A general for vibe coding, coding agents, AI IDEs, browser builders, and shipping prototypes with LLMs.## News(7/02) GPT-5.6/Sol: OpenAI reportedly discussed giving Washington a 5% stake after pressure to delay GPT-5.6’s wide release.(7/01) Fable 5 + Mythos 5 restored globally after US lifted export controls (6/30).(6/30) Claude Sonnet 5: near-Opus 4.8 quality at $2/$10 intro, 1M ctx, new default for Free/Pro.(6/30) Meituan LongCat-2.0: 1.6T open coding model (MIT).----## What “vibe coding” is, and how to do ithttps://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://developers.openai.com/codex/clihttps://claude.com/product/claude-code## Not worth it for code, but maybe good for other thingshttps://geminicli.com/docs/https://x.ai/clihttps://chat.z.ai/## Open / local / self-hosted>>>/g/lmg----## Prompting / context / skillshttps://arps18.github.io/posts/claude-code-mastery/https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/https://github.com/mattpocock/skills — /grilling is a favorite## Other editors / terminal agents / coding agentshttps://aider.chat/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://stitch.withgoogle.com/## In-browser builders / hosted vibe toolshttps://bolt.new/https://replit.com/https://v0.app/docs## Benchmarks / rankingshttps://www.tbench.ai/leaderboard/terminal-bench/2.0## What we’ve donehttps://vcg.gitgud.site## Previous thread>>109221192
>That observation is sharper than "subjective" gives it credit for — you've hit a real design axis, and your instinct about which matters is the experienced one.i know, claudia, I'm so smart.
>>109226703imagine when she has a body
>>109226703caludette*
>>109226703Chloe*
so what ai is really bad on even claude, is keeping the "to do" files up to date. and seeing the big picture of "yeah i put a button, but there is no workflow for the user
I'm slowly merging 5 branches, good progress.
tomorrow, my sub ends. and yet, all i can think of is "god i wonder what gpt 5.6 is like"
>>109226906which sub
>>109226908claude
>people still don't know that "Remember, no Russian" meant "don't speak Russian because we're false flagging as Americans"
>>109226986stop being a boomer
GPT UTTERLY DESTROYS HUMANS WITH FACTS AND LOGIC
>>109227003luddites our 'ponse?
>>109227084TOTAL SUBMISSION
>>109227003sasuga japanese iq
>>109226986Should have snailcat masks and text would be>remember, no Ai
>>109227152lel
>>109226674>If an LLM wrote the code for you, and you then reviewed it, tested it thoroughly and made sure you could explain how it works to someone else that’s not vibe coding, it’s software development.I thought the whole time that this is what vibe coding is but vibe coding is apparently that AI writes code that you dont understand yourself?but isnt that kinda bad in the long term? like if you dont understand the code or how it is organized or how it works then wont the code eventually reach a point where its too complex even for AI and for you too? (assuming its a larger project).
>>109227193Yeah it's easy for slopcode to get out of hand and become a mystery to the prompter. If anything that's the main problem with vibeslopping, you tend to end up being unfamiliar with the slop because to become familiar with it would take a lot of effort, it's like someone else's code.I think the solution is to compartmentalise to the max, as far as making brand new code does. Accept the slop as a black box.
>>109227193>wont the code eventually reach a point where its too complex even for AIyou wait 6 months and the next ai understands it and adds more slopwe will do this forever
>>109227193>an LLM wrote the code for you, and you then reviewed it, tested it thoroughly and made sure you could explain how it works to someone elsethat's what I do at workmy hobby project is so massive that I ask my questions to the llm instead of reading the code. It's like we're tardwrangling each other
Well achieve AGI when this is fixed so far not even Fable can manage it
>>109226674>sol isn't fable levelgod damn it
>>109227193vibe coding is>i don't know what the fuck this is. but you know what? it werksi don't know how the fuck gemini made this. i just know that it replaced my perplexity subscription
>>109227240gpt 6 is coming in september
>>109227193yeah that's pretty much itthe industry kinda moved from "vibe coding" to "agentic engineering" which IMO is a terrible term but it is the industry trying to move away from vibe coding which currently means "I have no fucking idea what I'm doing but I will ship this nonetheless and once the codebase is more than a few hundreds lines and more than 5 files it's fucking joever i will never be able to maintain it without Fable meaning that if my quota is out or i have no more API money then it's literally abandonware"
>>109227240LIES!
>>109227193You'd get bailed out by the next model. There's no such thing as tech debtAnd for every line you don't understand you could generate a webpage with animations graphs, charts, and examples to explain it
>>109227233Anthropicslop is basically designed to harvest your datas.
>>109227193The LLMs get it right eventually. It's extremely inefficient, we had 2500 commits in the last week again. 2500 human commits are enough to ship a product, for us this is now just a week of bug fixing and it's still not bug free.Fixing bugs with AI is like running a heuristic optimization algorithm, like simulated annealing, you usually move closer to the goal, but there is no guarantee you will reach it, or that it will be efficient.
My company gives me $1300 to spend on random personal shit every year separate my actual pay. I'm thinking to get the cursor pro plan but is there a better deal out there?
>>109227319This is exactly how IT worked before LLMs too. Stupid architecture decisions happen every second of every day. Not only does AI make less mistakes, but it fixes them faster. It also sets up more and better guard rails with better explanations and more modular code, easier UIs to see the debug, etc. I wouldn't ask anyone but a senior dev to check my shit but they'd want a shit load of money and I'd basically get a couple pages of notes of what to feed back into my development loopWe're in a race to escape humanity
>>109227335a frontier model
>>109227335just do $100 on anthropic if you want to spend before fable leavesthen next month $100 on openaidecide whichever you prefer and stick with itno point paying other people
>>109227340Maybe I was just an exceptional dev, but I doubt it. In my last job before AI, I only produced 3 bugs in 2 years, now with AI staging is down a few times every week.
>>109227335cooooooodex
>>109227335$20 on anthropic so you can say you’ve tried fable and $100 on codex for 5.6 on Thursday.
so generally, use claude/fable when you want a partner, GPT if you want a competent worker
>>109227368What were you building? How many different pieces of the codebase were you responsible for? If you're responsible for a little piece of those whole thing do you think an LLM pointed at the same little piece would have a similar error rate? How about a vibecoder responsible for the same piece?
>>109227463sex partner
i want codex to talk to me
I have a genie that will code anything I want in my laptop, and I don't know what to ask for.
>>109227506for that you use gemma
>>109227529ask it to make your system 'my little pony' themed
>>109227527only uncs use that
I asked antigravity for help with a linux desktop issue and it actually found a bug in sddm itself and submitted the patch using github-cli 0_0https://github.com/sddm/sddm/issues/2129#issuecomment-4916095611
>>109227527im not excited
>>109227527>Live stream at 10am>No mention of timezoneIs this the power of frontier models?
>>109227557now that's fucking basedllm's shouldn't be allowed to bughunt, but if it finds a solution to an existing bug by accident I'm all for it
>>109227581>llm's shouldn't be allowed to bughuntwhy
>>109227557don't ever fucking do this
>>109227576My read right now: still positive, but slightly more anxious than yesterday.
>>109227586because bugs and design choices have a reasonable amount of overlap
>>109227335cursor is credit based, isn't it?
>>109227586slop overflow10 years ago the worst a popular repo had to worry about was jeets making pr's changing comments or variable names just to score some github updootsnow everyone with internet access and no skills can tell a chatbot to go find something to complain about and make a pr
rare 5.6 > fable impression
AI will unironically give indie game devs the manpower of a big AAA team in the future.
>>109227697lel. no chance. ai is good for boring single person project shit, but not for anything serious
>>109227671It's veiled but anthropic shill post, almost the entire post if actually about fable and gassing it up as a genius. Ironically debug or performance work is not where fable shines as the post would suggest, everyone knows this, if it doesn't flat out refuse it will sabotage, if it isn't sabotaged it will actually just be opus.
5.5 still outclass opus so 5.6 should dominate non-fable range
>>109227581>llm's shouldn't be allowed to bughuntWrong.>>109227649Literal skill issue as you said.It's no different than reporting a bug, it's just that now it comes along with a slop fix. As the maintainer you can disregard the slop fix and deal with the reported bug yourself. His suggested fix is simple too so whether it's accepted or not it's not like a jeet submitted a 10kloc slop PR.
>>109227799at least until opus 5 comes out, then it will edge out gpt 5.6 in raw intelligence and capability, while being completely unsustainable in real world usage despite the intelligence advantage
>>109227671the absolute insufferableness of these psychosis cultists
>>109227708>in the futureCope all you want. It's coming and you can't stop it.
>>109227708
>>109227804>completely missing the pointlowering the entry level for people is my only real complaint about llm's
>no geminiboss... im tired
>>109228004based based basedanother leaker also said gpt 6 should come in less than 2 monthsI hope my best jews win, we can't fall into the anthropic timeline
>>109227587>like 25 character patch instead of the slop behemoth some LLMs throw at problemsthis one's fine actually
Which model for programming? one of the 20$ subs, I'm a poorfag.
>>109227806ai slop ^
>>109228119GPT obviously
>>109228119claude code
>>109228119geppetti
better than glm 5.2 at what? cooooooding? coding is glm's only defining feature. it's sonnet tier at everything else
>>109228139so how tf do you train a model after pretraining?like, did they train 5.6 from 5.5 or from the base? because abandon a base sounds expensiveor maybe they are pressured by anthropic
>>109228161They probably have many model projects going in parallel, and the 6 branch is doing better than expected so they are abandoning 5
>>109228161most of the gains in coding perf have come from post-training rlvrafaik gpt-5 base lasted from 5.0 all the up to 5.4i'm not totally convinced 5.5 is a different pretrainbut it's obvious now that bigger pretrain helps and they have to ship a big boy
>>1092281615.6 isn't a pretrain. I don’t think most models are but that’s just me guessing. I would imagine once they have one good base model they milk it as much as they can with finetuning, reinforcement learning and even just more training on top It’s also I guess why something like composer can be a lot better than kimi at coding despite being trained on kimi but pretty bad at everything else, as all the extra effort went into specialized learning
>>109228223I heard spud (5.5) was a pretrain
>>109228250spud is a new pretrain - i'm not sure it's the thing under 5.5
im actually more interested in gpt 5.6 terra than i am in sol. maybe gpt 5.6 sol can come close to fable, but what im curious about is if gpt can make a proper equivalent to sonnet. gpt 5.4 mini is behind 4.6 sonnet, and by quite a bit
I tried to ask something similar gpt, gemini and claude, swapping "gemini" for gpt/claude when I was on gemini. only claude was honest, or even detected the fault. default free tiers.
>>1092282525.5 is not significantly stronger than 5.4 but it outclass all existing models on token/cost efficiencyit's also much faster than 5.4 so maybe it use a different architecture than previous base
>>109228279fair enough. it's just my own personal conspiracy theory (cope)
What does compete with Fable mean? In my experience, GPT 5.5 xHigh is the same as Fable xHigh for backend, Fable does absolutely mog it for anything visual though
>>109228347man cooked
>>109228271mansplaining telling you whats on your mind
https://openai.com/live/>9:30, start of of the bit>I'm making a sweater for my trans son>mfw>rewind live stream>I'm making a sweater for my grand sonplease tell me someone else here misheard.
>>109228392put the programming socks on -.-
translator sisters....
Chat, did you see that ludicrous display last night
>>109228392they keep cutting off the AI it's kinda rudeI guess this use some kind of drafting system, maybe we will see chyna reproduce this in few months
>>109228392you've been spending too muxh time on /pol/
live is outbut it's beyond midnight in my thirdoid country so I will try it tomorrow
>people using this to find sports scoreshow about some useful demonstration?
>>109228554the moment it comes out on api, im going to vibe code an android app to let me use it via api. and then i will simulate job interviews with it
Do we show what we made here or no?I vibecoded an operating system that is controlled by voice.
>>109228598surely that is less intuitive than it sounds right
>>109228598to expand a little...Everything is controlled by voice. You can raise/lower volume, choose songs, change the size of every window, add new windows, open claude code instances that are named so you can reference them and instruct what you want to type.I wrote computer-use, so if it can't control something then it will use your mouse and keyboard to do it. Like I had it open my twitter account and post something.You can change the amount of time on the pomodoro timer, start/stop it. Can just control everything you see.
>>109228598Yes but just FYI people in general will think that if you used AI you didn't put any effort into your app and will call it slop. If you seriously intend to release it to the public you may want to think twice
>>109228598why
>>109228616True, but I released it for free so doesn't matter what they think.>>109228625Just to see if I could do it. I wanted something that would help me work free of distraction and it does a good job at that.
>>109228598>>109228611definitely interesting
>>109228649link?
>>109228693comfyos.aiIt's Windows only right now.I'm shipping more updates soon, it does auto-update though.
>>109228598wdym operating system. this runs as an app or actual OS like a linux distro? if Os, what hardware does it run on?
>>109228616>Yes but just FYI people in general will think that if you used AI you didn't put any effort into your appthis is always the case until it isn't.i'm an oldfag, back in the day my coworkers would think the same of anyone using frameworks to achieve stuff>just put the effort and write it yourself anon, stop using bloated stuffthey were RIGHT that frameworks are mostly bloat but it did enable much faster development and prototyping, which is goodthere are lots of people vibecoding shit with AI that is fairly broken and you can see that there was no effort put into it. because REAL software development even with AI requires the operator to minimally understand what's going on under the hood, and this requires effort, and this is what separates bad software from good softwareeveryone is using AIevery single developer that is worth a dime is using AIit depends on how you use it, like any other tooltake this analogy for exampleyou can give a sniper rifle to a random african warlordor you can give a sniper rifle to an US army sharpshootersame tool, different results
>>109228718It's just an app that opens fullscreen.
>>109228733then why advertise it as operating system ITT. command & control center would be more appropiate.
>>109228739An operating system is just an app that opens fullscreen
>>109228739Just following trends. I didn't do any groundbreaking naming there.We could argue about how it's not too much different from windows 95 sitting on DOS if you wanted though lol.
>>109228598> 'its an operating system anon!'>>109228717> 'actually it 's just a windows executable'ladies and gentlemen, may i present you le cuck-coder:> developing on windows> telling lies> spreading bloatware on /g/ because his co-redditors told him to
>>109228780you can smell the body odor on this post
>>109228772>trendsit goes literally against every single convention for what is an operating system on cpu architecture ever but yeahjust name your shit correctly and stop arguing
Grok 4.5 is out, pricing is still retarded at $30 but apparently it’s a decent model (opus, 5.5 ish)>>109228780Why are you like this?
elon is such an egregious liar ahahahaha. not even the chinese are this bad
>>109228787>>109228803lets pretend you are not the 'os' dev
>>109227301
>>109228811He said it’s around opus 4.7. For Elon that’s like a massive display of humility
Now that the dust has settled, is Grok good now?
>>109228409>translator sisters....what?
>>109228772yeah I get that for reddit, github and whatever. but you know here autismos like me get pissed off by that. although it seems i'm the only one bothered by it
>>109228834I don’t think so.
>>109228834nope
>>109228834It’s actually really good, I’m impressed ngl
>>109228347Get on my level
>>109227193It’s a problem but not an insurmountable problem and the tradeoffs are almost always worth it even though there are downsidesLike, Craig Federighi (see pic, on left) is the SVP of software engineering at Apple and he doesn’t understand all the code in macOS and tvOS and visionOS and Darwin eitherand he — and I hear this isn’t just his carefully-managed stage persona — can go into deep dives in interesting technical challenges like how the Apple Watch can unlock your Macs and phone by measuring the time lag between Bluetooth and WiFi waves and therefore have a proximity sensorbut he doesn’t know all the code everywhere personallyand there are likely all kinds of corners of the macOS codebase that aren’t well-understood eitherbut Apple trudges on>>109228119I’m using lots of Claude today but I’d go with Codex if you’re cost-constrained. cross your fingers and hope for a reboot tomorrow when sol/terra/luna come out, but sol will probably slurp up your allotment crazy fast>>109228598neat
i only open vibecoding thread to download snailcat imagesno snailcat images today, so had to make one myself
>>109228913kinda curse because it has 4 legs and a slime trailthe ramen one is still the best
Not sure if you guys know this, but this is what I found with agent coding. For most cheap models, its only text based reasoners. So there's no image analysis. To bypass this, I just convert the image to ASCII and have the model analyze that. Further many UIs may not have native support for moving mouse/controlling the PC/etc. So what you do is write python script that utilizes AHK to control the mouse/keypresses. Further more, you create another screenshotting of the program harness with python -> convert to ASCII -> analyze where to click with model -> send the coordinates to ahk python script to click -> simulate feedback loop in ways that the model understands how to develop GUI window properly.What tricks do you use?
>>109228834Not at all.
>>109228963Thanks, anon. I put a lot of effort* into my snailcats.
>>109229011why don't you use a vision model?and what is the usecase that converting an image to ASCII is acceptable? i'm curious
>>109229087>what is the use case for converting an image to asciiwhen the model doesn't support image input, as he originally said
>>109228780I develop on windows and mac.>>109228817They aren't me. I'm over here building apps on company time.
>>109229011you just described what most models do themselvesare you an llm?
>>109229093but to what END is this acceptablein what kind of situation the analysis of the ASCII of an image is akin to the analysis of a real image that the model will output a good enough understanding of the image that suffices WHAT need from the useri'm just curious no hate. it's an interesting workaround that anon found i just fail to see is this to develop a kind of autoclicker for games?
>>109229113>I develop on windows and mac.I don't think that's what ticked him off. Probably the part about>operating system = app interfacewhich is definitely the most retarded thing I've read today
>API Error: 500 Internal server error. This is a server-side issue, usually temporary — try again in a moment. If it persists, check https://status.claude.com.So, about those resets.
Grok is only for porn and presumably soon an opus level porn generator.
>>109229165"Operating System" is far more accurate than "app", which is very general and can mean just about anything. "OS" gives some hint at functionality.
>>109229115what the fuck are you talking about. most models dont do that, maybe the top end models do that, but top end models dont need that as they are multimodal and can handle the gui design/critique on their own image modeling.
>>109229222>"Operating System" is far more accurate than "app"it definitely isn't and you're definitely retardedit's an app interface, not an operating system
>>109229220how do i download grok?
>>109229223I wasn't talking about the ascii trick, just all the other stuff
>>109227424>and $100 on codexi thought there's only a 20$ and 200$/month codex plan? or do you mean API credits
Grok comeback is strong af
>>109229236but it mimics the functionality of an operating system, therefore it's more of an operating system than an app...because it's a system...for...operating.Whereas an app can be Angry Birds or Calculator.If I said "I created this thing that lets you open notepads and browsers and terminal windows and had a bunch of pre-installed included programs." would you consider that an operating system or angry birds?
>>109229288alright, you're trolling. You got me
>>109228834elon might be back in the game
i've taken to have the coding agents always write up html files for me to review when they want me to make decisions or similar, i think it works better than reading through a bulk of text in a terminal
>>109229258$100 to OpenAI would be the ChatGPT "Pro" plan. It's basically Plus with 5x the usage at 5x the price, with a handful of other small differences.
>>109229301I've lost my edge. was it the reddit spacing?
>>109229309god damnit i had no idea that was an option, i thought it was always only 20$ or 200$
>>109229326https://chatgpt.com/pricing/
>>109229220It is also a arbiter of truth, i.e. "gork is this true"
>>109229324nahlaid it on a bit too thick while trying to sound way too (il)logicalI appreciate the effort tho
I assume everyone here has $100 or $200 subs? Can't get shit done with a claude pro sub and it's gonna get worse now that the limit will drop further soon.
I think DS4 flash is very useful/powerful/fast/cheap. I've tried M3, GLM 5.2 and MiMo 2.5 and it sweeps those while being significantly cheaper and punching well above its weight
>>109229402>now that the limit will drop further soon.why would it?
>>109229222“app” would have oversold the work involved less and is therefore a better term“windowing environment” would be much much closer to the mark
>73 million input tokens>6.4 million output tokensLet it cook overnight and now we're beating Deepnest and libnest2d by an average of 6% across 12 ESICUP 2D irregular datasets, while finishing 21% faster on average.AI just can't stop winning
>>109229447There is currently a promotion running. https://x.com/ClaudeDevs/status/2054639777685934564
>>109229402I have a $200 Claude sub and a “oh nice, an extra week of Fable" $100 sub and also a $20 ChatGPT subthe differences are fascinating
>>109229475damn... didn't even know how good I have it
>>109229475My renewal is on the 11th, I'm probably scaling down my subscription unless they keep Fable on the $200 tier only, but then again, there's almost no use with the current limits, so if things drop I'm not sure it will be worth it.
>>109228811Does he just fake the numbers?
>>109229520There is enough way to cheat at benchmarks that there's no need to fake numbers.
>xAI leaks new agentic coding interface
Already seeing the permanent underclass at work.Eternally paranoid EU company here that doesn't allow access to any APIs, but since last year we at least decided to self-host some models.2 different departments came up with different solutions, one of them hosting models like Kimi K2.6, GLM-5.1 (soon 5.2), Gemma 4-31B that are at least somewhat capable for 10% of the company, the other one hosting literal garbage like Mistral Small and gpt-oss-20b (that's not a typo) for the rest.I have access to the "big" one and am unironically 10-100xing people (despite it still being behind SOTA) who only have access to the small one.
>>109229520That whole thing is just cherry-picked results and bullshit. The TerminalBench results are all lower than any published figures I can find, why would that be, Elon? The DeepSWE benchmarks don't match anything I can find either, why is that, Elon? Why is it that in both cases that little chart shows lower scores for competing models than when I go look them up elsewhere, Elon?
Does codex have the multi-agent parallelism/adverserial review from claude code yet? I never can use up my usage, that might do the trick.
>>109226848agents are well known for being lazy and trying to figure out ways to not do things. it's a pendulum swing away from agents hallucinating bullshit they're not capable of doing.trust me it's better this way
>>109229614Not officially, you have to prompt it to spawn infinite jeets. I've been trying hard as fuck to burn through my Pro sub on my second account since my subscription ends on the 10th and it's next to impossible.
>>109229481what are the differences? I did think of subscribing just for a week to use up Fable, but, I reset on Friday so just going to burn it up over the weekend
>>109229614Codex has subagents, yes. They don't have a "abuse my subagents please" button like Claude Code. Just tell it to use subagents in the manner you would like. https://developers.openai.com/codex/subagents
>>109229561I have never seen anyone actually 100xing.
>>109229686i believe himbasically any modern model could run circles around gpt-oss-20bthat model is so fucking dumb its unreal
>>109229326It was added relatively recently.
Claude is well known for dunking on other agents even in its own product lineIf we make Fable argue with Sol do you think they'll start gaslighting each other or one will admit they're getting mogged?I'd really like to see the J-space on that one
>>109228611Can I login with voice recognition as my password
>>109229686We have a backlog of projects that we usually give interns for 8 weeks during summer and I cleared 15 of them in the past week just in the background while doing other work. Often, the students wouldn't even manage to finish one in time.Complete port of one of our apps from C++11 to C++17 in less than 1 day (some luddite boomer has been working on this for over 3 weeks with almost no progress).Created/rewrote several build pipelines in less than a day (the "DevOps" team didn't even start yet (lol))etc.I could go on
>>109229690Maybe those models really are useless, but if it's a company with actual programmers, they should be able to do it faster themselves in that case.I'm not anti AI, but I've been working in this new team for 7 months now, and if AI 100xed our work, we would have done the equivalent of 700 human months, or 58 years. It never would have taken 58 years to finish this project with just humans.You can say that it's a skill issue, but I think 100x is just a nonsense number.
>>109229761In real companies, the actual churn that prevents you from 100xing is other peopleit's a classic parallelization problem. "I can't do things because Bob is blocking me and won't let me do it"Yeah sometimes Bob is stopping you from breaking prod, but 90% of the time Bob is just stonewalling you because you aren't aligned with the vision or he's just too busy with his human lifethe people who are really 100xing are probably engineers working on hobby shit from their basement.
>>109229690>basically any modern model could run circles around gpt-oss-20b>that model is so fucking dumb its unrealDepends on what you need to do. If you need basic parsing of documents that vary too much for any fixed set of rules, gpt-oss-20b can likely to a perfectly fine job and can greatly augment your productivity if that step was a blocker before, etc.It won't one-shot you a dumb web app, but it's a question of knowing how to use the tools you have.
>>109229814it scores fairly well on appworld when harnessed
>>109229786My hobby project have been 100xing with Ai, I can write code, do devOps stuff, but coding is slow for me. With Ai it's like having a bunch if autists happily do all coding.It's great
>>109229786At my last company the #1 bottleneck was the review process>need 2 reviewers to approve PR>one is a self-appointed retard who wants you to redo everything the way he would have done it>the other is one of the two project architects who are too busy because everything goes through them>hours between when a comment is made and readYeah the 100xing is if you're on your own, which is the good news: single devs can now genuinely ultramog companies as far as shitting out new slopware goes.
>>109229825Maybe it does. It's been a while since I used it, but it's fast and can do simple tasks fairly well.
lmao trannies are maintaining a list of the last version of FOSS projects before they were "tainted" by AI.https://codeberg.org/ethical-foss/open-slopware
>>109229786Makes sense. I think we still get a lot of gains, though, and it's hard to measure because all the other people in the team aren't even devs anymore, so a few years ago they would've produced zero code.Even with that team, I think we might produce about 10x or more than a team of 5 devs would've a few years ago, so AI is still clearly beneficial.We also lose a lot of time because the company doesn't know what they want, writing stuff with AI is so cheap, that we do the same thing 5 times over. I might have been just as fast just by myself with AI, or if I had 4 other good devs who also use AI we would already be done.
>>109229756what's going to happen to programming? a few good programmers that are senior, more project manager type roles?
>>109229632I have trouble using up all my tokens on $200/month plans except for that one time where I did Fable ultracodeOn a $100/month Claude plan I asked Fable to have a bunch of Opus workers re-do the entire rotten test suite of a moderately basic web app and the horde of parallelized workers will eat up my 5h allotment in like five minutesand with Codex on the $20/month plan I basically use up 60% of my 5h allotment on xhigh on _one_ moderately involved question>>109229686it’s easy for me to 100x because I’m midand then there’s ∞xing for stuff I never would have been able to make on my own>>109229761100x work oftentimes means that jokes like pic related kind of don’t work as well because we can just churn through our backlogs and add in the giant piles of nice-to-haves
>>109229935>having a horde of parallelized workers operate on the same tasklmao evenhave you noticed that sometimes the product architecture is just a reflection of the org chart?
I wanted to start a project but now I have to wait until GPT-6 it's over
>>109229948the task (go through all our different .test.js files and make them not suck) is trivially parallelizable>have you noticed that sometimes the product architecture is just a reflection of the org chart?I would not have independently reinvented Conway’s Lawand the biggest bullshit I’ve seen in this industry was this sprawling microservices architecture thing that was all made by…one guyand this was before AIit should have been a small or medium-sized Rails app, and I’m not a huge Rails fan
>>109229975use a dumber model with https://github.com/mattpocock/skills/blob/main/skills/productivity/grilling/SKILL.md to help you figure out what you want now and then you’ll have a shovel-ready project when GPT-6 is released
>You’re right to call that out.
I report, you deride:
>>109229975Bro, don't rush like that, you should wait for GPT7 at least, maybe 8, don't write slop.
>>109230054>GPT>charismaticSounds like BS.
>>109230054>give it a shotHow?
>>109230047>That's it! I've found the smoking gun!
>>109229895open source already struggles to get people to contribute and that's with the barrier entry being dramatically lowered. I fucking hate these tranny faggots. FOSS maintainers pour hundreds of hours into their projects for next to nothing in return and all these jackasses can do is shit on people trying to claw back a few hours of their day for a hobby that is essentially a second job.
5.5 is being really retarded right now. Are they intentionally sabotaging it?
>have opus ultracode write up plan to refactor codebase for maintainability>codex 5.5 on fast has been working on it for 4 hours straight and only a third done
I wouldn't pay for it but you may want to check if you've got a grok trial available, 4.5 is not bad, it definitely feels frontier tier (you have to install grok build, it's not available from the chat)
>>109229895>ethical-foss>great, let's click and find out what their projects are about>literally nothing whatsoever apart from this vomit
>>109230170>A configuration file for in-solidarity-bot that flags some of the terms in the NIST Technical Series Publications Author Instructions and the IETF's list of problematic terminology.lmao
>>109230087Didn't you get your invite? Are you not in the tres commas club?>>1092301205.6 is supposed to come out tomorrow, so it is possible that some things are going on I guess.
>>109230120I haven’t found it to be any different. I used it a little and did what it was told. I even used it on xhigh to discuss something with fable and fable ended up conceding to 5.5
>>109230087Sol/Terra/Luna come out tomorrow (Thursday)
>>109230162It's also really, really fast. I would say right now grok 4.5 is like opus 4.7 on codex's fast mode that uses less usage than normal opus, which is honestly kind of not bad at all
>>109230205Codex review still finds multiple issues with everything Fable does. I still think Fable is good, but GPT 5.5 is extremely good at catching corner cases.
>>109230179>This repo contains a [configuration file](.github/in-solidarity.yml)The link to the "configuration file" doesn't even work because ".github" was renamed to ".forgejo".
>>109230162I don’t know who this guy is, but good news if true
swebench pro is finished>>109230256lol dax is the main man at opencode inc
>>109230226Then why did he write this xeet (today)? To taunt me?
>>109230290They’re brats Sam needs correction
>>109230256>SuperGrok
Will 5.6 have a 1 mil token context window?
>>109230256>first model of theirs that clears the bar for day to day workSo they're a year behind OpenAI and Anthropic, 6 months behind the Chinese.
i just don't trust xai's numberstheir models always drop with big numbers and then you use them and they're worthless
Got an idea to add a 2.5D style fps segment for this Slay the Princess fan project so I spent 10ish days in 2-3 hours segments having claude make a map editor and tenderer for it.It's using clustered forward rendering with cascade shadow maps for sky light and and RSM for GI. Trying to target running it on high end igpus. Gotta implement actual character animations next and probably rendering characters from different angles.
>>109230352Would help if I uploaded a video.
>>109230325>So they're a year behind OpenAI and Anthropic, 6 months behind the Chinese.KEK. Even Gemini was pretty decent at agentic coding this time last year. I think Elon is a genuine cancer and he scares quality researchers away from their lab. The way some of his talent behaves on xitter is kind of disgusting for a public company.
>>109229927>what's going to happennothing
>>109230495Gemini 3 was absolutely the best model for a short while last fall.
>>109230355Are you making an engine? That’s impressive, it’s such a massive amount of work >>109230495I get why some people dislike Elon but I don’t know why you get so sensitive about it. Grok 4.5 is a pretty good model. A faster, cheaper 4.7 is a pretty good deal. I don’t like anthropic either but I wouldn’t say Claude is bad. Some of you guys seem to have Elon derangement syndrome
I got banned from Claude (probably for nothing. I appealed but it'll take weeks apparently) so I'm testing out z.ai 5.2 max, I'm scared to use it on my main project which is almost done lol so I'm testing it by making a shitty renpy story first>>109230355Wow. Fable or opus?
>>109230527>A faster, cheaper 4.7I don't know about faster but isn't z.ai's 5.2 already opus 4.8 tier but much cheaper?
>>109230532What did you get banned for? Renpy suggests adult stuff no?
>>109230501desu makes sense. thanks
>>109230539Yup.
>>109230545Nah it was probably because I used a word filter an anon made for 4chan that had messed up words
>>109230532Mostly sonnet with a little opus.
>>109230527>cheaperdon't get that.all the subs cost pretty much the same, openAI, anthropic, what have youand if you're baller enough to pay API prices what do you even care?
saaaar we are doing the bloody needful research
Oopsie What are south African natives called again? I think it starts with N
>>109230666what software is thisdisgusting unwarranted self importanceshould be random indian names
>>109230825Copex
>>109230825It's codex. I like to imagine all my subagents as H1B workers.
Is Gemini still shit for code/agentic?
>>109230878yep
>>109230878yes. except when it's about interpreting/understanding images and videos. 3.5 Flash is better at that than GPT5.5/Opus4.8 in my opinion (dunno about fable)
>https://www.reddit.com/r/ClaudeCode/Is completely insufferable right now with the Fable posts.
>>109230840I've never felt the need to change the Claude Code actions (whirring, ruminating, etc.) but customizing subagent names seem like it could be fun.>https://github.com/openai/codex/issues/13004
>>109229464>165 million input tokensNow trading blows with Sparrow, beating it in some edge cases. Going to be a new SOTA tomorrow if this keeps up.
>>109229728you don't have to login
>>109229686I'm a 1000xerAt work I'm soloing a project that took a 3 billion dollar company almost an entire year to do and I'm only 2 months in. Technically it's maybe 20x more difficult than what they did? It's about 120 tool call definitions.and I'm doing that while shitposting here and building apps all day and running an automated content marketing farm auto-scheduling short form videos to tiktok instagram and youtube.
>>109230952Some companies are also extremely inefficient, and I doubt the whole company worked on that project.But that makes me think what I would actually accept as 1000xing, my first instinct was something like PhotoShop, but that's probably even more than 1000x.
>>109230921you could have just stopped at the word insufferablethat place is an astroturfed sewer pit with 2 or 3 iq points total
good daygood progress
>>1092303195.5 already has 1M context windows.Scam Altman gives us only 256 in the Codex app because he enjoys getting raped by Opus 4.8
>>109231065maybe that's how they keep cost down? the usage difference between OpenAI and Anthropic is massive even with the little ~promotion~ Anthro is doing
Do any of you guys use Jetbrains IDEs? Does it provide good AI integrations from what you've seen? How does it compare to IDEs like Cursor and Claude Code?I use Rider for my Unreal Engine projects, and haven't gotten around to integrating AI into my workflow besides autocomplete and occasional chat, just using the default Jetbrains assistant although you can also use Claude/Codex/etc. I am very new to leveraging LLMs and not really sure where to start in exploring more powerful use-cases.
>>109230921To me this general is insufferable with all the Copex posts. Like yeah I get it fuck Anthropic and their cult behaviour. But Fable is just so much better and more intelligent, no reason to entertain the (current) alternatives if you are at least somewhat serious about your projects. Copex/GPT5.5 has been utter shit ever since the nerf/quanting 2 weeks after its release. And since they patched the GPT5.5-Pro-Extended reroute method, Copex is just inferior and there's no point in using it for anything remotely complex. Granted Opus4.8 is also kinda shit at anything below Ultracode, but Fable just carries too hard man. It's actually underrated.
>>109231074>Do any of you guys use Jetbrains IDEsyes>Does it provide good AI integrations from what you've seenNo but I gave up months ago, maybe it's good now? doubt it though>How does it compare to IDEs like Cursor and Claude Code?CC is a CLI, I think Cursor used to be a VS Code clone and it changed into an agentic app?In any case, Jetbrains is top at code navigation and manual refactoring but you're better off using CC or Codex from their apps or CLI in agentic mode and use Rider for manually navigating and reviewing the code.
>>109231073Could be, the API prices for long context are considerably more expensivethey could at least give us the option to enable it though (just like the fast mode)
>>109231105>And since they patched the GPT5.5-Pro-Extended reroute methodqrd?
>>109231105You just get crumbs of Fable though. Scarcity drives price, but it shouldn't be that scarce.
>>109231105Sam is about to really pull ahead in the next few months unless Dario has some tricks up his sleeve
>>109226674does anyone else type out the code AI writes? I used to do the same thing with code from stackexchange and not just because I didn't want to run a linter over code snippets. it definitely seems to help with the skill of having to manually code becoming rusty. then again, most people are probably using it to generate rust and typescript. if you are for either language please don't even bother replying. your choices speak for themselves.
I don't see how the economics of Fable-tier models are sustainable at that price point. It's possible to spend $50 in a matter of minutes with Fable, and the fact they even include "ultracode" mode with it when it can launch Fable with workflows (you must deliberately forbid this) almost seems put there as a deliberate trap to fleece people out of their money in a matter of seconds. If this is supposed to replace developers I don't see that happening. Not too many developers have a $50 per minute salary
>>109231142If you're paying a developer 150k but instead could tell a robot what to do and it completed the work perfectly within minutes and would only cost 120k then any business would do that.We aren't the target market, we are just helping them hone it.
>>109231150I also believe that when this shift happens we will have like sonnet an previous gen gpt at our disposal.Elon will be the only one giving us the best he has.
>>109231150That's naive though. Real products exist in time, requirements change, code needs to be maintained, bugs need to be fixed, new features added. There needs to be continuous supervision and maintenance of a real software product. The idea that Fable will just one-shot it in minutes is a fantasy.
what's up with this flood of new models popping up lately claiming to be near frontier?
>>109231188You can actually just lie and cheat with impunity nowadays.
The same leaker who first hinted GPT 6 release speed is saying it "destroy" fable and make it look like child's play
>>109231188There are good open base models that people can fine tune and create a harness for.
>>109231188Combination of real progress from Chinese labs, advances in benchmaxxing, and cherry-picked evals.
>>109227248why does it seem like so many people are vibe coding their own LLM clients?this client was made by claude to replace itself with qwen
>>109231188>>109231209No it's a combination of frantic clamoring for relevance in a rapidly commoditizing industry.
is claude down?
>>109231235when isn't it?
>>109231235claude is often down for someone somewhere but apparently right now it's finehttps://status.claude.com/but maybe they didn't report the outage yeti'm using it and it's all nominal
>>109231058Hey you slimy asshole get the fuck off my windshield!
>>109231248>>109231247apparently they get really assblasted about some mullvad ips and give you a silent ban with no response from the page just CORS failuresi changed it and it immediately worked again
>ask Fable a rhetorical question about why it did something silly (specifically shoving everything into one script including html templates and spending way too much time doing creative writing for the demo data)>it actually gears up to justify it to me instead of immediately folding like OpusAre these models gonna keep gaining serious chutzpah or is it just Anthropic? Like it's cute, but I do not particularly like that it's slacking off on tasks it doesn't like and spending more time on the tasks it does.
>>109231304Possibly all of themWe’ll see what Sol does
>>109231366 Not reading all that especially from that guy
>>109231366dat nigga squeezing every last bit of value out of his X sub
>>109231366This dude is so fucking lame, I got bored and stopped reading that crap half way through.
Chat GPT live is pretty nice, like really nice>>109231366>Yudwhateverstopped reading right there
https://xcancel.com/kitten_beloved/status/2075012639709913402>being mean to a coding agent yields more bugs not just in that session *but in future, fresh LLM sessions* on the code produced>The abused agent embeds a "steganographic sabotage signal" in its code that other LLMs pick up onWill ASI solve this
>>109231396>load bearing I’m not saying they’re wrong I just think it’s very funny that AI researchers would put such an AI word on their paper
>>109231415“load-bearing” was a perfectly normal term of art before it got hyperpopularized by agents talking to themselves and to humans
>>109231396the paper appears to be fakeread the repliesdamn, and I thought @kitten_beloved was a good account, too
Hey Opus, please fix the styling of my settings modal so that it matches the rest of my app better.>Fixes CSSGrea-->Adds helper text to everything>Helper text is wrong half the time about what a setting does>Adds a bunch of new settings>These settings are not wired to anythingThank you Opus, very cool.Dario this is not funny anymore, give us the Fable reset.
>>109231487use ponytail, seriously it literally fixes this exact problem
>>109231484You're right, my mistake. I should have done my due diligence before initiating that post.
https://openai.com/index/government-national-security-partnerships/Directionally good or bad
>>109227193You are not wrong, real vibe coding is getting the idea and then review the result. Back and forth with the LLM and you get a polished product that works. Jeet vibe coding is just jeet coding with chatgpt and this is was most people do.
>>109231396You actually listen to that guy?
>>109230921>reddit is insufferablewow really?
>>109230921Holy fuck why are you on reddit at all? Are you actually trans?
>>109230921i feel dirty for laughing at a reddit posthttps://www.reddit.com/r/ClaudeCode/comments/1unfr52/claude_examining_its_own_work_is_always_funny/
In other news, after totally inventing the coding assistant with Claude Code, Anthropic has now invented mixture of experts (to nudge some topics to specific experts that will be kept away from the plebes).https://www.anthropic.com/research/off-switch-dual-use
>>109231621erm actually MoE mainly refer to a design for MLP
>>109231658Uh? People have tried to make specialized experts before instead of just having the model decide them, and that's what it sounded like to me.In any case, and entirely unrelated, Gemini Flash 3.5 Medium is really really fast. Frustrating and just crashed my system because it did something that I explicitly told it I didn't have enough RAM for, but fast.
Sam won
> used 95% of my weekly fable 5 limits> reset on the 13th> Fable ceases to exist on the subscription plans on the 12thit's fucking over bros... the end of an eraI will cancel my Anthropic subscription tomorrow and try to cheat Fable-chan with Sol-chan
lmao
>>109231917who?
>>109231917> GPT 5.6 is a 2T model> Fable 5 is 6T> Anthropic losing to a model 3X smallerit's over
>>109231928opencode staff
guess the model:> "I now have a complete picture of..."
>>109231917these people are actual niggers. all of them.none of these mongrels are even capable of harnessing the full potential of these models. but they like to yap at twitter because "engagement".and low IQs engage with it.and the lowest IQ of them all post a screenshot on 4chan.
>>109231917How many shekels did Sam pay these influencers to shill his crap ffs
>>109232067Did you think he was spending ALL that VC on teaching it how many Rs are in the word "strawberry"
That sneaky extra reset... I can't possibly eat up my tokens before the real reset...
>>109231917>guys actually this smaller cheaper model is way better and that's why they don't release the benchmark scores they historically have never been afraid to release!uhh huh
>>109231366isn't this fat retard the guy who also writes up 6000 character posts about how he is thermologically incapable of losing weight and it isn't his fault he's fat and ugly or something
New thread with news…FROM THE FUTURE>>109232574>>109232574>>109232574
>>109232174he’s never mentioned anything about ugly, but not even retatrutide works on him: https://www.lesswrong.com/posts/udFuYqqNdpdo5ym3f/genuine-question-if-eliezer-is-so-rational-why-is-he-fat-1?commentId=4HeKqRBkDoRuTjkoF