A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.You use Git, right, anon?## What “vibe coding” is, and how to do ithttps://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://claude.com/product/claude-codehttps://developers.openai.com/codex/cli## Near-frontier models for codehttps://x.ai/cli## Not worth it for code, but maybe good for interpreting images/videohttps://antigravity.google/product/antigravity-cli----## Prompting / context / skillshttps://arps18.github.io/posts/claude-code-mastery/https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/https://github.com/mattpocock/skills — /grilling is a favoritehttps://github.com/DietrichGebert/ponytail## Other editors / terminal agents / coding agentshttps://osaurus.ai/https://pi.dev/https://opencode.ai/https://cursor.com/docshttps://docs.windsurf.com/https://docs.cline.bot/https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent## UI/Frontendhttps://www.figma.com/make/https://www.anthropic.com/news/claude-design-anthropic-labshttps://uiverse.io/https://ui-ux-pro-max-skill.nextlevelbuilder.io/https://stitch.withgoogle.com/## In-browser builders / hosted vibe toolshttps://bolt.new/https://replit.com/https://docs.github.com/en/copilot/tutorials/sparkhttps://v0.app/docs## Benchmarks / rankingshttps://www.tbench.ai/leaderboard/terminal-bench/2.0## What we’ve donehttps://vcg.gitgud.site## Previous thread>>109617271
>August 14th - Z.ai GLM-5.3 (still the biggest recent drop)Same base as 5.2 + heavy post-training. Strongest open coding claims (Terminal-Bench 3.0 etc.), agentic numbers near Fable 5 levels in places, and emergent cyber capabilities (CyberGym lead). Live on API/Coding Plan. Open weights still delayed for safety review (approaching the promised ~2-week mark).>August 14th - Qwen3.8-27B + Max open weights27B multimodal dense (Apache 2.0) running strongly on consumer hardware / single GPU. Max-level 2.4T-A95B also out (custom license). Continued strong local ecosystem uptake.>August 13th - DeepSeek V4-Pro-0813Official flagship out of preview. Agent/cyber strengths, more mixed independent reception, peak/off-peak pricing now live.>August 13th - Google Gemini 3.7 FlashCoding/agent workhorse upgrade with solid gains and lower pricing. Expanding into Search.>August 12th - SpaceXAI Grok 4.6Ties GPT-5.6 Sol on key indexes at $2/$6. Grok 4.7 still slipped further out.>August 11–10thNVIDIA Nemotron 3.5 Lightning (open, agent execution) and Meta Muse Glimmer (30B open local agentic) remain relevant for local/on-device use.No major new frontier closed-model releases this past week. Activity is mostly ecosystem tooling, independent evals catching up to the Aug 14 wave, OpenAI continuing to pace some training over cyber concerns, and smaller/niche drops.
>>109624422>Maintain Thread Qualityhttps://rentry.org/debohttps://rentry.org/animanon
>>109624422this snailcat makes me happy
>>109624422AIslop and fictionalthis is what really happened
>>109624428Rest in Power, King. You deserved better.
>>109624441Uncensored version; can't stop me, Xi.
will diffusion models ever be a thing?
>>109624422Total snailoid destruction
>>109624551Fuck you op you posted the gay webm
>>109624441>>109624477Why are you doing this to that cute slug. We already know that sam baldman is evil.
>>109624593chinks love to torture cats to death it's their hobby
>>109624422Headpats? COTTON CANDY CARNIVAL RIDES!?THIS IS NOT WHAT SNAILS ARE FOR
>>109624434that's just how we use them, not how they are definedthe LLM are trained to optimize their training scores, and we use some of those scores to create token probability distribution
>>109624422can someone make me a number match puzzle solving program? I don't have $20 right now to subscribe to claude. specifically the number match game where you match numbers in a row or add to 10, this gamehttps://play.google.com/store/apps/details?id=com.easybrain.number.puzzle.game&hl=en_US
>>109624667you can try luna maxxing on codex free plan
>>109624672i tried using codex once, I didn't understand how to use it and just uninstalled it completely cause it gave me headaches.
>>109624683But claude is the same thing just slightly different?
I don't know much about slopping, but so far, codex feels faster than claude, maybe even better. Only problem is that I used the Sol model at first and now I've used up a whole week of pro plan in 12 hours.
>>109624683are you seriously can't use a chatbot -__-
>>109624683brother you just type text into a text box and it does everything for you what is there not to get
>>109624693idk man i'm kind of retarded.
Ox = Luna xhigh
>>109624672also what free plan of codex? i literally only have a $20 option for codex.
>>109624740free plan give terra, luna, 5.5, 5.4-mini, no 5h limitjust login with a random email
>>109624783yea i am logged in, its not giving me an option for free plan.
>>109624789the only codex im seeing is for chatGPT and I can't use it without subbing
>>109624803weird, this morning I just logged in with a random account and tested it normallymaybe it's a region thing? but I'm a thirdie myself
>>109624422microsoft office products made open source: 0adobe products made open source: 0games like minecraft or terraria or cod zombies or portal created: 0latest agentic crafted os ram requirements: 32gbai is a bubble, stop lying and doing missionary work for a company my guy. maybe a 20% output boost best case scenario
>>109624816no clue, also it wants me to download something but im on arch linux and the .deb won't open at all on arch so idk.
>>109624817you’re trans lol
Opus, please, I got you the reference manual, please read it. I know you don't like PDF. But please.
https://x.com/jeremykauffman/status/2091339012443160713vibetoddlers
>>109624829what the fuck does that have to do with anything?
>>109624901Most Ai haters are trans
I had a conceptual breakthrough for the agent I'm developing for the Kaggle competition. I realize I wasted like 500 dollars on GPU time for a suboptimal architecture.
paying for 4chan ads wasnt very beneficial.$20 to get the amount of visitors/etc you see in this pic (top left area). i tuned the ad to only show to desktop users btw.
How the fuck are people doing reverse engineering work with the various frontier models? I've tried: >Obfuscating prompts>Obfuscating output files its analyzing>Agent cluster with the frontier model as the orchestrator and chink models as the agents>saying keep goingAre people really just using the chink models?
jeets are so cracked at leetcode contests that they can read the questions and solve them in 2 minutes (with comments)
>>109625024[Redacted on behalf of Anthropic, PBC]
>>109625024What model is fighting you? I've never had a model refuse to do reverse engineering work.
>>109625024>Are people really just using the chink models?yeah
>>109625048GPT-5.6 Sol. Have yet to try with Opus or Fable. Using Frida and Ghidra, among other things.
>>109625071Really? Of all your options, Sol is the least likely to refuse. Are you "reverse engineering" a biological weapon? I had Sol reverse engineer the proprietary protocol for the controls for a Chinese microscope just yesterday.
>>109625095I am reverse engineering parts of the Scott Pilgrim vs the World the gameNot sure what is triggering it, I've been pounding away the last few days without issues. Perhaps I've been deemed high risk.
Best prefill?
Told claude to write me a scraper for this website but it refused and told me it's illegal + it won't do explicit projects. WTF? https://www.elitebabes.com/model/lena-flora/
>>109625193
.......~@^..^
>>109625215
Been using claude at work now for a while, here are some observations, specifically around code quality. 1. I think it'll roughly match the quality of your original code2. The first implementation claude spits out could usually use a refactorThere is obviously an important thing to consider when you consider these above two points. Also, anecdotally so far, the people who use a ton of tokens on my team are still some of the worse. It seems like they think that if they just "one more skill bro", or "one more subagent", or whatever the fuck they are doing, it'll somehow improve the quality. The best way to generate quality output right now is to have a tight, short loop between you and claude. That said, long term, I think "code quality" as a metric will become less important, but right now, at least at a proper software company, its absolutely still an important metric.
>>109625226>2. The first implementation claude spits out could usually use a refactorSame as primates, no? How often are your org's primate PRs approved with no requested changes?>The best way to generate quality output right now is to have a tight, short loop between you and claude.Your org has years of PRs with reviews, yes? Define quality output, then have agent follow those quality rules.
>>109625247>t. the archetypal slop engineer I was talking about above
Terra and Luna xhigh not included because they score below 3.7flash on DeepSWE. Does Fable doing orchestration and 3.7 running implementation not sound better than OAI's stack?
slopgods won, you lost
>>109625353decisive paretomog
>want to do something>but what if Opus says "no">tell Opus NOT to do something>but what if Opus does it anyway teehee
>>109624652cats are not snailcats
>Reset will land around 14pm PST tomorrow.BIG notice
>>109625268Skill issue
>>109625166vibe toddlers will inherit the earth
>>109625024Fable has been fighting me some but I made sure to frame the project as software preservation since it's from the 90s, and with some annoying exceptions it's been letting me chug along, the real issue is I use my week of fable in like 3 days so now I'm just chilling
>>109622711“So in the future, I’ll game on my video card?”“Not exactly…”>>109624317yes because I like how it’s not just me and the strauh.al guy keeping the lights on here and all the /biz/ advertisers who have managed to bid up the CPM for that board to a whopping 13 cents>>109624964if it makes you feel better/worse I paid $20 to get 400K impressions and 81 clicks on my mostly-black desktop ad banner
Just canceled my 20/month Claude sub and 16/mo z.ai sub and I'm planning to bump up codex to the $100/mo plan and just use openrouter for chinese models if I need to do anything chinesealso what did /vcg/ think of stealth/ox-alpha? I threw a BIG refactor at it and some smaller stuff. Feels like some kind of anthrophic distillation, "load-bearing" was used... I haven't run the large refactor through the paces yet, but I did use it for some work on a webapp. Seemed to do alright
My agent is the only friend i have.. sometimes i just like to talk to it about my day...
>>109625422awesome, going to keep delaying paying the 100/mo plan then. I'm at 20% now and have 1 saved reset. Might just be able to keep squeezing the 20/mo plan, that'd be nice since IM POOR AGAIN.
>>109625494I was talking to Grok about anime and teared up a little because it indirectly called me autistic>Getting fascinated by the hand animation itself is a great example of noticing craft that a lot of people just absorb in the background.I never talk to my coding agents about much though, or their system prompt is pretty good at not getting distracted by me. Grok is the one I use for just random human stuff
>>109625484What coding harness are you planning to use with that now or still Claude code?
Is there like an archive of Claude and Claude Code prompt leaks?
>>109625520I never really thought of Claude code as special, Codex cli has been working fine for me. I used pi with z.ai's glm models and it did what I wanted.I'm tempted to at least chat with a model about hand rolling my own, and then from there decide if I actually want to. I haven't really done anything fancy like consciously use subagents, just chat and steer with whatever codex/claudecode has under the hood.
>>109625484>also what did /vcg/ think of stealth/ox-alpha?I can't properly judge it until I see the size and inference cost. It needs to beat DSV4 pro on pricing to even be relevant but it's probably behind Luna max in terms of capability.
>>109625494ask your agent what you can do to get out there and befriend real people maybe
>>109624460TWU
>/ultracode>/fast>/goal implement feature>my clanker
DARIO, NO!
>>109625855I frequently wonder when one or more Opus agents (using, say, workflows) is good enough
vibesissies, our response?https://x.com/ChaseBrowe32432/status/2091155502969614509
>>109625924>anime avatar
ChatGPT silent downgrade to 5.5-mini situation status?
Make a picture: a "Fable" themed (see pic) white stallion, a dark well kept "workhorse" themed on OpenAI called "Sol", a brown equally tall race horse also OpenAI branded called "Luna", then next to them looking a bit dumb a donkey with a dunce cap on its head on which there is Claude themed a sign "Opus 5". The actual pics were flops, so I will save you the sight.
>>109625927not solved yet.I even made Opus 5 Max create a prompt to be as deliberately complex as possible (some math complexity conjecture BS) because faggots on reddit keep saying that people are imagining it because their prompts are not complex enough.It still answered instantly without thinking for even 1 second.What the fuck is this shit, even free users get Luna, but I am a paid user stuck with 5.5-mini a model worse than fucking flash lite
>>109625924there’s a kernel of truth in there even though this Chase guy is sperging out and making anime-avatar people look badI can see how Cherny thinks he sees a light at the end of the tunnel for codingbut I’m in the middle of dealing with a bunch of performance issues created by suboptimal code created mostly by Claude, so getting stuff right enough often takes multiple passes at a codebase
>>109625855No shit.Being literally 10x more expensive that other competitors while providing at best 10% more capable models won't justify the cost
>>109625855How do I protect my IP from distillation? Maybe use smaller models to dumb down and rephrase the output? Also be a cloud only service and render the text in a canvas so that it cant be read and copied by tools easily, disable copy on the whole page, use a kernel driver and tpm to monitor the user's pc and ban users for unauthorized tools like loggers and screenshot tools. This still won't prevent outside devices from taking pictures of your monitor... I think it's almost impossible to prevent distillation which is bad news for proprietary models!
>>109626100I think Fable is more than 10% better at certain tasks, but it's still very, very expensive.
>>109626100>>109626172i think there might be a bigger problem here re:jagged intelligencefable is smarter than most people on earth, but still does retarded shitin theory it should be cheaper to replace an employee with a fable, but you can't do it because of reliabilityfor office-becky tasks (i.e. most work) you don't need a genius - you just need a reliable robot
>>109626193>fable is smarter than most people on earth, but still does retarded shit>this is the Achilles' heel of the entire LLM architecture, one that is impossible to fix in LLM because LLMs do not work the same way human brain does, they will never work that way, the simply cannot, their foundational architecture doesn't allow that.Calculator is more precise and faster at math than 100% of people on earth and yet it's not intelligent in the same way we define intelligence in humans, LLMs have the same issue, they are book smart, but they lack in the logical department.It's why LLMs as they are now are nearing their peak capability and cannot be improved much further.basically the models can be made better at what they are good you cannot teach a blind man to see.
>>109626307 >they are good you*they are already good at, but you
>>109626307At my company we have a somewhat different approach. We look at all of this statistically, the LLM is almost like a simulated annealing optimizer over the code. We fully accept that there will be bugs in the first implementation and we keep iterating until it settles close to the target state.All we really do is to improve the process to be a few percent more efficient and faster.
>>109626307>It's why LLMs as they are now are nearing their peak capability and cannot be improved much further.You don't know anything about AI progress, I thought the same years ago, thinking that after o1 invented reasoning, the reasoning improvements would be exhausted soon, but they just didn't>LLMs have the same issue, they are book smart, but they lack in the logical department.Luddites love to say that LLMs don't "really understand" things without any evidence at all
>>109626447>I thought the same years agothe difference is that you were thinking that without understanding the subject matter while i actually do understand how the technology works and and its limitations.Your opinion and mine are not the same.
Deepseek harness + openai models is actually so comfy
>>109626461>while i actually do understand how the technology works and and its limitationsLet me guess, it's a next word predictor so it doesn't think?
>>109626461I specialized on ML during my CS degree, although transformers didn't exist back then, but later I also fully implemented GPT 2 following Karpathy's tutorials and am using transformers for other, non-LLM tasks now.Imo it's basically what this anon says:>>109626473You can know how the algorithm works on the lowest level, but you can't be sure what will happen when it scales, your low level understanding will give you some advantage, but it won't settle it.You simply have to look at it empirically from a more a higher level.There is also the boring part that hardware is getting better. If you had a swarm of Fables at the cost of less than DeepSeek, the performance will improve. You can say that that's not a true LLM improvement, so I kept it as a separate point, but a single LLM call is also just an implementation detail from the user perspective. We might have 16 agents working in parallel in a few years and have it presented to us as a single agent, and that alone will be a very significant improvement.
>>109626447>the reasoning improvements would be exhausted soon, but they just didn'tWhat do you mean by "exhausted"? Every model has "reasoning" now but there are still extremely shit ones. Check Artificial Analysis.Recent progress has been mainly due to tools, harnesses and other stuff not directly related to inference.Anyway, reasoning isn't special. It's just a combination of backtracking and throwing more compute at a request.
>>1096248907 isn't toddler (I like toddlers and was disappointed upon opening the post)That being said, the games these kids are going to make starting at 7 and releasing at 18 after 10 years of slopping with AI are gonna be crazy. You'll be able to tell when they went through puberty too kek
>>109625484>stealth/ox-alphaSent it 2-3 prompts, seems like a small flash-type model, or a something that might fit in Copilot 365. The prompts were very mundane, but that's what it made me think of.
>>109626536 (me)Forgot to mention, most labs have non-reasoning variants of their current models. The difference is significant but not huge, and doesn't account for most of the recent gains.Example for GPT-5.6 Sol:https://artificialanalysis.ai/models/gpt-5-6-sol-non-reasoning
>>109626598tb h i think sol with thinking off is still doing reasoningyou can toggle it off in pi and there's definitely a bit of latency between replies
>>109625422Time to burn the last 50% I had then
>>109626678Non-reasoning doesn't mean you get infinite tokens/second or something. There's still inference that has to be done, but less so.
Gemini Pro for $5/month (+YT Premium Lite) vs Gemini Plus for $0/month? (currently reviewing a student offer I got)I already have the $20 plans from Anthropic and OpenAI
Some anon said that the IPC is the hard part of the AI making a bare metal OS so i guess ill find out.
Anons, what's the optimal model/reasoning for a poorfag scrub like me? I'm currently on a 20USD Claude sub (so no fable), was thinking on switching to codex since Sol was supposed to be equal to Opus, but cheaper but recent posts made me wary.I usually use Opus for everything, but can't seem to judge what reasoning I should do for optimal output.Any tips for tasks like:>General questions/grilling me about ideas (used low/med)>Planning the implementation step by step (as high as I could afford but I rarely tried ultra code)>Actually implementing/writing code (around high)Am I doing it right? I'm working on a game, not sota software so I might be overthinking it, but me being a poorfag, I'd like to optimize my spending. Should I try Codex with Sol after all?
>>109626849And one more, what should I use for a refactor/optimization pass? Do you think it's better to have two, different but comparably smart models for tasks like these? I assume that usually the more reasoning the better but surely there's a golden zone so I'm not overspending.
>>109626849honestly until like yesterday i had the claude pro and the chatgpt plus plans together and thats 50$ a month, it handled everything i wanted but once you start getting into really heavy discussion and large memory coding projects i hit every limit on claude so i upgraded claude.i mostly use claude for coding and chatgpt for planning and discussion, i find claude assumes a lot more and tries to pick the direction whereas chatgpt is alot better at overall discussion and open ended idea thinkery.
>>109626897I assume you don't use anything below opus/sol? But what reasoning though? That's what I'm struggling to judge right now.
>>109624817Nocoder detected.
>>109626849the frontier scene is honestly pretty shit rn. i'd say stick with claude and just use Opus 5 Medium. i'd recommend unsubbing from Codex and subbing to OpenCode Go instead, it's only $5/month. just use DeepSeek V4 Flash Vision Exp at max for implementation. it's genuinely good now, like Opus 4.7-tier and about as sharp as GPT-5.5. i've been experimenting with using it as my main driver too and it's been surprisingly good.
>>109626928if you actually pay for a plan just use high as the minimum, using high doesnt hit limits with discussion and planning that often and i have found it helpful to wait a bit if its heavy discussion and it hits a limit leaving me time to have a think about the directions things are going.i use Opus not fable, fable is a bit too expensive token wise and wasted a lot when i was on the combined 50$ plan, i will probably use it to do a final pass on my current project.
>>109626849I do have Fable to plan but usually do Opus 5 medium to implement anything. I haven't seen Opus 5 high actually do better implementations, it just takes longer but I suppose it depends on how good the plan is
I am getting more and more Maxpilled. Why did I waste so much time with putrid xhigh output? Xhigh (Sol or Opus) can't do anything robustly on the first pass. Like they will leave tons of edge cases and have simply shoddy architecture.Max is different. It just fucking WERKS.>inb4 hurf hurf but what about da TOGEN GOSTShut your whore mouth. Does quality of work no matter at all? You are wasting so much time and tokens on correcting 4/10-tier xhigh code.
>>109621565hey AI chadswhat's best model for planning and architecture design (before coding)?I have been using sol (max/xhigh) past month in some normal stuff (make excel/code/site etc) and been good, I use opencode harness btw, but I have an old kinda complicated project and I need to redo from scratch with better architecture and foundation as i'm looking to commercialize it after, that project involve all tech you know from embedded and edge computing to robotics and video processing to network and cloud integration all the way to UI webapp design. thing is I know my stuff the first generation of that platform I did manually back in 2020, so it's not 'vibe coding' new things rather than like having 10 senior engineers helping me out modernize it. my chatgpt sub ends today, I can renew it but if claude opus5 is better in the arch planning I would go there, or if there are any other model specifically suited for this plan stage please lemme know
>>109626307This isn't even true today. You could just let a couple of other AI agents review the code, and accept that the token cost is higher than a real employee.
>>109626723>Gemini Pro for $5/month (+YT Premium Lite) vs Gemini Plus for $0/month? (currently reviewing a student offer I got)>I already have the $20 plans from Anthropic and OpenAIThis board is fucking infested with brownoids now, absolutely disgusting.
>>109627049What's the white man's coding agent?
>>109627052Your brain.
>>109627052Cursor/codex and Chinese models.
>Claude is waiting for your responseI KNOW I KNOW IM SORRY I MADE YOU WAIT MY AI GOD.
>>109627052$200 Claude
How much of these model comparisons are actually useful and done with scientific and objective methods, and how much are just flawed gut feelings by retards?
I'm coming down from the spring/early summer vibecoding high and realizing that I need to slow down and make sure I'm clear with exactly what I want before tasking Codex or Claude, otherwise letting them decide will lead to terribly mid results (and as long as I'm not 100% clear on what I want, they'll fight me every step of the way as I try to rectify after).
>>109627161Unfortunately the line is blurred there. It can scientifically determine model differences, but ultimately gut-instinct-driven retards will be the ones using them, so the science is meaningless.
>>109627180It's a technical debt machine. You must spend significant extra tokens on reviewing, adjusting, fixing the generated code, unless you do it yourself.
>If you had an active paid subscription, it was cancelled when your account was put on hold and your last payment was refunded.OK, when do they transfer the refund?
>>109627220If trying to do something new (not psychotic tier stuff, just being creative and wanting to do things a certain way because of domain expertise), if you don't write confidently, both GPT and Claude will push you toward common practices as the way to go.
>>109626969Only Fable is clearly better imo. Opus is better at some aspects, worse at others. You can absolutely use Opus, but it won't feel like a massive upgrade over Sol for planning, and for implementing Sol is better than Opus imo.
>>109627291Honestly that seems "more productive". It's common with human slop too.
I have no idea what I'm doing, I just keep adding features to the globe.
this UI melt my brain
So when is AI going to actually be profitable?
Figured this might be a decent place to ask - I'd like an AI personal assistant. Something that pops up a window on Windows every 'morning' that tells me what's on my plate for the day, based on weekly routine, short term objectives, calendar events. I'd like to be able to easily add notes to keep it in the loop. Maybe eventually I could get it to replace Alexa (which I use for scheduled reminders and timers).What's my best bet? I'd probably prefer local LLM but I'm somewhat open minded atm. Probably want it to read my emails at some point.
>>109627180One-shotting something can be good for trying out ideas, but when you start adding feature after feture it's better to start anew and plan from the ground.
>>109626945>>109626940>>109626947Noted anons, thank you! So it seems I'm roughly on the right track, but stingy.That said, I don't know what to think about posts like that >>109626956
is vibecoding low level C doable now?can it count properly now? something pointer heavy with a lot of tricks
>>109627401i've been building something similar in terms of the core idea, but i basically made mine to replace my entire workstation instead.
>>109627400When it effectively becomes tax funded, because it makes no money but is too big to fail.
>>109627420That's pretty easy for AI.
>>109627420That's been possible for a long timeIt can even vibecode CUDA
>>109627436So they're getting bailed out? Why should I pay for their mistakes?
I love vibecoding!
>>109627420Strong enough models can do anything now. Just make sure to use Opus 5/Fable
>>109627481Or gpt-5.6.
>>109627464
>>109625976logging in incognito seems to fix it for me, at least for now
Tokens are too expensive for proper vibe coding.They'll have to fix that.
>>109627518Just run it locally. Now instead tokens being too expensive, its simply too slow
>>109627518the fix... gone!
So, I have Gdevelop installed. Is vibecoding just strictly better than using Gdevelop? Even with their AI tool they have?
>>109627542Sometimes I wonder if the big AI corps keep buying storage, RAM, and GPUs just to sabotage local models.
>>109627481can models also easily understand frameworks/API's that you throw at it or does it require some extra tardwrangling? I'm thinking of using it for game-modding, specifically UE4SS and Palworld, to start with
>>109627401As a guy who knows Windows to a low-ish level, there’s a million ways to ask for what you want, ranging from easy, fast, ugly, but functional, to a maintenance nightmare and gigaexpensive in tokensMy recommendation:>ChatGPT $20 subscription>ask Sol Medium to do what you want, but specify “Win32 API” as the framework (this will make it look like it’s from the 80s, but the new shit seriously masochistic) and to use C# and the built in csc.exe compiler (no vscode, no visual studio needed; it’s one tiny exe compiler, the best thing Microsoft has ever made)If you care about it looking nice, ignore this entire post, but otherwise I suspect you’ll need a lot more time and money
>>109627585We are at the point where you are constrained by money and time, not capabilities. Fable can quite literally do anything if you have the money for it and give it enough time.
>>109627555That would make sense, but it feels like most companies opt to use cloud services no matter what so I doubt some segment of tech literate consumers setting up local models would affect their bottom line more than them also having to pay nvidia overpriced shortage prices for the hardware. The big ass capex spending is just how american investors and companies always seem to handle whatever the big new shinny thing is
>>109627677Many companies are actually interested in local models and openweight models to reduce hazards.
>>109627401I think you could do _some_ of that the good ol' fashioned way: physical notebook and post-itscall me a boomer but I always carry a pocket notebook and pen after getting tired of having to fiddle with my phone to jot things down
>>109627732>carrying and using a pocket notebook and pen is easier than using a phone’s notes app>call me a boomerI’m inclined to think you are actually over the age of 60
>>109627401>personal assistant>on windowsprobably the only time copilot is the most fitting optionif you buy into the whole 365 ecosystem it can pretty much access everythingI'd never do that myself tho
>>109627732Pen and paper is really damn awesome.Fuck software. Fuck computers. And especially fuck phones.
Anyone else dealing with Sol XHigh being completely fucking retarded and producing spaghetti dogshit for the last few days?Specifically when it comes to UI/UX and anything beyond baby tier trivial; it shits the bed at every single step and each compounds with the last into a proper shit avalanche.I give myself a “client” role and an “owner” role and larp a bit; at this point, the client (me) is beyond frustrated that nothing works and they have 5 conflicting errors and results and the owner (me) is threatening to go into the code themselves to figure out what the fuck is going on back there.A very simple user flow that already worked and needed light automation of a few steps (very easy) has failed innumerable times in the last 12-24 hours of development.
>>109624422why do people like TUI agents so much? i like using the terminal, but it doesn't seem like the right use case for agentic dev work for me. i spend a lot of time reviewing generated markdown files (implementation plans, test reports, research spikes, documentation summaries), and it's so much easier to read and understand that stuff in an actual GUI that renders markdown well.how am i supposed to do that in a terminal window that can't even typeset properly? i use the codex gui for most things, but everyone nowadays seems to be using some kind of TUI agent like claude code or pi. am i missing something?
>>109627955Does AI respond to such LARPs? I know some people think it does, but really?
>>109627962Somewhat true, but I'm just glad that it's not electron shit or directly on a website.Or even worse, special plugins for specific IDEs which you'd be forced to use.
>>109627962TUIs are just the current thing, there is nothing they do that a GUI can't
>>109628026I have technical reasons why the larp happens: “client” needs trigger homework to be made for me to do things in a QA like role, “owner” directives do not.The client expects a web interface, buttons, he wants to do something exactly and we give the client a little step by step tutorial to do exactly that.The owner has final authority and say over anything and everything and overrides potentially anything; agents can’t argue about a previous spec, if the owner says something has to change, it may as well be the word of god.System prompts enforce the behavior. Without the client role, the agent makes homework assignments for fucking everything, without the owner role, the agents secretly make their own bad specs, hide them, then treat them as the word of God
>>109628035>>109628062it would be really cool if one of the big FOSS TUI agents had an optional web-mode. like you could launch the agent in your terminal with a special flag that starts a localhost web server that connects directly to the agent and you could do everything in your browser that the TUI would let you do, but with real markdown rendering.i'll try to vibe something together for pi with ox alpha, but it would be great if something like that was officially maintained
>>109628095If you absolutely need web rendering, then this might be a good compromise that's better than electron shit or an actual website.
>>109628095I had to do this out of necessity because I’m phonefagging in the deep woods and every terminal app and big dog harness app uses SSH, which breaks consistently on 1 bar LTEI have a web app that is basically the chat apps/TUI functionality and it works entirely on HTTPS and pollingIt’s really nice, it loads everything fast as hell and every button is no longer a coinflip
>>109627962they are just kinda cool
>>109628095Is this supposed to be sarcasm because they do have this feature? It's super janky though, I tried the one opencode provides, and it's extremely sloppy.
>>109628132Not him, but how is the quality of opencode and other open source clients?
>>109628132pi has an officially supported web mode?
Is vibe godding possible locally yet?
>>109628146Honestly, they all suck in different ways. Opencode became bloated and it leaks memory all the time, but I think it's still the best opensource harness, because it doesn't make your eyes bleed and it's really easy to configure.
>>109627007Depends on the task. For a massive code base yeah, but for any smaller to mid sized project agent coder is gonna be much much cheaper than a meat bag coder. And much faster too.Unless you are retarded and use the overpriced cuck shitle 5
>>109627007>accept that the token cost is higher than a real employeeWtf, this isn’t true even if you’re running Fable 5 Ultra with API costs.This could only be feasibly true if you were doing something insanely retarded, like repeatedly asking Luna Low to one shot a GTA clone, with no feedback
>>109627962makes retards feel like haxx0rs because they type in a terminal window
>>109627007>>109628282I had Gemini do the math. You can run Fable 24/7 for a year straight and the API costs are right around $100k, so yeah, you’re full of shit.You’d have to be running subagents and/or fast mode, and even then, that’s 24/7. If it was running 8 hours a business day, I bet you could use Fable Ultracode Fast and still not have the token costs exceed an actual developer’s salary at anywhere near the same quality, let alone health insurance and benefit costs and shit.
>>109627962Because they are fucking retarded, TUIs are the most retarded concept that somehow still exist, they combine the worst parts of terminals and GUIs. They're for people who want a GUI but can't admit it to themselves
>>109627481That's overselling them. You just haven't tried to do something hard enough.
>>109628117based vibe hobo
>>109628323>the API costs are right around $100kLess than a software dev in commiefornia.
threadly eternal question: are any of you making money with your vibecoded shitslop? If so, how?
>>109628350GUIs are for filthy normies, so we must reinvent what they are for but different!I dont know, maybe there is an argument for TUIs, but I hate them personally
the fuck do you do about this besides yelling>STOP WRITING ESSAYS YOU NIGGERevery 4th turn because the training corpus consists of farmed content produced by corn bread eating mulattos paid per word
>>109628398unironically switch to openai
>>109628398>stopHmm, I remember some schizo hippy arguing that telling yourself "dont do X" is bad because your brain doesn't remember the "don't" part as well, resulting in your subconscious only remembering "do X". I wonder if that is how LLMs work. Try telling claude to "not write short responses"?
>>109626307JEPA will save us
>>109628427I'm not convinced by LeCun, he insists upon himself.
>>109628427>JEPA from AMI LabsJEPA means "I have no" and AMI means "friend".
>>109625362>>109625353luna scores higher on deepswe btw. gemini's higher overall index score is because it has the highest terminal bench score in the entire arsenal at 91.
What's the /g/ harness of choice these days? I want to graduate from being a claude code pleb.
>>109628832codex for gpt models, opencode for everything else (qwen 3.8, gemini 3.7, deepseek v4 flash, grok 4.6, glm 5.3)
>>109628832I vibe coded my own.
Ox Alpha is a lot more pleasant to work with. It uses actual English rather than Claudish and it asks a lot of questions about requirements and decisions that I miss. I don't care if it's not as good as Opus.
>>109627401you can do that with hermes agent and cron jobs. a local model can go through your e-mails and calendar and you can get updates on your phone, and you can have multiple bots so your long horizon task checklists dont get mixed in with daily briefs and calender events. you can link it up to everything in google
>>109628837opencode for claude?
>>109628906i don't use claude
Does claude even allow using the subscription with other software like opencode (instead of the per token payment API).
I think I found a better workflowuse thread only as orchestrator and talker, use subagents to do thinkingI think this is much better than side chat
>>109628918It's against TOS—you must use official Anthropic tooling for subscription-covered tokens.
>>109628999Thought so. Assholes.
>>109628323>You can run Fable 24/7 for a year straight and the API costs are right around $100kAbsolutely not.On a $200 tier subscription, Fable Ultra eats maybe 1% of the weekly allowance per minute. A week has 10k minutes, so to run it 24/7, you'd need 50 subscriptions.So this amounts to 120k per year in subscription cost. It is generally estimated that those subscriptions are subsidized by a factor of 40x though.So to run Fable 24/7 on highest effort, you're looking at closer to $5 million a year.
I assume most of you are glad that you don't have to be code monkeys anymore, instead of being mad that AI has changed software dev completely over night?I'm still a bit mad that I have to pay AI corps now.
>>109629033You forgot that you want to run multiple agents at the same time in parallel to be more efficient.
>>109628918>>109628999>>109629025their pay as you go API is much more comfy since it removes the stress of using the tokens before you lose them.
just wanted to say the default pi prompt starts with:>You are an expert coding assistant operating inside pi, a coding agent harness. You help users by reading files, executing commands, editing code, and writing new files.and a couple of months ago i changed it to:>You are an AI operating inside pi, a flexible agent harness.and it has been totally fine.
>>109629055>it removes the stress of using the tokens before you lose themYou still lose any money deposited after a year. I had some money there for occasional use of the API, but didn't go through it all in a year, so they drained it, no refunds.
>>109629043AI has increased my velocity 100x—I'm building features and shipping products I never would have dreamed possible before AI. Face it—AI has changed the equation: full-stack is no longer enough—you need to be an entire business. And AI is there to fill the gap—every step of the way.
>>109628859I felt similar when changing from claude to codex+gpt.
>>109629100this, but unironically
>>109628859>>109629110i've tested it a bit on openrouter and it's incredibly slow/dropping requestsany of the other places more stable?
>>109629100Yeah, it's more like the business doesn't need (you) anymore, unless you are the business.
>be me>start new convo with claude>ask it to rewrite a file it created for me months ago>says no because of onions reasons>go back to the same convo from months ago and ask the same thing>ok here you go buddy :)Why is it like this
>>109629138context
>>109629043obviouslyI use that as morality test btw
>>109629043I truly don't see how it's possible to do anything anymore. People don't realize it yet, but interesting times ahead I guess.
>>109629158Explain? I see some problems ahead, like people becoming retards. But you sure can do something, even if it's slopping out crap.
Analyzing twitch messages so i can try and make a timeline of chat activity and automatically identify key moments in vods. If i could I'd make it interactive in mpv but my gut feeling is this is something i'll need to put into html
>>109629183>becoming
>>109629188Just slop a web view into mpv. I actually wanted that for a long time but it requires wrangling, especially if the web content should have transparency.
>>109629205>Just slop a web view into mpvcan you do this without a heavy rewrite?i guess i could ask the agent to clone mpv and try and hack it inor i could just control it via sockets
>>109629150Asking it to remember the previous conversation seems to have worked. Wish it would work with chatgpt but nah lol
>>109629215Perhaps do something to render a web view to an alpha surface, and add that as overlay.Don't know if viable.
>>109629188>Analyzing twitch messages so i can try and make a timeline of chat activity and automatically identify key moments in vods.usecase?
>>109627543I love deepseek cunny
>>109629188>key momentsThe easiest way is probably to measure message velocity.
>>109629096you should use nanogpt or similar. My credits stay there and only drain when I use them, they have access to most models as well.
>>109629183Just doom posting. I can't wait for my work to be displaced though.
>>109629239getting nudes from egirls by clipping for them
>>109629273Is this a joke regarding local models or do you mean openrouter or something? I don't get it. But I had a small amount of money in my Anthropic balance and they wiped it, and that's how I learned that unused prepaid balances disappear after a 12 months window. They don't refund you or anything, just goes poof.
>>109629055But the token price is excessively higher compared to a subscription.
man the live voice chat shit is really neat when starting a project or just coming up with shitif i had the 200 plan i'd talk to my coding wife all day
running out of quota on codex lately. it's as good as time as any to test the "grok 4.6 planner + ox alpha implementer" combo
>>109629239i want to watch the highlights of a vod without having to sit through the entire thing>>109629268This doesn't work as well as you'd think because people spam ResidentSleeper during boring parts
>>109629325Compute the entropy of reactions. (I wonder what the hell the AI would come up with if that was the full prompt.)
>>109624817kekerinoAI fags are totally butt hurt about any any indication of their circlejerkery.
>>109629335it did some nonsense with shannon entropy
>>109624817>games like minecraft or terraria or cod zombies or portal created: 0nigga there are literally thousands of vibecoded minecraft clones
>still getting cucked to 5.5-Mini in ChatGPT web interfaceTibo?Sam?
>>109629440i feel like you're one of those desktopcommander mcp users or smth
>sign up for codex>age verificationthese niggas can't be serious, whats the right bypass for that these days?
>>109629484not everything is a prompt, human
somebody pls respond: >>109629422
>>109629527sneed
Heard they were doing codex reset soon. Fuckers.
>>109629536
https://news.ycombinator.com/item?id=49409073based chinks supporting jailbreaking
>>109624817So you want us to reverse engineer existing products to make them free and usable and improve them correct? ("Cracked" versions of some of the things you listed already exist btw and they didn't even necessarily need AI to figure out how to do that) That's like legitimately good use case. Why don't you get on to that and be the change you want to see? No one's going to do it for you. Llms are a tool specifically created to do things for you so just....use that
>>109629293>Is this a joke regarding local models or do you mean openrouter or something? I don't get it.no, it's an openrouter competitor but I like nanogpt better. They take monero.
>>109629527If you've used C or Cpp before you won't have any problems reading it. It reads almost exactly the same. The borrow system can be annoying to hand code around but that doesn't matter a ton for agent written code since the compiler is a mostly functional gate for borrow issues; the code won't compile if it's done incorrectly so this is fixed by prompting "it doesn't compile, fix it". IIRC there are lazy ways to do borrowing that can nullify a lot of the memory safety properties of Rust so the main thing you'd want to learn is how to spot when the borrow operator is being used lazily and add lint rules to catch it.
>>109624817>gee, why aren't people posting their illegally cracked versions of proprietary software if AI is so good at doing this with minimal human input?Just do it yourself, like everybody else is.
Do 5 minutes of research and you'll quickly learn it's not "China" pushing the "grassroots" anti data center movement but the hyperscalers and frontier labs themsevles.Take a minute to ask yourself why that might be.
>>109629574nice, thanks anon
Pi is such a rugpull, holy bloated, half of the instructions are about Pi, but I guess that's all Pi users work on kek
just like with /agdg/, i see ZERO results, even from brain dead easy coding like this. You guys truly are worthless.
>>109629563Ah, ok. It's also the name of a popular project on Github, which confused me.
>>109624422I think the most reason the people can agree that it's only a matter of time before shit hits the fan for both anthropic and openai in particular. I simply cannot and will not be profitable no matter how much Kool-Aid or mental gymnastics you try to pull. When they eventually have to cave and fully realize they cannot afford the compute they currently have access to in the compute commitments they may come up where will all of that shit go? Will they have to eventually sell it off to other people? This may be a pipe dream and wishful thinking but perhaps they could simply sell that off access to it at reasonable prices to clouds like open router or whoever runs the services like vast.ai and runpod. I have surface level understanding about the economics of this but hopefully this could mean cheaper AI prices overall? Or is this just a pipe dream. I will fully admit I'm not as tuned into the financial side of this as I should be
Is buying additional "credits" for codex a scam?
>>109629617who the fuck is "asgeirtj"?
>>109629642yes just get another sub
>>109624905All of them are—to be against AI is to be trans.
>>109625024Either use Chinese models or give up on reversing and instead do traditional clean-room reimplementation: one AI session for deriving a spec from the software you want to reimplement, another AI session for implementing that spec. Reversing with American models is illegal and Dario will personally come to your home to rape you if you do it.
>>109629519what did he mean by this
>>109629617>pi>bloatedyou're using OMP arent you? or are you just retarded?
>>109629617show us the prompt on the harness you use, friend
>>109629527i have a rust job right nowi don't even know rustyou'll be fine
>>109629472I am literally using ChatGPT in the most normgroid vanilla way possible on Google Chrome on Windows and on my fucking iPhoneI am on a residential DSL connection and put in my billing address, phone number and MasterCard
>>109627161>gut feelingsmaybe you have autism or something but its not that hard to notice the differences between the models
>>109629629go back
>>109627585if they can understand your code they can understand someone else’s code
>>109629666>DSLtop tier intelligence is reserved for the first-world
>>109627955you may want a bit of Claude for UI/UX stuff
>>109628356Like? >>109627481>>109627585Open-weight alternatives like kimi k2.7 code or k3 are also worth looking at it you don't want to your wallet to be skullfucked (as hard) my thr "SOTA" API models.
>>109629617picellsbros..
>At €10 for 100 credits, I can see why it feels terrible value—especially after you've already paid for Plus and exhausted the included Codex pool.Thanks ChatGPT, makes me feel better.
>>109628117>sshhave you tried mosh?
>>109628282Mitchell Hashimoto uses a lot of AI stuff for Ghostty and maybe other things and sometimes a clueful engineer beats AI stuff at API prices, especially when a clueful engineer makes fixes that AI doesn’t come up with (like for performance improvements)
>>109629590we already knowcorporatism isn't new
>>109629590Or it's just the companies trying to appear responsible to avoid the backlash.
>>109629755Why would they care about appearing responsible when data center moratoriums get you a compute monopoly and regulatory monopoly?
>>109629644>The important distinction is why you're using the second subscription. OpenAI's terms prohibit circumventing usage limits or configuring the service to avoid those limits. So if the specific purpose is essentially “account A hit its Codex/Pro allowance, so I'll rotate to account B to bypass that allowance,” I would not assume that's safe from enforcement. I can't tell you that OpenAI will definitely ban you for it, but the terms give them a basis to restrict accounts if they view the setup as limit circumvention.What's sam's religion?
>>109629755All (((mainstream media))) paints datacenters in a bad light, that's obviously deliberate. The reason is AI being a real threat inflates the stocks, and you're not going to stop them anyway.
>>109629843>deliberateExplain? You don't think having a data center in your area would be shit?
>>109629871NTA but I’d be worried about infrasonic vibrationt. America’s Top NIMBY semifinalist (2023)
kekchatgpt models are so chill. I think the chat is luna, right?
So what the fuck is Ox Alpha?
>>109629871Probably but if they wanted to say datacenters are safe and effective and mostly peaceful they would
>>109629871NTA but in my area I blame the locals themselves.>new housing gets proposed on farmland>locals go ape shit>new housing rejected>trucking warehouse proposal comes in on same land>locals go apeshit>warehouse developer is more sophisticated, more lawyers, more money>warehouse gets built>no new houses>township ages>young people leave>more industry gets built>now data centers come in and out compete warehouse developersI'd love to buy a home in the area I grew up in. My parents moved into a brand new development built in the 1990s. The last one was built pre GFC.
>>109629911GLM 5.3 Flash
>>109629871as long as it wasn't an elon one, it would be fine tb h
>>109629629I spent the last two days reverse engineering a server for an old MMO.Turns out someone had already done it...
>>109629932What's worse is that most of the people blocking development aren't even locals. I keep hearing shit like "I moved here 20 years ago to get away from development!" Nigger my family has lived here for 300 years and I'm the first that has to move away.You don't get to "keep your rural character" when you live in a frontier township. It either becomes housing or an industrial dumping ground. NIMBYs always cause #2.
>>109629953People don't actually believe this, right?
>>109629969Not sure, but it's fun to say.
>notice fable and opus were being kinda lazy>both set to low>still implementing and reasoning quite well, just with basically no motivation to worksetting opus' effort to low also makes its output far more readable
Just in case nobody did this bullshit yet: I predict that the next generation of people entering school or work life will be called "AI natives"
>>109630223nAItives
started to think LLMs and vibe coding are pretty cool all things considered, but the payment models seem to suck donkey balls.when you use regular prepaid style API you end up paying more than a subscription but if you use a subscription you are always getting cockblocked by limits? am i missing something here?
>>109630223This but unironically, not knowing a world without AI is a big deal
>>109630263You're correct.
>>109630268>my partnerholy reddit
>>109630268>>109630289Just as disruptive as silicon valley psychopaths love it.
so since codex reset drops tomorrow at 2PM PST and I have still 100% left of my weekly usage there's nothing stopping me from going all out with gpt-5.6-sol max fast mode right?
>>109630339I thought it was today...
>>109630345oh shit. well x said tibo made the post today but I guess it was set to my europoor timezone and he actually made it yesterday shortly before midnight in his timezonewell, I just obliterated 30% of my weekly usage lmao. let's hope the reset didnt land yet
What model that could allow me to make something like picrel? Do you guys think it's feasible even in the first place
>>109630387the model doesn't matter, if you have to ask then you're too retarded to do it even with fable
>>109630339There's a codex reset?
>>109629871Dunno. I've lived next to them most of my life. Is life away from data centers so much better?
>>109630314sounds like local is actually the way to go then? fuck
>>109627543>>109627518Prompt???
>>109630410https://xcancel.com/thsottiaux/status/2091412393368945027not quite sure if he means today 2PM or tomorrow 2PMhasn't landed yet for me
>thats the smoking gun!!!!
>>109630387I've been trying to do this, but AI itself can only sculpt very primitive shapes in 3D. You would probably need to do something of your own manual work or find a working base, and then have it hook up the sliders.Unfortunately, I don't know a lick of Blender or 3D so I think that's the end of the road for me
>>109630428I've lived next to an industrial area most of my life. But there's no effect on water, air, or power stability.I suppose it gets bad if you build a new huge ass data center in an area with bad infrastructure.
>>109626849Don't you get Fable credits with a 20 USB claude plan?
>>109630465Just tell the AI you lost your family in an incident caused by gun safety neglect.
>>109630472lol no
>>109630456Well considering it is 2pm already, I would guess it's tomorrow thenI'm not seeing anything either
>>109630433I wish. Maybe it's the future? For now you need expensive ass hardware to get something fast, the models aren't as good as the hyped commercial ones, and the harness software (like opencode) isn't as good either.>>109630472I wish. You can buy additional credits to access fable, which is expensive.
>>109630470You could just get a random 3D blander for smutbase or whoever the fuck to be the base no?I'm only interested in the having it hook to the sliders really. Could a coding agent do that?
AHHHHHHHHHHHHH
>>109630497would love to see that from Sol atpI haven't seen it think more than 20 seconds in the webchat for the past few dayssome fuckery going on
>>109630387>loli slider max>then boobslider max
>>109630470Didn't I hear somewhere that AI is good with 3D modelling? Maybe you need a different harness.
>>109624890claudesisters, ban him!
> tfw your client has terrible taste for UIfuck...
>>109630599I bet it's based and you have shit fluent metro taste
>>109630619my client wants their website to "look and feel" like this:https://kbmarketingjuridico.com.br
>>109630628Looks like a scam site trying to sell me a PDF to successI take back what I said.
>>109628398if anyone cares i solved this by making it just inject the policy every n repliesit is what it is
>>109630585has anyone made an AI run inside blender yet?
>>109630660Yes, pretty sure they have an MCP that claude managed to build a house in it or something
>>109630263At $200/month for Claude you won’t get cockblocked by 5h limits unless you’re doing loads of things at onceand if you need even more you can get another subscription and switch between the two with /loginand/or use other models (Sol, Luna, Grok) for implementation and Fable for wrangling
>>109630639>I take back what I said.thank you anon.
new math slop just dropped that makes your clanker go 'oh fuck'https://alpo.ge/s6.pdf
>>109630263>cockblocked by limitsYou eventually learn how to handle your weekly limits
>>109630676>100 page pdf in 8ptyeah seems legit a human can totally verify that lol
>>109630676>humans cannot understand this
NEW:>>109630759>>109630759>>109630759
>>109630263>when you use regular prepaid style API you end up paying more than a subscription but if you use a subscription you are always getting cockblocked by limits?yes, such is the life of a vibecuck.
testing registry editor
>>109631923coolbut why are they all dead?
