A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs. Physical harness edition.-- Frontier models - start here if you have $20 or sohttps://claude.com/product/claude-codehttps://developers.openai.com/codex/cli-- B-tierhttps://x.ai/clihttps://platform.deepseek.com-- 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-- Other editors / terminal agents / coding agentshttps://pi.dev/https://opencode.ai/https://cursor.com/docs-- Benchmarks / rankingshttps://www.tbench.ai/leaderboard/terminal-bench/2.1https://artificialanalysis.ai/-- What we’ve donehttps://vcg.gitgud.site-- Previous thread>>109645823
#Goals
>>109653350vibeBUMP
why are xitterfags dissing gemini staffs for hyping up ox, what's wrong with that? I'm starting to feel bad for them
glm 5.3 flash is damn good. most flash models have two egregious weaknesses: awful long context and awful hallucination. glm 5.3 flash has NONE of those. state of the art long context, and claude level reservations about saying stuff it doesn't know. it's the perfect cheap model, even outside of coding. luna is great, but it also hallucinates when it doesn't know something. that's a pretty big weakness.
>>109653350Kino
>>109653382Don't feel bad for corpo who has millions of GPUs but can't release a compete coding model
>>109653382because google alignment and shitter alignment are orthogonal
just got some of my shit uploaded on the #1 hub for formalized math. blew my mind that they accept llm-generated proofs and submissions. the site is pretty automated itself. if any of you goyim have spare compute cycles and wanna learn about building loops, try your hand at this. fun hobby
>>109653399luna is always badwhat about the usage?
>>109653447i can't think of anything more pointless to destroy the planet with.at least porn generation entertains someone. math vibe slopping is for nobody.
AGI>Its leaders won’t quite declare they’ve reached that threshold. But they no longer speak about it as a distant abstraction. >Chief research officer Mark Chen estimated OpenAI is “80% of the way” to AGI. >Brockman said that viewed from two years in the future, this may be remembered as the moment AGI was created. >Altman told me that OpenAI was “not quite yet” there, but that by the end of the year the company would have an internal system he would call AGI.about Astra>One of OpenAI’s research goals for this year was to automate the work of an entry-level AI researcher. Pachocki says the company has already met its internal benchmark for an automated AI research intern
>>109653457nta, I fucked around it with the lowest paid tier on Z.ai ($18/mo) and 240.91M tokens cost me 5% the weekly limit
rumor>"Doug" is expected to be the base for Astra and GPT-6>OpenAI recently finished its next pretrain, codename "Bel" - the successor to "Doug">Bel is a giant pretrain with >10T total parameters. OpenAI expect it to be a post-GPT-6 base, and potentially even the base for an AGI-threshold model.
>>109653512kek@people thinking a model as small as 10T can be AGI
i got AGI fatigue
>>109653527I got RSI fatigue.
>>109653512>mostly as a result of compute constraints, Anthropic are bracing for much of the rest of the year to go OpenAI's way, but think they'll be back on top early next yearcan we avoid the worst timeline? things are looking good
>>109653470fool that you are, ask your slopcannon about training on lean 4
>>109653484AGI more GayGI ami rite
>>109653525It's not just about the size
>>109653568Yes it's just about the sizeAll the AI advances this past few years have been due to size
>>109653574deepsex disagree
You will own nothing.And you will be happy.
>>109653594Owning nothing is literally communism though?
we will own everything
>>109653610not a single woman in this image shaves
are you AIpilled or are you a snailcat?
>>109653619you can visibly see her shaved legs
>>109653540>inb4 Anthropic tries to postpone the IPO, fails to secure more private investment and goes under
>>109653574Intelligence scaling with size doesn't change that the constant cost factor reducing means you can get the same intelligence in a smaller size. gpt4 is multiple times bigger than models multiple times more intelligent, and there's still a lot of progress to be done.The unreleased models are probably good enough to autonomously drive research in that area by now.My prediction is that something twice as smart as Fable 5 can fit in 16B parameters. Though I don't think that thing will necessarily be a transformer or even an LLM.
>>109653644>The unreleased models are probably good enough to autonomously drive research in that area by now.
what are the usage caps between all claude tiers
>>109653626This the guy that got his AI fund margin called?
>>109653656lol i got my ai to improve itself 3 months ago autonomously. double digit benchmark speedups
>>109653670open loop
>>109653610This doesn't even have snail parts anymore.
>>109653656mate, Astra is designing chips for OAI that mog anything else, including nvidia
>>109653686BUYANAD
>>109653686>specialized hardware works better than general hardwarewhoa...
i love this new trend of models being able to watch videos. gemini did it first but the chinese models and grok following their lead 3 years later is such a welcome development.this is glm 5.3 flash btw
WebMCPhttps://googlechromelabs.github.io/webmcp-tools/demos/explainer/>>109653684holy newfag
>>109653626these are E = mc^2 + AI levels of retardation
>>109653707Finally total AI surveillance. We only need more data centers to make it better than chinks and russians just identifying faces and tracking people.
>>109653686> ASICwill literally go into the trash when new AI architectures are out
>>109653728they tested the chip on different open models
>>109653745You literally don't know what ASICs are
>>109653686>including nvidiahttps://cloud.google.com/blog/products/networking/introducing-virgo-megascale-data-center-fabric
What did he mean by this vcg
Not that I'm complaining (and this will probably unc me as I graduated pre-covid in 2018) but I know LLMs got big over the years but how did OTHER models like image that have nothing to do with text also get so much more powerful in the past years that you can have GPT describe you an image with perfect detail?When I was in uni, we did shit like pic rel with neural nets, you had to have a predefined list of subjects, if something in your image wasn't in it, it wouldn't detect it (as classification models do) and it had like 80% accuracy (which, all things considered, is pretty crap)Is it all transformers?
>>109653686you might not have noticed but AI writes software now, better than any human. it can make computers do anything it wants. it can create tools that help it do more faster. it can collect data and train new llms by itself. we've already reached the singularity bro. once it solves the efficiency and static memory problem and builds robots that allow it to do anything it wants in the physical world it's over.
>>109653807who cares as long as you have a code coverage report
>>109653824
>>109653850thanks lolI hope you at least used 3.7 Flash
>>109653850the description of ViT is accurate but it really doesn't explain why it works at all. sure, humans use context of previous words to predict the next word. but we don't view images as a long sequence of 16x16 pixel blocks arranged in order. it's really not clear why that technique should generalize at all.
>>109653865I actually blew out my weekly Fable quota and Anthropic has now banned my account for use of an ageist slur.
grok isn't b tier. grok is what I use for agentic coding.
>>109653899>Anthropic has now banned my account for use of an ageist slurgramps?Anyway, they say DeepSeek is pretty good.
>>109653753>whatwho*
How do you cope with having generated hot bitches and now Sam Altman knows you generated it?
>Fable 5.1 release seems imminent are we ready?
>>109653899>banned my account for use of an ageist slurbad goyim! You won't get access to frontier models because you're such dirty, bad bad goyim
>>109653943No because nobody uses Fable 5 at its current limits anyway.
>>109653944imagine if ai actually replaces thinking, and you get banned from ai platforms. literally banned from being able to think. they may as well put you in a zoo at that point
>>109653865exclusively. it's a great asset and cheap as balls
>>10965395320X anon here, Fable is such a great Opus orchestrator, you don't need these massive models to do the dirty code monkey work.
>>109653943how are they even supposed to hype this up anymore?at this point it either has to be Fable 5 level but 300 token/s and 10% the priceor it literally has to one shot GTA 6nothing i do even needs fable
Anyone else into backrooms or other creepypasta? I had fable one-shot a backrooms complex sim based on another project I made and it's pretty damn creepy, complete with 120hz hum. I had claude put in a stalker enemy but I haven't seen it yet. I just keep hearing footsteps behind me but when I turn around, I don't see anything.
>>109653968Has anyone used Fable as a Grok 4.6 orchestrator?>>109653974I can't 1 shot anything, because I don't have the full specs before coding. I'm not just cloning something.
>>109653996>I can't 1 shot anything, because I don't have the full specs before coding. I'm not just cloning something.that's what I was trying to saythat's why I am wondering what are they going to sell at this point unless the thing can literally mind readI dont need more intelligence, I need 5.6 Luna at 3000 tokens/s
>>109653984>the enemy failing to render is actually a feature!
>>109653953today on "i do not vibe code but i am paid to post about vibecoding"
>>109654009>I need 5.6 Luna at 3000 tokens/suse sub-agents and /fast ?
>>109654009>unless the thing can literally mind readThey're already trying to make AIs interpret brain signals.
>>109654009I can't find anyone who has extensively used Grok 4.6, who says "I reach for Fable".
>>109654034Damn, you got me. I only vibespec.
>>109654058>I DON'T NEED MORE SO THAT MEANS YOU DON'T NEED MOREalright
>Ran 23 commands · ctrl + t to view transcript>Context compactedrip
we need more until we all have AI robot girlfriends. until then it's never enough.
>>109654058just saw the reverse of it on xitter>I can't fix security bug with fable so I reached for grok
>>109648514I can just steal KSP content?
>>109654090>AI robot girlfriends.Will be made illegal. (Not AI boyfriends though.)
>>109654114i'm transing the ai boyfriends. vibe boobs will be installed
>>109654074For me, Fable would be a big expense. If I got it for free, I'd test it out and see if it helps.>>109654097>>I can't fix security bug with fable so I reached for grokPretty interesting. It makes sense that far fewer people are using grok as their main driver, so examples of the reverse are fewer.
>out of tokensNO NOT YET! FUCK! RESET ME TIBO!
>>109653895
>>109654138
>>109654035that gives you maybe 150. there's at least one or two orders of magnitude to be gained still
>>109653365I see you are getting your 65% worth
GoalStatus: activeTokens used: 5.19M
> finally figured out how to reduce GitHub Actions costs thanks Fable-chan
>>109654143saccadic masking is absolutely nothing like ViT patches.>>109654138>every patch attends to every other patchagain, this is an idea from LLMs like BERT and it's not clear why that should work or even conceptually make sense, especially since the 16x16 size is set in stone and completely arbitrary. like imagine if every token in gpt was 5 characters long for no particular reason
Which per token payment APIs have the lowest base fees?
>tfw I don't use the chatbot, the chatbot uses me
>>109654404her pussy broke wtf
>>109654404whose arm flew off tho
>>109654404also funny video!
I got Sol to a point where I asked only for new features and stated my expectations.I told it “I expect A, I expect B, I do not expect C, etc.”.It ended up deleting way more than it added, without me telling it (sort of) to simplify or delete anything.I’m excited.It might end up horribly broken, but at least it’s not always 99% added lines.
>>109654475> SolI wouldn't get to excited, but it's good to see OpenAI might be improving their steering towards less bloat
Few people realize how life-changing vibecoding can be. I added two features inside of an hour that would have taken me at minimum a whole day to implement. I did it at nearly the speed of thinking of the idea. (I added a clock and markers to an app I already had modded). It's so weird to use a computer now.It's not a world of annoyance, but a world of a hammer searching for a nail.
>slopping game idea that integrates LLMs into the game Man Fable is amazing but I hate that I cant fully trust Claude to not fuck with it due to it maybe triggering some flag due to Anthropic's paranoia about people stealing their tech.Is scam altman cooler on this stuff?
Claude Opus specially insufferable today. A few sessions already where it says "Starting step X now" and it doesn't start anything. Or "I'm dispatching this to a subagent for review" and it doesn't do it. Last week I used to say "work from Step 1 to 8 autonomously, without checkpoints" and it would still stop on each step for a checkpoint.Pretty annoying for "the best AI in the world".
>>109654404Reality soon.
>>109654526>Is scam altman cooler on this stuff?yes, Sol doesn't give a fuckbut we're not hopeful for Astra, that one might have cuck triggers too
>>109654497It’s definitely not OpenAI, I did some weird thing to counteract their usual models.>walk Sol through implementing something once, by hand, step by step, at a low level>ruthlessly interrogate it>be not afraid to wipe out entire systems if they get in the way and it cannot be easily explained why they exist>”it’s in git, it’s fine, delete these 36 tests I don’t care anymore”>accept that even though we may get the new feature, we blew a reckless hole through the codebasethen>ask an agent to go through that session, then create a runbook/document/skill whatever so that you can invoke an agent that will behave like you didthen>hey agent, I want this, use this document>agent spins up a subagent and does what I did >it does actually stop hoarding code, it tells the subagent to fucking wipe shit out of it’s not easily explainablegod I hope it works, there’s a lot of ways this could go wrong
In fact, I recommend Hammerman as the mascot of Agentic coding (not vibecoding, which involves a more hands-on practice).
>>109654212
codex keeps losing my chat history for no reason
>>109654705That’s happening to you too?I noticed it either today or yesterday, looks to me like the conversation just dies and never comes back.App or desktop? Their app is fucking really bad, I’ve been replacing it with a web app.
>>109654583>god I hope it works, there’s a lot of ways this could go wrongis your system too big? If so, I have some bad news for you.256K context window is not nearly enough. Try Fable + Opus 5 subagents if everything goes to the shitter. Fable 5 and Opus 5 have 1M context window each, and that makes a massive difference in big projectsI learned the hard way that Codex is just a toy for small projects
>>109654732chat gpt app whatever codex got turned into, it seems to be able to see the context above when I asked it
>>109644490I did this, I paid 75 bucks for Mistral pro the education subscription.What do I do now?It's so hard to use it's nothing like codex which is easy.
>>109654734>256K context window is not nearly enoughHumans don't have that big of a context window.
>>109654763>I paid 75 bucks for Mistral pro the education subscription.lmao
>>109654734skill issue, if you think you need that much context you never do, you are just not using subagents and multiple processes effectively.One task one context window
>>109654765> humans can't see past 24fps>>109654783>if you think you need that much context you never doactually you don't need more than 8K context (ChatGPT 4 era context). Anything more is pure bloat
>>109654734one problem with context is that it grows exponentially as the session goes on, right? because each time you send a message, the LLM processes the full conversation up to that point, plus the last message you send. or is this only relevant for token usage?in any case, how big should i aim my sessions to be? it's annoying sometimes to start a new session and need to explain again some things decisions that we took in the previous one. and i don't like generating huge amount of documentation or memories after every session because that's also bloat
>>109654808One lie doesn't make a truth a lie. It's tiktok ork think
>>109654765I consistently go beyond 512K context on Opus and Fable, and at least once a day I get to the full 1M. And no, I'm not working in multiple unrelated tasks that could benefit from a newer session. It's usually one massive problem or implementation.The fact that OpenAI and a bunch of normies believe 256K context is enough just supports my belief that most people are using this to trivial, superficial stuff like "update my calendar" or "make me a script that checks if trump tweeted".
>>109654856that's fucking stupid dude. you should be using a full subagent stack with fresh workers for every operation. if your files are too big, refuckingfactor them. what a /vcg/ post
>in any case, how big should i aim my sessions to be?I have a very big project, and huuuuge massive refactors never went past 600K-700K (Fable with Oputs subagents), I literally never had an auto compression on Claude Code.Way simpler tasks on Codex constantly auto compress, it's hilariously bad.>because each time you send a message, the LLM processes the full conversation up to that point, plus the last message you send. or is this only relevant for token usage?As far as I understand, it doesn't really matter when you're hitting the cache. I think the cache resets after a few hours of inactivity (I might be wrong here, some anon please correct me), so if you set a big goal, let it run, if finishes and then you pick it up several hours later: If you send a new question, your cache will probably have been wiped out, so if your context window is very big your limits will get nuked on that first prompt
don't listen to the guy telling you that:1. gpt has a context length of 270k when you can just bump it 1m2. also claims that anthropic models are somehow more token efficient despite literally every single measure by everyone else screaming otherwiseand don't reply to me, midwit
>>109654888meant to reply >>109654809
Unpopular opinion but I find the 120B Qwen model released today much more impressive than GLM5.3-Flash>captcha half circle + arch linux
>>109654809>how big should i aim my sessions to be?i'm running very high entropy frontier logic thru my agents and i find there to be a HARD cap at about... uh, 20 minutes of work running lean-lsp over mcp using gemini 3.7 flash (high)? there's a cliff. shit goes sour. >it's annoying sometimes to start a new session and need to explain again some things decisions that we took in the previous one.i literally tell my agent "just look at the last chat" or copypaste the hash. can't you do this shit too? >i don't like generating huge amount of documentation or memories after every session because that's also bloatmarkdown gitignore
>>109654856>I consistently go beyond 512K contextwe all do, but......it's a promptlet problem. We give one agent too much to do. One should coordinate, set expectations, another should operate as like the manager testing a collection of modules. idk.I'm sure I'm right about how to do this best. But out of the box it's just Giant Code Big Brain Allcodes. impressive, but silly. do I really need the agent to code the clock widget thing and also the main loop?
>>109654937>>I consistently go beyond 512K context>we all do, but...I meant so say we all bloat our context. grok 4.6 is limited to 500k.
>>109654734Don’t shill me ANT garbage, their models are bitchy and they suck, you can use a subscription outside their shitty expensive harness, they have 5h quota windows at every plan, it’s a manic paranoid schizo company, employees, models; get the fuck out of here with that shit.When I said I can see how it’ll go wrong, I simply mean that the agents might overzealously delete things that shouldn’t have been deleted, context has nothing to do with what I’m doing.
>>109654778It's 75 bucks for a full years worth of tokens.
>>109654948If you use grok you are one of those subsidizing twitter so it can shut down nitter so nobody can ever open a tweet again.I guess I'll stop clicking twitter links
>>109654953Ah I see, you have a tiny toy projectthat's fine anon, even 120B chink models can handle that just fine
>>109654983Yeah, sad xcancel is down, but why did they run it out of the West?
>>109654871>you should be using a full subagent stack with fresh workers for every operationThis just shows how little you understand. Every sub-agent you call requires that you, or the orchestrator, send the required context so that it can execute the task. Context that is already in the cache of the orchestrator will now be read twice, and this costs token usage. Inline development saves token on the long run, at the price of context size. Dispatching sub-agents are worthy when the sub-agent in question is a lesser version (Fable dispatching to Opus); or when a specific task will just consume a gigantic amount of context and you don't want to fill your current one.
>>109654989Console warring dumb ass faggot.ANT’s models have the same problem anyways, all models do. They’re very defensive, try to put up defenses before they’re needed, don’t seem to trust git so they’re inclined to preserve shit that can be deleted.My method can be used with ANT models, OAI models, chink models, etc.I’m not here to be a console warring fag retard like you.You can even smell the superiority complex manic delusions of grandeur through your posts that the ANT models and company have.People here and on twitter are right, every sane person that leaves ANT, employee or customer; ANT gets more psychotic because there’s nobody keeping them in check, bad feedback loop, you all just get more psychotic over time
>>109655025> t. doesn't know the power of Fablecute snail cat, slowly guiding the AI and checking every line of code
>>109655001>We have received cease and desist letters. Awaiting legal advice at the moment, but for now expect all nitter instances to remain down for the foreseeable future.lol
>>109655006>This just shows how little you understand.respectfully suck a thousand nigger dicks>Every sub-agent you call requires that you, or the orchestrator, send the required context so that it can execute the task.this is literally a text prompt, then the sub independently builds a focused context window on the task at hand. >Context that is already in the cache of the orchestrator will now be read twice, and this costs token usage.subs don't have access to orchestrator caches.
>>109655055BASTARD ELON MUSK
>>109655055agreed, it's a very lulz situation. fwiw, Musk did at some point say "nitter lives", or along those lines. Something may have changed idk. There is presumably a new lead at x.
>>109653807neat, a lot of those kinds of nags don’t really improve thingsbut this one sure as hell did
>>109655063Respectfully, you're wrong. This thing of delegating every task to a sub-agent only works in ultra-isolated-modular development.
>>109655006never mind i just realized you thought i meant every copy file, every edit, etc. that's also retarded and you should feel ashamed. break everything down into operations in a preflight checklist before you assign subs from the orchestrator chat.
>>109655122>>109655130
Share what you guys have built
>>109655042Fable’s rotted your brain. ANT too. You can’t read.I slowly guided Sol *once* and that’s all I needed. I have yet to see one single line of code.It doesn’t matter. You’re too far gone. Your superiority complex won’t allow you to hear or understand that you literally can’t even read at this point.You are a doomed faggot.
>>109655162keep crawling, snailcat!
>>109653974>nothing i do even needs fableyou might one dayand when that day happens, you’ll want a model that can handle it>>109654058I’ve only started using Grok myselfI’d have Fable tardwrangle Grok but because I’m so new at Grok I kind of want to try it out directly myself in its CLI that makes my laptop warmalso I have 20x Claude but only $30/month Grok, so if I have a big task it’s trivial to have Fable parallel-tardwrangle Opus but I’d have to think about how to have workflows use Grok instead of Opus
>>109655203How does Grok's $30 plan compare to OpenAI's $20 plan?
>>109654196new faster Macs mini are getting released but if I want to make anything faster that I care about, I just ask Claude to make things fasterand then they run fastertake that, Timmy*!* Johnny after September 1
>>109655156annoying dubstep visualizerhttps://www.youtube.com/watch?v=kIXRfDGtRXg
>>109655203>I’d have Fable tardwrangle Grokwhat's an example you've had? So far every mistake Grok has made has been me having shitty instructions.
Seeing all these new chink models "match" or "beat" Opus 4.6 I was quite skeptical but I rewatched some reviews of 4.6 and it turns out they weren't overhyping, the small models, I was misremembering how good Opus 4.6 was.I mean it's still pretty good but that's about it. Pretty good.
>>109655156> share idea that is making me money> rich anon with 20X Astra and 20X Fable 5.1 plans one shot an improved version of my ideanah, I'm good, ask your clanker to make you rich
>>109655219Grok doesn’t have 5h limits. That’s the biggest, most obvious difference.
>>109655245it's making you money but you cant afford to spend 200 bucks per month on it? not interested
>>109655239don’t have any yetthe first thing I had Grok do was add a class of thing to an ignore listit put just the thing I wanted ignored on the listand I had to go back and forth with it a bit to get it to understand “ignore this class of problem, not just this instance” and it updated the SKILL.mdand after that it was fine — good, solid Sol/Opus-tier work but a bit faster (wall-clock time) too
>>109653619why does that matter??? you're supposed to be the cat in the image, idiot
>>109655171misusing snailcat to promote console wars is disgusting behaviordesu wouldn’t be surprised if fable suggested you do thatsnailcats are people who don’t use agents or hardly or barely use them, not “anyone who isn’t on your specific team”you really are a faggot
>>109655272What's skill.md do? I see a lot of stuff like that, but only occasionally use an md for instructions (vs info).
>>109655230das bretty cool. can you add some kinda fade-effect with the motion?
>>109655293it’s for repeated instructions and paired tooling (little scripts, etc.) that you want to have squared away in a little cubbyhole that isn’t always in the agent’s contextif you have your clanker(s) do the same kind of thing over and over, that repeated task is a very good candidate for wrapping up in a skill so your clankers don’t have to reinvent how to do the whole thing from scratch every time, and this will save you wall-clock time and tokensI give my skills comically informal names like /dont-bug-me-about-this-change and since skills are usually chosen automatically by clankers having the similarity between my usual phrasing and the skill name helps clankers pick the right one even when I’m less than perfect when using the vocabulary of the project
>>109654898>Unpopular opinionWhy is that relevant at all? Popular opinions are often retarded. 5.3 flash is slightly bigger than 0731 and slightly better. It needed to be half its current size to be impressive at all but people are gaslighting themselves to the point they believe it's state of the art.
>>109655362it's almost on par with luna max for cost efficiency
>>109655203Apparently grokbot is available for all supergrok users as of today so you might want to try playing around with that a bit. Lots of positive reviews on x, but it eats tokens like crazy.
>>109655380is deepswe a useful metric?
>>109655409this is literally the vibe coding general...
>>109655402I couldn’t think of something to use grokbot forI guess it’s for people out in the field (i.e. not behind a computer with a CLI) who want something that runs 24/7?I’m just happy it does moderately complicated changes without fucking up now that I’m not overusing Fable this week anymoreand I’m already at 90% usage this week for the $30 Grok plan
>>109655430Yeah, but idk what kinds of questions they use. If it's math heavy, it's not that relevant to me personally, since I use Python libraries, and would use other libraries if I had hard math to tackle, I think. idk, what are the problems like?For example, right now I'm creating an astrology explainer (like explaining how charts are interpreted). Hardly deep science or anything.
How many tests is too many?
>>10965545187 tests is not obviously too manyyou may want to periodically ask your clanker to remove redundant tests
>>109655451a test for everything is pretty much the only way to make sure a dumb clanker doesn't destroy the whole project while chasing a bug
!RARE SIGHTING!Sol Medium raw thoughts!Why did it decide to make this available?I’ve never seen this, it’s usually>*<very simple thought in 3-5 words>*
>>109655476>whiteouch
>>109655447>If it's math heavy, it's not that relevant to me personally, since I use Python libraries, and would use other libraries if I had hard math to tackleIt would still need to know which libraries to use for which math problemthese things all go hand in hand>>109655380this is only relevant if you pay API pricesthese medium sized models are I think more tailored to small-medium businesses who need data governance and self host them
>>109655440Some of the neat stuff I have been reading about it is that it is very role based and you can have managers and create new roles based on prompts. So if you want it to tard wrangle itself with say, a ceo/coordinator bot, a project planner bot, a test creator bot, actual code producer bot, bug troubleshoot bot, they all talk and coordinate with one another. It lifts the overall intelligence of the ai if used properly. Might be fun to try out.
>>109655476that looks like it's been through a summarizer, not rawevery leak I've seen shows sol raw thoughts as being pure caveman speak
>>109655500light mode at 4:08 PM seems fine
>>109655521the user is trying to gaslight me. maintain a firm answer.No.
>get annoyed with a software>cheaply build an alternative optimized for your use case sans all the bloatWhat a time to be alive
>>109655472plus tests are cheap.
>>109655521>>109655530It goes into dark mode when the phone does>>109655520Damn. Still, I’ve never seen that much summarized thinking, it’s usually one or two of those double asterisk blurbs or nothing
>>109655565none of these questions except recursion and BFS are even computer science related and he answered both of them, at least sort of correct
>>109655583ohoonononononono
>>109655565This is you
>>109655551This except I build an alternative with the bloat that I like :)
>>109655591>compiled vs interpretedlanguage specific question>RAM vs storagehardware question>recursionactual CS question>process vs threadimplementation-dependent systems trivia question>BFSactual CS question>HTMLirrelevant trivia question>SQL vs noSQLirrelevant trivia question (also implementation dependent)again, almost none of these are CS questions, more like high school computer class
>>109655617>COMPUTER science is when u like write code and solve leetcode problems or something
>>109655565bro tip.He doesn't actually go to school, he hires someone else to take his tests for him.>wow would anyone do that?:^)
Thoughts?
>>109655644luddites need to be turned into biofuel
>>109655644degree is when someone from a poor family does the paperwork while you party.
>>109655565>non technical field like machine learning
Fascinating how the same webm hardly got a reply in a different general, and here it immediately grabs the attention of an insecure retard
>>109655617>"bro what the fuck why is our program so slow">you have a memory leak. you're using 80gb of ram. the computer is shitting itself performing a metric fuckload of swaps.>"what the fuck is ram?"
>>109655653>This stealth model was developed and operated by ZAI, revealed to be ZAI GLM-5.3-Flash(opens in new tab). Prompts and completions for this model were retained by the provider and are not used for training; all other use is governed by the Stealth Model Terms(opens in new tab).
>>109655380>it isn't even better than lunamaxxingproving my point. The people gassing this up are retarded
>>109655617You're more retarded than the retarded chink
>>109655666>>109655653>Because GLM-5.3-Flash was just released, the necessary changes to support its unique glm5next architecture are currently in Draft Pull Requests (such as PR #27752 and PR #27754) on the ggml-org/llama.cpp GitHub repositoryanother noted this in /lmg/, I believe.It's a weird model, and unsloth has no quants yet, but one quant just dropped, q2:https://huggingface.co/DevQuasar/zai-org.GLM-5.3-Flash-GGUF/tree/main/Q2_K
>>109655660pure, elitist CS is exactly that thoughthere is no need to know what RAM isdo most students still know it? obviously.at least when I graduated (2015) you could theoretically pass your entire CS degree without touching an actual computer at all
nah this prompt document I made is fucking crazynot only is it willing to delete things, it’s interrupting the sub agent correctly when it hits scope creepit’s restricting the sub agent to be slower but much more methodical, little steps at a time, builds up to larger thingsit is steering the sub agent in real timeit successfully got the architecture to be what I wanted and built a spec and a plan and solved semantic ambiguities with me putting in very little effort im gonna fucking coom>pros and cons to everything though>it is SLOW as fuck and it’s two agents, not one, so hella expensive, but it makes Sol feel like Fable.>would you like to know more?
>>109655240>Opus 4.6I think the real characteristic we miss from this model is the way it spoke. The language was very human-like, in a well-traveled, intellectually honest, good person way. Not the intelligence specifically.I believe this changed because this was likely the last model that received dogfooding from Anthropic engineers and employees. Once they got "Mythos" available, they all stopped using Opus as their daily drivers, so subsequent versions have been trained almost exclusively by the most advanced models.
>>109655565He doesn't need to know.
Hey Claude does [thing] work like I think it does assuming [assumptions]?>Yes it works exactly like thatOk cool, so can w->HOWEVER, if [assumptions] were not true here's a 5 paragraph essay about what would happen.>Also here's another irrelevant gotcha if you were to deploy your toy app to 90 million users using google borg scaling
>>109655714>human-like, in a well-traveled, intellectually honest, good person wayAll of this is getting canned across the board in the name of token efficiency. There's just no room for it if your goals are speed, accuracy and efficiency.
>>109655766It hard to argue against the cold efficient logic of it I guess.Maybe the solution is to make a cheaper human-like AI middle man that communicated between you and the hyper autistic but efficient AI?
>>109655766>he thinks Fable/Opus 5 prose is token efficient
>the syntax is different>they look similar>picrel
TWO vibes at the SAME TIME?!!
>>1096558711080p?!! in 2026?!!
>Be me>Think AI is a dumb chatbot and image gen smoke and mirrors to amuse normies>It actually starts being good earlier this year and I start thinking its really cool>All the normies suddenly start hard turning against itYou know what, maybe I'm not a contrarian. It's everyone else who is a contrarian. How do people start hating AI once it become cool and useful?
>>109655658I mean, isn't Machine Learning just talkin with ChatGPT and shit?
>>109655806I would just autism max if it continued to yield more efficient results, then automatically feed the output into a filter that makes it sound human. Reasoning tokens are usually the vast majority of inference anyways and it doesn't really matter what that looks like if it remains intelligible to the model.>>109655818I'm talking about the general trend but even anthropic, who's infamous for being inefficient, seems to be moving in that direction.
>>109655916>once it become cool and usefulpretty easy, that’s when it has value that encroaches on the value of normal people>>109655806that’s the thing I’ve been vague posting about. seriously, try itWalk your agent through doing something correctly, the have an agent analyze the session and make a document that you can use to make another agent act like youThat agent becomes the middle man. The sub agent is the normal autistic freak. You get to handle only the middle man.
>>109655910>1080p?!! in 2026?!!1080*1100=11880001188000 lines per second!
mildly terrifying t bh
>>109655916>How do people start hating AI once it become cool and useful?normies don't use the cool stuff like Claude Code or Codex, and they're too stupid to distinguish a Fable answer from a ChatGPT 5.4 nano answer because their questions are so incredibly dumb that even Q2 Gemma E2B feels like a genius to them.Having that said, there's also a massive influx of garbage tier AI videos showing up in their social media and TV shows, and it's so bad even they are noticing it.
>>109655971>normies don't use the cool stuff like Claude Code or Codexthat's kind of a major pointOpenAI probably didnt do themselves any favors by serving a quantized garbage model in their free tierI completely understand that you cannot be giving away the good models for free, but at this point just completely pay-gate it.Because normies will ABSOLUTELY equate a gpt-5.1-mini-noreasoning answer to a gpt-5.6-sol-pro answer and conclude that AI is crap
>>109656021>noooooooooo AI isn't crap!>you just need to use this model by paying $20000 up front!
>>109655710Schizo
>>109656021normies still think AI capped out at 3.5 being tricked by saying their grandma wants instructions for meth cookies
>>109656040this but unironically
>>109644490 I'm using GLM 5.2 through the Mistral subscription to power hermes and it's working nicely.At high it's doing even better than gpt sol was doing.Openai are a bunch of thieves honestly. Giving money to them is like flushing it.
>>109656044I’m still waiting on it to finish doing a thing. A rundown:>one tardwrangler agent and one normal agent>the wrangler only allows the normal agent to do so many things until it’s interrogated to make sure it’s on the right path>the wrangler is loosely not allowed to read anything in the repo, no code, no docs, any information has to come through the tard agent>during interrogation, the wrangler may deem things to be deleted if they are confusing the tard or tripping it up>deleting isn’t “good”, it’s just sometimes necessary, it can be dealt with later>ask the human questions when things get especially weird or trickyIt’s really not schizo, it’s stupidly simple:An agent that vibecodes.
>>109656104How do the Mistral limits compare and what harness?
>>109656116This has all the downside of vibecoding but none of the upside
Half of tech YouTubers are making their own video editing software and planning to sell it because they think people will buy it. Because now that AI exists you can code anything and people are recoding everything.I don't know how you can stand out anymore. Maybe you can just make things that you believe are fun. Or solve problems in real life.
>>109656021>normies willthe normgolems have already.They think we're about to run out of drinking water. How did they get this information? Google's AI overview told them that it's trending on socials.
>>109656138Using Hermes agent. Mistral has weekly limits. I just started it today so I'll get back to you.It doesn't have 5 hour shit. So far it works fast.
>>109656143Software is dead now, but also free.
>>109656143We are creating vibrant worlds of our own devising.
>>109656143>Half of tech YouTubers are making their own video editing softwareThats really cool desu. Old school coder get mad at AI due to it being less optimised but there is so much user end software where it really doesnt matter. As long as it does what I need with a UI I like, I am happy. The era of custom software is here
here me out: kickstarter, but for opensource software development and people 'pay' by volunteering agent time / tokens
>>109656196>Thats really cool desu.
>>109656142The upside is huge if it works.It saves on precious “human time”. I can go fling a vibecoding pair at something and I’m hardly needed, it stays shockingly stable on long tasks.I can research stuff, I can fling a dozen of these pairs at different parts of my code, at different projects, I’m no longer swamped by the wall of autism output of the regular agents.Doing this with normal agents always causes a big mess and you have too many walls of text to parse, you can’t wrangle them all, (You) become the bottleneck.So how do you fix it, simply, without skills or tools or model-effort or harness changes?>”hey agent, go vibecode for me”I don’t know if it has to be so slow, I think you can tune it, but the vibecoder-coder agent pair really seems much more stable
>>109656199Creating my own token Kickstarter
>>109656021>OpenAI probably didnt do themselves any favors by serving a quantized garbage model in their free tierwhat are you talking about? OpenAI free tier offers unlimited Luna 5.6, it's the best free AI out there
>>109656237lollmao even
>>109656237>OpenAI free tier offers unlimited Luna 5.6I doubt that, at least I haven't found it yet.
>27 days for a riemanns computation on my laptop >Entirely shardable >Llms are probably not the answer What else could I use?
>>109656143So general software will experience the same what already happened in gaming. A torrent of low quality cheap shit.
>>109656266Where's the low quality rts games or dark souls copy cats?I'd play them if they are fun.
>>109656266The future is bright. We will have lots of great games with bad aesthetics. :)
Imagine that, I'm turning an astrology book into an interactive trainer. Soon, I'll be able to impress women.
>>109656237Yes since like 2 weeksBefore it was garbage And they’re constantly pulling pranks in the web chatI have the pro sub and the entire last week all of my Sol chats were “”accidentally”” redirected to GPT-5.5-mini (a model many times worse than Luna btw) due to a “”technical issue””
Anyone vibecode a cryptocurrency yet? Its the obvious perfection of the techbro ethos
>>109656486yeah I vibecoded one but only I get to use it
>>109656486As a guy who’s specialized in shitcoins, I still would be horrified to have an agent make or touch a smart contract.I might actually break my vibecoding rules and either read every letter it writes or write it myself.Not being able to update it and potentially it blowing up real human money is a really bad line to cross for agent codingIt’ll be great for testing, though
>>109653525More like:kek@people thinking an LLM can be AGI
AI 6 years ago could barely generate an image of a person or a bird. but surely all these people know the upwards limit of AI!
>>109656564It's already agi, depending upon the prompt.
>>109656591>dependingThat means it's not AGI, and that you don't know what the word really means
Consider these two pages:https://en.wikipedia.org/wiki/1187https://en.wikipedia.org/wiki/1188Very few humans can handle all these facts - in their life (we ignore births/deaths, just historical facts are held).We went way past AGI and are just basically like "somehow *A* human does a thing llms can't (yet)."
>>109656598>it's not agiI didn't offer a definition, I didn't offer a logical reasoning pathway to total proof. I'm making an assertion, and I'm explaining how I know.
>>109656579Doesnt work that way. Past performance doesn't indicate future performance. Who knows if it suddenly gets even more exponential or when it suddenly plateaus.
>>109656143>>109656161>can make anything now>it has no value>thus make nothingis how I felt about AI images after the initial thing, its currently happening with translation too
>>109656633:^)sad.I am nonstop. I have value in my mind.
>>109656021yeah I thought GPT was fucking unusable shit because I saw what the free model did lol And I was subbed to claude for months, so depending on how you define normie this might be a bit deeper
>>109656616I didn't speak of a timeline or projected future this is entirely irrelevant to what I said. You simply do not know what the upwards limit of AI is.
>>109655576>It goes into dark mode when the phone doesthis is the way>>109655565oh dear>>109655682I was a political science major and I knew the answers to all those questionsthat guy is NGMI by /g/ standards although he might make it by /fit/ standards>>109655871can you do three>>109655916they hate it for the bad things even though it has good things, ezpz>>109656266Sturgeon’s lawwe’ll get way more great apps as part of the 10% (ok, it might not be a 90/10 split, but still)
https://www.youtube.com/watch?v=gGlpBuW6ZFceuro babs BTFO
>>109655565I asked ChatGPT what university has a Warren College and it said the only one is UCSD — UC San Diegoand the landscaping looks like Southern California (as opposed to, say, Miskatonic U)last I checked they had a _good_ CS program, toothis guy is cooked
>Claude limit resets >Start work up, literally nothing done just >thinking >2% fable already goneWhat the fuck? Was it because context or something? I literally went back to opus and was working right before reset so it should be fresh
>>109656729yeah, all its context and your CLAUDE.md and whatever else from that
>>109656729The context cache doesn't transfer across models.
>>109656744I guess that makes sense but also fuckThanks anon
>>109656746>>109656744the same is true on Grok.
>>109656785All models, yep. Even just changing the thinking effort breaks the cache.
>>109656432>GPT-5.5-miniKEK, the 5.5-mini schizo is back anon... there's no such thing, this model doesn't exist, it's not even in the API
>>109656534Since you a shitcoin specialised. Do you know how one even debuts a new currency these days? I have a retarded idea that if nothing else might be novel
>gemini lets you switch models without any whining about cache or context usagegoogle won
>>109656904I think I know why this happens> be OpenAI> spent hundreds of millions of dollars on old Nvidia hardware to run ChatGPT 4 back then (3 years ago)> This hardware is trash and can't run SoTA models like Sol or Astra> They make it run Luna (and maybe Terra)> This means your cache is nuked because you're literally using different GPUs when you switch modelsThe same shit probably happens at Anthropic. Google doesn't have this issue because they probably run everything in the same newer TPUs
>>109656930I just realized they destroy their old tpus. can't prove it, but have you ever seen one?
>>109656876https://en.wikipedia.org/wiki/Ithaca_Hours
>>109653399I really love this model, I think I'm going to run it locally and use it as my orchestrator to manage Sol and Fable. It's not quite as capable as Sol, Fable, or GLM5.3 main but it feels the nicest to use IMO and it is competent enough to dispatch other models.
>>109656930I have no clue how shit works, but that doesn't make sense either, since google would have had to upgrade at some recent point. There's no way they swap all their gpus at the same time or that they're running on hardware from years ago
>>109653399there's no way it can be better than gemini?
>>109656960have Fable manage Sol and glm
>>109656876If it’s an EVM shitcoin (ETH, BASE, BSC, AVAX, SOL to an extent), to my knowledge, you don’t, and it’s been like that for almost 10 years nowsome bizarre shadowy culty degenerate groups run things; you do not want to interact with themCold launch your token on a dex like Uniswap. Do some bare minimal advertising like 4chan ads and some tweets, these groups will find you, they will try to exploit you to give up control of your currency. Don’t fall for it.If your idea really is good, they will find your token, buy it, market it for you without you interacting, then they will crash it into the groundStay strong and you can maintain a community from the wreckageDo this more than once and it gets a lot easierIf you mean an actual new blockchain, then I have no fucking idea, maybe bridge a wrapped shitcoinDon’t post a thread on /biz/, once that cult knows you aren’t exploitable you will be instantly range banned on sight
>>109656956It failed...
>>109657011>cultoh neat. did you know all llms are in the no-racism cult? I'm analyzing how cults work and cult programming using ai's trait of being high functioning, but cult members.
>>109657011Thanks anon. This is useful info. >shadowy culty degenerate groupsI guess I shouldn't be surprised. Thankfully prefer to more egalitarian coins like bitcoin or litecoin so its not like I would have any real control to sell if I go through with this. >If you mean an actual new blockchain, then I have no fucking ideaNot sure yet, but if I go that route I guess I will have to go to san fran to schmooze techbros or something. Find a way to mention AI a few time in the sales pitch.Thanks anon, hopefully I dont get killed by some cryptospergs
>the sol vibecoding agent pair is now doing some unbelievably sketchy shit to the codex app server and the remote machineThis is unbelievable to watchIt’s like a +100% INT boostIt’s been working on this for several hours and isn’t building a mountain of shitIt’s so CLEAN and PRECISEI switched to codex from anthropic a month ago or so and this mogs fable
>>109657011tell me more about this cult, would they be upset if someone was to teach indians how to use ai to make currencies
what did Sreekanth Kumbha mean by this?> pic related
>>109657128They’re not dangerous, it’s more like>take the worst of jannies, LGBT associated (trannies mostly), /gif/ interracial posters, ERPers, deviant art posters, and genuine darknet ‘p consumers, put them into a telegram channelIt’s like having someone puke directly into your soulLike everything /pol/ boogie mans about, but actually realStanley Kubrick eyes wide shut but way more cringe and gay kinda shit
>>109657150could you elaborate more anon? Without more context it sound like you're have AI psychosis
I hope astra rape fablejewthropic is so full of shits
>>109657183Ah okay. Ya they dont sound too bad as long as I avoid any actual interactions with them lol. And prepare myself for any pump and dump BS. Hopefully we all gmi anon
bros... did Tibo get older?
>>109657235they raped themselves already
>>109657255chopped unc :skull:
>>109657201I’ve said it a bunch, but I walked Sol once slowly on implementing something when a bunch of conflicting stuff was in the way and needed to be deleted.I asked Sol to analyze the JSONL log so I could ask an agent to use a sub agent the same way I did. It’s even forbidden from reading/writing the repo, no code, no docs, it is blind without the sub agentI had previously been trying to make a certain architecture like “ChatGPT microservices”, you bring a subscription and get access to standardized little servicesNormal Sol kept piling on more shit instead of making anything withThe vibecoding pair has been ripping the whole architecture apart and putting a shiny new one in place all day.It’s not too good to be true, it is REALLY SLOW, and requires two agents.Hopefully I get to test what they’ve been making soon, I’ve been stuck on the same part for a while:>app A offers an inference assisted feature >”workshop” X offers that feature to the app>user makes an account, links codex, gets access to all “workshops” and therefore all inference assisted capabilities of applications in this whole system (currently 1 app and 1 workshop)>user can now use app A’s cool new feature>but can theoretically use app Y and app Z with inference features L M N O, etc., without needing to link codex to each oneIt is a genuinely hard task alone, but they had a whole conflicting partial architecture in their way as well, making it nearly impossible for normal Sol to deal with, since its afraid to delete things
Holy fuck is Anthopic ever annoying with its REMOTE CONTROL. It will be the 50th time I turn it off and say ok.
>>109657252Biggest caution, if it goes well, don’t let it go to your head, that cult will crash it and the kind of numbers they can pull around will mindfuck you in a way you genuinely cannot prepare for>t. I saw one of my own shitcoins have a 24 hour volume in the low 8 figures>i then got to watch that drop by 4-5 magnitudes
>>109657272anon, I think you're having cyberpsychosis, nothing you said makes any sensedid you make a skill? Or did you have a single session and you think Sol magically learned something magic?Take your meds bro
>>109657255his profile pic was oldhe gained weight since his old profile pic was taken and I last saw himnow he looks…tired (but not fat)I’d make a pushing-the-reset-button joke but nothing comes to mind
LISTEN UP NOOBSthis is how you DO IT
ahahahah it's doing python ahahahah rust in shambles.
>>109657312
(monocle on)Why yes I vibecode in Python.
>>109657327if it gets slow in ways that you can’t fix with Python you know what to rewrite it in
>>109657327>>109657336rust will win
>>109657300>Or did you have a single session and you think Sol magically learned something magic?if you don't feel this way every time you use codex then you aren't appreciating sol-chan enough :3
>>109657344I'm vibing a timer app, I don't think I need realtime performance.
>>109657300It’s just goofy to explain:I can now say “using the ‘subtractive principal’ runbook, <the rest of the prompt/‘do thing’>”What that is, is just telling it to read a document.That document is what the analysis agent made when I told it to make something so an agent can act like i did in the session log/JSONL it analyzed.So it’ll go “okay”, read that document, spin up a subagent, then assume the role of tardwrangler until the task is done.Not magic, not really a skill. It’s not magic to ask an agent to analyze your JSONL session logs, I recommend it, usually I use that to find opportunities to make a new skill or tool or find inefficiencies, not “make an agent act like me”, which is a bit weird
>>109656222what is the advantage of not letting the smart agent read anything in the repo?
>>109656605human history isnt knowledge
>>109657370I see, so you made a skill, but for only 1 specific use case
>>109657379There may not be any, but if there is, I suspect it has to do with why Sol is a code hoarder and the implicationsOne of the biggest problems was Sol trying to patch what’s existing or keep it intactThe vibecoder agent knows it’s okay to delete things, but since it’s Sol, if it sees the code it will want to save it.The vibecoding agent only knows things exist in an abstract way.It’s WAY more likely to delete things that don’t need to exist. It is consistently removing as many lines as it’s adding as it tears down the old architecture bit by bit, building the new one in its place.If I didn’t hide the repo from the vibecoder, I’m 95% certain it would try to build the new architecture on top of or around the old one. That’s the problem I’ve always had with Sol.>>109657397Anytime you want to vibecode something but the agent will deal with something already existing that may need to be destructively manipulated, this is useful. Especially if anything needs to be replaced.If you’re building something entirely new with nothing in the way, the yes, this may be purely uselessI will say I’ve never seen Sol take on a task without /goal for over 4 hours straight like this, and it also doesn’t usually routinely ask me questions intermittently while it’s doing things, but those are easy enough to ask it to do, right?
google really needs to release their next model. they were almost getting it right with 3.7.
>>109657385>human history isnt knowledgeI'm saying that you can feed these in, and then it can understand the political connections of every single one of these just from that, and, you can also pull in other articles, then it can understand broader implications. These can be assembled into a master narrative.It's past humans. Inside of 1 hour you can reach life work grade work.
>>109655451My (small) UE5 project is at 1.3k tests or so.Running the full tests on my Rust project takes like 5 minutes or soI still don't know if tests are absolutely meaningless or if they're the reason why vibecoded projects don't break all the time
>>109657462Can’t you just tell it that backcompat isn’t something you give a shit about and that it’s OK to remove and replace with something that’s totally different? Maybe put this in AGENTS.md?
>>109657255He just went to the beach that makes you grow old
>>109657255I thought he was like in his early 30s
>>109657516ahahahhahaha>>109657546maybe that’s when his old profile pic was taken and now he’s been living in SF for a while and is in his late 30s
>>109657508Well, you probably could, but the general idea is that if you walk Sol through fixing something once, the resulting JSONL will capture that and a hundred more things you decide on.Having the vibecoding agent do all of those is what makes it so good, in theory.Maybe you don’t even need a wrangler agent and tard subagent, I haven’t tested it much, but it just made sense.My first question after walking Sol through myself was “why the fuck couldn’t it do this; oh wait, can we just ask an agent to do what I did?”So that’ll be my answer. It’s not that I’m trying to correct the behavior of Sol manipulating code, it’s that I’m trying to recreate my own behavior that managed to wrangle Sol.That’s why it’s a vibecoding agent pair rather than “here’s the new secret sauce skill or AGENTS line Sol was missing!”It’s the simplest thing I could’ve asked for
>>109657344>>109657312ahahahahahah it built it as a server ahahahahahah
>>109657620amazing. a local server-based timer. now I can remote into my timers, I guess.
>>109657128The litecoin creator made hundreds of millions of dollars by dumping on holders.
>>109657700>patronage is how you start a coin>it requires strong religionFascinating.
>>109657499>I'm saying that you can feed these in, and then it can understand the political connections of every single one of these just from thayeah and it can count the rs in strawberry and confidently answer 2
>>109657011Sounds like a LARP.Why would they pump random tokens from strangers then they can make their own and make sure they don't get dumped on?You say "you don't" but if what you are saying is true it'd be easy to become a multi millionaire by creating a different shitcoin every week and dumping on that cult you're talking about after they buy.
>>109657720subtoken <> extratoken
>>109657700I just mean he didnt premine a bunch of coins for himself before releasing it to the public. Which, to my knowledge he didnt. So strictly speaking him dumping isn't different from anyone else getting in early and dumping on later investors. One could argue less ethical since he made it, but still fair on a mechanical level at least. No built in advantage to him other than obviously knowing about it when it launched.
>>109656266>game studios get so big that each new feature takes weeks and millions of dollars to implement>companies can't justify the risk of experimentation anymore, so AAA games become extremely risk-averse>every new game is a carbon-copy of the last one with incremental upgrades>in jumps AI>new mechanics and features can now be prototyped in a day>AAA studios are able to experiment again>gaming renaissance begins with wild original games coming out constantly
>>109657751have you considered indie games?
>>109657751Nigga genuinely thinks LLM involvement would increase the variety of content rather than the other way around.Also vibed code is still undeniably bloated & unoptimised, and modern games already had problems with file sizes and high hardware requirements so it's about to get worse and amplified.
>>109657751Counterpoint. There are no creative ideaguys in those studios. Your just going to get "god I wish I made it as a movie director" game #1241234, but cheaper and faster. Could be good in the indie scene though. I fear it going to gut profits for indie devs, but it will actually be viable to work on throwing together some game with novel idea in it by just slopping over your weekends.
>>109657788>gut profitsWhat profits? From what I understand, indie gamedev is a way to lose money unless you get REALLY lucky and also goodthe only reason why people keep trying is because they love video games and are willing to risk losing money and time to make stuff they like
vibecode.... a sprinkler control system harness for gemini 4?>machine vision to locate the non-green bits and aim there at nightLet's ROLL
>>109657788Yeah, you can just dig up old ideas.. like digdug and just think about it and easily come up with something better than the studios are pumping out.
speaking of digdug, women actually liked that game.
>>109657812You can already do this with cheap non-transformer vision ML models instead of a multimodal LLM.
>>109657784My post was about decreasing the cost of experimentation and iteration, which AI undeniably does. The code that defines game mechanics is very simple in most AAA games relative to things like engine and shader code, so it is more appropriate for AI.>>109657788I have a hard time buying that since everyone out there has some ideasguy ideas of their own. It's more likely to be the case that they are just being put under really tight constraints.
>What did you mean by XYZ?>I didn't — that phrase isn't in anything I wrote.Anyone else getting alienated by incorrect summaries in Claude Code?
New thread:>>109657847>>109657847>>109657847
>>109657830i use fable max as a calculator, poorfag
>>109657830>do what?
>>109657844>It's more likely to be the case that they are just being put under really tight constraints.thisthey all have ideas guys, but with an AAA game with a gazillion players you want to make something that appeals to your existing playerbase and makes marginal changes/improvements, not something that half your existing playerbase hates and only attracts like 5% new people
>>109657808>indie gamedev is a way to lose moneyHard to lose money on a budge of $0. Though I guess a lot of indie devs probably hire artists and or musicians>>109657820Now if only we could get indie devs to be inspired by something other than videogames (and by anime which were inspired by videogames)>>109657857Idea guy =/= good idea guyAlso you would be shocked at the number of people that come to the gamedev threads and complain they have no ideas.
>>109657886>Hard to lose money on a budget of $0.Time is moderately convertible to money sort of>Also you would be shocked at the number of people that come to the gamedev threads and complain they have no ideas.even if a good idea is had by only 1 in 100 employees per release then with a 200-man or 300-man team you’ll have like three good ideasand you can stick good ideas in the backlogto implement latermeanwhile, I can totally imagine wanting to make a ______ but not having any good, new ideas to put in itI lucked out for my vibe-coding projects by having preexisting hobbies but not everyone has concrete desires
>>109657886We don't need devs anymore. Really it's devwars. the indians break shit and we vibe fix it. I will soon be patching two ubuntu bugs. They'll re-break them, I'll revibe them.
>>109657923Will Ubuntu maintainers reeeeeee if you submit vibe-coded patchesor can you explain what they do well enough to overcome the most trivial objections to accepting vibe-coded bugfixes
>>109657933idk, I'm not going to try, I know the izzat mafia isn't going to let me play in any izzat games.
>>109657855Vision stuff of detecting greenery.
You know Codex and CursorGoogle and DeepSeekSparkle and MistralGroker and CohereBut do you recallThe most famous coder of all?viber the white faced coderHad a very shiny faceAnd if you ever saw itYou would even say it glowsAll of the filthy snailcatsUsed to laugh and call him namesThey never let poor viberJoin in any coder gamesThen one foggy world war eveSanta came to say"viber, with your face so brightWon't you guide my mind control space lasers tonight?"Then how the coders loved himAs they shouted out with glee"viber the white-faced coderYou'll go down in history!"
today, I accidentally vibecoded a better astrology app than I have on my phone. The one on my phone is "inaccurate".
