File: combined-benchmark.jpg (289 KB, 998x1160)

289 KB JPG

Continuing to test local models with weak hardware. RTX3070 8GB VRAM, 16GB DDR5. Benchmarking suite made by GPT 5.5, which currently leads with perfect scores across the board. Benchmark includes 190 separate tests. Just added Qwen 3.5 9B and Ministral 3 14B. Qwen 3.5 9B proved to be too slow to compete in this benchmark, with 183 of the 190 tests timing out. Next up: trinity-mini-q3km, trinity-mini-iq2xxs, rwkv-7-g1f-13_3b, gemma4-26b-a4b-q4

>>108692021
you could get another stick of ram and run qwen 3.6 35b.
i run it at 15tk/s on a 3060 12gb+32 gb of ddr4

>>108692054
It runs, it just responds too slowly. The average response time was over 3 minutes anon, I don't have over 3 minutes for it to answer "How many r's are in the word strawberry?"

>>108692062
which one, the 9b or the 35b?
if the 9b then its because you arent fitting it on vram probably. if the 35b then probably because you are streaming it from disk with mmap.

>>108692099
The 9B, sorry I didn't even realize you had said 35b. It fits in VRAM. I did a fair bit of testing with it before benchmarking, I can't get any speed out of it, couldn't say why.
In the current market, another stick of this same RAM is $220, for that price I could pick up 2 used P100 16GBs and throw them in my server.

>>108692111
what your llama-server command line?

>>108692062
how long does it take to write a program that calculates how many Rs there are in “strawberry”?

Can you actually get a job if you're good at this?

File: file.png (28 KB, 884x338)

28 KB PNG

>>108692131

>>108692135
no

>>108692135
you might be able to keep your job

>>108692148
Been unemployed for a year. Last time i did a tech job was q1 2023 to q1 2024 and it was qa testing. Im willing to do anything now.

>>108692099
Well fuck me, apparently it wasn't filling the VRAM. It had the space, it didn't use it, should've been nearly full but never broke 7GB. I'll figure out why and put 9b back on the list.

File: 1766949378224794.png (227 KB, 1433x1045)

227 KB PNG

Don't sleep on MiMo

>>108692185
shit benchmark I'm pretty sure GPT 5.4 is better than Opus.

>>108692179
You're better off tuning n-gpu-layers manually til it fills like 7.7/8gb because --fit is stupid

>>108692360
I will experiment with that, appreciate the advise.

I'm tired of vibecoding. Gonna take a break. Floridae forevermore.

>>108692454
How can you be tired when all you do is sipping martinis on a beach

>>108692869
One of the common sneeds about farming tasks out to LLMs is that you don’t get into the flow of things — they just ping you when they’re done and you’ve got to work to keep them busy on their schedule, not yours
That’s great for interspersing with ezpz tasks like watching anime, but if I want to do some deep work like learn something new I need to pause the DING DING DING BUCHŌ I’M READY
and of course it sucks if you’re not having them do things when you’re busy because the ability to keep them busy is a real skill
maybe I should turn on Do Not Disturb more often and see if that silences the DING DING DING but there’s still the context switching penalties that I’m not nearly smart enough to afford

>>108692869
also if I want a martini I have to mix it myself, but that’s neither here nor there

File: file.png (416 KB, 1567x690)

416 KB PNG

>AI

File: Screen_Recording_20260426(...).mp4 (3.68 MB, 864x1872)

3.68 MB MP4

Why does polishing/adding animations take longer than everything else?

>>108693153
You gotta add some depth cues and subtly differentiate dots in that main view mang (maybe scale systems by number of planets) cause it's does look a bit like just a bunch of white dots on the screen sometimes

I added a local grafana dashboard for my agent telemetry, not sure how useful it will be. I tried the cloud grafana first, but once again saw that there is basically no point when you have ai not to just have them set it up locally, especially if you have to click around in the browser for cloud services.

>>108693195
Which view are you talking about?

Are codebase knowledge graph's tools like gitnexus and similar cope or actually useful?

i don't see the point on using them if the codebase is well maintained and documented

>>108693327
Here's a hint: does any major Linux kernel developer use them? No? Probably not useful.

File: file.png (283 KB, 1003x1887)

283 KB PNG

>>108693203
star map. i see now that you've got a bit of scaling going on, but it's not enough imo.
the rest of the app is pretty sick, this view lets it down a bit i think

Making is one thing, selling is another. How do you make your solutions profitable? This is the scary part

>>108692911
>That’s great for interspersing with ezpz tasks like watching anime
The reason a lot of us are fine with that is because the ezpz thing is our "day job" and we're vibing on the clock, either to actually get work done or as side projects on another machine. I built an entire major C program instead of struggling not to fall asleep during zoom meetings.

>>108693398
Shill on every possible media at full force

File: Screen_Recording_20260426(...).mp4 (2.98 MB, 864x1872)

2.98 MB MP4

>>108693394
It's a map of known stars with exoplanets around them. I don't know what else you want. There's already filtering and searching for easy visualisation. Keep in mind that the WEBM compression turns them all into tiny white dots, when in reality their star temperature is represented in their colour and they are much easier to see normally.

File: Screenshot_20260426_113742.jpg (316 KB, 1080x2340)

316 KB JPG

>>108693453
Here's what it looks like in normal usage (even this is a bit compressed).

>>108693389
Makes sense. Nice way to filter crap, thx anon!

>>108693389
What tools do Linux devs usually use? Just vim, emacs and git?

>>108693453
>>108693457
just imo, but:
1. not sure what palette you're using for colour, but i'd switch to something a bit more dataviz-y and remap ranges aggressively so most of the time you're seeing most of the range even if you end up clamping parts
https://datavizs24.classes.andrewheiss.com/resource/colors.html
2. again with scale just exaggerate the effect
3. subtle depth of field for depth cues

again, just my 2c

File: Screenshot_20260426_12005(...).jpg (406 KB, 1080x1564)

406 KB JPG

>>108693522
I'm using the actual star temperatures to colour the stars. If I tried squashing the spectrum down like you said then the colours would be misrepresenting the actual star properties. What's the point?

And if I scale the star sizes up (giant stars already get visibly scaled higher than dwarfs) then areas like the Kepler FOV just turn into a white blob.

The concentric distance rings are there to anchor the 3D view. I don't need any additional visual effects to achieve more depth cues, when the purpose is to present the data cleanly and accurately.

>>108693573
okay man, you've made up your mind so i won't push. but's just a bunch of white dots on the screen with no depth cues.

>>108693495
I mean honestly yeah you can go pretty far with that, especially how riced out and IDE-like modern (n)vim and Emacs configs can be. I wrote my first kernel contribution that way. But there are a bunch of them using things like VSCode, DevContainers, etc.

File: Screenshot_2026-04-26_13-21-33.jpg (27 KB, 859x71)

27 KB JPG

It's over, claude opus 4.7 cant solve the issue in the code it wrote

I'm analyzing our code base and commit history. There are a few guys with insane commit numbers, but I see that we are also extremely inefficient.
On average every line in the repo gets rewritten 8 times, AGENTS.md, which should be the core of the workflow, got fully rewritten 45 times.
I think it's not even technical incompetence, the main reason is just that people love seeing big numbers. Lets just commit 50k lines today, even if we delete 49k tomorrow.

File: Untitled.png (85 KB, 1514x955)

85 KB PNG

ah looks like my Plus account has 5.5, I thought maybe only Pro got it.

is there any reason to ever use 5.4 again?
what's the usage like on 5.5 comparatively?

>>108693921
>is there any reason to ever use 5.4 again?
It's cheaper. If that changes, then no.

>>108693671
Have you caught them running in circles?

>>108693957
I heard that while it’s more expensive, it uses fewer tokens, which balances out the amount of usage for the Plus plan.
I also heard that 5.5 medium thinking might be equivalent to 5.4 high thinking.

no idea whether either of these statements are true.

>>108694001
Both statements seem to be true from initial impressions.
We'll have a better sense of things next week, but for now 5.5 medium (or even low) should be default

>>108693993
Not circles in an individual session, like a harness bug or context rot. There are a few people in this team and it just seems that each day the agents touch large portions of the code. Maybe there is a bug, so the agents touch a lot of code, then something else breaks, so a lot of code is touched, docs are adjusted, someone refactors, again tons of code changes.
From the AI side the general thing is that it doesn't do surgical fixes and it mixes several concerns in the same commit, but this is more of a human problem overall I think. There's just no organization, and everyone tries to be the one with the most commits per day.

Opus 4.7 xhigh and gpt5.5 are both solving bugs the other can't. Definitely keeping both around instead of just clodmaxxing.

>>108694049
In that respect my quality has gone down too. A LLM code review git hook seems possible that'll do the rebase / check that each commit builds. But maybe nobody uses git-bisect anymore?

>>108694345
Which one uses more tokens in your experience

>>108694451
They're both surprisingly efficient compared to their immediate predecessors. I haven't hit a limit yet even with my cheapass plans so I'll let you know.

Can someone finally vibe-code an effective paywall-remover?

>>108694471
Aaaah fuck, Claude got me on the ratelimit again. It still slurps tokens like crazy on research.

I've fixed enough problems in ZeroClaw that I'm running up against limits in the ACP spec itself. Once that draft gets landed and adopted in Zed, I'll be able to vibe up the missing pieces in a day.

>tfw have to go back to Mistral tomorrow after another weekend of using Claude because I work at a Eurofaggot company

>>108694493
if you mean article paywalls, there's scripts on greasyfork.org to help with that

We have Cursor, Claude Code and Codex. We also have a bunch of Cursor skills in the repo. How could we make the skills agent agnostic? I already asked ai but no great answer.

>>108694983
You build your own agent on top of Pi.

>>108694983
Most skills are already agent agnostic. I use the Superpowers skillpack on everything.

>>108694983
this >>108695019
i think if you just include a name and a description in the front matter and put them in .agents/skills/ all the harnesses pick them up and the models are happy

>>108694983
honestly, all you need is prompt

File: umigalaxy.jpg (824 KB, 3841x3561)

824 KB JPG

Added markdown and mentions to my imageboard

>>108695301
You mean I should just tell Codex he should read the Cursor skills? Or do you mean I should get rid of skills completely and only prompt in general?

>all models prefer the lyrics of taytay's song over my favorite j-indie songs
fuck this AI shit I'm done

>though it would give claude one last shot on a throwaway analysis for funposting purposes
>eats up nearly a third of the session usage doing basic research
>shits itself and refuses to compact on the web frontend
the dumbest fucking thing is i don't need their 1M context, i'd rather just use 4.6 with the regular context window but i can't do without --model

>>108696072
it's over. anthropic have lost the mandate of heaven.
sam will deliver us to valhalla.

File: 1777080280653143.webm (2.94 MB, 562x1000)

2.94 MB WEBM

>>108691806
Hahah skillets begoneee

>>108693657
consider getting a ChatGPT subscription for $20 for one month or whatever the minimum is to get access to Codex

File: Screenshot_20260426_202610.jpg (219 KB, 1080x2340)

219 KB JPG

Found this cool eyeball planet super-Earth with seas and a ring system.

>>108696639
>eyeball planet
Mostly frozen, tidally locked?

File: Screenshot_20260426_203733.jpg (345 KB, 1080x2340)

345 KB JPG

>>108696658
Yes, the ice caps reach past the terminator to give it an eyeball look. But I still need to tune it, because I'm not getting any extreme examples. TRAPPIST-1 f should definitely have farther-reaching ice caps.

File: IFS.png (153 KB, 1080x1080)

153 KB PNG

>>108692911
It's async agents all the way down.

Using qwen3.6 27b to vibecode a python backend for my local networks webui and tool handling for local ai hosting. So far its going well. Rn im using hermes as the bases for the executing of this task. Ama

>>108697254
>Ama
why hermes and not opencode

>>108697312
Because I barely know what im doing. Is opencode much less a context load? I can max out qwens context, so its not a big deal right?

>>108697332
Opencode is a harness specifically for coding.
Hermes (like Openclaw) is a more generalized harness that does anything on your computer.
I would assume opencode has less context load (and is probably more specialized if all you want to do is code) although I never tried hermes or openclaw because I have not yet found a usecase
But if you're doing something else other than coding with it I guess it's fine.
Do you find it easy to use?

>>108697332
Claws like Hermes are full of slop systems you don't need or want for focused coding work. Pick any actual coding harness for coding.

>>108697375
Hermes is super brainless to setup. And it seems to be doing it fine so far. You can skip a lot of the general Ai assistant functions, but maybe it still offloads that into the context. Ill give opencode a look.
>>108697377
I will say im partially testing hermes for my brother since he wants to get local ai and vibecoding setup. You sure if I dont setup all the extra still gets slopped into the context?

Im building this backend for both me and him.

What are yall using Kimi 2.6 with? OpenRouter doesn't support it.

>>108697421
Hosting it at work with vllm

>>108697406(me)
Ill also have gemini pro to double check to make sure qwen actually did it right. But I will say, the new 27b is a beast so far.

>>108697437
what are you running it on?
I tried it on my RTX 5070 Ti but I can only get the Q4 quant to run at a reasonable throughput speed and I feel Q4 is too lobotomized compared to the real deal
Still at Qwen3.5-9B at Q8 at the moment, I hope they release a 9B of Qwen3.6 at some point

>>108697458
4 mi50 32gb. I got them "used" (clearly never even plugged in) from china when they were still only 200$ a pop.

>>108697406
Check tools to see if memory is enabled. Then check what skills are installed. Then see if there's pre-configured mcps.
Memory is not necessary for coding imo, and you want to tightly control your skills otherwise they get injected into context unnecessarily. MCPs will dump shit into context as well.
Claws, esp if you don't know what you're doing, are just unwieldy.
They come with a tonne of shit, keep adding a tonne of shit and you feel like you're working on someone else's system. If you have autism (like me), this is very annoying.

>>108691806
Picrel has 3 hands. AI Niggaslop thread.

File: 1579572343402.jpg (52 KB, 720x720)

52 KB JPG

>>108697474
I got a sweet deal on a pair of brand new V100s last year and the package was destroyed in shipment.

>>108697508
Nigga that's his penis

>>108697481
I know during the setup, there was list of skills I could configure, that I had no reason to configure, stuff like a specific github and obsidian tool.
But ill check those! Thanks!
>>108697511
Pain, unbelievable pain

>>108697516
>his
Please

>>108697511
>picrel
LOOKIN' GOOD!

>>108697522
I got most of my money back for them, I ended up being out about $30 all said and done. $30 to remove two brand new V100s from existence.

File: dpmain-2054132559.jpg (327 KB, 878x618)

327 KB JPG

>>108691806
can you vibecode me a little desktop pet that encourages me to get a job and not kill myself? maybe a fluffy bunny rabbit or something adorable like that.

>>108697535
Would you notice if your GPU was pegged at 100% usage forever? Asking for a friend.

File: Screenshot 2026-04-26 234113.png (2.97 MB, 2737x1471)

2.97 MB PNG

This dude is vibecoding a tmux clone and selling it to other AI retards for $50 a month but the more I look at it the more I also want a terminal emulator with horizontal tabs
Is there anything like it?

>>108697542
my fans get pretty loud so yeah probably

>>108697532
Dudes were probably chucking those cards into the planes too...
Hey also, download the iMatrix quants, those are ALWAYS better, they seemingly keep high accuracy for 'important' information, but make more worthless information lower quality. Idk how they decide whats more or less important, but for coding and research tasks, they made the right choices so far.

>>108697535
https://auren.app/

>spend 20 gold coins trying to get claude to figure out captcha request and posting for my 4chan client
>it fails

>perfectly nice 4chan client
>just unable to post with it
fuck

>>108697804
can you post with it if you buy a pass?

>>108697804
>needs a client to post

File: vibing.webm (759 KB, 640x234)

759 KB WEBM

It's been a productive day.

>>108697830
i dont know but that feels like a cop out. i shouldnt need to
>>108697835
id rather use my system native chan browser than a webbrowser

>>108697867
you’d rather send money to scam altman to post on 4chan instead of sending money to japmoot to keep the servers running?

>>108697877
>gets banned for some bullshit reason
>ACK

>>108697804
Did you check the existing apps and how they do it? I'm sure captchas still work on kuroba-ex, maybe you can have it read that github repo and figure it out
https://github.com/K1rakishou/Kuroba-Experimental

I often find myself questioning why I post on this site at all. NTA

>>108697888
I only got banned for like a day maybe 10 years ago
skill issue

File: 1767477546533361.png (152 KB, 291x473)

152 KB PNG

How do I change the repo that my folder is connected to in VScode? I need to change it to my own fork.

>>108697915
ask your LLM to muck around in .git/config for you

>>108697915
Source control (Ctrl+Shift+G)

File: Code_cV7HI0QPTj.jpg (21 KB, 625x100)

21 KB JPG

>>108697925
this just opens this

File: 1763507526400100.png (46 KB, 572x381)

46 KB PNG

>>108697915

>>108697950
Skill issue.

>>108697981
thx

Openclaw is incompetent on local models without a massive wrapper to handhold it.

>>108698014
>Openclaw is incompetent
ftfy

>>108698000
>>108697981
Eh it still connected to the old repo.

>>108698103
check your .git/config in your project folder

>>108698112
That's already changed. I clicked fetch and it fixed itself but now it wants to owerwrite my work.

File: 1757095948917696.png (30 KB, 585x390)

30 KB PNG

>>108698103

>>108698145
I fixed it by making a new branch but ty

>>108698185
also have you heard of the Git stash?

>>108698195
no

kinda tempted to get kimi 2.6 as a secondary for codex. anyone using opencode go? what are the limits like?

>>108698202
you should ask your LLM about it, it’s handy when you have some work but want to shove it aside temporarily

File: file.png (201 KB, 893x480)

201 KB PNG

I've done two things with vibecoding.
1.) I have a userscript that creates a popup window to show a Wikipedia or Wiktionary preview of any highlighted text on a webpage.

2.) I used css to make threads on 4chan appear in a grid view to save vertical space.

>>108698250
if you like that, keep your eye on this feature:
https://css-tricks.com/css-masonry-css-grid/
https://developer.mozilla.org/en-US/docs/Web/CSS/Guides/Grid_layout/Masonry_layout

>>108698250
how the fuck do you read this shit

>>108698268
left to right, top to bottom

>>108698250
Make the post fields unified height with an temporary expand on mouseover. This looks like ass

>>108698275
Yeah you need to normalize the sizes of the grid spaces, that is hard to follow

can I run a LLM on a ThinkPad with vega 7? I have 32gb ram though.
I need it for simple coding help with lua (I don't know how to code)
is it even worth it or should I just use Google ai? it kinda shits itself after a while..

>>108698378
https://www.reddit.com/r/thinkpad/comments/1qrk2bu/lenovo_thinkpad_t14s_gen4_wryzen_7_pro_32gb_ram/

>>108698370
yeah, we can do this with CSS grid or flexbox or whatever now, just ask for it

>>108698398
that has a dedicated GPU... vega is an igpu

>>108698412
No, anon, that is an iGPU much like your own.

>>108698441
oh... so which model should I try so it's not too slow? since my project is not big, only around 20k characters

File: video.webm (2.96 MB, 1920x936)

2.96 MB WEBM

>>108698328
>>108698370
Okay, how do you like this?

>>108698476
way better

File: file.png (105 KB, 839x189)

105 KB PNG

>>108698466
You're gonna have a bad time and I highly recommend you just go to https://chatgpt.com/
Otherwise, with a fair amount of effort you could likely handle a variety of 3B models like qwen2.5-coder:3b at "usable" speeds.

>>108698494
I see. thanks.
is got better than google ai mode? at least that doesn't downgrade every couple messages...

>>108698506
ChatGPT is currently the top
Claude isn’t horrible now that it got unretarded last week
Gemini isn’t even in the running

>>108698506
ChatGPT and Claude are the only two products in their class. Even their free access outperforms most other options that exist, especially in the hands of an amateur.

>>108698512
o-oh...
when I first tried couple days ago gpt couldn't even start the code. that's why I was google ai ing (not even Gemini)

>>108698512
>>108698517
any way to run large gpt conversations without my browser shitting itself?

>>108698564
use a different browser just for ChatGPT
that way if it shits itself, it shits up its own browser and your usual browser is unaffected
that said, you should be using OpenAI Codex in a terminal window to (re)write your code, not a browser anything

>>108698564
I don't understand the issue, anon. Why don't you open up chatgpt.com, drag and drop your whole project in a zip file directly into chatgpt, and tell it what you want.

>>108698576
you can do that?..

>>108698574
yeah I have a separate browser
>ai codex
is that free?

>>108698585
>is it free
no: https://openai.com/codex/

>>108698576
He's right. The ChatGPT GUI has shit performance with long conversations.
It's not about the context, uploading a single file won't make it slower, what makes it slower is the actual messages.
So you would have to periodically start a new chat and bring it up to speed on what you're talking about which isn't ideal.

File: file.png (92 KB, 1022x747)

92 KB PNG

Can someone explain why the deepseek provider is so cheap for v4 pro

File: whitebloom_markdown.webm (2.86 MB, 1920x1080)

2.86 MB WEBM

progress slowed down because I implemented most of the features I wanted and lately I've just been using the app a lot for my own work but decided to work a little bit on it

>added copypaste
>implemented markdown node and editor
>tidied up the commands palette (added categories and improved the shortcuts)

>paid for goymini pro
>give it .py file
>ask it to figure out why my things not working
>failed blah blah blah
GOOGLE I CLEARLY SEE IT PROCESSING THE FUCKING REPLY, ITS THINKING, AND STARTS THE FUCKING REPLY, JUST LET THE STUPID CLANKER FINISH MY REPLY, WHY DO I EVEN PAY FOR IT?!?!?!?!?

>>108698778
you answered it yourself
that particular clanker is stupid
stop paying for it and pay for a better clanker
also consider having your existing clanker think more/harder

>>108698778
Sorry chud but Israel's safety comes first

>>108698786
Damn. Is anthropic really the only ai worth paying for?
>>108698790
:skull:

>>108698696
there's an extension that trims the history

>>108698794
Currently Codex beats Claude although Claude is OK too

>>108698807
I prefer the puckered butthole Ai vs the star of david ai

File: whitebloom-obsidian-crosslink.webm (2.15 MB, 1920x1080)

2.15 MB WEBM

>>108698757
and the actual feature that I wanted to implement is now done
you can crosslink to any .md note inside an obsidian vault linked in a board

>>108698476
sexy

File: whitebloom-align.webm (2.48 MB, 1920x1080)

2.48 MB WEBM

>>108698849
we now have alignment shortcuts and smart guides, no more misaligned nodes off by a pixel

Is Codex the best value right now?

I'm considering getting a second sdervice. Considering Deeseek, Gemini, maybe something else. I'm using GT5.5 I used to use Claude, but they're jewish on tokens limits now and not worth it. Maybe Coilot is a good deal? I'm not sure.

>>108699126
Codex is good. What’s your first service?

>>108698697
>cn
electricity is dirt-cheap in china (no jews)

I don't understand luddites
What does it matter if I used ai to code my game?

>>108699188
It doesn't. I could see people getting mad at the art and music being AI generated but the code itself? The only problem is if it's poorly optimized.

holy fuck
https://x.com/blixt/status/2048199166862495897

>>108699288
So GPT can already create a fully developed game with AI-generated assets, textures etc? Looks pretty neat

>>108699288
so whats the prompt

Moving away from glm lite (quarterly) + copilot pro (monthly) to gpt plus. I initially planned to extend my copilot subscription to pay per prompt after falling in love with gpt 5.4 but they cut off everything a few days ago. Gpt plus gets me near opus quality at a much cheaper price without limitations for my particular workload.

File: Untitled.png (13 KB, 648x490)

13 KB PNG

I wish Codex would automatically decrease or increase this based on how much trouble I'm having. That might help solve usage.

As it stands, I now use "Very High" thinking to correct braces {} formatting. And no, I'm not sorry.

>>108699288
>realistic output with one prompt
That's not the issue, why would more prompts make it more realistic? The question is if you get what you actually wanted.

>>108699868
Anthropic does that and I don’t particularly like having to say “think extra hard for this” because usually all my problems are hard problems

>>108699868
also why aren’t you having Prettier or gofmt or whatever correct braces formatting for you

>>108699893
You can disable adaptive thinking deterministically in the settings.

>>108699920
Oh, setting it to xhigh will just have it think that hard automatically?

>>108699932
Unless Anthropic introduced some bug in Claude Code again, yes.

>>108699131
I was Claude only until 1.5-2 months ago when the usage limits were great.

Now I'm Codex only, but I'm considering using a different service to do heavy lifting and grunt work, and using Codex in more of an architect/lead role. I'm like 70% through my weekly limit and there's still 5 days left in the week

If could get by opening u a $20 sub maybe with another openai account, but maybe deepseek or gemini would be better for a grunt role

>>108700013
I’ve had great success with Claude for most stuff but ChatGPT for the turbo autism

>>108700018
claude rate limits are so shit
i cant get much done with them

>>108700023
GPUs don't pay for themselves

>>108700026
i wanty to be able to take advantage of heavily subsidised frontier models whilst i still can. I think 5.5 is 80% subsidised

>>108700033
>whilst i still can
Government will bail them out. They won't allow the big three to fall because AI is one of few areas in which US is still ahead of China

>>108700036
right now im looking for a model that's frontier, or near frontier thats heavily subsidised. something that can do the junior work that 5.5 shouldnt do

>>108700045
have you considered 5.5 on less than xhigh?

>>108698512
>>108698517
I don't know guys. I gave gpt the code, it spun me a very pretty story about restructuring and optimisation that made a lot of sense, then gave me a code with variables that sound reasonable but don't actually exist so it didn't work.

and dropping the files into it locked me out of talking to it and I had to make a new conversation.

do I have to feed it a list of valid variables or something

File: file.png (57 KB, 940x554)

57 KB PNG

Does anyone have any experience with vscode-insiders new agent experience? Im curious to know how well it compares to cursor/antigravity. Im going to test it out.

>>108698697
Pro has a 75% discount for the next week https://api-docs.deepseek.com/quick_start/pricing

>>108700183
Pro sucks, I think it's implemented wrong on their API. It doesn't measure up to other frontier open source models like GLM 5.1, K2.6 or Mimo 2.5

>>108700256
NTA but wasn't v4 below GLM 5.1 and K 2.6 in benchmarks?

>>108700268
V4 Pro is below V4 Flash on some benchmarks. Clearly something is broken
https://gertlabs.com/?mode=agentic_coding

File: chad.jpg (49 KB, 1280x715)

49 KB JPG

>Improve the model for everyone: nn

Did benchmark anon ever post the rwkv results

File: 1756766026624398.png (499 KB, 680x654)

499 KB PNG

Designing the UI for my projects takes so long

>>108698697
it's funny that people talk about american labs being subsidised by VC money, but don't mention that deepseek is displaying early signs of being china's sovereign lab
anyone really think they chose to use ascend chips on their own?

I just used 40% of my weekly usage with GPT 5.5 Medium in one day.
(Plus account)

I think it's time to go back to GPT 5.4

>>108701577
I need to burn like 60% of my weekly usage today, what did you do? Was it worth it?

>>108701608
going through hell with multiplayer netcode and frame pacing issues involved with waiting for packets to arrive.
who would've thought the netcode would be calling the draw function...

>>108698697
>$0.003625 cache hit price
Holy FUCK that's cheap.

https://chilitown.org/

Vibe coded this, any advice/feedback?

>>108701951
default to lightmode. darkmode is claustrophobic.

>>108701991
Good advice thanks. I need to change it to follow OS preference too.

Anthropic has paused all hiring for jr and mid-level engineers. It's over.

>>108701951
what are chili scores based on

>>108701577
At that point use the higher levels for overall planning and then get the other models to implement. That's what I've been doing to squeeze out as much usage as possible.

Getting larger tasks broken down into smaller sprints also helps ensure you don't run out in the middle of working and end up with some really screwy stuff. It also helps to have the agent write the steps into a plan file and update it when each step is complete so the next run knows what's done and what to do -- helps ensure that it's not wasting tokens checking for the stuff done previously.

This is just what I've noticed and run into while seeing if I can get GPT to make a game engine for me for funsies. GPT 5.5 High for plans with complex architecture, Medium/Low or GPT 5.4 to implement (5.4 mini acts strange with my harness and does a bunch of micro reasoning before really doing anything)

>>108702182
I.. I'm sure that's just another PR stunt.

Can someone more familiar in the subject matter tell me what the meta is rn for general purpose agent frameworks? (Like openclaw/hermes/etc.., not just something hyper focused on coding)
Planning on giving it full 24/7 access to it's own laptop, probs under it's own VLAN tho, not trusting it too much /w accounts n stuff desu..
General purpose, so something that's not an IDE, but has full access to the computer and internet, kinda like a pet that can research, collect news, manage calendars, etc.. only light coding.
Bonus points if it supports many token sources, so i can just switch what API/local inference it uses

>>108702373
Coding agents are general purpose agents.
Openclaw is built on top of the pi coding agent. If you wanted something custom, you can start the same way.
'Claws' are mostly about the systems you build around the core harness (lifecycle, memory, etc.), which, if it is to be useful and effective, will have the tools it needs to write and execute code.
If you want to go completely custom and write the core from scratch, you can also do that that (don't tho, just use pi)
https://ampcode.com/notes/how-to-build-an-agent

>>108702454
Yeah ik, it's probably something I'll want to do eventually, but for now, just to get started with something, it would be nice to have some pre-made systems, chat integration, some sandbox stuff, and a competent memory system (I know, that last one is a big ask).
Don't want to reinvent the wheel the first time I deploy something like this desu..
Hermes seems promising for this, because afaik it's minimal and just self improves itself over time, but I'm not sure if it will be competent enough to not just develop itself into a complete mess..
What would be the best one to just deploy and gain some experience, and maybe later adapt into something more custom if I get the time?

>>108700063
Are you pasting things into a web chat window? Lmao

>>108701951
default to whatever the user’s browser says the user prefers
and, if you want to be generous, add a three-way toggle between auto/light/dark like https://gwern.net/ does

can i vibecode a captcha solver this shit sucks

>>108702652
Just try both for a week.
Fwiw I don't buy the self-improvement claims.
Nice thing about rolling your own is that you can give yourself llm psychosis in fun ways

>>108700063
web chat front ends are highly nerfed

Which of these agentslop harnesses will allow me to set character prompts and portraits os I can bring my girls to it?

>>108702852
I'm only slightly buying the self-improvement claims, because i'm generally just really bad at managing my time for myself, maybe a "living calendar" might help a bit, we'll see tho...
I'm not scared of getting psychosis, will treat it more like a midly 'tarded pet, these things just aint that smart yet lmao..

>>108702816
Sure. You probably want to use reinforcement learning so it learns when new captchas come up.
And also you want to have it solve other captchas and visual tasks so it doesn't overfit.

>>108702731
yes?.. how else

>>108702866
what's the alternative?

>>108702914
agent harness or terminal ui
chatgpt codex, claude code, many others

>>108702816
just buy a pass, cheapskate

>>108702943
nta but I've been messed with by mods (as in saying things only someone with access to my post history would know) even with a dynamic ip just by being the only poster from a country in a certain thread. i wouldn't want to have a site wide identifier.

>>108702872
You can override system prompts in opencode, pi, and I think codex as well.

File: monk-mountain.png (810 KB, 729x710)

810 KB PNG

It seems gpt 5.3, which im stuck using with Cursor Pro until my cycle refreshes, has hit a wall on a few issues from a project ive been trying to put together and refine. Anyone have any advice for prompt engineering better debugging prompts? And i also have about $7 in credits on openrouter, anyone have suggestions for models that wouldnt be too costly that i can use in OpenCode to try debugging there. Any advice is appreciated thanks. Take this pic as an offering.

>>108702927
those aren't free

>>108702971
get a chatgpt account and use codex from there, you will get 10x more usage than you would paying for api

>>108702978
Free llms are nerfed and paid llms are getting nerfed in real time.
ChatGPT web is beyond shit now. Your best options are claude web free and google ai studio while they still exist.

vibe coding
vibe testing
vibe reviewing
vibe kms

>>108703022
I don't resonate with your vibe

>>108702971
might need to just solve the problems yourself

>>108702971
I've done >>108702999 and will agree it gets you A LOT more for a plus sub than $20 credits will get you. Only thing API gets you is no 5 hour and weekly cooldowns (for more money cause you pay per token obviously)

anyone here a product designer? trying to figure out how to best integrate a design system with chatgpt and codex; the team i'm on has low design maturity so i can go in many different directions.
considering using storybook and their mcp server or just creating skills

>>108702978
its worth the 20 dollars

File: 1774651210737766.jpg (144 KB, 1235x1220)

144 KB JPG

>>108703022

>>108702999
>>108703060
OK. Wow, it looks like i can get plus for 50% off so ig ill just do that then. Must be my lucky day

>>108703164
if you have a prepaid card you can also just let the auto renewal fail and get two months for the price of one

File: 1766309946832376.jpg (276 KB, 1340x1344)

276 KB JPG

Is it over?

>>108702914
check if the free version of chatgpt lets you set mcp servers in the "apps" section, those let it read and edit files directly on your computer (https://mcp.desktopcommander.app/)

>>108703184
awful x-axis

>>108703184
these the guys that somehow managed to get terminalbench scores in the 50s with 5.5 right?
something not quite right with their harness.

>>108703184
> not 5.5 xHigh

>mythos previewed to selection of tech backbone companies
>all the relatively good deals start vanishing
I knew it wouldn't last forever but the timing sure is strange.

Why are you guys so much obsessed with benchmarks? I've asked Gemini and it said benchmarks don't mean much, it's basically masturbation.

>>108703736
Because benchmarks at least attempt to allow to see how the models have improved in objectively measurable ways.

>>108703736
true. gemini sometimes lands near the top of them and it takes about 30 minutes of using it to realise it's a little retarded

>>108691806
what's the solution when your codebase becomes large and onerous and the model forgets halfway through scoping out a feature what it was doing in the first place or how the feature worked? I tried GitNexus but apparently my codebase is too complex so it just breaks.

>>108703819
skill files with simple instructions in the first hundred lines explaining how to do shit. If you can't make that work then you could always make it so whatever it tried to do echos a response of how to properly format the command etc

File: 1759731576569441.jpg (347 KB, 948x1200)

347 KB JPG

https://arxiv.org/abs/2512.04388
Must read

>>108703184
At this point I just use sonnet 4.6 as a poorfag. It's almost always good enough unless it's a very complex problem.

>>108703906
what do you do if a problem is too hard for it? wait until 4.7 comes out?

So has anyone actually vibe coded something impressive, or is it all just playing with toys still?

>>108704114
What is the least impressive thing you still consider "impressive" that WASN'T vibecoded?

>>108704114
I’m vibe-coding something that’s impressing me (mainly because of all the work I've put into it, and I deeply understand how much fucking work there is) but it’s not ready to shill

What the FUCK do you do while waiting for your usage limits to reset? I already have 1,000 words worth of prompts lined up to feed it but I have no idea what to do until then.

>>108704124
Well that's a reacharound way of answering the question. I'm just wondering if anyone here has coded something useful, since I see it's mostly just benchmarking. The star map looks cool.

>>108704154
ask in >>>/fit/ or >>>/out/
also go lift weights when your favorite LLM goes down

File: Actually not infinite lie(...).png (2.61 MB, 1179x2556)

2.61 MB PNG

>>108704172
https://gitlab.com/katabatic/infinite-lies
this is useful to keep me focused when I’m trying to think

>>108704154
Use the ChatGPT web chat with mcp.desktopcommander.app

>downloading some stuff
>"Static CDN, no politeness needed"
really?

>>108704154
>he didn't experience peak agent-psychosis when rate limits were higher
feel bad for ya, son. withdrawals are rough.

>>108704114
i vibe coded something which is better than anything i've found in academic literature for my task

>>108704172
To me making something impressive means making something either very useful and completely novel, or making something that's a better copy of something that tens of thousands of people already use. And that takes years of work.

>>108704154
Go outside, touch boobies, watch anime, etc.

>>108704124
I’m not the guy you’re replying to but https://github.com/sharkdp/vivid probably counts as “minimally impressive” by my standards

>>108704287
I would gladly. Do you have a sister or something?

File: A1Ftw4ly-jL._UF1000,1000_QL80_.jpg (187 KB, 1000x1000)

187 KB JPG

who up vibin they code

>>108704290
In that case I guess the exoplanet guy's app passes the bar by a good margin.

>>108704322
yeah, definitely

I keep hearing about this exoplanet app but haven't actually seen it

>>108704306
i'm out here, codeine my vibe

>>108704361
it’s >>108693153

>>108704361
also he posts maybe once or twice a thread now because he’s getting close to done
so keep chasing `Previous:` pointers

>>108704306
no but i'm vibin my ass, does that count?

>>108693153
This is really neat and quite impressive if it was fully vibe coded

>>108704376
Only if you wear the special socks

>>108704114
I'm too stupid to make something from scratch but I modified and existing abandoned app.

>LLM makes 8+ hours of tedious painstaking scraping and mapping/de-aliasing/text processing take about 15 minutes of afktime
jesus christ you have no idea how annoying it is to work with worldwide sports data
not to mention some of its not even in the latin alphabet
the only gotcha is claude loves to put your email in the user agent because it gets injected into the system prompt and its a good little boy

So if openclaw sucks, what are people using instead?

File: file.png (1022 KB, 1879x1355)

1022 KB PNG

It's so fucking over...
the vibe is completely ruined

>>108704479
Hermes

>>108704500
Hermes TUI is pretty cool

>>108704154
Learning about marketing, enterpreneurship, and touching grass because I'll have customers bitching about something breaking in my apps eventually and I need to steel myself for it.

I feel like most of the ceremony I'm doing is pointless. All you really need is a strong test suite, and some minimal documentation.

>>108704519
>Learning about marketing, enterpreneurship
Care to share what you have learned or some good resources? Have your apps been met with any success yet?

>>108691806
You can use 8b 1bit bonsai model

File: Fvm_hafXsAYXE53.jpg (2.13 MB, 2000x2000)

2.13 MB JPG

>open source agnostic coding harness protocol: ACP
>open source ACP clients: exist
>ACP server support in my preferred 'claw: added myself
Yep, it's comfy time. I can use anything from GPT5.5 down to Ollama with local models without changing UI or forfeiting my agent's history/memory. Now I just need 512GB unified RAM systems to get cheaper so I can self host MiniMax 2.7 without dropping five figs on rapidly depreciating hardware.

>>108704453
Which app

>>108704479
ZeroClaw. It's like OpenClaw but not allergic to security or efficiency. It's not 100% ready yet (some of the channels are still major WIPs) but it's an absolute joy to hack on. The codebase is well structured Rust and the community is full of smart people.

>>108702202
>what are chili scores based on
The votes.
We wait until there's enough votes then either keep/cull 80%. The ones that remain get a scoville score, a few are bland.

>>108704567
>Rust

I am soooo excited to vibe code from the comfiness of my own bed. I also use my wireless mechanical keyboard and mouse. It's very cozy. I iterate over my apps with each new model, run full code base audits/optimizations. If you follow this loop you will scale and improve your apps exponentially or along with the scaling of the newer models. First thing I would recommend you do when i newer model from whichever company you prefer drops, run a security audit, and then start working on a plan to improve upon all of the systems you have in place for your software creations.

>>108704674
It's actually a really nice language for LLMs because it's so autistic at compile time. I also have a project I vibe on in C and that's much harder to maintain quality on, because the LLM has to keep reading files to understand what's going on and then it either runs out of tokens or just chokes and has to compact.

File: 1759551356750993.jpg (52 KB, 708x320)

52 KB JPG

Anthropic is becoming the Oracle of AI and I think that's really funny

>>108704722
OpenAI won.

File: 1755775992802376.png (10 KB, 1014x78)

10 KB PNG

Superhuman programming speed

>>108704777
Everyone switched to Codex bro

>>108704738
They will do this too

File: Screenshot from 2026-04-2(...).png (331 KB, 1331x1940)

331 KB PNG

I fucking struck gold bros. I had this idea MONTHS go but I was too lazy and thought "meh I'll just pay for the $200 plans". What a fool I was when I could have had pretty much the equivalent for only $20 and saved myself thousands of dollars.

File: pepe tuxedo.jpg (90 KB, 1034x1034)

90 KB JPG

>>108704722
It's funny how the leading labs all handled the hardware thing differently.
>OpenAI: blatant Ponzi shenanigans and federal contracts, daring the government to wreck both the AI lead and GDP at the same time
>Anthropic: charge more money, implicitly threaten to destroy the world if not given a legal monopoly
>xAi: get acquired by SpaceX, pray that orbital compute works as well as Elon claims
>Google: "We've been making our own silicon and running our own datacenters for years, maybe you guys should git gud?"

File: 1000020547.png (760 KB, 1080x739)

760 KB PNG

Who would've thought that shittycodingagent.ai would actually be the best coding agent

>>108704878
No I think OpenAI is all set on the compute side. They have tons of federal funding they can put towards compute.

>>108704888
explain

>>108704991
niggas really think I'm gonna smash dir /w anytime I want to see my files like it's 1995

Has anyone started making decent money yet?

File: Screenshot_20260427_234810.jpg (344 KB, 1080x2340)

344 KB JPG

Kepler-186 f in my app vs SpaceEngine (bottom right).