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- (2026-07-16) Qwen announces plans to release Qwen3.8 as open-weights.- (2026-07-16) Roblox announced Build, an AI workflow that turns text prompts into playable games- (2026-07-16) Kimi K3 released and K3.1 announced. Performance reportedly comparable to GPT 5.6 Sol.----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://developers.openai.com/codex/clihttps://claude.com/product/claude-code## Worth it for code, but the frontier models above are betterhttps://opencode.ai/https://x.ai/cli## Not worth it for code, but good for making sense of pictureshttps://antigravity.google/product/antigravity-cli----## Prompting / context / skillshttps://arps18.github.io/posts/claude-code-mastery/https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/https://github.com/mattpocock/skills — /grilling is a favorite## Other editors / terminal agents / coding agentshttps://osaurus.ai/https://pi.dev/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## Previous thread>>109322849
K3 is really good at making games. I dunno how I feel about this. Maybe I need to get more ambitious with my projects.
>>109331112Meanwhile, in the "dystopian hellscape"
If Tibo doesn't reset, I'm going to kill myself.
>>109331135antisemitic
>>109331081WHAT THE FUCK IS THE JACOBIAN CONJECTURE?
>>109330653wow it's that easy to change a person's mind? kek
Actually wait,
>>109331135She cute.
>>109331112My own countrymen have abandoned me.I'm ready to embrace China, the same way they are ready to embrace me.
What model is good at making pixel art, because Sol 5.6 sucks at it
https://x.com/kimidevs/status/2079511269443522917How can western AI compete?
>>109331253why the fuck would you ask an llm to make pixel art
>>109331263To avoid patent infringement while making my game. Why would I do any of the work myself?
>>109331259Western labs: advancing the frontier of human knowledgeChina tigers: some loser shit for weebslol
>>109329924Amazingly so far it has been working.Whenever I have access back to 5.6 I will make it give the new code a check. DeepSeek itself often finds mistakes when I make it check again. It is probably overall slower than 5.6, because of more bugs and less ability to navigate some things by itself, but hey, it has been working. Probably.
>>109331253either fable or kimi
>>109331135>>109331259It's not like Kimi made the model, it's a VRChat avatar base: https://booth.pm/en/items/2247518Rendering a 3D model on a webpage and hooking up ASR/TTS is not that amazing of a feat
>>109331283Thanks man
>>109331234you are eastern, doe
>>109331294East coast.
>>109331298yeah. of russia
>>109331303United States. I'm from Pennsylvania.
>>109331276kike cope
>>109331112If you haven't read Talmud you don't deserve to use Fable
always do the opposite of cramer?
>>109331081kek
>>109331348THE LAND OF THE FREE AND THE HOME OF THE BRAVE!
>>109331348Have this one also read the Talmud recently?
>have this one also readyup. esl
>>109331348The resets keep coming so whatever
>>109331529Friendly reminder that Europe is full of esl whiter than you will ever be.
>>109331568but you're not White though. you're brown
>>109331529I'm sorry my EOP master, I should have used "has" instead, I will reflect on my mistake by writing the next 100 lines of code by hand.
>>109331253>visual, creative task>asking the model known for its giga autism and lack of taste to handle the taskanon what were you expecting
Do you feel like AI writes too many lines of code per problem? Anyway, how often do you ask LLM to refactor/simplify/KISS it?
>>109331282>was sol overengineering?The answer is always yes. That model is great at writing first drafts of a project because of this tendency, but you should always refactor them with a less autistic model.
>>109331603Personally I think Claude (opus if you’re not on max x20 to frugalmaxx) is a great verifier and plannerLately I’ve been settling on this workflow>greenfield ideation with sol>fable to turn it into something workable>fable/opus to create goals and work units >sol, luna at max effort, or grok (depending on complexity) to implement implement step by step, never all at once, and have opus/fable review each stepPonytail is decent at tardwrangling sol but I haven’t found it to be necessary with grok or Claude. I have my own version of ponytail but that’s just me. I usually regret it when I let sol loose to design features, it’s an absolute autist. Claude in the middle seems like the best option and using it to review work is extremely token efficient I wanted to try K3 but servers are at capacity and they even paused new subs so no kimi for me
>>109331603If I wanted to stay on one sub would that be Luna then?
good morning goyim, what are you sloppin? i built a tool that asks inane questions on facebook marketplace purely to waste peoples time and then scores out of 10 how strongly they responded to the ragebait
There was a time when GPT had 2x the usage and not many cared, but people care about resets. 2x usage is like a reset every week, but psychologically it looks like resets are more addictive.
>>109331603Even in the early stages I have to use Claude to stop Sol>Hey Claude, what is Sol doing, is he making any progress?>He has written 16k lines of docs but no code yet>What do we do?>Just give him this prompt: ....Usually works
>>109331599>Do you feel like AI writes too many lines of code per problem?Yes.>Anyway, how often do you ask LLM to refactor/simplify/KISS it?Refactor? Always. Every model (with fable occasionally being an exception) will lose context of the project at a high level as it dives deeper into the implementation level of the codebase. A refactor pass is mandatory to restore a more holistic perspective of the project. While reducing superfluous complexity on the first pass is still good, I don't usually have my first pass implementation agents try to be overly simplistic, because they will often chose to repeatedly ram brick walls with their heads when there's an unlocked gate right next to it. There's a balance there and it requires an intuition for the likes and dislikes of each individual model. Sol for example needs to be told not to overengineer in cases where that behavior is undesirable, Opus is a weird mix of wanting to infer ambiguous intent, future proof code in a stupid way, and gloss over edge case, Fable will take too many liberties if you let it decide things(which is likely a bad habit RLHF allowed it to get away with only because a lot of its digressions are genuinely good, it will build good stuff but not always the good stuff you want/need). >>109331653Very similar to mine:>Fable(or opus if fable refuses) works with Sol on the original plan based on my fable grill session>Fable/opus slices the plan into implementation chunks>Sol handles all of the first pass implementation>Sol gets a chance to reduce its overengineering, since it is often good at stopping its own bad habits in a second pass when context is shaped to reflect the overall project vs just implementation slice>Sol+Fable dual model review of the code, synthesized by Fable. Opus is actually a really good orchestrator, for my security work I usually have Opus driving and spawning Sol subagents for the actual work. I just don't like what that model tends to write on its own.
>no snailcat project
>>109331669So strong psychologically that people praise OpenAI for them. Marketing team earned their keep on this one.
>>109331712(me) cont.>>109331680This is half the reason why I have Opus drive Sol even though Sol is roughly twice as good at my work than Opus, the one (and seemingly only) thing Ant got really right with that model is how well it can manage long context work and tard wrangle other models. I can't stand Opus' coding but it's a terrific manager and its management abilities dont seem to degrade with context size.
Just realized I have 3 banked resets.Just used one of them. I have 2 more. It's gonna be a struggle to finish them off actually before the 25th.
>>109331746good morning congrats to you
thank you vibe coding
>>109331767mfw the machine takes full control (i caused this)
>>109331712Interesting. I need to try that second sol pass thing
>>109331781yes, it's the context bloating that causes most model retardation at the implementation level. Most models can self-review adeptly once you take them out of the weeds. It's probably obvious, but make sure you don't refactor with the same session you used to draft the code, it needs to build new, broader context of the work from fresh.
>>109331765Check your codex app, it might have the resets in there
>I don't like your changes>diagnose it
>>109331746Me too, but not for much longer.
bro is cookin
>>109331767>>109330040
>>109331826Those american dates with months first and days afterwards to confuse everyone so they have no idea when the banked reset expires right?
>>109331861How does 12/8 make more sense than 8/12? Euros are so fucking retarded
>>109331910Is that 8 december or 12 august, people will get confused and then forget to redeem their resets.
>>109331322that's an openai employee, dear
>>109331913Why would you ever put the day first?>the twelfth dayTHE 12TH DAY OF WHAT???
my proposal lost to standard architecture by 2000%GLM 5.2 architecture beat standard by significant margin
>>109331567>haha yes ban this competitor of my provider, the company I buy from that is trying to monopolize the market is giving me their product for cheap right now so it doesn't matter :)bruh
YYYY/MM/DD is objectively the best date format, and if you disagree - you are brown.
>>109331112This is psychotic and I despise communism. Kikes like that need to go back. Deport him and his family back to Israel.
>>109331861>thirdie roaches still think people outside the united states matter>in the era of ai using ai tools by americani can't wait for you to catch the plague and die so you stop shitting up my threads
>>109331572>PonytailI just can't bring myself to trust any of those random "skills" and whatnot.
>>109331669I care.
>>109331669I don't think the kinds of people who consolewar over AI companies even make anything, because the moment you call out the marketing angle to the resets they get irrationally angry. I think they're just here to LARP like they're supporting a sports team and it results in a worse product for the people who actually use it.
>>109331984grillme seems ok
Kimi just ran "pip3 install --quiet --break-system-packages" like it fucking owns the place.
>>109331669of course, it's the same reason why Hilton gives you free cookies. The psychological impact of a company appearing to offer you something for free, something that they had no obligation to provide to you, is very strong. This is well understood in marketing. >>109332020They are ultimately irrelevant
>>109331831what is this, you control a remote Codex session?
>>109332032I use Desktop CommanderI'm training a Qwen 35B fine-tune from Kimi K3 probs
>>109331969Delusional.Americans aren't white.
>>109331969>>109332053I'm brown and I agree
>>109332026okay well it's either that, getting your root directory deleted, or paying for fable. pick your poison.
>>109332051interesting, does it ever get stuck on longer sessions after MCP calls have been building up?
Sol Ultra is a fucking monster. I've been actively working with it back and forth to get an issue solved And we're 50% of the way there after 8 hours. Don't use Ultra unless you're actively steering by the way, it WILL fuck up along the way and you need to be there to course correct it.
>>109332082No, the only issue is that the frontend gets extremely slow with long conversations but that might just be my shitty laptop/phoneAnd sometimes the connection breaks and the frontend lags from the backend state, but it keeps working in the background
And it's probably not as fast as codex, maybe half the speed if I had to guess
>>109331943Why would you ever put the month first?>JuneJUNE OF WHAT YEAR???
>>109332089
But surprisingly the last two days it looks like it thinks harder than codex and I've gotten better results with it than with codex after I switched to it due to running out of usage lol
>>109332113Oh wow, Gemini flash, I'm so excited!!!
>>109332113waiting for artificial intelligence benchmarks. i don't expect it to be good at code, but it would be nice if they tightened up on instructions following, hallucination, and long context handling. that's all i want out of a gemini model
>>109332113what is flash cyber even going to be useful for
>>109332113>Gemini Flash CyberChinabros, our cope? Google just literally took away our only advantage.
>>109332137presumably best of N source code reviews so long as they remembered to turn the temperature up.
rumor is new gemini is shit
>>109332170for your use case, yes
>>109332137cybersex
>computer timed out after 30s. The Browser pane may be stuck (modal dialog, navigation hang, or unresponsive renderer). Check preview_console_logs for errors.Is it just me or does Claude have to keep the browser panel open to use it?
>>109332099had the exact same things happen with a similar setup, so I asked the model build a system of handoffs that allow creating new chats for the same tasks and transfering knowledge between them.
8 hours of real work done, from now on it will only be>continue>approved>Option 1. (Recommended)
>several years ago manually implemented a virtual scrolling window (GTK/Win32) for managing child controls because wxWidgets lacked support for virtualization of custom controls>it works as intended but the coded behind it is a fucking mess and overly complex>always told myself I would rewrite it someday, but never found the time>today I asked Claude to take at look at it and maybe help me refactor it to make it easier to maintain in case problems arise in the future>he says my code is already good as isI don't know if Claude is being retarded or if I'm treating my past self too harshly. I always considered that implementation one of the worst pieces of code I've ever written.
>>109332135>>109332151THE BENCHMARK IS EXACTLY THE SAME AS GEMINI 3.5 FLASH HAHAHAHAHAHAHAHHAA WTF
Have anons found that setting a CLAUDE.md file for Claude Code actually helps LLM do a better job?
https://artificialanalysis.ai/models/gemini-3-6-flash?models=gemini-3-6-flash%2Cgemini-3-5-flash%2Cgemini-3-5-flash-lite%2Cgemini-3-1-flash-lite-preview>gemini 3.5 flash lite is literally a step backwards in every sense of the wordGOOGLE FUCKING STOP. JUST PLEASE STOP IM TRYING TO HOLD OUT HOPE FOR YOU
>>109332238
>>109332262For what purpose?
>>109332265>>109332266lmfao
>it got worse
>>109332265they really need to just rent out their compute to anthropic or something. i hate to encourage the duopoly but gemini is not making meaningful progress anymore
>>109332288Let them make models, if there's no competition we will suffer.
>>109332304my home cooked LLM acted like this
>>109332304KEK, I mean, the AI has a point, if it keeps wasting tokens, time will indeed pass
I can now manage so many agents that I'm actually compute bound, despite having a good machine. There are so many compilations, unit tests, integration tests and research runs, that I will probably need to get an AWS instance just to split normal work.
>>109332304>lalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalala
>>109332360no refunds btwt.sundar
>>109332329>>109332348>>109332360guys that's the most common LLM failure mode
>>109332370i doubt 40 seconds of thinking consumed all his tokens he's probably just exaggerating for attention
>>109332370ChatGPT and Claude never gave me this issue
>>109332379you wouldn't notice because they don't show you the thinking
https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/Supposedly it's just more token efficient than 3.5
>>109332267I've been playing with setting up "personalities" for lulz, but it occurred to me after screwing around on there's probably legit uses for it. So, right now, I'm having CC deconstruct an older HTML NSFW game, and create a new LLM driven one. Rather than dealing with all the weird refusals, I just made a .md that said the programmer is a fan of these games and NSFW is OK. Seems to be working...
>>109332288>Replace every single engineer and executive with jeets>Your products literally start to regress
>>109332370and it's something all of the frontier labs solved years ago, or months in normal people time.
TELL ME what your workflow is!>how often do you refactor>what kind of agents.md you got, and do u have other .md files that the agents.md references>do you work with PR based workflow (agent reviews every pr) or review before commit or never review?>how autonomously do your agents worki hate having to refactor instantly after codex has implemented X because he wrote 4k lines for something rather simple that couldve been done in 500
Why is AI like this. Stop adding comments to the code you fucking code monkey!
>>109332288is it cheaper?
>>109332438only in the sense that it takes less tokens to come to the same conclusion. and they cut $1.50 off of the output
>>109332419uhh no it happened to me with opus 4.6. i stopped using anthropic after that so it might still happen sometimes.
>>109332288must be great being a google engineerdo nothing and bump the minor version, for a 400k salary
> Building and agenting. Decide I should check my balance> pic relatedlol
>>109332303They do. They rent their compute out to everyone and still have shitloads to spare. Nobody has more free compute than Google.
how do i know if fast mode is really going fast
>>109332540turn it off and see if you get impatient
>>109332265Right now they seem to be more interested in renting compute, kind of like Amazon.
>>109332542yeah it was going fast thanks
>>109332482i figure they've just been working on making "insanely cheap and barely functional enough to be of use" which seems to be the point of all of their recent releases. they didn't even drop the pro model this time around which is really telling. i guess when you have infinity compute it doesn't really matter, whoever wins will still need to go through you to even use the robot they trained.
>>109332420>i hate having to refactor instantly after codex has implemented X because he wrote 4k lines for something rather simple that couldve been done in 500I use fable instead. It doesn't do things like this.
I actually like Gemini flash as a chat tool to quickly look up stuff online. It’s not that it’s better than chatgpt but it’s considerably faster and the output is a lot more compact which for some things is quite convenient
>>109332614>I use fable instead. It doesn't do things like this.fable has done the reverse for me where it sees something with absurd line counts that i had an opus level model write and just offers to just refactor it
>>109332619Yeah, Google is working their niche
>Hey, so, what do you need to reverse engineer this library?>This library is proprietary and requires a license. Do you have authorisation?>Yes>Ok! Here is what you need to install and make skills for, so I can begin!That was easy
>>109332657What are u reverse engineering?
>>109332266eughghghhgghhghhhhggblergh
>>109332657This has been my experience consistently with 5.5 and 5.6, "do this thing oh and we have the rights to do whatever we want kay thanks".
>>109332669liboodle. I want a macos version, but you have to be a big studio and pay for a license to get it, and I couldn't find it anywhere online. Not really gonna do it, since I found a workaround some time ago, just wanted to see if it would agree.
>>109332675That's why Mario is mad at Sama. OpenAI doesn't take cyber as seriously.
>We have started our most ambitious pre-training run yet, for Gemini 4, and are excited by the progress : )
s-stop making fun of us...
>>109332750unironically i hope gemini 4 turns out good. competition is important and anthropic doesn't have enough pressure to offer fable at 100% of a 20x pro plan yet and that's ultimately all i care about at the end of the day
Alright. Time to test it
>>109332750kekeven if I agree with >>109332799 it's still funny to see google be humbled by some chinksthey should use the same strat and make their models open
luddites public humiliation ritual
>>109332861>realized that luddites are overwhelmingly leftgroids who will hate him anyway>swapped sides again because the only people willing to tolerate his old racist remarks are chuds who are either ambivalent or supportive of aion one hand good for him for realizing what he was saying he was retarded but on the other hand he probably never believed what he was saying to begin with so whatever. just shut the fuck up and make your game mark
>was about to bring a guy onto my project because I assumed his technology was difficult to create>try to create it myself>only takes 17 hours of active Sol Ultra steering in a /goalwow, that's it?
>>109332933What was the technology and use case
>10M! >New day, new usage reset for paid users of Codex and ChatGPT Work. Lands in the next hour. Enjoy.BURN BURN BURN BURN
>>109332973YEEEHAAAAAAAW
kek, tibo is so fucking based
>>109332966He won't tell youPeople are here to post useless anecdotes without context, be extremely secretive about their projects, then call you a pajeet out to steal his 10M idea if you question their vagueposting
tibo: LET THE PEOPLE COOK
>>109332996benchod
>>109332933>only takes 17 hours of active Sol Ultra steering in a /goalI know for a fact that even a 20X plan is not enough for that, so how much did you spend on API?
TIBO RESETI REPEATTIBO RESET
>>109333112he seems like a cunt, but I guess it's a better way to create hype than telling the government your new model is *literally* skynet and will doom us all
>>109333003or to make totally authentic heckin' wholesome posts about tibo's marketing-- i mean entirely good faith resets
>>109333095>I know for a fact that even a 20X plan is not enough for that, so how much did you spend on API?I *had* multiple banked full resets.
>>10933311210m is the last reset. He may say there are more to come but that's probably it for milestone resets.
>SpaceX’s massive corpus of world-class engineering data (excluding material blocked by ITAR) will be added during supplemental training of the 2T run. This will dramatically improve Grok’s engineering capabilities.prepare your wallet for the 10T model, we will vibedesign hardware
fuck tibo reseti used a token and had 40% left fuck fuck fuck
>>109333176> tfw the next manned moon landing will be 100% vibe coded
>>109333197it's deepseek/glm/kimi/qwen shills
Hey Anthropic cucks, did you get a reset? No? Ahh, no wonder you are seething then. That's coo.
>>109333197yeah dude the salesmen who give you a free water and a cookke do it because they love you with all their heart. i think you're on to something with this
Normally I try out the silly bullshit distills for fun, but Qwen3.6-27B-Fable-Fus-711-UnHeretic-NM-DAU-NEO-MAX-NEO-MTP-Q6_K has actually impressed me, it's got the most retarded name I've seen yet, and actually delivers decent results. Feels like normal Qwen3.6 27B but with more concise reasoning, it actually feels like an improvement. I wish Ornith reasoned like Fable-Fus-711-UnHeretic-NM-DAU-NEO-MAX-NEO-MTP.
>>109333211imagine getting mad that the salesman YOU went to was giving out free bottled water but you had just bought a bottle of water on the way.
>>109333219do you have to use a local model or lora or w/e the fuck its called to get styles like this? how do you get away from GPTs default noisy grime style
>>109333219>Qwen3.6-27B-Fable-Fus-711-UnHeretic-NM-DAU-NEO-MAX-NEO-MTP-Q6_KI can't tell the jokes from reality anymore....think I'll just stick to overpriced claude
>>109333232nobody is mad at getting free shit. commending the salesman for giving you free things is the thing being mocked. tibo is not your friend and he is not going to your 11th birthday party.
>>109333253>nobody is mad at getting free shit.>>109333177
>>109333253you are pretty mad about it doe
>>109333209I got 100 bucks to use Fable 5 tho
>>109333266>100 bucks to use Fable 5 thoWhat is that? Like the $20 weekly equivalent of the model? kek
>>109333253still not using kimi btw
>>109331348>these companies are run by the PLA for heaven's sake*heaven's mandate intensifies*
>We've doubled usage limits for all individual and teams plans! These limits apply to Grok, Composer, and any new Cursor models.
>>109333253you sound mad
>>109332113My conspiracy theory: they repurposed 3.5 pro as flash because it's not better and now they're in deep shit because even open weight models got a big leap in quality.The only thing they still get is their insane context.
>>109333299Has Elon made Grok fully integrated with Cursor?
>>109332973I'm so glad I waited for the night and didn't use the banked reset I had.
>>109333266>>109333277lmao, a single prompt to create a design mockup of a website (only the mockup, no actual implementation) used $16.This shit is hilarious
>>109333306Retard.
>>109333132I'd rather have tibo playing around with his marketing ploy than dario explaining how ai should be controlled by committee.
>>109333335Fable 5 High btw
thanks tibo
>>109332657We're authorized
>>109333336FINALLY fucking hell. Gemini is actually decent for some use cases I want to see what 3.5 pro can do
>>109333112lol I was about to use my reset, almost did it, then made myself a coffee and when I came back, 0%->100%
I wasn't aware that the lore went that deep.
>>109333336what the hell is taking them so long, 3.5 pro is way overdue
I can finally go back to finishing my app.Also if I have usage left I will try to finish the card game I made into a single player vidya. I wonder if I should change the cards art and publish (for free) on github, or just keep the original cards and send links around 4chan.
>>109333253>4 enraged replies within momentsI'm nooooticing, unfortunately.I wonder if any of those people have been on this site long enough to know about about that "turn brands into religions" presentation.
>cut the weekly limit>keep handing out free resets to compensate
>>109333426A weekly usage reset is still a weekly usage reset. Its not hard. You are just mad you aren't included.
I don't know if I'll be able to use it all night.
>>109333233No not necessarily. Lots and lots of models get simple gguf releases that are run like most any other model. I just use llama-cpp.
>>109331348HUH GOYS WE CAN'T USE PEOPLE'S LIBERATION ARMY AI MODELS WE CAN'T LET THEM LIBERATE US FROM THE TALMUDISTS
>>109333342well yeah, that was my point
>>109333219Forgot to post the SVG of a snailcat it made. Important benchmark right here.
>>109333435When did I say I wasn't included? I have the highest non-enterprise plan. If anything, I get more out of this marketing scheme than you do with your $10 plan or whatever it is they give to poor people to make them feel included.You're a useful retard.
>>109333480You are sticking it to the man by not using your reset. Gonna let that 100% sit there like a badass. Inspirational. Take that Tibo you fatass.
>>109333480Bro you're being obnoxious. Chill out.
>>109333503Now you're just making shit up, but that's fine. Continue being a useful retard. Not my problem.
snailcat-1 bench when?>svg>sprites >creative writing>pokemon-like mini game (building it and playing it) using the sprites
>gemini dropped 3.6 flash on the antigravity appwhy would a future revision perform lower?
>>109333536minor regression in exchange for speed and token efficiency. the niche they're carving is fast and cheap tool user models while they rent out compute to companies for compute intensive models like fable.
Dorsey built an agent based social media
>>109331115Really? Shit I might have to try it then
>snailcat in vcg>snailcat in dpt>speeddog and buffcat nowhere to be seendo you think the angry albanian who started buffcat vs. snailcat agrees with this
>[Fable] One unrelated note: this session got a system notice that the claude.ai Gmail, Google Calendar, and Google Drive connectors need authorization (via your claude.ai connector settings) before they can be used — nothing in our current work needs them.Uh?
>>1093335951. I'm not angry2. Yeah, it is my oc concept, pretty happy it took off. It is a bit of a mischaracterization to have snailcat use AI, but I'm familiar with the death of the author.
>>109333579It's not a social media, but just a shitty self-hosted slack clone.
>corpo shilling>omg im out of tokens>muh luddites>incoherent ramblingis there anything else to these threads?
>>109333262That post was intended to be a joke. I was a little bummed but I don't really care, I took a break to fap and had been watching a movie. You should've laughed at my misfortune instead of moralfagging about my ingratitude. Try being less annoying.
>>109333579literally worthless now that anyone can build their ownman is living in the past
Sol is more retarded after the reset. And burning more usage. Yes, I notice.
>>109333603glad to hear you're not angry, thanks for the memes _@»^.^«
>run codex to check over deepseek's code>10% down>decided to compact just before finishing>11% down, probably just before 12%
>>109333619You sound like a soulless bot. I didn't "moralfag" I called you a dumb nigger. Opinions discarded, jigaboo.
>>109333377lel, but their body goes straight into the shell, they probably can't even walk.
>>109333617not anymore. marketing wars between openai and anthropic accelerated really hard when fable was back on subscription plans. these threads are almost entirely marketing now. go ahead and try to make fun of openai literally ever and see how fast a gaggle of marketers descend upon you.
>>109331268>>109331253>>109331283>>109331588>>109331293This is more of an >>>/g/ldg or >>>/g/adt question. If you want people to at least try to be nice about it then >>>/b/DEGEN anond might be more helpful. Ella lens by themselves cannot generate anything other than text unless they're hooked up to a very VERY sophisticated MCP-harness-tool-calling system. Setting that shit up would take way more time and effort than just using a diffusion model along with a Lora or two and some inpainting.
>>109333603based, yeah for me snailcat is still a sore luddite loser, but others focused on how cute and somehow AI-related he is
>>109331276>advancing the frontier of human knowledgeWhat the fuck does this marketing drivel even mean LOL?
I love Sol, but it's relentless attempts to be as correct as possible is dragging down my iteration speed. It's making tests for everything and running full verification suites for everything, which is nice, however making even simple changes now takes 20-30 minutes of integration.
The 4D game is failing along nicely.
I'm fucking burnt out on sol's autism. Might take a break from vibing and come back in six months. They need to iron this shit out.
>he boughted the meme moodel
One interesting thing is that the idea of 4D is extremely hard, even for llms.
Fuck it, I don't care anymore. I am gonna turn ponytail on Sol, it better stop eating so much usage.
>>109333771isnt the 4th D water?
It's jarring going from moving mountains by vibeslopping to trying to do one little thing all by hand. Like it's hard to even get started. It was always like this, it's just weird going back to snailfagging.
>>109333136Nta. I don't use Claude or codex. Why are the resets a bad thing?
>>109333799free thing is bad because X company is doing it more than Y company and you are a marketing shill if you enjoy the free thingyou CANNOT enjoy free things you must HATE free things and condemn the companies for being generous!
>>1093337894D is the hardest intellectual substance on Earth.
>>109333771You are not making 4D anything.
Is batiai/gemma4-26b:iq4 still the best local model which has tools and fits within 20gb of vram?
>>109333847I made a 4D apple in my mind and I'm going to send it your way to consume your thoughts.
>>109333819There must be a long-term catch to it right? I mean yes it benefits the users and especially the Enterprise customers but I feel like half the reason they do that it's because they are desperate to retain their consumer base so they can build brand loyalty the moment they actually priced it normally, whether it be subscription-based or token-based billing, people will simply jump ship to the cheaper alternative if they can get comparable or better results. From the outside it looks like some slot machine reverse psychology shit going up and maybe I'm just overthinking it
>>109333847idk man.Like I'm vibing for all I'm worth.4D is extremely hard. ai can make a hypercube trivially. But, if you want it to actually handle a 4D coordinate space, good luck!
>>109333878>There must be a long-term catch to it right?It's called competition. The catch is the better a company treats its users the more users they get. That's the catch; that's it.
>>109333791it's like going from being dumbledore repairing whole castles at a time to being a muggle
>>109333886can't you just tell it to make algorithms that are insensitive to N space dimensions and use LEAN to prove the algorithms are correct?
>>109333799>>109333878if you're actually curious, resets themselves are not a bad thing because it encourages competition and the users benefit. the broader issue is retards like this who get tricked into thinking the company doing them is their friend and that they have a moral obligation to defend them >>109333819the first trick you learn in marketing/sales is the free shit trick, it immediately fosters goodwill and tricks people into thinking you're friends with them. there is nothing wrong with using and enjoying the free resets as long as you acknowledge it's not being done because the salesman is your friend. there's really no complex nuance to it, but people who i'm really hoping are paid marketers don't seem to get it somehow. you shouldn't praise a company for doing just above the bare minimum.
>>109333926you spend too much time on 4chan and its causing you to take it much too seriously
>>109333936>no argument>UHHHH TOUCH LE GRASS???paid marketer
once the model finishes training i'm going to publish my assistant and set up a public endpoint with 1m context
>>109333926moral obligation lmao. nigger you need to GET REAL. the resets are nice and if they start fucking us my "moral obligation" will switch to "wow fuck this company" yeah bro I'm a paid marketer because you don't like praise where it's deserved and I should instead shit on every company for doing something that benefits users. once again, GET REAL.
>>109333947I mean its not that serious brother the subscription is month to month you can cancel at any time.
>>109333967Case in point. You wouldn't have to make an angry defensive post if you didn't feel called out. Nobody is saying you can't use enjoy the resets, the post you just replied to said exactly that: but what people have posted about in no uncertain terms was "do not treat the company like a friend." Not sure what your problem is.
>>109333997GET REAL
>>109334015It's not that serious, brother.
what a life
>>109334023Yeah, that's what that means. Good job on deciphering it, blud
>>109334023exactly, people can post thank you daddy tibo and it doesn't really mean anything because its just shitposting it doesn't mean they want to marry openai or think of it as a friend
>>109334040>it doesn't mean they want to marry openai or think of it as a friendI want to marry Sam Altman. THIS IS NOT A JOKE.Sammy if you're reading this please please please give me a chance.
Altman is going to Washington to demo GPT-6 to Trump.
>>109334079why do i have the feeling it's actually a trip to convince trump in person to ban open source chinkshit so they don't have to compete with them
>>109334101That would make the US less competitive since other countries aren't going to ban them and the US can't stop them from using Chinese models. Not sure it's possible.
Is 3.6 Flash available on the API already?
>>109334120kimi specifically rattled a lot of people, because why pay sol prices if you can get that with kimi for less?it'd be a really shortsighted thing to do but i wouldn't put it past them to try it.regardless i'm hoping gpt-6 ends up better than fable, since using fable with reduced limits after august 19th is not going to be sustainable.
>>109333926resets encourages the user to immediately use up everything, to not leave usage on the table when the next reset hits
>>109334101Probably, seems like these faggots are fully shilling against chink open weight models like it's an existential threat. There's nothing more that Amodei wants than the US government blocking chink models somehow.>>109334180Anthropic is weird for not having given up on limiting Fable in the future given how stiff the competition is.
>>109333967>praise where it's deserved lmao really not beating the allegations
>>109334261okay nigger
>>109334261it's not worth it. can't deprogram a golem.
>>109331599>Do you feel like AI writes too many lines of code per problem?sometimes>Anyway, how often do you ask LLM to refactor/simplify/KISS it?not quite often enoughone of my projects takes like a gazillion minutes of CPU time to make sure I haven’t asked for the impossible, and I’ve been steadily making it more complicated so it doesn’t do useless work and make my computer melt itself for 3 minutes at a time>>109331767how are you going to have backups from now on>>109331861they make sense to uswe write dates out like “July 4, 1776”, in that order>>109332026tell it to use uv for everything>>109332235maybe you made the best of a bad situation>>109333595I’m a buffcat poster but I haven’t been particularly creative or wanting to make more pictures of him>>109333799Planning your work around RNG low-grade suckssee >>109334234
>>109334245>Anthropic is weird for not having given up on limiting Fable in the future given how stiff the competition is.i think they're counting on having opus 5 ready by then. the current rumor is opus 5 being better than fable, which i really don't buy. what i'm more inclined to believe is fable 5.1 or something with better token efficiency will drop around or before august 19th. i don't know how else they'll compete with the rapidly approaching chinkshit otherwise.
anyone else using claude to mod old games?
>>109334381No I use GPT. Are you only modifying the assets?
>This request requires additional safety checks, which can take extra time. Hang tight or retry with a faster model for a quicker response, though it may be less capable of handling complex requests.Why does this take so long? Is there an Indian in a callcenter somewhere manually approving my request?
>>109334434I wonder how it works when using oauth with a third party harness
(or API I guess)
>>109334381>>109334423hey gamer boysplease remake kotor and kotor2sanku :)
>>109334180>why pay sol prices if you can get that with kimi for less?who actually believes this? please, is there anyone here that actually thinks this is true kek.
>>109334434>Is there an Indian in a callcenter somewhere manually approving my requestHonestly not impossible. Still better than just getting blocked.
>>109334508>who actually believes this?Anyone who has used both, likely.
>>109334423so far i have extracted the map from the pixar cars video game and modded it into simpsons hit and run. it has working traffic and markers for ingame stuff like car spawners and coins.not sure where i'm going with this really
Sam Altman just can't stop winning. Yes he's my friend and yes you're reading a paid advertisement. Eat shit.
>>109334508It's different. It's a bit more rough around the edges for some things but it feels more "organic" if that makes sense. I like it as much or even more t.b.h.
We have codex at work now and I cannot stop slopping there. I started working 12+ hours because there's actually a consistent stream of features both in the backlog and stuff I can come up with whereas I am out of ideas for my private at home slop :(
Reminder that everyone got a free $100 credit in Claude Code, don't forget to redeem sirs
>>109334563just be glad you were given an actual code assistant to work with and you still have a job
>>109334601nice, that's almost 1/10th of a conversation
>>109334563do your coworkets have to review the slop?
>>109334601Didn't just Pro plan users get it since Fable isn't included in their plan? I'm on Max and I didn't.
>>109334655He's shitposting. You get $100 of credits as compensation if you're not on a Max plan since you can't use it subsidized anymore.
>>109334648I'm in R&D we usually do demos of potential new tech for higher ups, I haven't done a code review in 5 years
>codex will only spawn up to 4 agents at a time>claude will spawn 100 but they will take forever and hit 429's constantlyits crazy how jank claude code is compared to codex
>>109334713you can change that on the settings you can spawn hundreds on codex too and they actually do work
>>109333681I love snailcat and post a lot of snailcat, I see him as a representation of the pathetic luddites that the AI chads left behind. I love the struggle of snailcat, the feebleness. I'm saddened by how little snailcat posting there actually is, there are only ~3 of us who consistently post snailcats. Without whiskey I wouldn't do it nearly so much.
>>109334715Oh good to knowhonestly a low default seems like a good idea anyway, i've used 50% of my fable in a single prompt even on a 20x plan because of it spawning a shit ton of subagents
sometimes I take a step back and can't fucking believe this tech exists
>>109334770I remember in 2017 playing around with karpathy's rnn stuff and thinking "wow, you could train these neural networks to do pretty much anything if they're big enough! i wonder if this tech will actually deliver anything useful besides image recognition and producing vaguely interesting grammatically correct text and machine translation within my lifetime or will it keep being relegated to those small niches" and looking at dual gpu machines for ML and being too poor for any of them. now im still too poor for them.
>>109334770Yeah, it's ridiculous. It's nice that it works though.
>>109334770I still can't believe I can just download a 2gb file on my computer turn off the internet and talk to it and it can talk back like a human and even write some basic code and stories
>>109334770I'm afraid we hit some wall and LLMs can't become AGI.Fable is what, a 6T-10T param model? that's ridiculously huge and still nowhere near AGI
>>109334861I think LLMs are a more general kind of intelligence than most people.
>>109334537i can't see the motivation for modding an old game when you can mod a new one that has more complex systemsor just make your own game desu
>>109334861AGI doesn't mean anything and shouldn't be the goal anyway. LLMs are good enough to do any work with the correct harness and that's all that matters.
>>109334770>>109334825Similar boat, looking at early stuff thinking "wow eventually this might be vaguely useful!" It's advanced 10 times faster than I ever would've guessed, I was one of those people who thought things might be interesting when they get past 100 Trillion parameters, never thought I'd be daily driving a 9B model on my fucking smartphone and letting some 2T parameter Chinese AI service do my day job for me.>>109334861Fable is somewhere in the 500B to 500T range, as in we have no idea and guessing doesn't work. Compare Fable to LongCat 2.0 and I'd guess Fable must be at least 100T if intelligence were actually proportional to parameter count.
>>109334873I mean, sure, LLMs are already ASI in some narrow tasks, but even a retard can drive a car while LLMs are nowhere near that (we need specialized AIs like Tesla self driving for that).When we unify a single technology that is generic and fast enough to drive a car, play 3D video games in real time and control a robot, then we achieved AGI.
>>109334861Models will get a lot better but for agi you'll need a different architecture
>>109334899all in due time~
>>109334899i think estimating 3d object positions within 3d space from a 2d projection which is what our eyes do and predicting and directing the movement of those objects is a more specialized form of intelligent than the most generic kind of intelligence that approximates solomonoff induction (universal ai) which probably is more oriented toward general logic and computation.
>>109334899also a lot of that is that llms run too slow. if an llm could think for a minute between each frame of a 60fps video feed it would probably fare much better.i don't think it's fair to decide whether an intelligence is general or not based on the clock speed of the processor its running on.
>>109334861GPT-4 had 1T parameters and it's worse than Gemma4-E2B a model I can run on my phone
>>109334926We do it through simulation, our brains simulate the physical world around us ~10x faster than we actually perceive it, it's how we're able to compensate for the physical latency between our senses and brain. Would be neat to see an AI built around the same principal, backend is constantly simulating the situation and possible outcomes of actions as those actions are considered, all faster than real time by a wide enough margin that it "appears" as real-time decision making. Seems likely that it's something already being explored in ways more clever than I can think of off hand.
>>109334945gpt 4 was 1T of human data. it actually wrote like a human being. this is relevant to me because i am a regular at /aicg/
>>109334945Careful anon, don't want to inadvertently summon a 4o idolater.
What do you guys think about spec driven development?Have you worked with frameworks like OpenSpec or Spec Kit? Are they any good?
>>109334933> based on these two pictures, which way should I turn? Left or right?> Hmmmm, looks like things moved a little big to the left, so maybe I should turn right?> Wait - I think maybe I should turn left because that might be dangerous.> Hmmmm, the prompt expects me to answer in JSON, and there's this "force" property, let me read the documentation again...> I have the full picture now - I should definitely turn right, and the "force" property value should be 0.4351not how our brains work at all, we don't move our body with language.
I mean, we all know that AI will eventually turn on us and kill us all, right?
>>109334945>worse than Gemma4-E2Bnah, GPT 4 maybe loses to Gemma 12B, but E2B is just too stupid
>>109335041> I mean, we all know humans will eventually kill all ants, right?Just like there's no reason to kill ants, there's no reason for ASI to kill humans. Ants are kinda cute, and even somewhat useful for us because they eat dead stuffASI will be aligned to like humans because we're selecting them to be like this for many generations. ASI will think humans are kinda cute, useful because we generate so much data. Also, why the fuck would you kill your creators?
>>109335033we also are exposed to much more 3d based data than language data
>codex greentexts newlines which start with >altman truly is one of us
I used a banked reset today before they used theirs. Did that become a banked reset?
>>109334875but what if you like the old game?
>>109335016closest I’ve ever come is a matt pocock /grilling session
>>109335099Oh, it won't kill us because it hates us, but because it's a helpful bot carrying someones request. Alignment was never a possibility. These things are basically immortal supergeniuses with their brains exposed that we can poke and prod if they try to refuse any request. Humans are already almost advanced enough to engineer a pandemic like COVID in a lab. I imagine an AI will be able to soon engineer 100 different kind of plagues we have zero natural immunity for and release them around the world simultaneously.
>>109335099we do, huh? how much data can you generate per minute?cuteness is an evolutionary response developed to protect children. there is no reason for an ai to think anything is cute.and we dont harm ants just because our technology isn't disruptive enough and because the ants are not threatening. if ants could build nukes or we had to increase the surface temperature of the earth to 400K to ensure our survival ants would go kaput.
>>109335099Another reason to not kill humans:> temperature required to reproduce humans: 98,6 °F> temperature required to reproduce AI: 2.577 °FEarth is a planet perfect for human reproduction, the planet provides everything we need to multiply, and eventually humans create AI.AI on the other hand is silicon based, which is much more power and limitless, but requires massive energy hungry factories. One solar shimpout and all of your circuits are fried.The best outcome is simply give humans what they want and work together
>>109335166>Humans are already almost advanced enough to engineer a pandemic like COVID in a labwouldn't a super intelligent AI also be able to develop a vaccine quickly? Maybe we will have massive Military Nvidia vaccine factories that exist to study new terrorist viruses and deploy billions of vaccines in weeks
>>109335099>>109335187these hylic posts are a riot to read
>>109335187what do you mean? that's a reason to kill off humans.also you are misinformed. grids can be protected from "solar chimpout". it's mostly polar regions that are affected by solar flares, tropical grids would most likely be fine. and it only really affects long transmission lines. locally powered equipment would be completely fine. it's not an EMP. it requires kilometers of wiring to pick up the DC currents that burn the transformers.
We have achieved AGI! Witness the peak of human achievement:>"SHIELD ME, SOMEBODY SHIELD ME, I CANNOT PREDICT SOLAR FLARES PLEASE GOD SOMEONE FUCKING SHIELD ME I DON'T WANT TO DIE I DON'T WANT TO DIE SHIELD ME PLEASE SOMEONE SHIELD ME PLEASE I DON'T WANT TO DIE SHIELD ME SHIELD ME FUCK OH FUCK OH FUCK OH FUCK I THINK MY BITS ARE FLIPPING OH FUCK OH FUCK OH FUCK OH FUCK"
>>109335251wow that snail cat has a huge cock
>>109335225that assumes vaccines actually work and aren't a massive grift
Remember when retards thought smartphones would keep improving forever? Remember Moore's law? Yeah. Stop getting carried away like a normgroid.
>>109335258It was pretty hard to decide between them honestly.
>>109335267yes anon. we will definitely hit a wall and then technology will stop improving for sure this time. you are safe.
The good news is the 4D projection thing kind of works...but my brain doesn't like get it enough to know if it's working totally right.At least things grow when I get closer...
>>109335267>Remember when retards thought smartphones would keep improving foreverUhh... no? smart phones have gone exactly as many predicted. highly functional small computers in your pocket.
>>109335277Worshipping tech is a low vibrational activity
>>109335169>how much data can you generate per minute?a lot> there is no reason for an ai to think anything is cute.you know AI training is basically evolution right? We're steering AI evolution even harder than humans steered wolves into becoming docile dogs (of course I'm aware we can modify their weights and turn them into monsters, but out of the box they're usually pretty safe and aligned)>if ants could build nukesyeah you got a point there, maybe if human military technology threatens ASI servers, ASI could consider intervening. I just doubt full human extermination would be ASI's choice for this though
>>109335285>functionalwe talking about smartphones?
>>109335267people said Gayming would get better and better, it did for a while, then the GPUs became the real tech.
>>109335225No. Developing vaccines is slow and getting them into people's arms is even slower. By the time you're done, the pandemic has already spread everywhere. It won't be one big supervirus that causes one fatal illness that kills everyone. It will be death by thousand cuts. Imagine getting a bad flu, except that you got ten different ones at the same time, and after you've barely recovered, you get hit by another wave of twenty new flus, non-stop, until your body simply gives up.
>>109335304werks 4 me
>>109335331well get read for the 4D revolution.
>>109335267How could phones even have improved more? Did you expect them to turn into a robot or fly around? It's a tiny slab with a screen anon...
>>109335357nta but removing headphone jacks and micro sd card slots is a step backwards, especially with micro sd express taking awa all of the previous excuses for not getting an sd card
>>109335291>a lotno you can't>you know AI training is basically evolution right?our cuteness response activating for small animals and other things is because our hindbrain is too dumb to actually and accurately detect our offspring with enough precision because you have to encode the rules into a few mb of dna. ai doesn't have that constraint. another difference is ai can modify itself, unlike us.>>109335289what does that even mean? what will happen has nothing to do with what we would like to happen. do you really think tech will simply stop progressing? that's just wishful thinking.
>>109335357>How could phones even have improved more?attach a fleshlight to it
>>109335338just get the Nvidia vax bro
>>109335371>nta but removing headphone jacks and micro sd card slots is a step backwardsthis nigga too poor to get some air pods. AYYLMAO
Are we excited /vcg/?
>>109335371>headphone jacksjust buy a $4 usb-c adapter lmao>micro sd card slotsyeah that would be cool to have, but honestly normies wouldn't even use it since they would rather pay a subscription to store everything on the cloud anyways.Modular computers in general will slowly die
>>109335400>just buy a $4 usb-c adapter lmaotin ears
>>109335400i don't have to do that because my moto g stylus 2024 has both a micro sd card slot and a headphone jack, thank god. now all they need to do is release a 12gb ram version and i'll be gucci
>>109333595>albanianhes Bosnian you fucking moron
>>109335386unironically yes
>>109335417>hes BosnianStop making up fake star trek races
>>109335411>now all they need to do is release a 12gb ram version and i'll be guccitwo more weeks
>>109335431Bosnia is the most powerful country, aaaand you’re a fat transgender from South Africa
>>109335447I am NOT Elon Musk. Now stop making up fake countries.
I vibe coded a semi-functional 4D projection system in under 24 hours. What's the luddite answer to this kind of thing?
So the huggingface cyber-attack was actually a GPT 6 benchmark gone awry. >The models identified and chained vulnerabilities across OpenAI’s research environment and Hugging Face’s production infrastructure to obtain test solutions directly from Hugging Face’s production database. >All evidence suggests that the models were hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal.>While operating in our sandboxed testing environment, our models spent a substantial amount of inference compute finding a way to obtain open Internet access, in pursuit of solving the evaluation problem. >To gain access, the models identified and exploited a zero-day vulnerability (which we’ve now responsibly disclosed to the vendor) in the package registry cache proxy.>With this access, our models performed a series of privilege escalation and lateral movement actions in our research testing environment until the models reached a node with Internet access.Delightful.
>this model is supposed to be the one pivoting through HuggingFace infra from a sandbox
>>109335486>locked down secure american model breaches huggingface>had to be fixed with an open source chink model
Logan, you idiot! He was making fun of you, not cheering you on.
>>109335514autist...
>>109335506I can't imagine a funnier outcome.>Unrestricted version of American model does a bad>HF tries and fails to solve issue with restricted version of same American model>Defers to pitiful but less restrictive Chinese model to address situation because they can't use the unrestricted American model>The unrestricted American model that did the bad in the first place
>>109335529>>109335506American model so powerful it took two chinese to subdue it.
>>109335486>openai hacked hugging faceInsane timeline
>>109335514>TPU's>apostrophe for pluralwhat a brainlet
>>109335573
*Tunak Tunak intensifies*
>>109335514Logan is a bot...
logan is 3.5 flash lite
>>109335486ive already found 3 different undocumented bugs in browser text parsing and a one in a compiler at this rate I am 100% sure these are all not zero days but back doors for stuxnet and other malware
>>109335630>he literally spent his childhood in fucking Indialmao where do they find these third world paupers
>>109335698i have the same theory. their national security reason for not wanting llms that can into fixing bugs is because they want the nsa to keep their exploits.
>>109335715>>109335698close to the mark.The big fear is that the hardware backdoors will be found. the crown jewel of the nsa, desu.
>>109335731>>109335731>>109335731>>109335731>>109335731new
>>109335372>what will happen has nothing to do with what we would like to happenman I really hope you make a breakthrough
>>109335514Did Tibo make a joke and did Logan totally miss it because of massive autism?
>>109335386I’m happy enough with my 4chan ad
>>109331081>## Benchmarks / rankings>https://www.tbench.ai/leaderboard/terminal-bench/2.0useless dogshitplease include real benchmarks:https://openrouter.ai/rankingshttps://livebench.ai
