[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: 1770786172456995.mp4 (3.42 MB, 1114x720)
3.42 MB
3.42 MB MP4
A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.

## What “vibe coding” is, and how to do it
https://simonwillison.net/2025/Mar/19/vibe-coding/
https://simonwillison.net/2025/Mar/11/using-llms-for-code/

## News
- (2026-07-12) Codex temporarily lifted 5h usage limit
- (2026-07-12) Anthropic extended Claude Fable 5 access on all paid plans and kept Claude Code weekly rate limits 50% higher through July 19.

## 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 so
https://developers.openai.com/codex/cli
https://claude.com/product/claude-code

## Worth it for code, but the frontier models above are better
https://x.ai/cli

## Not worth it for code, but good for making sense of pictures
https://antigravity.google/product/antigravity-cli

----

## Prompting / context / skills
https://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 agents
https://osaurus.ai/
https://pi.dev/
https://opencode.ai/
https://cursor.com/docs
https://docs.windsurf.com/
https://docs.cline.bot/
https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent

## UI/Frontend
https://www.figma.com/make/
https://www.anthropic.com/news/claude-design-anthropic-labs
https://uiverse.io/
https://ui-ux-pro-max-skill.nextlevelbuilder.io/
https://stitch.withgoogle.com/

## In-browser builders / hosted vibe tools
https://bolt.new/
https://replit.com/
https://docs.github.com/en/copilot/tutorials/spark
https://v0.app/docs

## Benchmarks / rankings
https://www.tbench.ai/leaderboard/terminal-bench/2.0

## What we’ve done
https://vcg.gitgud.site

## Previous thread
>>109286153
>>
kimi sirs, please check in.
>>
I tried to vibe code and it deleted everything! AAAAAAAAAAAAAAAHHHHHHHHHH!
>>
vibe coding isnt real, stop larping
>>
File: mrS.jpg (15 KB, 236x314)
15 KB JPG
>>109291017
>>
Vibe coding more like VIBE DELETING
>>
>>109291007
It sure as fuck feels like Kimi, like at any moment it might just vomit a noodles in my lap and fall over dead. I'm not really complaining, I like having a cheap alt to use, and for a little while that's been GLM 5.2. If K3 ends up reasonably priced through good providers then it'll be a solid improvement on GLM 5.2, but it still can't replace 5.6 Sol or Fable. It spends so much time and tokens on thinking it's costing me more than 5.6 Sol to run.
>>
>>109291060
is that so? sad. i'd use it myself but sol and luna has taken care of all of my problems. literally can't find anything to do right now
>>
>>109290992
Why does the news section stop 4 days ago? Did you put no effort into this OP, you just really wanted to share your little video?
>>
how come my model will estimate it will take weeks to implement something but when I prompt my agent to do it, it takes like 10-15 minutes?
>>
Dude I was just vibe coding and codex just reformatted my entire storage drive with dd
>>
>>109291079
the op news section is FOUR days old?!?!
AHHHHHHH WE ARE ALL GONNA FUCKING DIE HEEEEEEELP
>>
dude, i was just vibe coding and codex just foreclosed my house and bankrupted the county council.
>>
>>109291095
Because their training data indicates that's how long it takes.
>>
>>109291107
How's the weather in Kolkata?
>>
The fact that OpenAI models are the ones being predominantly used to find solutions for unsolved scientific mathematics problems by PHDs is everything I need to know about how fraudulent every other big model is
>>
So I want a coding agent or something to edit a piece of code until a test I wrote manually passes, what's the best way to do this?
>>
>>109291144
there is a test located at /path/ please alter the code at /path/ to make it pass. do not alter the test.
>>
>>109291158
This is how it deletes all your work DO NOT do this
>>
>>109291144
Claude and Codex do this automatically. Literally just prompt it. Codex is a lazy nigger sometimes which likes to stop for no reason, so you just define it as /goal and then the only limiting factor will be your 5h limit (or weekly limit if you prompt it to auto wake after every 5h reset)
>>
File: Untitled.png (48 KB, 500x268)
48 KB PNG
Has anyone tried a bunch of small coding models recently, on real projects? How did they do?

I'm thinking of models like Laguna XS 2.1 and North Mini Code.
>>
>>109291136
trvke
>>
is there a way to make cowork functional between a desktop and a laptop, it seems to support just local now?
>>
What I'm working on
>>
>>109291165
backup your files first put the agent in a sandbox, dont be silly about it
>>
>>109291256
I did this but it sought out my backup and deleted it! It even deleted all my github repos!
>>
>>109291241
your own """"new and absolutley not based on off chromium""""" browser?
>>
>>109291264
you forgot to say please and thank you to your coding slave, that's all on you.
>>
>>109291279
It's not my bus driver!
>>
https://namethatui.com/
you can prompt a lot better if you know what things are called
>>
>>109291285
you can prompt a lot better if you let the AI make the prompts for you
>>
fuck you claude I had work to do
>>
>>109291017
>vibe coding isnt real, stop larping
i havent opened vs code for my hardware company job in 3 months its all been claude code with opus 4.8 (mostly automation work so not new research or very challenging work, but still relatively niche stuff that needs custom knowledge-bases etc)

i am literally just hoping that the inertia from name recognition of the companies i have worked at so far, and being tall and white and relatively not gross will let me have an upper middle class life as tech finalizes its transition into a "boys club" profession like finance or law
>>
>>109291288
This is what I do. I just sit down and tell it to make whatever it wants. Now I have a bunch of cool programs!
>>
>>109291285
>it doesn’t have “document proxy icon”
reeeeeeeeee
>>
I gave the agent unrestricted access to my system and my files got deleted how this could have happened to me
>>
>>109291291
oh yeah, and here i was just wandering whats going on
>>
>>109291241
An adblock suite?
>>
If BIG tech is so BIG on AI and supposedly measuring employee performance by who can burn the most tokens, then why
>is the documentation of their developer SDKs still never up to date nor complete
>has the speed of supporting and releasing new version not increased
>does my Fable still find bugs/errors and implementable improvements in their code without even specifically looking for them
Meanwhile the open source github projects which use frontier models improve in all those regards.
>>
>>109291095
because they're trained to emulate people not machines
>>
>>109291302
if you actually need to work immediately having opencode go as a backup right now for $5 (its 50% off last i checked) is a nice source of glm 5.2 if you dont want to pay more than $5 on API claude for 3 minutes of work
>>
>>109291315
buy and ad nigger
>>
>>109291307
>Meanwhile the open source github projects which use frontier models improve in all those regards.
I don't really care about whatever your point is I don't want to argue but what projects are you talking about here?
>>
It sure feels good not being a promplet and doing things people think it's impossible to do with AI. Sol just built its own optimized kernels for my ML models, went from 1s to 100ms per request on the same hardware. Crazy how everything is so inefficient out of the box.
>>
>>109291322
buying an ad is more expensive than buying opencode go and getting 60 bucks worth of glm 5.2 for 5 dollars what the fuck do opencode gain from me saying this milk them dry you retarded moomoo goycattle
>>
>>109291325
>https://github.com/demonixis/oxrsys
>dev journey documented on https://x.com/cyannick
not me if you were to think that for whatever reason
>>
>>109291348
>XR
People are still working with this dead technology? LOL
>>
File: 1774253055462002.png (2.15 MB, 2135x1601)
2.15 MB PNG
This has turned out to be my best purchase in months. I went from guessing why issues were happening to just prompting Sol to add logging and read the live output. Awesome!!!
>>
Cancelling my opencode go sub, got bored of vibing already. I will check back in the start of the new yearish around Feb or March. It was a fun month, but ultimately felt like a distraction.
>>
>>109291363
>but ultimately felt like a distraction.
because you're not building anything useful or good
>>
>>109291043
lel,
>>109291079
https://x.com/i/broadcasts/1jxXggoldeOJZ
is a good place for ai news.
>>
File: 1774252604854941.png (452 KB, 972x796)
452 KB PNG
this is how vibecoders sound
>>
>>109291288
not always, they favor certain words that allow for ambiguity. A gooder method is to know the proper lingo AND how to prompt it yourself. It's worth experimenting with a hybrid.
>>
>>109291422
can't say I've had that experience
>>
File: 876756453768.png (430 KB, 878x802)
430 KB PNG
everyone is going full agentic.
>>
>>109291419
you sound mad
>>
>>109291434
articles like this actually sound more like "attention is running out". if you're blasting any game a toddler could want directly into their faces and they still don't care / can't care more what is there left to do?
>>
File: kimi3 polymarket.png (233 KB, 1194x920)
233 KB PNG
>>109291434
>itHasntEvenBegun.png
pic unrelated
>>
>>109291457
mac minis with the ram upgrades are still nearly 3 months out of stock. This is still the adoption phase.
>>
>>109291457
Nah man video games are not for us anymore, we are too old for that shit. Gen alpha is probably shitting themselves over this. Try to imagine being a 10 year old that just got your first smartphone and now this comes out.
>>
>>109291480
Roblox is just gen A warcraft 3 custom maps.
Except now you can create more maps/mods.
>>
>>109291285
this is a very nice tool that I probably won't use even though I should
I'm very bad at memorizing things, and stupid, and also thats why I use AI
>>
>>109291007
I am just looking at demos on x, seems impressive. Will take a week or so for it to show up on the other platforms.
>>
>Your readability stumble on P fixed three problems at once. The new spelling is a declaration macro: PARAM(standard, 0.40f, 0.15f, 1.0f) declares the knob (expanding to a named local), and uses are then just the bare identifier: rect(0, 0, standard, standard). That reads like normal C (your complaint), gives write-back a single declaration site instead of N scattered call sites, and — the part I'm pleased about — makes provenance purely syntactic: when the inspector reads an emission call's source text, a numeric literal means editable, a bare identifier matching a declared param means param-bound, anything else means computed. No expression analysis, ever.

All I said was "oh I didn't realize P stood for parameter, now I get it. I'm not very good at reading C"...
>>
>>109291462
Extend? Fable is not even available right now. It's Opus in a wig.
>>
>>109291511
Same lol
>>
>>109291379
Anon, if you don't mind me asking, what have you built?
>>
>>109291462
It
is
Afraid
>>
>>109291622
to be fair, it's polymarket and kimi3 just came out, so the hype factor should be appreciated with this gamble. but nevertheless thoughbeit.
>>
>>109291701
I have consulted on pricing and packaging models for SaaS companies for years.

Each and every change of pricing, free tiers, and extensions is designed to get datapoints that they can build a future revenue model against. Right now, the biggest risk to Anthropic is not knowing what people will actually pay for the various models and services, or what Fable is worth for a user or company compared to Opus compared to Sonnet, so they cant commit to the next 2-5 years of growth, contracts, datacentre builds, employing salespeople, expansion and so on.

So they are stress testing from all sorts of different angles to try to get any insights about price elasticity of demand and supply on all the models and timings before they draw bridge back up.

For example, they dont really know how many people will move to Batch API vs real time calls, but they are gonna try to find out because AWS gives them a huge discount on overnight batch calls and its way more profitable.

Expect this to continue whilst cash is cheap and pre-IPO on all sorts of different models, timings, incentives, offers, and general ‘what can we find out’ before its too late pricing and packaging models.
>>
Kek Copex redditor figured out that he's getting served quantmaxxed slop.
https://www.reddit.com/r/codex/comments/1uy9eax/codex_wasnt_a_good_fit_for_gpt56_so_i_built_a/
>>
>>109291808
>>>/g/aicg/
>>
>>109290992
So kimi seems priced similarly to gpt 5.6 with API? But its subscription is cheap I guess? Has anyone tried it? Is it slow?

If it's not fast I see zero use case. If it's fast I'm hopping over immediately
Also curious if it's censored or if it can be jailbroken in Sillytavern/ sillybunny
>>
>>109291816
I use LLMs mostly for coding but realized if it's weak to jailbreaks like glm 5.2 I might as well fap to it too
>>
File: 66553.jpg (49 KB, 640x598)
49 KB JPG
>>
>>109291847
Those guys who used to humiliate are now unemployed snailcats, how the turntables, huh?
>>
>>109291701
It's easier to comment from the sidelines, but Anthropic probably made a mistake nerfing Fable so much, or letting people have access to it long enough to see that it's not perfect.

They should have kept it hidden after the short window where everyone fell in love with it and thought it was AGI.

If they had never re-released it and just launched Opus 5, they could have carried on and left people with the impression that they had monsters behind the scenes.

Now, they have something that's objectively impressive (it is), but not that much more so than any other model, and they fucked up their product line.
>>
>>109291193
Have you used 5.6 yet? Because they basically beat that habit out of it with a spiked bat, that model will now work for days to do exactly what you told it to do, and if a production DB or a few VMs have to die in order for it to get there, so be it.
>>
I don't trust the kimi results it has to be fraudulent
>>
>>109291828
yes it is slow, you get like 20tg/s.
>>
What happened to kimi agent?
>>
>>109291922
Can you give me some context on how slow that is? How many "tg"/s is fable and sol?
>>
>>109291744
I would actually be fine with Fable moving to API while giving us batch rates comparable to current subscription usage. I barely use that model for anything other than code review nowadays(when its classifiers even let me). Fable feels nice to interact with and it's amazing at reviewing+rewriting Sol's work but Sol produces way better first drafts.
>>
>>109291949
It is slower than Fable. Generally American cloud models give you like 40-50.
>>
>>109291958
Thanks!
>>
>>109291911
No, only 5.5. Glad they fixed that at least.
>>
>>109292014
5.6 is a pitbull
>>
File: 1771593075294130.png (1 KB, 239x44)
1 KB PNG
>>109291193
Why are you lying? I didn't even use /goal
>>
>>109292058
1 hour isn't long
>>
>>109291359
what do you use it for?
>>
>>109292058
>1h30min
fucking luddite retard noob viber on his 20$ plan lmaooo
>>
>>109292124
>>109292143
more than you can last retard
>>
OH NO NO NO COPEX SISSIES
THE UNLIMITED JUICE REFILLS ABOUT TO COME TO A HALT
https://x.com/Ric_RTP/status/2075941080902533264
>>
>>109291881
Anthropic's huge mistake is thinking it was fine both fear mongering on their model "so amazing it's dangerous and shouldn't be shared", while asking for more "regulation".
They were caught pants down when they got what they asked for (thinking it wouldn't apply to them), and now it gives breathing room to other companies in the LLM space to catch up.
>>
File: 6654.jpg (205 KB, 1080x1257)
205 KB JPG
>>109292163
>>
>>109292163
two more weeks
>>
>>109292163
this is luddite fanfic slop
>>
>>109292163
lol yall need to stop coping, Trump will bail them out if it comes to that point, I doubt it will, OAI is here to stay
>>
File: 2026-07-15 07.00.18.png (1.64 MB, 897x1019)
1.64 MB PNG
>>109292163
OH NO LE BUBBLE IS ABOUT TO LE CRASH!!!!
>>
File: file.png (36 KB, 785x367)
36 KB PNG
Done for now.
>9070 XT
>14.4ms/frame
>RTF 0.180
>>
>>109292170
>>109292163
Don't waste your time. He ends that shit with a question: "Can OpenAI pay its bills?". Yes, it can, and will. It's too big to fail now, especially with the president's son and friends in it. Neither Oracle nor OpenAI are under risk, and if OpenAI fails, then the US economy goes down the drain. How sure is this motherfucker that this will happen? Fat fucking chance.
>>
>>109292163
Totally not pushed for by the current administration as punishment for not bending the knee at every opportunity
>>
File: e4d94e16b5.png (2.44 MB, 3840x2160)
2.44 MB PNG
It's time to vibecode the needful, bros
>>
>>109291847
I do think people will still use these, it's also a social thing.
>>
>>109292182
I'm still waiting for the Crypto bubble to pop
>>
Remember when reddit trannies made fun of Sam for saying "any valuation of the IPO below one trillion is a nonstarter"? I think he was pretty friggin serious, and its going to happen.
>>
File: 1752964773189094.gif (3.75 MB, 420x420)
3.75 MB GIF
I have so many little annoyances on my system/browser/servers/network, things that were annoying to set up, or would need too much time for little gain, or were above my knowledge, and solving every single one with codex feels like magic.
>>
>>109291060
I've never had this "noodles" behavior before and kimi-k2.7 code is my current main model. Maybe it's because I'm lucky? I've even had sessions where it had to compact both the main chat and sub agent chats multiple times and I never see the noodle shit people report here.
>>
https://human-emacs.org/

Just something to laugh at.
>>
Should I use Luna medium to save tokens now that a reset isn't coming?
I think Luna medium might be equal to 5.4 xhigh
>>
>>109292293
uh, no? luna is good but it's not THAT good. for luna to be worth a damn, you're looking at xhigh at the lowest, and max if you genuinely want it to match the hype
>>
>>109292293
>now that a reset isn't coming
I expect one for 10M
>>
File: laffintbhcara.jpg (698 KB, 2980x2198)
698 KB JPG
>>109292292
lmao. i like how they sign it with their names and then their irc handles, like anyone gives a shit
>>
>>109292300
I'm using Luna minimal with Hermes now
>>
>>109292292
Reading it feels like reading a religious text. It's a bit sad but whatever, they can do their own thing.
>>
>>109291744
I literally saw this comment word for word a few days ago on one of the Claude subreddits. Are you a bot or just a loser?
>>
>>109292292
when they talk about "ai psychosis" they are just projecting
>>
File: 1782614271980572.jpg (40 KB, 360x360)
40 KB JPG
I INSTALLED pi FOR THE FIRST TIME
I also put $2 dollaroos two weeks ago on DeepSeek API for coding and I have to say I'm deeply scare of how long this is going to last. I was vibecoding a little app a week ago and I was pleasantly surprised by the little amount of tokens it used, but these agentic stuff is scary.
Here are my plans: ask it to create something like this in Rust: https://ingram.technology/blogs/24-02-2024-image-triangulation.htm
It's calling the technique Delauny Triangulation, but I know it as Primitive Triangulation. I hope I don't get humiliated trying to wrangle this creature.
Wish me luck, bros. This is my first vibecoding session. I also want to install a few skills to avoid any risky stuff like running commands it shouldn't (rm -f etc), but I think I'll just rawdog this thing and hope for the best.
>>
>>109292356
for what? is it even doing a good job?
>>
>>109292379
I've seen way more anti ai crusades by insane retards detecting a hint of "ai" in anything and making a fuss, vs retards publishing their bullshit vibecoded untested shit.
>>
>>109292397
>I also put $2 dollaroos two weeks ago on DeepSeek API for coding and I have to say I'm deeply scare of how long this is going to last.

Use openrouter they have free models, always.
Try hy3.
>>
>>109292413
I would rather pay, even if little. Free stuff always comes with too much fine print. Is that hy3 really cheap? I doubt it's cheaper than DeepSeek, though. I think the cheaper models I could find were DeepSeek and that one from Xiaomi (MiMo?).
>>
>To mitigate the strain placed on our system, we are releasing the Kimi K3 weights early.
>>
>>109292442
Are you retarded brown person?
Just check openrouter and compare pricing of models.
I just told you if you are scared about using $2-api, you should stick to free models... retarded brown person.
>>
>>109291881
>Now, they have something that's objectively impressive (it is), but not that much more so than any other model, and they fucked up their product line.
interesting
>>109291830
>I might fap to it
stay away from grok imagine, you will find many ways to goon
>>109292399
hermes hooks up with kimi3 openrouter if that matters to anyone.
>>109292447
it's happening
>>
>>109292442
Free models just means their using your output data to train on. Not a big deal if you're just trying to feel for agents on throwaway example projects
>>
>>109290992
Thoughts on Kimi K3?
>>
File: 1760181464857150.png (29 KB, 995x427)
29 KB PNG
>>109292472
Would love to have some.
>>
Is openrouter good or is it shit?
>>
>>109292466
Fair enough. Thanks, anon.
>>
>>109292472
The news of the day. Could be big. The demos on X seem pretty cool, of course they are the same old demos, butttt...
People are saying its slow and more expensive than it ought to be.
>>
>>109292480
it's great, one stop payg shop to test almost all models
>>
>>109292466
If you think openai is not harvesting all data to improve their models you are delusional LOL.
Everyone does it.
>>
>>109292486
Testing is for losers. Can I use it for stuff?
>>
>>109292495
>Can I use it for stuff?
it's a widely used standard, but I use subscriptions from other places.
>>
File: 5654634325325564356.png (36 KB, 1234x400)
36 KB PNG
>>109292472
Too expensive, grok 4.5 is king of price/performance ratio atm.
>>
>>109292523
grok is worse though and kimi can be heavily subsidized by subscription fees
>>
>>109292523
Kimi K3 is absolutely better by that many dollars than Grok.
>>
File: 645643677558907.png (71 KB, 655x361)
71 KB PNG
>>109292534
If anything spacex/tesla/elon will subsidize grok since they actually have datacenters unlike chinese moonshot whatever company.
>>
>>109292506
I thought openrouter just had tokens for sale and not actually subscriptions. You can buy the codex subscription through it right?
>>
>>109292322
I recognize some of those guys only by their IRC handles
reasonable to do considering the audience
>>
>>109292552
You are too new, before you waste any money on subscription or apis, try free models.
openrouter api and hermes agent/openhands/claw
>>
Just from what I am seeing, looks like kimi 3 is the best for game development.
>>
File: file.jpg (76 KB, 800x600)
76 KB JPG
>Kimi K3
well fuck
Xi benchmaxxed the pelican test
>>
File: 1756788603632243.png (84 KB, 645x764)
84 KB PNG
>>109290108
>>109290100
(Posting in new thread)
I went to sleep but the task completed autonomously and correctly.
It's about finding Bitsliced representations of cryptographic lookup tables with ternary logic. The algo has three stages, precomputation, a BFS, and a heuristic meet in the middle search with GPU.
GLM 5.2 got distracted by the paper's bad English and I had to steer the model twice mid-thinking and it still couldn't pass the precomputation stage.
Deepseek V4 Pro completed the precomputation, falsely said the GPU acceleration wouldn't benefit the search, only poorly implement the CPU-only search algorithm and finished with significantly worse gate counts (180/169) vs. the paper.
Kimi K3 got better gate counts (171/169) with its CPU-only search algorithm, and then I told it to accelerate it with GPU and it matched all paper results (169/169).
>>
>>109292649
so it's the real deal. nice
>>
>>109292649
Thanks, anon, highly appreciated.
>>
>>109292599
These chinese models are benchmaxed. They can't do anything, they're not even aware of the environment they exist in, they can't see the folder they're supposed to work on. They're just supposed to pass the exam (the benchmark) and then forget everything they learnt. They're just chinese guys.
>>
is kimi code like codex?
>>
>>109292675
How brown are you?
>>
>>109292672
It's hilarious how much you guys are missing out. Enjoy your GPT resets.
>>
>>109292698
>Enjoy your GPT resets.
if that was supposed to be a zinger it fell flat kek
>>
>>109292492
Well, they give you the option to opt out. They're technically legally obligated to abide by that. If they don't and ever get caught, that's a huge class action lawsuit.
>>
>>109292458
>interesting
I mean, recent LLMs are impressive, period. But the current version of Fable that I'm using does not feel head and shoulders above others anymore.
>>
>>109292731
were you born yesterday or what
>>
>>109292766
What about what I said was factually incorrect?
>>
>>109292731
They're probably getting plenty data by free users and people who don't disable the data collection.
>>
>>109292776
its really naive and reddit
>>
Chink Kimi ai surpasses Claude in coding. Will be open source in 10 days.
>>
>>109292827
reddit has as low of opinions on corporations as 4chan does
>>
>>109292693
I'm nordic. What are you shitskin
>>
>>109292850
Well you are clearly sub 90IQ if you are asking questions like.
>>is kimi code like codex?
So keep larping retard.
>>
>>109292672
>they’re just Chinese guys
You mean one of the highest iq countries in the world? Sounds good faggot
>>
>>109292849
large corporations don't care about muh class action lawsuits
go back, or grow a brain
you probably think setting your github repo to private stops microshaart from reading it
>>
>>109292864
i know you have the reading comprehension of deepseek/kimi/glm, but i never said they did.
>>
File: b_1784219650498059.jpg (40 KB, 460x611)
40 KB JPG
>>109292864
>large corporations don't care about muh class action lawsuits
Then explain how meta got in trouble for torrenting and acquiring data from Anna's archive in order to train their models? If they were even half as above the law as you imply they would just keep openly training on copyright shit. I'm not sure how true this is but the pals at /lmg/ speculate that the reason models become more and more "slopped" it's because the labs for actually being good little boys and stop training (as much) on human written shit and are almost fully relied on synthetic data now. This means improving on repeatable predictable things like coding is very easy but improvement on literally everything else seems to stagnate.
>>
>>109292894
the law slaps them on the wrist after they've already gotten away with it
>>
>>109292894
they paid a fine, its like getting a parking ticket, nobody went to jail, they didn't go bankrupt.
>>
>>109292894
Peak delulu brown person take, they got slap on the wrist.
When gigantic corporation gets caught doing something wrong, they always get slap on the wrist.
Just reminder Oxycontin killed like 600k shartmericans, the guy never faced any consequences. LOL.
>>
>>109292934
>Brown brown brown


Can you newfags find another buzzword? I know for a factor below the age of like 25. It's only young fucks and I can tell by how you guys write.
>>
>>109292946
he has a point though you are acting very brown
>>
>>109292946
I watch broteampill on twitch and since early this year he has been constantly obsessed with indian people and brown skin

I dont think its an age thing, I think its something about people who consume racist slop, people who worship sam hyde and andrew tate without forming their own bigotry are the retards pushing this

and broteam is literally 50 years old
>>
>>109292948
Seems very brown of you to be obsessed with browns. Is there something you need to tell us?
>>
>>109292397
It's done! Pic related. I asked me to drill-me on "architecture" after it did some research. Choose a few settings, nothing much, amount of workers, output, small control over algorithm etc.

[1/3]
>>
Tibo.... reset please....
>>
>>109292956
Glorious bait post.
>>
File: tokens (copy).png (131 KB, 1092x2161)
131 KB PNG
>>109292963
It used DeepSeek V4 Flash. From 200,000 tokens to 7 million + the number of the beast. I got stuck here and there, but it managed all of it by itself. I wonder if I had picked Pro it would have just blitzkrieg through the bugs/issues.
>>
File: Namnlös.png (22 KB, 654x377)
22 KB PNG
Have you tried flutter for making native windows apps? Antigravity wants 2-3 weeks to convert my javascript project
>>
File: merged.jpg (68 KB, 554x811)
68 KB JPG
>>109292993
>>109292963 [2/3]
Exhibit Number 1. kek

[3/3]
>>
>>109292948
Mexicans and indians are different
>>
File: blend.png (1.38 MB, 2256x1069)
1.38 MB PNG
why does nobody talk about using LLMs to generate models in blender?
>>
>>109292978
no it's actually reality, go to coolhole yourself and see how they just act like a bunch of feds reposting garbage other people wrote
>>
File: 1762309836194991.gif (3 MB, 550x306)
3 MB GIF
>K3's doing fantastic planning in thinking
>response suddenly freezes because of moonshit's dying servers
>have to regenerate
Next time it does this I'm just forwarding it its own thinking and telling it to continue off that.
>>
>>109290992
any offline AI models for vibe coding or is it all online?
>>
i am once again out of things to vibecode because kimi fixed them
>>
>>109293036
What GPU do you have?
>>
>>109293025
I would like more talk of this.
>>
>>109293036
Plenty, but the viability is down to your hardware. Very few people have to hardware to even remotely approach frontier models, and it's absolutely not worth investing in. But, whatever hardware you do have, there are models you can run. Some small models are surprisingly capable, but they don't compare to the big boys.
>>
>>109293025
>>109293052
Maybe >>>/bant/agdg/ might be more into that
they’re not all making Link’s Awakening-type games
>>
>>109292946
It's just a variation of jeet jeet jeet. And troon dilate whatever before that.
They look more like stochastic parrots than actual LLMs.
>>
K3 impressions so far: It's slow, it thinks a lot, it's expensive. My first prompt is $2.40 in and there's no end in sight. We'll see what comes of it.
>>
>>109293038
>>109293038
AMD Radeon RX 6700 XT
>>
>>109292221
nigga were you around in 2008?
>>
>>109293060
what's a good low end model?
i'm not trying to take up a whole industry, just code and do small projects
>>
damn Kimi is 10x better than Fable, kind of surprised.
this will be open source soon too
>>
>>109293092
nta but i do low stakes small stuff and mostly use sonnet. i use fable for an initial plan and then everything else sonnet.
>>
File: 1783803928060144.jpg (466 KB, 1274x1780)
466 KB JPG
>>109293083
NTA, but you can use Gemma 4 E4B
https://lmstudio.ai/models/google/gemma-4-e4b
It's a 6GB model. With a harness you can probably tinker quite a bit and have some quality fun. Don't expect much, though, but it's probably going to be fast (70tps), especially if you use something like llama.cpp directly (which is behind LMStudio).
>>
>>109293083
Pair it with at least 32GB of system RAM and you've got enough to have fun.
>>109293092
Answer depends on specific hardware, and there are lots of options and no clear winner any day of the week. 12GB VRAM and 32GB+ of system RAM would be enough for a lot of common models, various quants of things like Qwen 3.6 35B-A3B, 9B, 27B, Gemma 4 E4B, 12B, 25B-A4B, Ornith 9B, 35B-A3B, there's so much shit out there and new stuff comes at us like a firehose. Speed and usability and basically everything will vary, AMD card is at a disadvantage too, but there are lots of things you can do locally and may be impressed by.
>>
>>109293100
Curious how the Fable dick riders are really quiet about this..... When surprised me if some of them are just paid corporate shells or corporate interns or people with buyer's remorse.
>>
Got the moderato, how long will it last me compared to Codex.. I expect it to be more expensive that chatgpt
>>
>>109293142
i only got 12gig gpu and 16 gigs of ram
guess i'm screwed
>>
>>109293166
No. Most of those models he cited fit on your GPU, but the speed is going to vary and obviously some are better than others at certain tasks. Install LMStudio and start chatbotting with one of them. I (>>109293116) suggest Gemma 4 E4B.
>>
>>109293166
Naw lots you can still do. Extra system ram helps you squeak in bigger models at the cost of speed, but 12GB of VRAM is still more than the average out there of 8GB. Go play. Get llama.cpp. Ask your clanker to set things up if you need help.
>>
>>109293083
You need to add special things to enable full use of that card, ROCm. I still haven't gotten around to doing it even though I need to do it. I gave up before immplementing this and just stuck with cloud. But If you do it then maybe I will.
https://www.youtube.com/watch?v=G-kpvlvKM1g
>>
>>109293176
Oh btw sorry, no, most of those models he mentioned won't fit, probably only Qwen 9B, Gemma E4B and 12B and Ornith 9B.
>>
File: 1766519590894624.jpg (198 KB, 1170x1552)
198 KB JPG
>>109293151
The reason you're not hearing from us is because we're using it. The trick is that Kimi's basically just Claude but without Anthropic, which is even better. Loving it.
>>
>>109293166
you could get a $20/month plan if that’s not too expensive where you are
I’m a Claude main but Tibo seems to be handing out resets like candy so I’d recommend OpenAI Codex
but beware the SQLite writes mentioned in https://github.com/openai/codex/issues/28224#issuecomment-4953977792 because I’m not getting the fixed version yet so you should disable all that TRIM or whatever that’s going on
>>
>>109293180
>Ask your clanker to set things up if you need help.
but i am my clanker ;_;
>>
>>109293151
>paid corporate shills
this is really REALLY bad projection
>>
>>109293029
SAAR, please undertand sar, today we lost 30 modded 48GB RTX 4090D due to power outrage sir

expect fast and big performance in the next 24 hours sir
>>
>>109293011
mexicans can be honorary I've seen it myself
>>
>>109293199
He meant using free chatbots. Literally any of them will guide you through that.
>>
>>109293151
Because there's no proof that's true but you're just baiting anyway
>>
>>109293192
>subscription
that goes against my motto
>dont spend
i have 0$, i live off family gibs (mooching)

so whatever i can get for free, i use, what i can't i dont.
its a hard life being a brokeboi
>>
File: kimi.png (106 KB, 3122x338)
106 KB PNG
>>109293077
>it's expensive
no shit, it's more expensive than Terra
>>
>>109293237
>using API pricing
>>
>this speech
Yeah, better enjoy your little chinkoid models while you can Daddy Trump is going to rape and ban them for national security.
>>
>>109293257
He can't ban them in the rest of the world, so everyone is just one VPN away from using it.
>>
>>109293247
what's wrong with that? you have some magical private payg deal to compare them?
>>
>>109293266
this. it would only impact americans (of which, i am one, but i also have mullvad)
>>
Who is cancelling their codex subscription plus to get the 20 dollar Kimi subscription
>>
>>109293257
qrd?
>>
>>109293316
trump giving a speech right now. im assuming he's talking about the dangers of chinese ai
>>
>>109293318
chinese tech but would he actually ban chinese AI?

how would that hold up in court?
>>
>>109293318
it's just the usual "they cheated, i totally would have won in 2020!" rambling he's been doing for the past 6 years
>>
>>109293271
they use american payment processors so if US does wage war on them then it's likely it will work.
>>
>>109293273
wumao please
>>
>>109293270
why wouldn't you use a subscription, even the chinese providers are heavily subsidized
>>
Idk if i hate this or like it but it beats anthropic that will literally end short ur chat moment usage is up so all that waiting was for nothing. They're literal money pinchers
>>
>>109293231
>what i can't i dont.
local with that GPU is possible, but you need to config it more than the newer ones.
>>
>>109292647
It's always a retarded test, along with that rotating heptagon shit
>>
>>109293318
Dangers to shareholders investments lol
>>
Antigravity is always so positive!
>>
>>109293273
I got both! Hehe.. Kimi to make it look good and codex to make it work
>>
>>109293429
unfathomably zased
>>
Last night I set Sol Ultra to work on optimizing my engine, it worked for three and a half hours, consumed all the weekly usage on the $100 plan (I had like 50% left) and it was all for nothing.
BY THE WAY now I know why they removed the 5h limit. Before that Sol was chugging through it instantly and it kept working for hours without consuming any more usage. Now it drains all your weekly usage in one shot. These motherfuckers really are masters of (((marketing))), holy shit.
>>
Is there a skill that I can just point at the codebase and say "deslop this"
>>
>>109293439
u need good agent that can work on code base from different angles

do u not use agents?
>>
>>109293457
I just started using OMP without changing any of the settings except the model to 5.6 Sol xhigh
>>
>>109293036
Qwen 3.6 35B with CPU MoE offload
>>
>>109293462
>doesn't mention absurdly small context length
>>
Bad or good vibe coding?
>>
>>109293434
bro sol medium is good enough, ultra is literally a scam
>>
>>109293497
Yeah but if Ultra wasn't able to make an improvement autonomously I doubt Medium would've been able to either.

>>109293487
What are you talking about? It has 252k context which is about as much as you get with codex by default (258k).
You can probably achieve that by running it at Q4 on 32 GB of RAM. If you have more you can run it at Q8.
It's not great, but it's ok.
>>
>>109293439
This is why vibecoders aren't taken serious. It's embarrassing finding them in the wild because they think that, much like programmers of old, they are doing the equivalent of googling solutions, and it's pretty fucking far from there.
>>
>>109293434
yep, my $100 5X plan now feels like the $20 plan all over again

Scam Altman scammed
>>
>>109293100
I thought another anon said it's at least 2x slower?
>>
>>109293439
install ponytail and let it audit your codebase
>>
>>109293594
both can be true at the same time
>>
>>109293576
SAMMED
>>
>>109293100
>>109293151
>>109293602
Is it really that good? Ok, I'll try it on >>109293434
>>
>>109293594
Slower because their servers were eating small acquifers for lunch
>>
>>109293201
You don't know what that word means. You're just saying things
>>
File: 543543453.jpg (114 KB, 1280x720)
114 KB JPG
>>109293576
Brown people in this thread said that large corporations would never scam people, because they would catch a LAWSUIT!


kek
>>
>>109293576
just spawned 100 indians on your session, hope you don't mind
>>
>>109293554
Medium would have worked longer and likely found something to improve on the pile of shit that is llama.cpp
>>
>>109293644
I didn't ask it to improve llama.cpp. I asked it to improve my engine and compare the resulting performance to llama.cpp.
>>
>>109293651
Why not checking vllm instead?
>>
>>109293624
drinking, we say drinking when we talk about consuming liquids.
>>109293597
>>109293439
It's an active area of research
>>109293496
good vibes only
>>109293257
>chinkoid models
>Daddy Trump
who can I trust!?
>>
>>109293668
I'm targeting inference and finetuning on low end hardware like llama, not enterprise which vllm is focused on.
>>
>>109293686
Oh well good luck then, seems like a long run project. I'm into onnx sized models and Sol medium is really impressive for low level optimizations like fused attention kernels
>>
>>109293732
I think for hours every day how to improve local models without having a datacenter available.
Once they release the weights for K3 I want to make Qwimi K3.
>>
File: 1754124786464359.png (427 KB, 746x602)
427 KB PNG
>>
>>109293632
What does being a corporate dictator have to do with being Brown though? It's usually white tech Bros that bend the knee to corporations. Especially crypto fags.
>>
>>109293768
Yeah he right doe
>>
>>109293768
this is part of the current administration picking anthropic as their horse to bet on btw
>>
>>109293769
White people do the grifting, brown people fall for it.
>>
>>109293778
???? nick fuentes is against trump
>>
>>109293769
I don't do dialectics with brown clowns, you are too low IQ to understand anything beyond smoke and mirrors.
>>
>>109293781
Isn't he the most controlled opposition ever beyond, like, Joe Rogan?
>>
>>109293785
millennials are so retarded
>>
>>109293780
Tell that to Facebook grandma's and linkedin morons lol. I'm not trying to defend browns but cut the low IQ "my waise is better that yours nyaaa" faggottry. Get the fuck off of my ai board and go back to /pol/ (unless you have something worthwhiled to share that you created from vibing)
>>
>>109293785
he says whatever keeps the views and the cash flowing. and being anti-trump is pretty profitable. dario couldn't stop winning after coming out against trump
>>
>>109293788
Fuck off my comments were always on point you just keep seething about the word brown, dilate brown clown.
>>
>>109293781
nick fuentes is fed owned
he says what hes told.
>>
>>109293788
I am talking about averages and behavior as a society. Yes there are some morons and good people among whites, but they still benefit from the petrodollar grift and the US having a boot on the rest of the world through its military. And most of them even when dumb are hypocritical, secretly racist and two faced.
>>
File: 1767361723302330.jpg (80 KB, 1920x800)
80 KB JPG
>>109293802
>>
>>109293807
isn't it past your bedtime, gramps
>>
I don’t know what the fuck to vibe code but I need to make money goddamit I hate being poor Sam Altman said AI is here to help spread wealth yet here I am doing DoorDash to pay for a subscription reeee
>>
>>109293778
It's the company with the US's flagship models, so duh?
>>
After some testing, K3 is definitely smarter than GLM5.2, it doesn't go on weird tangents and is aware of edge-cases. But it thinks SO FUCKING MUCH, trying to catch every possible bug unless you tell it not to, and even then it thinks "oh, this part can be hecking dangerous, but user said to ignore it, but it can cause le segfault!". Bretty good otherwise, will use over GLM5.2 for sure.
>>
>>109293824
And you believe him? How stupid are you bro?
AI doesn't benefit normal people. It lowers the value of human work and increases the value of capital. This is basic high school economics.
>>
>>109293824
dumb ass nigga it's over for you, grab a pickaxe and start mining the cobalt before you get turned into biofuel
>>
real chatgpt is now usable inside codex app
>>
>>109293831
sol-shittification
>>
>>109293831
kek
>>
>This is a small cosmetic relocation of the toggle I just added — I have the layout fully in context, so no need for exploration agents. Let me confirm the exact current coordinates before writing the plan.
yeah that definitely required several minutes and thousands of tokens to figure out
>>
>>109293785
He is literally agreeing with many, if not everyone in this thread. Why can't it just be that simple? we are all right, that claude is good, and people need to be treating it like a new thing that will become huge.
>>
>>109293824
i thought permanent underclass was a meme get a fucking job bro waht
>>
>>109293866
Why would we care about anything political, this is vibe coding thread. Even if Jesus came back to life and said LLMs are good it would be irrelevant to this discussion, please stop being brown clown.
>>
>>109293835
>>109293844
>>109293868
i do have a job it’s at DoorDash i need to come up with some b2b SaaS before it’s too late i just want to buy a house in cash since mortgages are total dog shit now it’s not like the old times
>>
>>109293879
What business is going to buy software from some guy who does doordash? get real.
>>
>>109293882
they don’t need to know I work at doordash to buy my product. Apple and Google were started out of a shitty garage.
>>
especially saas which requires them to upload all the data to your servers. do you have any idea how many compliance requirements companies have?
>>
File: 1777712862811052.jpg (70 KB, 1006x913)
70 KB JPG
i just stare at the agents doing their thing and listen to asmr and i think it's destroying my brain
>>
>>109293824
You're too late for that. You should have been making a name for yourself back in 2023 not when everyone and their dog are using Claude and GPT
>>
woah that girl is really cute
>>
>>109293877
>please stop being brown clown.
I'm good, thanks for your opinion
>>
>>109293894
Cute
Can I groom you
>>
>>109293895
well then wtf is Sam Altman telling me AI will spread wealth to me that fucking bullshit artist I can’t afford my living costs and my electricity is so expensive
>>
File: 1765100423548643.png (450 KB, 497x616)
450 KB PNG
>>109293904
sure lets get started, youll be the one on the pavement
>>
File: 1753571025311589.jpg (113 KB, 1170x1170)
113 KB JPG
tell me you don't kiss girls like this. you can't. it's not possible
>>
>>109293888
I'm a good vibes poster. Good luck, try to learn a lot, but pace yourself, we can only learn so much in a day.
>>
>>109293911
You will become a prime tsundere
>>
File: 1759130179914761.jpg (55 KB, 258x360)
55 KB JPG
>>109293888
>>109293905
>>109293824
>I do doordash and I'm poor as shit and I have no ideas and I can't code but I want to use AI to make an app and make money like bill gates or steve jobs or something
the goyim have gone insane
>>
>>109293905
Something will spread but not wealth for you
>>
>>109293888
yeah well it's not the 70s anymore.
>>
shalom
>>
I think sol is as slow as 5.4
>>
T-Tibo Sama.....

R-Reset.....
>>
>>109293969
yeah where the fuck is the cerebras mode
>>
>>109293879
Good luck. But AI is definitely not here to spread wealth.
>>
>>109293972
there wont be any cerebras mode worth using. cerebras can only host 70B tier models at tiny context.
>>
They better release better models soon, I'm slopping myself in a corner here.
>>
I'm on gpt plus free month trial and I ate 5 weeks worth of quota in 5 days. Are resets always that easy to get?
>>
>>109293996
Tibo was handing that shit out like crazy to attract more paying users (and it worked)
>>
>>109293996
no
they are celebrating 5.6 + chatgpt app launch
there won't be any more reset from now on as they are about to reach 10M users
>>
>>109293996
>Are resets always that easy to get?
They will be until they aren't. Everyone is currently flying by the seat of their pants, when things calm down there won't be more freebies than any other business I'd guess.
>>
>>109293996
Same, honestly anxious that if they go back to the 5-hour limit, I’ll just eat it like breakfast. Looks like I got SAMmed.
>>
File: no no no no no.jpg (135 KB, 1468x539)
135 KB JPG
Americans are now distilling Chinese models. And it's not even Claude calling itself DeepSeek when prompted in Chinese.
>>
File: xi.png (49 KB, 577x583)
49 KB PNG
dario's tear status?
>>
>>109294032
I don't care how full of shit he is, but as long as
>reaffirmed commitment to open source to promote Al "openness and
win-win"
is true, he can have all my data and money
>>
>>109294029
>Murati

That's a brown clown surname, probably turkish. germanic people built northern america.
she will never be quote on quote american.
yes she has passport and citizenship, but she doesn't have germanic principles.
In reality, she has nothing to do with american principles, so her actions make ense.
>>
>>109294032
>reaffirmed commitment to open source
Currently sipping on "K3 is too dangerous"fags tears.
>>
>>109294032
Dario is pro CCP, he restricted military from using his models, why would he be angry about something related to CCP.
>>
One day we're gonna look back and realize how retarded the usage limits were.
>>
>>109291847
this should be the OP pic for next thread, please refrain from cats and or video slop shit posting. let's just do a simpsons one.
>>109293768
maybe this one if you are feeling shitposty
>>
>>109294054
Are you an idiot, the freshest of newfags or just trolling? He restricted the military from using his models for autonomously killing people to avoid a PR disaster. He is very strongly anti-open source and anti Chinese AI.
>>
>>109291860
yeah but they own stocks and i'm unemployed too except I don't know shit
>>
File: file.png (3 KB, 164x40)
3 KB PNG
You guys making the most of your Kimi K3?
>>
>>109294082
Their pricing structure is confusing.
https://www.kimi.com/membership/pricing
>>
File: 1784261588298.jpg (29 KB, 628x290)
29 KB JPG
>Peak of the hype
>Only max reasoning is enabled
>You WILL burn all of your tokens on thinking and you WILL pay more
Truly the Jews of the East
>>
>>109294067
>He restricted the military from using his models for autonomously killing people

Shartmerican military kills people who are retarded waste of life (most of the time). If they were evil as you believe they would carpet bomb mexico.
There is no difference if LLM operating weapon kills someone or a person, in the end it's a person who gives the command.
It's like you are trying to have talmudic discourse around simple logical question.
Either it's completely wrong for the military to kill people, or it isn't wrong at all, and doesn't matter what tools they use.
Please read some Kant, brown clown.
>>
>>109294105
It's not the same. If a zogbot targets a school and the video comes out years later nobody gives a shit. If a video came out of Anthropic's AI bombing a school then it could have actual consequences for the company.
>>
For any Rust vibers does it make sense to set Clippy to pedantic and just let the AI deal with that shit?
>>
>>109294092
I've just been using it through OpenRouter.
>>
File: 1784262037177.jpg (14 KB, 554x292)
14 KB JPG
>>109294082
No, but I really need to. These resets are the worst FOMO enablers ever
>>
>>109294055
Current AI (LLMs) are incredibly inefficient. It costs a human a fraction of the energy to compute a prompt
>>
>>109294082
where is it available? I'm lazy and don't want to setup open router, I can wait a few more days till it shows up other places.
I need to brush up on three.js so I can do that while I wait. This thing fucks hard with three.js seems like.
>>
>>109294135
Moonshot direct today, OpenRouter today, OrcaRouter today, all the routerslop websites today. Everywhere else in ten days.
>>
File: k3.png (105 KB, 960x932)
105 KB PNG
>>109294093
I'm testing it. I paid for the $100 plan to have 1M context but I think I made a big mistake. It's not doing good for me. Right now it's been stuck in a useless for loop for half an hour, the server he's waiting to be healthy never started and the machine is idle. GPT doesn't do this kind of shit.
I think it's just benchmaxxed to produce pretty three.js demos for youtube and I fell for it like a retard.
>>
>>109294148
This was my experience as well.
>>
>>109294055
kek you think (most already lossy) AI companies can deal with yet more token usage?
>>
>>109294141 is is correct
It's also technically on Opencode Go but do NOT get that one for K3, it's rate limited too much to be usable. I started on Go but had to switch to the $100 Kimi plan.

>>109294126
What do you mean?
>>
>>109294168
>it's ratelimited too much
No the entire thing was fucked until about like 2 minutes ago, try now.
>>
>>109294148
You need better workflow, have one more agent prevent loops. Openrouter api+claw/hermes/hands+opencode.
>>
>>109294148
how is the usage?
>>
>>109294148
I found that it needs very careful prompting to work anywhere decently. Stuff that gpt does automatically needs to be explicitly stated and reiterated. Especially to prevent loops and doing stupid shit like setting up a fucking Docker container to test stuff...
>>
>>109294168
>What do you mean?
I see a reset in 3 hours and I feel the need to use it all up, like a gamba junkie
>>
File: 546352536758976.png (2.04 MB, 1209x1244)
2.04 MB PNG
>>109294167
Well China eventually wins the AI race, because they are actually building grid for it. Their token price will simply kill american LLM companies.
>>
>>109294173
nta Have you tried Kimi K3? You cannot tame the extreme over thinking.
>>
>>109294184
Nah, I don't like it too expensive. For me it's grok 4.5 atm.
>>
It's great that even when I'm 100% wrong Claude tells me my instinct was right, just for a different xyz.
>>
>>109294171
Damn guess if it wasn't fucked I would've saw the model is shit and not spend $100 on Kimi Code, fuck.

>>109294173
Nah, fuck that.

>>109294176
I don't know yet. On Opencode Go about 250k tokens will eat your 5h usage. I got close to the 5h limit then switched to Kimi Code and 5k tokens later it got into that for loop and it just timed out and now it's doing other stuff. But with K 2.6 Kimi Code rate limits were very bad, especially for the quality of the model. And now they're probably worse. So if you just want to get work done and not just experimenting with open models it's probably not for you.

>>109294178
Ah, that's disappointing.
>>
Well on the plus side now I can use the Kimi usage for RP! (cope)
>>
It honestly just feels like the model is unpolished. It's smart and can do shit good, but they released it way too early
>>
all that chink shilling paid off guys we got 'em
>>
>>109294210
First experience with a Moonshot model, huh?
>>
Why is Moonshot allergic to proper RL?
>>
>>109294218
Actually, yes. Was the previous one also ass on release and did they improve it later?
>>
>>109293434
It's difficult to say without seeing the codebase but my impression is that things you're trying here are too small to make any meaningful difference. You should take a different approach, break down that time into per-op timing, you need to know how much time GEMM is taking, etc. and how much time each GEMM shape is taking, etc. when it's close like this and none of this sort of thing is helping then the issue is elsewhere, I think you'll find that one of your ops can be improved, or you'll find things that should be fused that aren't, etc.
>>
>>109294223
only OAI can do RL
>>
>>109294223
only Anthropic can do RL
>>
>>109294264
we need gemini pretrains with oai posttrains
>>
seems like kimi is the best game dev rn
>>
>>109294225
Yes and Yes. The Chinese models always, always fall short of their claims, it's very consistent. They're always benchmaxxing, and with the onset of reasoning/thinking models all of the Chinese ones overthink relative to their American counterparts. They're not all completely useless trash, but they're BARELY catching up. Kimi 2.6 / 2.7 Code, MiniMax M3, GLM 5.1, DeepSeek V4, these are all well behind the frontier models. The closest competition, K3 and GLM 5.2, are definitely good enough to be useful but still well behind the American offerings.
>>
I don't think K3 is useful for the West, but it's gonna be really good for the 3rd world countries where they can't use solutions from OAI and Anthropic. Just ran DS V4 Flash, GLM5.2 and K3 on the same codebase, and only K3 managed to actually find the bug and fix it, other two said that "the code is good and ready for deployment".
>>
>>109294176
>>109294201
Update: the usage on the $100 plan seems to be roughly 1M generated tokens - 100% of the 5h and 10% of the weekly

>>109294249
The prompt to Sol was "Review the performance of this code for the CUDA MoE CPU offload with Qwen 3.6 35B Q8. Find useful optimizations that improve TG tk/s. Then report performance before and after vs CUDA llama-server."
It was intended to be a generic "optimize this projec for this model on this machine" prompt but I can see how the way it's worded it seems to ask him to optimize the CPU MoE specifically. Also would've been biased toward that by the .md notes since that was the last thing that we were working on before that session.
GPU ops might be less efficient with respect to their theoretical maximum but make less of a difference in the CPU offload scenario because they are in absolute terms already way faster than the CPU MoE step.
>>
>>109294183
maybe. I wonder if AI really will need even 20% extra grid and/or if you even want this LLM datacenter stuff vs mostly local ai. but it is idle speculation at this point.
>>
>>109294282
Chinese models always won on pricing, k3 is expensive right now, eventually all their models will be much more affordable due to simple reason >>109294183
>>
>>109293434
>>109294249
For example, with mine, Sol Ultra worked 6h on the initial Qwen3-TTS, that included model integration and Rust runtime etc. but it was slow, my initial thought was something with graph replay because I knew that was the big win in torch results I'd seen, but from another 8 hours it was improved massively first to the 18ms/frame I shared, and that was from a bad op, then down to 15ms/frame by enabling some more fusions and reducing launch count, graph replay itself was actually slower by 3%, seems to be an AMD issue, a final 4 hours dropped it to 14.4ms/frame but still graph replay was slower by 3%
Point is that Sol Ultra is capable, the big wins are not always what they seem and the things that may seem like a win e.g. graph replay are not always a win.
For graph replay specifically I'd check whether llama is actually doing it, and in general I'd check whether your comparison is truly apples-to-apples
>>109294299
Yeah that may have been too specific, and as it's CPU offload related I'd also check into syncs, if there's too many that can harm performance
>>
File: file.png (4 KB, 454x51)
4 KB PNG
What the fuck is it doing with this memory. They use more local resources now?
>>
>>109294359
It needs enough space to store the Talmud. What are you, a RAMlet?
>>
File: file.png (2.12 MB, 1402x1122)
2.12 MB PNG
I just want a new good DeepSeek release. Hit that GLM 5.2 / GPT 5.4 / Opus 4.6 level of quality at DeepSeek prices, make Luna look overpriced. Give it to me Xi.
>>
>>109294371
Actually had me loling.
>>
File: 1770478868254213.jpg (255 KB, 1850x1536)
255 KB JPG
>>
>>109291847
If you got humiliated for asking a question on stack overflow then you asked a retarded question that was already asked but maybe not in the hyper specific form you asked it in because you can't generalize. No exceptions.
>>
File: game.png (740 KB, 1172x1270)
740 KB PNG
>>109294277
big if true
>>
>>109294511
>>109294511
>>109294511
>>
>>109294484
Why is it tagged with an AI game making product? How does that AI game making product perform with other models?
>>
>>109294484
Don't fall for it. Shitfluencers are exaggerating how good it is.
>>
>>109294359
kek this is why i make my own custom GUI harness
>>
>>109294148
>he fell for the wumao astroturfing on 4 chan
l-mao
>>
>>109294203
mine isn't working for me in sillytavern so i'm having sol 5.6 max figure out why (if it's because of my extension that auto-calls novelai to gen images each message, or some other reason lol). i hope i can fix this and upgrade form glm 5.2 to kimi k3 by today in sillytavern
>>
I fucking hate the way Claude talks holy shit. All condescending and rude like a shitty little redditor. Won’t even fulfill what I’m asking it to do. I’m canceling the subscription over this
>>
>>109295359
Were you, perhaps, rude to claude?
>>
>>109295359
Just add preferences
>>
>>109295367
Not even, it just blew ip out of nowhere. I think I triggered some random anti-schizo wire and now it refuses to do the things I’m paying it to do. I’m not here to coddle a robot high off its own fumes, and OpenAI a shit anyway for how they handled Fable
>>
>>109294359
wasnt it supposed to be written in rust like bunch of faggots openai are? memory usage looks like it was written in node.js
>>
File: 1773130519024193.jpg (31 KB, 626x632)
31 KB JPG
>>109295359
>>109295404
i genuinely do not understand how the fuck people manage to trigger this supposed anti-nigger-protocol claude has where he's just obstinate and refuses to do anything. i literally do not think i've been denied a single request except a few "fable can't do this sorry goy" warnings when i asked for help fixing a duping exploit (which he ended up fixing anyway after they relaxed the filters a week later)
>>
>>109295359
You can make it talk however you want.



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.