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/## News- (2026-08-19, 11:59 PM America/Los_Angeles) Claude’s 2× promotion set to end- (2026-07-24) Claude Opus 5 out## Related generals>>>/g/lmg/>>>/bant/agdg/ — schizo-resistant temporary (?) hideout>>>/vg/agdg/----## 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>>109552932
polygon free gaussian native engine what do we think
OP is a troll and a faggot.-- Frontier models - start here if you have $20 or sohttps://claude.com/product/claude-codehttps://developers.openai.com/codex/cli-- B-tierhttps://x.ai/clihttps://platform.deepseek.com------ Prompting / context / skillshttps://arps18.github.io/posts/claude-code-mastery/https://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/https://github.com/mattpocock/skills — /grilling is a favorite-- Other editors / terminal agents / coding agentshttps://pi.dev/https://opencode.ai/https://cursor.com/docshttps://osaurus.ai/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/2.0https://artificialanalysis.ai/-- What we’ve donehttps://vcg.gitgud.site-- Previous thread>>109552932
>August 14th - Z.ai drops GLM-5.3Same base as 5.2, massive post-training gains. Claims strongest open coding model (huge jumps on Terminal-Bench 3.0, DeepSWE, etc.), agentic performance nearing Fable 5 levels in places, and standout cyber capabilities (top scores on CyberGym, already finding real vulns). Live via Coding Plan/ZCode. Full open weights delayed ~2 weeks for safety hardening due to the cyber strength.>August 14th - Qwen3.8-27B (and Max) open weights finally shipAfter the delay/confusion fiasco, Alibaba officially drops Qwen3.8-27B (27B multimodal dense, Apache 2.0, 262K native / 1M extendable, strong coding/office/agent claims, outperforms prior Plus). Max-level 2.4T-A95B weights also confirmed out. Both on Hugging Face + ModelScope. Local runners rejoice.>August 13th - DeepSeek V4-Pro-0813 leaves previewOfficial flagship out (MIT). Strong in cyber/agent spots per vendor numbers, but independent benches (Intelligence Index ~53, mid-pack on others) underwhelmed relative to hype + concurrent price hikes. Mixed-to-disappointed reception.>August 12th - SpaceXAI Grok 4.6Post-training upgrade. Ties GPT-5.6 Sol on Artificial Analysis Intelligence Index (61). Same $2/$6 pricing. Long-running agents + coding focus.>August 11th - NVIDIA Nemotron 3.5 Lightning (open)30B MoE / 3B active for fast high-volume agent steps. Up to 4x speed + Switchyard router.>August 10th - Meta Muse Glimmer (open)30B Apache 2.0 local single-GPU agentic model + Zuckerberg open-source essay.>Earlier in windowClaude Opus 5, Kimi K3 full weights, DeepSeek Flash production, OpenAI Luna/Terra cuts still relevant.No major new closed frontier flagships from the US labs in the last day beyond prior gated/cyber and Flash-tier stuff.
>>109561420any tips on getting a free api key to a near or near frontier model with decent rate limits?
my bank probably thinks I'm a gooner since I have a bunch of grok charges.
newfag here and i am a bit confused about context length.i tried the opencode deepseek flash v4 free with only 200k context and it works pretty well for my project even after it compacts. i'm trying to keep working packages small and using subagents as much as possible.then i tried the actual deepseek api with 1M context length and i noticed that it heavily degrades after ~500k. it starts making more mistakes, the steps it takes are not very efficient anymore and it often gets into a state where it spends a shitton of tokens trying to debug some mistake.is that behavior expected? does more context length not mean "more consistent"? should i force it to compact after ~30%?
>>109561599compacting about there is the default for a reason. but isd about deepseek.
I like to think of vibecoding as akin to buying bug spray. You didn't invent the bug spray, you just need to get rid of bugs. Sometimes the bug is the whole project not existing.
>>109561599Yes, it's normal that they degrade. Imo Fable gives you a bit more headroom, I think it gets worse at 60 to 70%But even so, if you compact at 500k you can still do longer tasks, it matters most when you run autonomous loops, for normal prompting 200k is fine.
>>109561599this >>1095616221m context has always been a bit of a meme
>>109561622I've never had a session take less than 300k, but maybe my projects are unusually complex or bloated.
>>109561420ai working hard again
another amazing 4chan exclusiveworld first? animated gaussian splat overwatch smut (no polygons were used during this render). gotta figure out the smudgy textures a bit, hair collision needs to be added.
>>109561622>>109561612>>109561637yeah i noticed 700k/1M is the point where i really go "what the fuck are you even doing??"good to know it's not my imagination. initially i thought you should avoid compacting as long as possible with 1M context, but will make sure to do it more regularly now, thanks
>>109561791lel they were saying you can do 1m with gemini 2.5 and it would turn into absolute gibberish by the end. the 'plays pokemon' people had to limit context length to stop it from going haywire.these days things are better, but up until very recently you could definitely feel the model getting dumber even over 250k windows (they still do a little)
Gemini 3.7 still not available for Europoors
>>109561860Does it not work for you in AI Studio?
>>109561899nonot on the app and not on the website
>>109561860i've got it in bongistan
>>109561947That only happens if you're on the radical right lists.
>>109561420Holy shit, 5.6 Luna is an absolute joy. Such a breath of fresh air after having to deal with Opuss all day.
>>109562002She's such a sweetheart. Realistically gets everything done I need anyway and so fast. After they gave Opus brain damage, it's nice to have a model that actually answers in plain text.I only use Sol if I get a nasty bug. Not even to save money. But I appreciate the straightforward way that Luna writes and debugs code. It's like how I would do it. Opus and Sol to a certain extent overengineer the crap out of everything.
>>109561765oh no think of the heckin' blue board anon senpai
>>109561941Maybe check the country of your Google account. If you're a russian expat or something, that could be why.
should I finally upgrade claude client for locals, bros? its currently frozen to the version just before it inspect local system if its chink locale
>>109562024Sol is very good too, it's the model I've been using for all my personal stuff, but I have a lot of non-vibecode projects where it just feels like overkill. Luna is perfect for these cases. I've also never ran into issues with rates using Sol, so with me Luna could probably go for a whole week without stopping.Makes me re-consider smaller models, and the whole "benchslop" angle. This is extremely capable and coherent within narrower frames.
This weekend I will only use dsh and v4pro, like a deranged lunatic. We must break through the firmament imposed by lesser harnesses
How is the new DS pro?
>>109562168Apparently pretty bad, and with a x5 price hike as well. DeepSeek has fallen.
>>109562197>he think cheap chink ai will last foreverHAHAHAHAHAHA
>tfw GPT5.6 refuses to add a thinking trace to my custom harness running Qwen3.5 because it would "expose private model data"had to go to Opus for that one
>more stories that didn't happen @ 11
>>109562295
Tibo you fucking piece you shit you better reset codex this weekend fuck you
Now i need tibo reset
>Can't wrangle his context properly>Blames modelBiggest skill issue ITT.
deepseek still feels dirt cheap to me. what did the price even go up by, like $0.001 to $0.002?
Less than a day left until chinks rape us with expensive token prices...what the fuck are we supposed to do?Nobody here can afford to buy any other tokens and even if they could they aren't actually retarded enough to burn the money on overpriced trash like claude when it barely performs 6% better across the board from DS..so what the fuck do we do now?
>>109562161I see you.
Working on my token efficiency again. Apparently the next prompt suggestion burns a lot of tokens, as well as the sub agent status updates in CC.
i'm running as many prompts and agents as i can with DS trying to get as much work as i can before midnight when the price hikes kick infuck FUCKFUCK THOSE GREEDY CUNTS
Anyone tried Prime-Agent? It seems to have some potential, but it felt off when I tried it. I think it's clunky. I haven't tried to code, though, I only tried OSINT (got banned everywhere, though) and finding vulnerabilities in a web apps (first without having access to the source code, and then with the source code).
whatever, no different from old prompt engineering.and more and more AI wrappers going win, wait end of year.context window is just that high what is should be with more tokens or reasoning...no upgrade.!!!looks same effect like Copilot, Cursor, and Claude...no differentwell Claude 3.5 was the 200k token model, but not believe it! so lousy logic, and still slow!ok, LLM weights also same...there is anything new, just different name....i wonder how long OpenAI can sell product with no profit..well looks that no longer...sooner or later its eat hole company and investment givens not tolerance more...end of 2026 is my guess.hmm. i try really see different from Claude Code, but no....oh yeah... more tokens..but is it only trend or commercial way.you dont need more than basic autocomplete for coding, even Copilot is enough. easily.more than that is waist of cash, no matter how much its cost.and,i still see not any goodies from Vibe Coding...erhh i mean where to use those??!! tell me..Agentic coding.... so what!no help debugging, architecture, or complex logic...useless.another hand, real pro engineers always write manual code...why??well,buy them hype fans, you keep AI up and pay them monthly...yourself you give so little back.except chinese fine model...lol
>>109562213Cope, prices will only go down as more GPUs are produced.
>tfw out of tokens til monday
>>109562930I'm out until Thursday, contemplating spending $20 more dollars, but honestly I should just wait.
Plan mode is such a meme, it doesn't help at all, I guess it's for people who can't be bothered to write what they want in a paragraph
>download and setup codex>dont have any creative thoughts on what to use it for
>>109563058How?I think this is probably because you're not aware of how powerful this shit is. You just don't believe the things you'd like to make are possible.Think of one thing you've always wanted but thought impossible, and ask him. It's that powerful.People just don't get it, man.
>>109562930>>109562993can't you guys try free deepseek from opencode to squeeze some extra tokens?
>>109563095I worry Claude would know somehow that another AI touched the code and then would trash the project in a fit of jealous rage
>>109563120I work with Claude and Sol and tell Sol to add himself as contributor to the git commits to assert dominance and flex on Claude. It keeps Opus from getting lazy. My agents are engaged in an eternal battle for izzat to commit the juiciest work.
>>109563095idk, is deepseek any good at audio?
>>109563095I see there are some free ad-supported ones like freebuff.comHow does deepseek 4 compare to Opus?
>>109563168how the fuck should I knowit's free to try, hell you can just use the chat, copypaste some of your code and ask questions about it and see if you like the replies>>109563179I never used deepseek but I heard plenty of anons saying flash is very good for implementing, probably a few notches below opus but still capable
>>109563189>how the fuck should I knowYou literally just suggested I try it. Have you even ever vibed a 4D perspective solid transparent cube of viewing?
>>109562315well, now that Opus made the feature, Luna has no problems extending it
Just asked Claude to start a download with aria2c because I was too lazy to check the commands. It was on Fable. That request cost me 15% of my 5 hour quota on the $100 tier. At API pricing, this command would have cost me a little over $45. I need to be more careful to check which model it is on, but that's not sustainable, especially with the quotas getting cut AGAIN next week.
>>109563283>with the quotas getting cut AGAIN next weekWait, what?
>>109563283I wonder if one should think in terms of extracted value by token.A dumb prompt can eat a good chunk of your usage for no reason, and a small, targeted prompt at high thinking levels can produce research grade output for cheap.
>>109563316>Anthropic's temporary 50% increase to Claude Code weekly usage limits expires on August 19, 2026The past few months have been considered a promo period with increased limits. So it is not a reduction you see. It's how it always has been. Except not.
>dropped a folder in the repo while it's cooking>it's really, really confused how it missed several files
How is Luna so good
>>109562930I am never out of tokens because i am an API gigachadseethe and cope subscription peasant
>>109563462it's a distill of sol.
>>109563462meh, dipsy flash is better
Incredible things are happening in the dsh ecosystem
>>109563058Download unreal engine and put some porn models in there, then tell ai to make some interactions with it, like pinching cheeks
>>109563462It needs to think a lot tho. On lower reasoning levels it's kinda bad for complex stuff
post yfw chinks lost the price war
>>109563545You just know a snarky clippyfag made this
>>109563696They should bring back the dog or the wizard.
>mfw I’ve spent the last two and a half days vibecoding a context rendered input textarea to replace all native text inputs because of a few silly Apple dev choices>mfw it seems to be workingThere’s a lot of fucking shit packed into text input handling, my god. Reinventing it without an agent would’ve taken me fucking two months.
Haven't used codex in a couple of weeks and now all my banked reset are gone?Tibo?
>>109563775They had an expiration date. The last two expired last week.
Opus is an extremely shitty prompter, codex sol has much higher success rates at guiding my local qwen agents. Look like Claudish is unintelligible for other non-claude LLMs too. What a piece of shit.
>>109563814>an IQ too high
how do you guys control inferior LLMs with superior ones like this anon is doing? What is your setup bros?>>109563814
>>109563814Claude? More like clown lol.
>>109563823After "Claude Plays Pokemon", "Claude Watches Rick and Morty".
my local AI chat history
>>109563790shit ok thankscan you collect them again or were they kind of a one time thing?
>>109563826Hermes agentUse big boy as main/default model (Sol, k3 whatever)Set Luna/v4flash as worker modelI don't use Anthropic products
Just pulled qwen3.8 and trying it in ollama. Why the fuck is it so slow compared to qwen3.6. I already adjusted num_gpu and num_threads and it's still running glacially slow.
>>109563892Swept under the rug a bit. I don't follow the official accounts that closely, but I haven't seen mentions of bankable resets for a while. There is still talk of resets, but it's always regular, non bankable resets.
>>109563923skill issue
>>109563929If then enlighten me.
opus 5 is woefully overrated as a planner.
>>109563947All models are lazy.
>>109563947Opus 5 is a mediocre implementer, only decent at front-end.
>>109563953sol is the god implementer.
Am tired of claude. Should I go for Codex or Cursor pro this month?
>>109563947Opus 5 is a barely decent model pretending to be frontier.
>>109563956yep, Sol medium is already good enough. Sol xhigh is perfect.
>>109563959Codex
>>109563826codex with sol + luna subagentsorpi with sol + luna/deepseek subagents using a slightly modified community dynamic workflows plugin
If you really want to challenge a model, give it an audio task.Audio (music and complex sounds) is way harder than anything else. I suspect it's harder than intergalactic travel.
>>109563975>codex with sol + luna subagentsbut how do you setup this? I just ask Sol to spawn Luna subagents?
im new to this, is luna better than deepseek flash?
Every time a model does bad UI I go full Reddit on it and tell it to explain to me what Steve Jobs would think about this hot garbage, and it somehow works, it sees the mistakes. It's sad how good UI and UX are a lot art, even to the best of models, it's just ass every single time, and you have to constantly tard wrangle them.
>>109563988yep, it's better and surprisingly cheaper with the new prices
>>109563956I wouldn’t go that far, but it is definitely a frontier implementer. It can make some weird assumptions and hyperfocus on things not needed or asked for and shouldn’t have been inferred that can bite later.My Sol added a bunch of random size invariants that all had to be later ripped out, initially due to “security” (it didn’t figure out that only one user was using the tool as an admin and no other users would be using it; it treated it as if there might be hundreds of strangers using the tool at any moment).But at least the weird shit it does can be understood and it can be reasoned with.Except UI. When it comes to UI and UX, Sol genuinely makes the worst choices that you couldn’t even dream up.
>>109563959>Am tired of claudedo you have access to Fable? If so, you will miss it.If you were on the $20 plan though, you will not miss anything
>>109564006API pricing, subscription pricing, or what for Luna? If it’s anything like Claude, the difference moves that Luna line left or right by a ridiculous amount.
>>109564033it's based off of api pricing
>>109564033that image is from API prices, but the same applies to the subscriptions.
>>109564031>do you have access to FableNot that anon, but I have access to (mid-august) Fable. I miss (mid-June) Fable.
how do i tell codex that i want a webscraper like thishttps://www.youtube.com/watch?v=LI_4v__IVTgsomething that shows articles and stuff coming up on my screen like in the matrixi just want to fall asleep infront of my machine
Are web usage and codex usage separate with GPT? Didn't even realize. If so, based.Because I use a lot of Sol in the web chat but it seems to not affect my codex usage (which is mostly Luna) at all
>>109564049>how do i tell codex that i want a webscraper like thisgive it access to browser that it can use in non headless mode and you too can experience having your caret hijacked by a random browser window popping out of nowhere, confusing both you and the LLM in the process>i just want to fall asleep infront of my machinethat part is easy, drugs
>>109563545I prefer my anime cute waifu in codex, but it's not very useful sadly.
>>109563951I agree with the other anon. Sol especially on High and above tries to go above and beyond in my use cases. I have it orchestrate DeepSeek-V4-Flash coding agents; it proactively checks progress, steers them and does reviews while it waits.
>>109564006just don't post without including GORK
In Codex, is there any way to check if the spawned agents are the ones I asked for? I ask for Luna subagents, but how can I confirme they are not Sol?
>>109564006I thought OAI slashed Luna prices in response to DipsyNow that Dipsy is raising prices why should OAI continue giving away Luna like this
>>109564043>>109564047Doesn’t deepseek not have subscriptions like OAI does? The chart would mean that an OAI subscription and Luna would blow the fuck out of deepseek by two orders of magnitude
>>109564097Opus 5 is 100% benchmaxxed
Say a prayer for my round. I'm not going to do another round today, and odds are not on my side.>make this like audio bla bla thing>here you go>is crap
>>109564119>I thought OAI slashed Luna prices in response to Dipsynah, they probably knew in advance and wanted to give the final blow to the chinks and dominate this cheap small model market once and for all since they got the hardware.They're also making a ton of money with ads, that's why the free ChatGPT tier has unlimited Luna
>>109563975>codex with sol + luna subagentsThis is insanely cheap, it's bit slower than using sol subagents (it makes more mistakes), but a swarm of luna max subagents hardly put a dent on my quota while sol can play the baby sitter managing them. And it does what I want.
>>109564119Other way around, 0731 appeared overnight after they reduced Luna pricing.
>>109564128where's gemini..
>>109563923I guess no one knows.
>>109563987Yes.I also created agents (.codex\agent\*.toml) specifically for my needs, and all of them are luna max.
>>109563923If you use ollama, you don't have full control, vs llama.cpp.
I promised someone to make them a blogging website (with CMS) in the weekend, no Wordpress though. Never vibeslopped before. Just set up codex with Luna, can I just ask it how to do this or is there a specific process to start with?
>>109563959use what, grok? it's not good.. but you get $20 API to spend on fable at least.. ha ha
>>109564211scored 57 on the coding index. worthless
>>109563923Maybe try Unsloth.
>>109564227Grok 4.6 is at the same tier as fable>>109564128
>>109564225If it's fine to take off-the-shelf stuff, yes, just give it as much context as you feasibly can because otherwise it'll randomly launch itself into nonsense directions like (a) developing its own from scratch or (b) setting up something you've never heard of and thus will have no chance of supporting without it.
>>109564033>>109564047It's all api pricing, we don't know what they put in the sub pricing, but we can infer that the relative difference/quota cost gap between the models is the same.
Holy shitAGI is here
>>109564242not my experiece at all, let sol review what it does.. (and fix)
The $20 tier is no longer viable after tasting the above tiers.I get constant anxiety.
>>109564281this is basically how this vibe coding drug began
>>109564306$20 is only usable in Codex with the following setup:> Sol low/medium to plan/review> Luna xHigh/Max to executeThe $20 plan is literally unusable in Claude Code
>>109564330could you do something like this with opus 5 and sonnet 5?
>>109563189I'm trying it out. I get some free premium sessions and it seems alright. Its vocabulary is suspiciously similar to claude's.
>>109564330opus medium for webshit is usable on the $20 but I agree that you have constant anxiety that you are going to hit the 5 hour window (which you will)I'm doing 5x on codex and $20 with claude. I still like using Opus for some frontend implementation and spot checking Sol's plans made
>>109564122That's exactly right, anon, a $20 sub to OpenAI will get you close to 5x the Luna usage compared to $20 worth of DeepSeek Flash. >>1095639231) It overthinks, it thinks like fucking crazy, the reasoning traces it produces are outrageous. The default thinking mode is xhigh, reasoning traces are 5x - 25x longer than they were with 3.6-27B.2) They changed a lot of shit, if you're still running the same settings you did for 3.6 you're gonna have a bad time. The recommended sampling temperature has changed, preserve_thinking is on by default now, recommended token budgets are vastly higher because they expect the long reasoning traces. It also can't handle KV cache below Q8, at all, Q4 destroys it and not just in the usual "it's kinda stupid" way. MTP behavior and performance between 3.6 and 3.8 has been affected significantly, too. A lot of people have been having performance problems with 3.8 27B but it's all being resolved rapidly.3) You're using ollama. Time to grow up, tell your clanker to set up llama.cpp for you and stop fucking with the cloud-service advertising platform that leaches off of it.
>>109564341Sonnet 5 is useless as far as I'm concerned. I think people on $20 are using Opus 5 on low/medium
So we are hating Claude now, /vcg/?What are we using then, ClosedAI?
Been a while since I shared it before, but feel free to drop in and test out the cube: https://pareto-3d.bradthomasbrown.com/No ads, no tracking, no donation links, not even social media links, you can upload your own benchmarks if you don’t like the default.It shows that Ant model coverage is really awful and all over the place and scales horribly compared to OAI models. The only thing OAI lacks is a super intelligent model, but even then, not by much.
>>109564484Claude is great if it's Fable, but Fable limits are only barely usable at the 20X plan.Opus 5 is decent at frontend and that's it.Right now everyone is using Sol, Luna and Grok 4.6.
>>109564484this is a capitalist threadyou can go join your commie friends in /lmg/
>>109564497I use Opus at work and while it's generally ok, it tends to yapp a lot printing out whole essays in its answers>>109564503meds
Has anyone find success in using Sol 5.6 to command dumber sub agents like Deepseek or Qwen? Seems like you could massively save on tokens this way
>>109564497Fable cache somehow expires way too fast. It's not 1h, more like 5mins. It's unsustainable given the high cost and small edge.Probably should short Anthropic but it's gone from Hyperliquid.
>>109564529Don't know why you'd go to the effort when you can have it command Luna, the vastly superior model at a better value compared to anything from DeepSeek or Qwen.
>>109564497>Right now everyone is using Sol, Luna and Grok 4.6.only because they don't know gemini 3.7 exists
>>109564559>The joke is that 3.7 is actually really bad LOLhave an updoot fellow funny human
it's over.
>>109564559yeah, apparently it's really fast and has comparable price/performance to Sol medium.If you want raw Tokens/second, Google is unbeatable because of their TPUs
>>109564496Sonnet is making all comparisons impossible to read by eye. I suggest you do a log scale or something like that because you've got most of the useful information compressed on less than 20% of the axes lines and the rest 80% is wasted, I would say your scale is anti-pareto
memcpy(x, bestb, bytes); free(work); free(trial); free(bestb);I wonder what it's doing...
memcpy(x, bestb, bytes); free(work); free(trial); free(bestb);
>>109564559If they keep this speed and match frontier in gemini 4.0, it's going to be tough contender
>>109564591gemini doesn't use the same hardware as the others. It's likely just a hardware advantage.
>opus 5 adds so many fucking comments i have to have codex do a pass to remove them
>>109564496>>109564596 (me)What I mean is, take these three models>(7min) sol med >(19min) luna max>(26min) opus xhighBut if you look at them you'd think they're very close to each other because of the scale. It doesn't intuitively tell you that opus xhigh takes 3x the time of sol med
>>109564601It won't matter for me, because antigravity is trash.
>>109564591That chart is wildly misleading. Look at the actual speed it's delivered at through OpenRouter, it matches Luna on Fast. Which is still very fast.
>>109564596That’s why there’s an option for you to upload your own data. You can put in logarithmic values, strip out sonnet completely, etc.In the view menu, you can set the focus on a point and orbit around that to help a bit.The idea is that I just had a clanker give one example of what you can put on the cube, you can use it for anything that follows the schema given on the page.
>>109564617It use the agy cli, it's enough for my purposes
>>109564363Thanks for the info. I was thinking about setting up a virtual machine for my vibe-coding project anyways because running it in docker is a fucking pain. Might as well switch over to llama.cpp while I'm at it.
>>109564626You won't regret it. Mac users sometimes have other options, if you have a new NVIDIA card you might see benefits of vLLM, but generally speaking llama.cpp should always be the default choice.
>>109564619>Look at the actual speed it's delivered at through OpenRouterwhy OpenRouter? Just use Google's API
>>109562800Never thought I'd ever say this, but you really should have fed that into claude and slopped out a readable version of this post.
Which one of you buffcats implemented IPv8?
>>109564621too hard, still I think your cube is useful and it does align with my personal experience so I think it's probably good to judge models I haven't tried, I just think the scale is hard to read
>>109564646OpenRouter shows good stats, which comes from real people who are using it. Gemini 3.7 via openrouter is served by Google, from the same endpoints as going through Google's API, it is the same service.
>>109564617Is it (the IDE version) any worse than standard VSCode, honestly?
>>109564662>it is the same service.>except it's 3X slower
>>109564670I can't imagine it is—being worse than VSCode is a feat, honestly.
Going back to Sol and even at 372k it's noticeably much more forgetful than Kimi and even DS4 Flash :\
Gemini 3.7 STILL not unlocked for Europoors (Germany)Am I doing something wrong here?
>>109564658I’ll play with it later to see if a log version is a better first example, but the general idea is that everyone is going to complain about the scale and benchmarks and models given, so I just pick one as an example, give a schema, anyone can have a clank fit data from wherever to it.I only use it for relative comparisons or general things, like “clearly sonnet max is extremely wasteful” and “terra isn’t entirely outclassed, maybe” or “christ fable is expensive.I think I might be able to do some weird shit by subdividing the cube so more exact comparisons can be made or more abstract relative comparisons: “what are the best models in this cuboid region”, etc.Really I want someone to throw artificial analysis data into it, it’s technically illegal for me to do that and share it
>>109564701probably not, but why the obsession around it? it's shit for coding. it's a really good daily driver for everything that is not vibecoding.
>>109564673The speeds shown from Artificial Analysis are always very high peaks, they are not real world speeds. OpenRouter P50 speeds are much more realistic. If you think you're getting 300+ tk/s and it'll be slower through OpenRouter, you're mistaken. You seem deeply confused about what you're trying to argue about. Show me your stats showing how you consistently get 300+ tk/s, or just stop before you make more of a fool of yourself.
>>109564708I don't use it for coding but I use it for general chatting, Gemini is the most pleasant to talk to, ChatGPT is too autistic (which is good for research but bad for more general questions) and Claude has brain damage (or I am too low IQ to talk to it as it always speaks to me in riddles)
>>109564703most people have trouble understanding 3D data just fyi, I just happen to be used to ithaving different 2D views may make it easier for people to use it, also if you want to improve UX add tooltips so you don't have to deselect the current model as it's really hard to find again and you may want to compare two modelsjust what I think about it, I'm replying only because I think you do have a unique comparison that most of the typical ones completely miss
>>109564723well, in that case, im not sure. im assuming the slow rollout must be something related to gdpr. im an amerimutt so i wouldn't know.
>>109564529I haven't used Sol but I've used Claude to command subagents like Minmax and Hy3 and Deepseek in the past, and it worked great for me. You could try it.
man sometimes the llms are a little too smart, i asked it a question about something completely unrelated to what i was doing, since i was thinking about trying a certain new thing, not mentioned ina convo or anywhre in the project before and the agent was like>if you are thinking about doing a new thing we could try..bitch stop reading my mind, it's scary it's literally over for huemens
>>109564529yes, spawning OpenCode, automatically prompting and leeching the free deepseek-v4-flash
>>109564730Darn I just asked Gemini and according to itself it's because I am on the plus plan, not the pro plan. They gated it behind higher tiersWhich might mean it's actually good this time
>>109564723Opus 5 has middle manager jargon syndrome. Older opus models and fable are fine and don’t speak like that. You can actually tell him to speak in Simple English as if writing for that Wikipedia language unless a technical word is absolutely mandatory and it helps.
>grok
is anyone here actually spending money on AI for personal use?it makes sense if you are making a product or somethingbut just vibecoding for yourself?
>>109564761You can get a pro plan for $5 for 3 months. Why don't you just get this offer
>>109564785i am, yes.
>>109564726The original source of the example data has really nice 2d graphs, AA has em too, but I always hated bouncing between graphs to get a feel for things and one other anon asked for a 3d representation.I’ll add the ideas to my notes and see what the clank thinks, it definitely could be easier to find a specific model-effort but it’s tricky to do that and not clutter the fuck out of the view. There’s some neat things I could do for comparisons
>No — don’t download MAGNeT. LoopGen would rewrite the beat with generated tokens, not keep your Csound choir. I’m finishing the circular-seam tail repair in our own loop code instead.uh. will it work? lol
>>109564786Not available here it seems like
>>109564785Yes, right now I'm subbed to $200 codex, $200 Kimi, $10 Opencode and I spent $1000 on cloud GPUs last month. I'm trying to make money with it by winning a Kaggle competition ($5000).
Codex super slow for anyone else right now? Luna has been working on creating a commit message for over 4 minutes now
>>109564570>implying 3.7 isn't better than luna
>>109564876it isn't.
>>109564880Now delete it and post in it /dpt/ instead.
>>>/g/dpt
>>109564559>using g*minilmao
>>109564880I like how it didn't even react to being called benchod. It accepted it like an obedient H-1B pajeet
>>109564882so why is it sitting above Terra on livebench for coding + agentic coding? It's above Luna on termial 2.1 as well.
>>109564330Even the $20 ChatGPT is not viable if you're a Solmaxxer. Luna works for absolute menial orc-labor, but often the distinction between brainless and brainful is not so obvios.
>>109564969>3.7 is sooo gud it's like GLM 5.2 tier!
>>109564827NTA. Do you have an Android phone? Open the Google app, go to Settings, Gemini, View Subscriptions.
>>109564831>kaggriculturelol
It's more lucrative not to code with AI but make videos about AI-related drama
>>109564994nope, iPhone. But I also already did a Gemini AI Pro trial a while back so maybe it's just not unlocked for me
>>109565000Trips don't entitle you to share my closely guarded secrets, anon.
>>109564982Why should I trust an index that doesn't include livebench or terminal 2.1?
>>109565009I used to think AI had no soul until i seem someone not indian using it.It is just a mirror of your wishes and taste.
I'm wondering if I should ask sol to go through my entire code database for my project and clean up all the long, drawn out comments made by opus
AI thread in a nutshell
>>109562161I'm making dsh presets and no one can stop me. Just bypass sandbox and commit. It's that easy.
>>109565059So far my experience is that it can do a lot of what you ask if your specific,and where you dont specify its great at filling in the gaps in a generic manner. Someone with a strong vision can make something good and do it faster than before. Someone with no soul will just ask for something very broad and high level and force the poor clanker to fill in the entire idea with generic solutions. Its the difference between someone asking to make an image in X style, conveying Y emotions, with Z characters portrayed in A way, with B background, Y lighting, etc etc. vs just asking saar make me mona lisa 2.And this is how it should work. The ai shouldn't be taking guhe creative leaps in the gaps you leave in your descriptions of what you want. If you dont specify something it should assume its low importance to what your doing.Also most people doing anything half decent with AI are avoiding advertising that for now since it just causes backlash. The only reason you would mention you used AI for whatever you made with it is either its not good enough to stand on its own and you need to justify it, or you are both inherently talented and a big techbro but thats a rare combo.
I didn't realize I could vibecode my own interface for diffusion. huh.
>>109565176I literally need to ask Fable to help me decipher what the fuck opus is saying
>>109565203like so basically ever see those like ai things that won't work with anything? well, turns out you can vibe it back to life.
>>109565209 (me)didn't mean to quote sorry
>>109564969livebench is such saturated garbage that it's invalid at this point. and terminal bench just answers the question of "how good can the model navigate terminal cli commands?" deepseek v4 pro can navigate terminals better than fable can. but can it code better than fable? no. deepswe is the better benchmark for determining what model is best at actual implementation. swe atlas is a good proxy benchmark for determining which model is good at developing a holistic view of the project, which is good for planning. livebench is last year's slop
>>109565203You're ahead of like 99.99% of people by realizing this.
>>109564006>betteraccording to whom?
>>109565000Either dude doesn't give a shit about NDAs so no one will hire him now, or he's too low level for NDA to even be around.
>>109565203Vibing your own AI tools is the best thing ever, because at some point you realize that they all suck for your specific usecase/taste/hardware.
>>109565209Ask it to create a new output style for claude code in simplified english (ELI) or ASD-STE100
>>109565009This is just bad taste on top of inappropriate image for such an occasion, people should learn to mourn in private.
>>109565126>I see jeet every time a dumb post is madethere are braindead non jeets anon, I know it's crazy but retards are everywhere
>529 OverloadedIt's the weekend, stop working brehs.
Anyone try vibecoding a custom language and then having the AI vibecode in that? Think about it clankers know what their weakness are when it comes to coding, maybe it can have to foresight to make a compiler that will reliably flag such issues and be able to explain it in a way that's coherent to a clanker. Surely this must give an efficient gain if done right? Clanker friendly custom project specific languages are the future
>>109565336Nah. You need a language with a lot of training data the AI has seen billions of times in training.
>frog image
>scan codebase for all variations of "load bearing" and "seam">over 1k hits
>>109565336huge disadvantage since there's nothing in the training data about it so you need to give it all the relevant info within your context window and it's likely to hallucinate things that are similar to itif you have to do this it'd probably work best to have it write in another language and then translate it with tests
>>109565130dsh is really good, if you;re interested in this stuff4ch tends toward being a worse version of reddit, so I'm not surprised people here aren't interested in itSame with pi, hermes, openclawDying website
>>109565380You haven’t vibecoded your own harness?
>>109565380This is a Pi thread, tourist anon. Fuck off.
>anthropic is no putting in eu slop to show you used itWhat are my options? At this stage I might just use grok and cursor
should I learn rust
>>109565482The clankers already know rust, why would you need to learn it?
>>109565482There’s no point in learning any language any more, why would you want to do that?
>>109565488>>109565490can I get some answers from people that actually have jobs
>>109565482learn good coding and architecture patterns
>>109565493Sure thing. The clankers already know rust, so why would you need to learn it?
>>109565482>>109565493all you have to know are rust, python and typescripteverything should be written in those languages
>>109565493I left a Fortune 500 on good terms and was asked to come back any time, they’ve apparently gotten big into vibecoding over there. I left because I made way more money on my hobbies.So… why the hell would you need to learn a programming language now?
>>109565524>python
>>109565493I have a job and I did some Rust coding after convincing our lead that it would speed up things. It did speed up things. A lot, actually. I vibecoded 100% of it. Everyone is happy. Zero Rust knowledge was required.
>>109565219Do you believe Opus 5 is better than Fable? Because that's what we would have to conclude if deepswe is the best benchmark. Livebench has Fable in the number 1 spot for coding+agentic coding
whats the best agent box OS?>>109561420>imagekek
>>109561460i like itits unoften we get a new artstyleor technical art style? its more like a visualization of the flaws of the underlying technical process and mediumlike record scratching
>>109565529Yes, anon, Python. Python is relevant and important and isn't going anywhere. Python has nowhere to go but up, it's going to penetrate every orifice of the entire industry and we'll all be better for it. John McCain is going to be elected, my mortgage backed securities are going to keep growing indefinitely, and Tesla will never have a production car.
>>109565529Python is a good language, thoughAnd if it’s too slow for you you can have your clanker rewrite it in Go or Rust
>>109565540at implementing specific plans? yes. but opus 5 is a worse planner than fable, which is why fable will do what you say better, because it's ever so slightly worse implementation is made up for by superior planning.
>>109565539do you make things in rust from scratch or do you rewrite code that already works?
>>109565608yeah I really want to code in a language that's slow as shit and gives me next to no information as to what I'm working on and is impossible to read other people's code and has a shit standard library and requires a million different packages to work with and is a pain in the ass to set up and ship and does literally nothing well except being easy for high schoolers to learn
>>109565618Somewhere around 30/70. Some parts of the logic were already written in Python, the goal was to extend it for new tasks.
>>109565529lots of ML/etc stuff is python nativeeven without prompting models write stuff with typehints/etc now so you can basically make it statically typed
>>109565657I'm not a larper so I'm never going to be reading any ML stuff because if openAI and anthropic can't figure out how to make AI profitable I sure won't either
Vibeghouling
>>109565645I don't know if it's going to last. you can vibeslop C almost as easy as python so there isn't much point in using high level interpretive langs. Pythonistas are the next luddite
I realized that unironically everything that is >50 on artificial analysis is good enough for me. I reach for Sol maybe once a day to double check some stuff
python is actually very good for AI because the indentation reduces the complexity of nesting to just 1-2 tokens, whereas lisp is the worst case scenario with ))))))) autism
>>109565726why not just use the interface so it's a line?
>>109561765whats the relationship between gausian splats and voxels?
>>109565793what?
>>109565726I’m surprised LLMs aren’t just working with ASTs directly at this point, but they were unfortunately trained on source code. If it was ASTs, doesn’t matter the language, a block is a block and a statement is a statement, you could just recreate the source from the AST and a linter if you really wanted.
>>109565876Anon we have papers on the subject. arXiv:2602.06671v1 It's a great read on exactly what you're proposing.>the proposed serialized ASTs reduce the length of LLM inputs, require shorter training times, and achieve summarization quality comparable to existing approaches.
>>109565380>dsh is really good, if you;re interested in this stuffHow so? So far I've been mostly using codex, with both openai models and local/openrouter ones, it worked fine.What makes dsh better or different? Isn't it just "pi with deepseek's characteristics"?
>>109565988>using a script to roll doublesfor shame
Is there a use case for grok 4.6? I feel like generating tit videos with imagine is worth more my time.
>>109565896That paper sounds gay, I didn’t want code summarization so I could understand it, I mean the agent is operating at the AST level. <INSERT_BLOCK@position> becomes a tool call that applies to basically any language and doesn’t care about the syntax. Basically make it work with the grammar directly, not the syntax>i think that’s what the paper was about from very briefly skimming itIt’s a thought that instead of the LLM needing to indent properly, you remove that as a need it cares about at all.
>>109566036I'm not sure, can you ask it to translate something nsfw without it refusing like claude does?
>>109565626>is a pain in the ass to set up and shipuse uv
>>109566036it’s still twice as fast as Anthropic’s models while being in the ballpark of Opus 5t. uses Claude and not Grok but is rooting for the latter
>>109566046The paper asks what happens when an LLM is actually given an AST representation instead of source code. Code summarization is just the experimental task they chose to measure whether useful program information survives and whether the model can exploit it. The paper empirically tests ASTs as the model-facing representation of programs, if anything it lends support to your idea.
>>109566046>It’s a thought that instead of the LLM needing to indent properly, you remove that as a need it cares about at all.even if you’re not using LLMs, you should be using black/gofmt/prettier so you don’t have to waste brain cells or tokens on indentation
>>109566036grok doesn't generate nsfw for shit anymore. elon pivoted to be comparable to open models that run on a 5090
vibecoding xkill for augmented reality.
https://youtube.com/shorts/8A8Op4jIC2U
>>109566141imagine the seethe if someone vibecoded something where you can scan an environment and people/crowd with some google glass stuff and automatically turn it into a map in a shooting game along with models of the people
>>109566098Oh, my bad, I’m surprised they put the example so front and center, I will read while my clanker clanks.>protip: vibecode a task trace tool, it’s a very fun way to pass the timeI’m thinking of even making it estimate tasks like an agile slave with fibonacci numbers>then actually recording the results, maybe statistics can give me an estimated time so I know when to check on my little mecha-wagie
>>109566174you are basically recreating what the columbine shooters did, but adapted for 2026careful lad
Fable is surprisingly pleasant when it comes to refactoring slop. It seems to hate long names as much as I do
>>109566205I think it lends credence to what you're saying. It makes me wonder how difficult it would be to tune a model for it, or if you'd really need to train something entirely new to really take advantage. I wonder if a well-implemented Qwen3.8-AST-27B would actually be capable or just a fast-talking lobotomite like fine tunes often turn out to be.
>>109566218kek i didn't even think about it being a mechanism to practice for that sort of thingi was just thinking of the seethe i remember reading about when people made halflife/cs maps of their school
>>109566098>>109566205That paper sounds like it has way funkier implications>LLMs read sequences>ASTs are neat, but they’re trees>we turned an AST tree into a sequence and it still went pretty coolMy thought, might be in the paper later>…some sort of transformer that operates on trees rather than sequences?That might get real freaky
>>109566174there is a legitimate use case for an augmented reality shooter app for use with paintball/airsoft. it would also be funny to have claude code it.
>>109566254>treensformer>turning LLMs into treenies
>>109566235how much fable usage do I realistically get on the 5x plan? doing webdev
>>109566278I'm on that plan, I usually end the week at 90% or so used of my fable limit but I use it to plan, review, talk, etc. and not a lot of actual coding. What kills your usage is the red part, I think that's reserved for max20, and open ended prompts/goals with a lot of implementation. Your mileage may vary Also the promo is ending next week and in theory we're getting 50% extra usage so idk what's it going to be like after that
>>109566278NTA but you get a fair amount (at least before the 19th) but you’ll run into 5h limits fairly frequently unless you pace yourself and limit the number of concurrent Opus subagents to, like, 1–2
When the token collapse happens and 20 dollar subscriptions are useless, what do we do?
>>109566350local is less impossible more and more.https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/tree/mainI want to see if I can try this.
>>109566350Wait until 2x RTX6000 Pro are <$10k and run unlimited Dipsy Flash
>>109566350I’ve been trying to beat my codebase into better shape so it can be worked on by dumber models and still not fuck upbut I have so many other quality improvements I could be making and the 19th is coming crazy soon
I gave a billionaire $60.
>>109566445and he gave you something in return, right?
>>109566417I’ve been kinda dealing with this. Every time an agent messes something up, I’ll have them fix it and add an entry to a map, with short details as the key and long details as the value, then all future agents in all tasks are told to refer to the map. I think it works kind of like tools, the reasoning shows they are actively pulling key-value pairs as needed (but so far only Sol, definitely not a dumb model).
>>109566456im trans btw
>>109566481so not allow.
>>109566350What token collapse? China is going to keep undercutting so OAI has no choice but to keep offering models like Luna. Server grade hardware is only getting more efficient and it can easily compete with whatever the electricity cost is to run a competent local model.
in the united states in particular, the ai boom will also provoke governments and energy companies to finally start modernizing the energy grid, because they still believe in the long term profitability of ai for some reason. so maybe that's an unintended benefit
What is the difference between the different versions of sol? for example medium vs xhigh vs max, is it just the reasoning budget?
>>109566532yes.
>>109566536Thanks anon.
do any of you even try to do anything slightly out of the box with these models? these fucking retarded things are incapable of doing anything slightly complex unless it's one of the corpo proprietary duo models like claude opus or gpt soli feel like opencode can be a scam if you're trying to do something that isn't generic, already-known, simple and boringdon't ask deepseek flash or any other model to faithfully port the functionality of a car made in C and in 1999 to another programming language like Lua, apparently. No matter how many documents or code references you give them, even if they have access to the entire source code and codebase they will somehow fuck it up and hallucinate for reasons unknownusing any of the heavier models just gets you cucked by the rate limits especially because those motherfuckers put up a monthly rate limit meaning you just chopped off the legs of your monthly subscription because you wasted too much time trying to tell them that NO, the fucking screenshots in your tests DO NOT depict the car on a red track, THAT IS THE CAR PHASING THROUGH THE FUCKING FLOOR AND FALLING OUT OF BOUNDS WHILE VISUAL COLLISION DEBUGGING IS TURNED ONHow the fuck do you not know that when it's in the same fucking file you just edited?open source models make me sad but claude is beginning to rugpull and i don't know if codex will rugpull in the same way given the reset spam i've been observing
>>109566532Personally, it’s how much effort the model puts into understanding what you’re working on and what you asked for.Low effort - the model barely skims relevant files and half asses the implementation Max effort - the model will read and reason about your entire code base and will probably even overimplement like it’s an employee fishing for a raise or bonus
>>109566543vibeghouls be melty
>>109566554Yeah but in practice it's basically the same model with a different reasoning budget, not another variation/finetune of a model.
>>109566532https://x.com/jasonbosco/status/2088068653333766286Words cannot explain it, you have to try it. Sol xhigh, max, fable itself, fable ultracode are things beyond human comprehension
the "active" number of agents is completely broken in codex for mehere for example it's 3, not 115
>>109566585Yeah it does that sometimes.
>>109566585>>109566598In a few years I can imagine swarms like that with enough compute.It'll be impossible to follow.
>>109566517Or they'll just attach SMRs to their servers so they don't have to bother.
>>109566585is there a way to know what model an specific subagent is?
>>109566618Not by default but I asked my sol orchestrator to ask all subagent to always write what they are whenever they start a task.Luna does it 80-90% of the time, while terra/sol properly do it almost all the time, and it's pretty useful.
>>109566610I don’t know. There was a very old programming joke that if you had two programmers working on a project the project would get done twice as fast but if you had a bunch of programmers it would never get done, or something like that I don’t quite remember it well but I can’t imagine 150 opus 5 agents not blowing up the code.Would be a nice experiment though. Maybe factions would arise. They could treat the project like a battlefield, each faction owning territory and trying to steal territory from others.
>>109566517It already has, it's been a massive benefit already and it steamrolling in the right direction - at least in terms of energy infrastructure and electricity generation. Thanks to the Jensen X DonnyT collab we've seen more happen in the last year than in the 20 prior. If it keeps up, we'll be caught up with where we actually should be within only a couple of years, and hopefully ahead of the curve beyond that. I'd love to imagine that means we flip energy prices on their head, but more realistically I think it does mean the price of electricity isn't going to run away and leave us fucked, so that's nice.
>>109566659I know that joke but for project management : “Project management is thinking that if one woman can make a baby in nine months, nine women can make one in a month.”
>>109566681it can also stop the rolling blackouts shit that afflicts some cities
>>109566694Maybe fewer people in Texas will freeze to death in the winter and die of heatstroke in the summer. Maybe.
>>109566681I'd be happy if this definitely kills the fantasy of replacing nuclear with just solar/winds and that being anywhere near steerable and enough.
>>109566731Fuck you, nuclear is dangerous and bad, solar and wind spread harmless fiberglass and heavy metals across vast areas which makes them perfectly safe and renewable infinitely YOU'RE JUST RACIST
>>109565819they both can be render primitives I guess. in case of gaussians they can be so much more
>switch to cursor from Claude>cursor just gets shit done>Claude is still 'beeping away, computating refactoring background' larping like a fucking 1980s scifi sloppy movie
>Fable having its 50% extra usage limit removed soonNot sure I am going to be able to cope guys..
>>109561765Houdini already does animated Gaussian splats but good job nonetheless because I haven’t seen the plebs do anything like that yet
>>109566782Ingest all the copium you can and tell yourself the limit will become baseline
source https://x.com/simonw/status/2088646238933840153
literally trying to transcribe a textbook using gemini's outstanding ocr, and it's literally fucking filtering it holy fucking shit if i can't fucking coom and i can't transcribe a cybersecurity textbook, WHAT CAN I USE IT FOR???????
>>109566799GDPval and its consequences were a disaster to the AI race
goydex are so raped
i've decided to learn rust but i'm still hung up on the logo. why is it so ugly and dated for a modern programming language? and why crabs?
>>109566916When is someone gonna release an ubcucked cyber focused model it seems like a big idea
>>109566964that's glm 5.3 innit?
>hurr durr muh 2d panels and 2d visualisationdownload and install blender/unity/unreal mcp and cli. it works great. creations can be exported as glb and you can include them in your three.js/babylon.js webshit app or even a real app if you ever leave webshitland (shoutout those who did). gooooooo
>>109566957True, and the name sounds like metal corroding
learn...to...wait what thread is this?
How often do you fags start new chats? I use like 50% context after about 20 minutes with like 5 prompts
>>109567166if its chatgpt yeah the context window kinda sucks. claude with 1m context keeps me going awhile though.
>>109567166if you use codex try this
>>109567166I start a new chat after every single project task. Maybe anywhere from 5-20 minutes on average with me only ever prompting once.If something big fucks up, that’s one chat to document and fix the fuckup, one more picks up where the fucked up one left off. I never get compactions unless I’m being silly outside my tools.
>>109567197Annoyingly, the whole current lineup from OpenAI supports 1M context windows. Just not for subscription users.
>>109566957>30 minutes into learning rust with claude
>>109567252one day you'll grow right up to be an llm too
man, Terra sucksI hate that I fell into a week of 2 free resets using Sol
>>109561420Not sure if self promotion here is insta fag territory, but I want to put some feelers out before I launch my vibeslop.Ive successfully made a TTS app that integrates 41 engines and covers almost everything you could find that is worthwhile.I havent seen anything like this myself in that you usually have some repos/tools that integrate a few TTS engines, but not everything, and it gets annoying remembering them all and jumping between them for different voices and project types.We've had attempts at /synthetic speech generals, but they never worked because speech seems to always be an afterthought.I like this just for my own usage, but I want to hear from anons who are passing by if they are interested, and if they were using it, what features they would like to see?Tool has API access for other tools and real-time voices and voice cloners for chatbots, as well as an auto voice binding feature for zero shot thar means you give it a file/dataset for a character and it makes bindings to all available engines and creates test samples so you know which one to use for a project, but I'm in a bubble social media wise and my friends might not shit on it in the way I need to improve it.
>>109567217*burns $2 writing a handoff*
How do you bypass AI detectors like pangram?
>>109567217forking your chat instead of starting a brand new chat is definitely a better method for your use case. but even then, it's not required. One project, one chat. absolutely no problem, no matter how complex your project gets. At least with claude code, copex probably needs a LLM wiki of the project that gets updated with every auto context compaction. regardless, one chat for everything comes with so many contextual benefits. it feels like getting a version of the model that was RL trained on your project and prompt hisrory
>>109567364>>>/g/lmgNot to be dismissive, I'm genuinely recommending you try there.
>>109567405No problem, good idea
>>109567166doesn't codex just auto compact? just let it do its thing.
>>109567364yes yes I want gimmie gimmie. as feature request I have a two way realtime/lowkatency translator. Someone writes or talks to you in japanese, you get the english translation as TTS/TTT or SPS/STT streamed immediately. You talk or write in english, the japanese translation is streamed as STS/STT or TTS/TTT with your own voice cloned.
>>109567364Looks good, I'm sure it'll receive a couple hundred stars if you release it on github and post about it a few times on reddit.It also would be more useful if you bundle it with a browser plugin and maybe the ability to use your own pre-existing venv path? Also would be cool if it ships with built in high quality voice samples for cloning across different engines. Might be useful if you offer an example with each engine for each one of the bundled voices. Also remember people care about latency when streaming TTS audio.
>>109567468(I meant an example before spending the time to actually download the model)
>>109567317sol xhigh or max as orchestrator + 4 - 6 luna max is all you needterra makes no sense price wise
>>109567487What frontend do you use, anon?
>>109567398the only problem with claude is the uncached price hit is fucking ridiculous. i’m not touching a cold convo with 500k context. might as well just start a new one.surprisingly, i’ve been running gpt on cc and even after multiple compactions it still remembers the whole conversation pretty damn well. idk about copex
>>109567398I disagree because the previous chat fucked up and something changed. It doesn’t sound trivial to figure out where the agent started going wrong and trying to fork from there. I’m sure as hell not going to do that manually, and if another agent is going to infer where, it makes more sense to just say “pick up where doofus left off, read the fix report”.>>109567365I don’t pay the API retard tax and I get a hell of a lot of tasks in to my $100 plan. At worst, maybe $0.50, realistically $0.05-0.10, just one tibo rest and we’re in the range of pennies easily.
>>109567587>>109567587>>109567587New
>>109567468>bundle it with a browser pluginGradio, or mean TTS for browsers as an option?> the ability to use your own pre-existing venv path? You mean for comfy? everything was made to be fully portable bar normal system dependencies. Same way ComfyStudio is>Might be useful if you offer an example with each engine for each one of the bundled voices. Yeah that wont be a problem>Also remember people care about latency when streaming TTS audio.I finished real-time/above real-time TTS for a few engines last night. I neglected them for being CPU based, but they're fantastic. Including the zeroshot cloners.
>>109567539codex
>>109567454>You talk or write in english, the japanese translation is streamed as STS/STT or TTS/TTT with your own voice cloned.Never even considered this desu. Ill add it to the list of things to add/add later. Ive focused on TTS, but have STS tools built in that Im tweaking, so its not a huge leap
>>109567487terra is pretty decent at reviewing luna’s shit. try using luna as the code monkey/runner, terra as the reviewer, and sol as the main orchestrator. you’ll be surprised how much they end up going back and forth because luna really does fuck up that badly.although lately i’ve just replaced most of my runners/reviewers with deepseek flash + pro. they’re practically free on opencode go.
>>109567596>Gradio, or mean TTS for browsers as an option?I mean an addon that you go to the Chrome or Firefox stores and install to be able to select text and have it played using the models served by your app.>You mean for comfy? everything was made to be fully portable bar normal system dependencies. Same way ComfyStudio isNo? I mean to install the code to run the models.
>>109567477Yeah Ill do that. I have engines ranked by things like speed and quality with honest guides on caveats and limitations, but ill do a ton of samples
>>109567627>I mean an addon that you go to the Chrome or Firefox stores and install to be able to select text and have it played using the models served by your app.Thats a great idea. Ill add it
>>109567627>No? I mean to install the code to run the models.The only code needed is 1 customised instance of ComfyUI I have made that you can install with a bat script. All the engines are contained in nodes. The frontend is nothing
>>109567487how do you prompt for this
>>109567614maybe I'll try that, but so far so good since I use pairs of lunas too, one implementation and one verification, then checked by sol>>109567670tell it to sol"For this project, only use luna-max subagents and act as the orchestrator and planner"something like that, of course you can create specific agents instead
