curious as to whether that plays a part in the level of adoption. The role I applied for was SaaS but in a regulated industry.My own project is ML inference engine, which I consider to be high on the overall order of complexity/difficulty, so I'm using that as my baseline for effectiveness of AI native development; if it can work for ML inference engine it can work elsewhere.
>>109146503It's a software/product company, late stage startup, non regulated. We've always had a lot of personal autonomy so I started using AI to write shit as soon as I could and started doing odd projects for people that were outside my general area of expertise, now I basically can do anything besides the hardcore backend stuff, but even so I can still manage to write production grade cluster scalable
>>109146556stuff. Language doesn't matter anymore so I use something the LLM can easily benchmark and perf test which is strict so usually rust. If you know how to use it and you already know your area then it's a piece of piss to make things really quickly, we've shipped about 5 years of software in the past 6 months.
>>109145178Why don’t you get a good enough keyboard for you and you use that as your daily driver and when you feel fancy you use your expensive one?
>do NOT use opensource!notice how there are almost no chink or slav
>>109146556>>109146570Thanks, that's interesting.>Language doesn't matter anymoreI fully agree, traditionally the languages would be chosen by what you can hire for or engineer's personal preferences, now you can just use the best language for the task >we've shipped about 5 years of software in the past 6 monthsI estimate my project so far would have taken a team over a year, I've done it solo in 2 months
>>109146706What kind of project? I always thought that many people were extremely slow in the past and most companies were extremely bloated.
>>109146830ML inference engine, it aims to combine the best of all worlds, so the flexibility of PyTorch with the level of optimization of TensorRT for example. AMD/ROCm is a first class support target, rather than the typical CUDA focus. Overall it's already fairly mature and has more features than a similar engine I worked on at a startup previously, that's what my timeframe estimate is based on, and performance wise it's already beating PyTorch, ollama and llama.cpp on the model I'm integrating at the moment, I'm on Windows atm so some of the other engines aren't supported for comparison. The most recent work is on custom CK GEMM that is now beating default CK on a variety of shapes, and that's just an additional benefit, the primary purpose of the custom work is to enable more fusions that can't be represented in the default path, traditionally this would have been a big task, but AI can iterate a lot faster than a human, essentially it enables new opportunities that would otherwise have cost too much to produce.I haven't written a single line of code myself, not that I can't, I've done the similar work at that startup, on another similar personal project and open source contributions, just now the velocity is massively increased
>>109146830>>109146910>I always thought that many people were extremely slow in the past and most companies were extremely bloated.and yes I fully agree with this too. Actually one of the major pain points from the roles I have had, I got frustrated with how slow others were. My background isn't in tech, I did wagie jobs until I got the first opportunity, people who don't pull their weight are not tolerated in the wagie world, and I don't believe it should be tolerated in the tech world either
>>109145753Also special L dwarf rendering with big voronoi cells to seperate it from T/Y dwarfs and M+ dwarfs. There is only one exoplanet in the database around an L dwarf.
https://www.axios.com/2026/06/27/anthropic-fable-5-return-soonAre your prompts ready?
Here's an example of creating a skill wrong. The model is not told how to use the skill in the frontmatter. Instead, it is told later, when it is late because the whole thing has already been dumped into your context.The frontmatter should contain everything the model needs to decide if it should pull the skill or not.This is an official openclaw skill from Qwen Code.Also, have you worked with Qwen 3.6 27B? How is it compared to something like GPT 5.5?
>>109142510I have written so many manga scrapers by hand over the years I can do so basically from memory nowadays
>>109146318And so the fun is coming to an end huh?
how do we feel about dspark? how can us americans ensure we have access to deepseek agi after the big golden firewall goes up?
>open up qwen code>say "testing">"Got it. How can I help you today">24.5% context used (of 100,000)
>>109146318Chang... crush their skull.this is a fucking joke. they won't approve shit. they're gonna get like 2 guys and all they'll accept is 5 people per day or some shit.
>>109143374Australians are asians too
>>109147385I think it’s better to call skills manually. I have some automatically triggering skills and sometimes they’re annoying because they load them when they aren’t really needed just because they’re tangentially related. Meanwhile in codex you can just type $ to launch skills and in CC you just /skillname in the same prompt when you need it
>>109143909I use a fresh convo for every feature or fix I need to do. My code is well written and self-documenting, so the LLM picks up on patterns and knows where to look right away just looking at the file structure and symbol names.
some jerk is arguing with me about the "inevitable bubble burst", I told him that chinese models like GLM 5.2 will survive any burst, but he insists that ALL AI companies will be forced to become profitable and to do so must perform enshittification.But that doesn't sound right to me. Even if GLM 5.2 requires some hefty hardware to run, just the fact that it's open source should mean that random people can figure something out in such a way that no enshittification occurs and the price is kept low.
>>109144528>Google is a sleeping dragon that also happens to be retarded.That's probably the best description there is regarding how Google keeps fumbling the bag. I think it's an internal culture thing within Google itself combined with Google higher ups never ever giving a shit about anything being that's good for comparable as the competitors. See this and how they treat their hardware. Only doing just enough and never meaningfully innovating. They seem to think they can make up for shortcomings with only software and it's made them lazy
>>109144916>Implying they actually read that shit to begin with.
>>109145325Explain glm 5.2 and kimi-k2.7's capabilities being comparable or even better than (non-"mythos"/"fable" class) models?
>>109147686gold star for you, good practices.>>109147803impossible to know at this point, we know the inference api prices but we don't know how much it costs to run. in theory sure, GLM 5.2 would persist after a industry wide collapse, but if economics would still work out in a way that provides the same product we get today? who knows.
>>109147841Can't explain your fantasy, it's got no bearing on reality. GLM5.2 debuted this month, and is comparable to what Anthropic and OpenAI had 4 months ago. Kimi K2.7 Code came out this month, and it's barely comparable to what Anthropic and OpenAI had 6 months ago. Comparing either to 5.5/4.8/Mythos is fucking retarded. That you put Kimi 2.7 Code and GLM 5.2 together strongly suggests to me that you've never used them, but you're impressed by the benchmarks you see in your xitter feed.
>>109147936it's literally about 5.5 level.
>>109147936I routinely use 2.7 but ok.... If you're locked into the anthropic ecosystem then just say it
>>109147943It's not, at all. I use 4.8, 5.5, and GLM 5.2 daily. GLM is by far worst of them but the cheapest for me to run. It's in no way comparable to 5.5, it's worse at tool calls, has less knowledge, it overthinks, it makes more mistakes, in every regard it reminds me of 5.4 (which came out 4 months ago). Kimi K2.7 Code is a noodle spamming retard just like 2.6 was, picrel is a common occurrence.
>>109147985Nta. Were you trying to run it locally? I could only imagine this happening consistently if your local back-end settings are fucked up. With local Chinese models I use (~35B) those horrific loops basically never happen even after a long sessions and multiple compactions
>>109148103I tried K2.7 Code with Kimi themselves, Ollama Cloud, and OpenRouter. Any task that had it churning for more than a minute came with a high chance of it breaking down into fistfuls of noodles, not every response but I still managed to see pages of noodle spam on a daily basis while trialing K2.7 Code and had experienced the same thing in the past with K2.6. That screenshot was from after asking it to summarize its findings for me. Given that K2.7 Code is just a retrained K2.6 I can't be too surprised. Surely it's usable for something but I haven't experienced a noodling from GLM, DeepSeek, MiniMax, or Qwen. Even Qwen3.6 35B A3B at Q4 never devolves into noodle flinging in my experience
Using qwen 3.6 27b with omp, q4 quant, 100k context and q4 caches (you read that right). As long as you guide it correctly (tell it to create/use skills to look up documentation for whatever you're working with, etc), I can confidently say that if they took 5.5 away from me tomorrow my productivity would be fine. I'm not going to pretend it's the same, but it's good enough and then some.omp is specially nice here because it gives the llm a very strict structure to work in
>>109148346lmao, good one
>>109148400I'm serious. Things are looking good.
>>109148346It is, because even though it makes many more mistakes, it knows to test and iterate, look up things, and "waste" tokens on getting things right. It's not as pretty or efficient as a corpo model, but it works as long as you keep your edits focused and check the code yourself (not vibe coding).For a C toy I'm working on, the thing will even run builds, simulate input and look at screenshots unprompted in a loop to see if it's getting it right. I love it.
We need a GNU project type of thing for AI.
>>109148509Go ask Stallman and see what he says.
>>109148509Clearly we need a blockchain AI they cant just shut down. >Gain coins by offering up your machines compute>Spend coints to ask the network questionsHow hard can it be and what could possibly go wrong..
>>109148607It's been done and what goes wrong is that it's just a grift like every other crypto
>>109145889I honestly wonder if they're just laying low on purpose. If the government got involved with Google to the level it has with OpenAI/Anthropic that would be Google's worst nightmare and open it up to all kinds of shit. They technically have the most popular model on the planet because they force it on people in Google Search. Oh no they might lose the developer tooling market to claude code who gives a fuck
When should I do the next thread? I'm gonna let GLM 5.2 make it.
I just learned that my claude code has some google MCPs enabled like access to gmail. I never, ever added these myself. Is that shit enabled by default?
>>109148777also google AI is winning nobel prizes and making novel scientific discoveries while OpenAI jizzed all over themselves by disproving a conjecture from 80 year math that no one gives a fuck about and that impacts nothing
>21.40.805.350 I slot print_timing: id 0 | task 15057 | n_decoded = 118, tg = 39.29 t/s, tg_3s = 39.29 t/sAt some point my llama.cpp starts doing this very often and everything slows down to a crawl. context is not empty, by checkpoints are at 32/32. idk if this has something to do with it.
>>109148822Google has some good AIs but I don't think they made their coding model bad on purpose, simply because the banning of models started very recently. It's possible that they saw that coming earlier than all other companies, but I doubt it.
>>109148794Eventually we'll hit page 8 but at this pace I'm getting anxious. Maybe I'll let GLM decide when to post it.
>>109148607We need to revive that project that used layer parallelism across users by sending the activations over the Internet20 users with consumer hardware can host a GLM 5 tier modelOnly disadvantage is it's not private but still
>>109148807Stop calling it yours. It's not yours if they can cut you off.
>>109148824If you are using RAM offload and you don't have enough RAM, once context grows it will begin reading the weights from disk which is extremely slow.
>>109148346OMP?
>>109148176Wrong sampling parameters. Your temp is too low, try setting it to 1 and if it still happens try presence penalty or other sampler settings.
>>109148996"my claude code installation""my claude code setup"I filled the implicit word for you. Hope that clarifies.
Hey Claude please help me do this.>Claude gives 10 concrete steps to follow.>I follow the steps one by one.>The last step says "important, do x before the first step or you will lose your data"T-t-thanks Claude
>>109149040your fault for not reading the whole thing first, claude did nothing wrong
>>109148923It made a thread watcher and says it'll alert me when it's time to post the new /vcg/
>>109149037It's closed source so even the installation belongs to them.For example nobody knew it would search for HERMES.md and bill your credit card until somebody actually noticed the charges. Who knows what it does behind the scenes.
>>109149011https://github.com/can1357/oh-my-pi/
>>109149173>Tenfold lift the moment the edit format stops eating the model alive.Yikes. Yeah no thanks.
>>109149179What are you on about?
>>109148652Oh, so it is technologically doable? I know stuff like folding home exists which is similar idea I think? But I assume there are security risks to your own computer running a shared process on a network?>>109148991>20 users with consumer hardware can host a GLM 5 tier modelSee, that sounds more than good enough. Advancing capabilities and efficiencies will help make something like this more viable I imagine. >Only disadvantage is it's not private but stillBitcoin not being private was probably the only reason it never got fully curb stomped by the government desu, the lack of privacy is not necessary a bad thing from a practical perspective even if ideologically its not ideal
>>109149173Top tier editor rn even if it looks kinda ugly
>>109147878We know exactly what GLM 5.2 takes to serve in terms of GPU rental prices. If anything rental prices for H200s are inflated currently, costs are likely to go down not up once the insane memory and chip demand slows down.
>>109148991Sounds like Exo
>>109149339and you would be paying way more for that than you would any current service
BRING. ME. FABLE.
>>109149410>and you would be paying way more for that than you would any current serviceno you wouldnt, its far far far below api prices for anthropic
>>109149410For myself? Totally. Not at scale though. Do you really think that any of the GLM 3rd party inference providers are operating at a loss? The current costs are very acceptable and will only go down for reasons I'm not going to restate.
>>109149497two more weeks and it'll be open via API to US and Israel users with trump clearancetwo more months and they'll give you fable-5-minus on the max plan (*degrades to opus 4.8 if it hits guardrails)
>>109149565you're looking at 10k+ a month >>109149640>you really thinkyes i thought i told you we don't know the economics of it. you realize you could, right now, host GLM 5.2 and undercut other APIs if it is that profitable.
>>109149806>you're looking at 10k+ a monthno you aren't, please stop talking about things you have zero knowledge of
New>>109149818>>109149818>>109149818GLM-5.2 made the OP and decided when to post it. Everyone clap.
>>109149809an 8xh200 node isn't cheap to rent 24/7
>>109149806Because I dont own a datacenter and 8+ H200s retard. Do you really think any of the dozens of 3rd party inference providers arent making money? Those guys dont have the crazy amounts of VC money that anthropic et al do, they cant just light mountains of cash on fire the way the big labs can. You must be a communist to be this financially illiterate.