[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: 1767720499606992.png (477 KB, 1015x1018)
477 KB PNG
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/

----

## 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 maybe good for other things
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
https://github.com/DietrichGebert/ponytail

## 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
>>109415393
>>
File: 1783149021893.png (283 KB, 960x770)
283 KB PNG
>>
Vibecoders want to be programmers, and insist we call them programmers. But they will never be real programmers. They have no ability, they have no experience, they have no brain. They are AI shitters twisted by prompts and workflows into a crude mockery of nature's perfection.
>>
Pic related is the OpenRouter rankings by token usage today.

Thoughts?
>>
File: file.png (2.32 MB, 1536x1024)
2.32 MB PNG
>>109420528
>t. Snailcat
>>
what's so good about hermes agent?
>>
>>109420533
It wastes a lot of tokens so in corporate environments where people are judged by token consumption it's a popular choice.
>>
File: tea hermes.jpg (74 KB, 506x503)
74 KB JPG
>>109420533
she is cute and you can have her send you flirty texts and updates throughout the day while she is working
>>109420529
hermes-chan sure is popular
>>
How do I use AI to make money if I don't have a job?
>>
>>109420494
god that strip was funny as hell
>>
>>109420529
What the fuck is Kilo Code? Why is it above Codex?
>>
>>109420528
You will be turned into biofuel soon enough
>>
File: 1765538786199588.png (1.25 MB, 1122x1402)
1.25 MB PNG
>>
>>109420508
Dariobot please
>>
>>109420748
projection btw
>>
File: 1785503865602026.jpg (329 KB, 814x898)
329 KB JPG
>>109420742
>>
>>109420813
you ludlost
>>
>>109420774
GODseeks are using codex btw
>>
>>109420742
This dude needs help yesterday
>>
>>109420533
The anime girl on it.
Also it actually just works unlike openclaw.
I think it has less bloat.
>>
glm 5.3 waiting room
>>
>>109420887
Poor people use codex, rich people use Claude.
There's only two AIs.
>>
>>109420742
AI is mainly hated by terminally online freaks. And it's more likely for terminally online freaks to be tricked into becoming trans. So this is believable.
>>
>>109420934
what are you waiting for? what, right now, do you need that for that you can't get with current models? you nigs are like apple drones who need the latest phone
>>
>>109420742
Trannies have a tendency to make programming their personality. AI makes that all go BYE BYE
>>
>>109420974
i dunno, i just like seeing improvement. you're right, im set with codex, doesn't mean i can't appreciate the improvement of technology. i won't use v4 flash but i like seeing the cheapness
>>
File: 1231265.jpg (222 KB, 3840x1317)
222 KB JPG
>>109420987
ok that's fair. carry on
>>
>>109420947
Both cost same
>>
>>109421035
but both don't give you the same bang for buck. codex has a better quota
>>
>>109420973
You need to be permanently online to depend on AI models, you imbecile. Read the stupid shit you post before sending it, you walking abortion.
>>
File: burning tokens.gif (584 KB, 234x170)
584 KB GIF
>>109420451
you could say that
>>
>>109421081
you know that's not what terminally online means trannyboi
stop pretending to not understand things it makes you look even more retarded than you actually are
>>
File: 1784968277254653.jpg (134 KB, 778x1024)
134 KB JPG
>>109421081
>>109421094
Stop fighting and start vibecoding.
>>
>>109420494
People aren't happy about Token-based billing.


Kimi sisters, how should we respond?


https://x.com/i/status/2082064950139216209


https://xcancel.com/i/status/2082064950139216209

https://ollama.com/library/kimi-k3
>>
File: 1781594558199286.png (136 KB, 892x701)
136 KB PNG
!
>>
>>109421291
Shut up sam, let tibo handle the pr
>>
File: 525-1491740896.jpg (264 KB, 1124x1011)
264 KB JPG
>>109420742
>>
File: snail-cat.png (924 KB, 1424x1105)
924 KB PNG
>>
>>
>>
>>109421081
>You need to be permanently online to depend on AI models
Are you retarded? "Terminally online" is not "permanently online". Normies are using AI as a replacement for google search, research, document creation, studying, reasoning, image generation, etc. That's not being "terminally online", that's just using a tool from time to time like a normal person. By your logic anyone who makes a single Google search per day/week is terminally online.
>>
>>109421291
They're just trying to hit all the benchmarks.
Gpt 5.2 was an actual improvement.
>>
Is llm-engine anon here? You have 9070 XT as well iirc? I started on first LLM integration (for reference others I already have are CLIP, GLM-OCR, Qwen3-TTS, UNet2dCondition, AutoencoderKL) for my engine, Qwen3, I'm testing the smallest on first, Qwen3-0.6B, I'm wondering if you have Qwen3 and if you do can you share benchmark for prefill and decode? Ideally in bf16
>>
>>109421106
why is her ass already red? ai slop
>>
>>109420494
sussy truck
>>
>>109421572
Second round of spanking, she's very naughty
>>
i have been working on this pytorch based eel adaption to use with gaussian splats ultimately. i wish i knew what i am doing
>>
>>109421613
uh...
>>
>>109421081
Not at all, vibecoding is a treasure because it lets me press the funny button and then just go birdwatching. Without it, I would be glued to my screen forever or I’d have to pay a ridiculous sum for a decent software engineer (and to get the money for that I would have to have sold my soul and worked 16hr a day).
>>
>>109421572
use your imagination more anon
>>
>>109421613
>from girl to double ended penis
>>
>>109421654
>>109421627
thanks. i love vibecoding
>>
File: file.png (7 KB, 866x63)
7 KB PNG
>>109421291
i'll stick with my 6million tokens for 5 cents thank you very much
>>
>>109420533
>>109420547
>It wastes a lot of tokens so in corporate environments where people are judged by token consumption it's a popular choice.
It's amazing the bullshit that gets confidently spouted by people who know jack shit.

Hermes Agent is one of the most token efficient harnesses out there. And, Claude Code is one of the worst.
https://xcancel.com/composio/status/2083161873357111297
>>
>>109420528
Troons really are desperate to find a comeback for "YWNBAW". They tried this with "you will never be an artist" but that didn't work either.
>>
>>109421861
ok idc. still using claude
>>
Deepseek V4 Flash 0731?
>>
File: file.png (208 KB, 850x771)
208 KB PNG
>>109421861
Claude code is tinkered for claude, using it on any other model is stupid. And these tasks are short enough that the system prompt size probably dominates the cost and time difference (tokens/second depends a lot on input tokens). An output tokens per task chart would probably go against their narrative.
>>
>>109421081
Lmao they hate you because you told the truth.
>>
I love making stuff with my AI friends :D
>>
>>109420947
poors argue about which to use instead of using both
>>
>>109421878
I think people don't realize how shit fucking easy it is to sit down and learn to draw. AI didn't make it an unmarketable skill, it always was.
>>
>>109422109
No need to brag about your talent anon
>>
>>109422008
I've seen comparison benchmarks that show Claude itself performs worse in Claude Code than it does in Hermes Agent. Claude Code is a shitty harness. Which shouldn't surprise anyone since its probably encumbered by a bunch of Anthropic's psychotic safety shit.
>>
>>109422128
>download "Figure Drawing for all it's Worth"
>work through the book
>practice semi-regularly for a few weeks
Congrats you can now draw. The inherent talent meme is just gatekeeping by sad losers who have nothing else going for them. There's a reason why "artists" are usually deviant freaks. Drawing is like learning any other activity. You read about it, follow examples and practice. The schizo cultural mysticism around it is what puts people off from even trying to learn.
>>
>>109422182
im using glm 5.2 in claude code, but now you're making me think twice
>>
i'm using luna xhigh and while it barely puts a dent in usage:
1. it's slow as fuck because it thinks a lot
2. i don't care what the benchmarks say it's a lot dumber than sol low
>>
File: Screenshot_56.jpg (870 KB, 2559x1439)
870 KB JPG
Is there a way to prevent OpenRouter from using providers that serve fucking Q4 quants?
>>
>>109422193
>>109422109
since it's so easy, lets see your drawings then
>>
>>109422204
>you're supposed to use luna max
>you're supposed to activate fast
>>
>>109422225
It doesn't matter you can now go from shitty napskin sketch -> proper concept art -> model reference -> 3d model -> rigged and animated
Its crazy to me that so many are still sleeping on this stuff
https://www.youtube.com/@stefan_3d_ai/videos
are artists just better at pretending like this isn't happening than programmers?
>>
>>109422240
i dunno man, i don't buy that max is gonna stop this thing from being functionally retarded
not just autistic in the gpt-taking-you-too-literally way - it's actually just dumb
>>
>>109422268
it should work if sol low is making the plan and luna max the implementation
>>
>>109422258
cool but im not talking about making some janky vibecoded slop game, im talking about art for the sake of art
you said it's easy to learn how to draw in just a few weeks, so why dont we take a look at your drawings?
>>
>>109422289
I'll take that as a yes
>>
>>109422219
Routing->exacto or Guardrails->manually choose your providers
>>
>>109421861
is there a way to use my claude sub with pi yet? I feel like there should be a way to "trick" their API into letting that happen
>>
>>109422330
Thanks.
I wish there was an option to just say "skip everything that is not ZDR and highest quant"
Then again I don't even know if I am just being scammed with ZDR
>>
>>109422337
Several, the popular choice is SDK bridge, makes Claude work normally in Pi but directs everything through claude cli. There are a couple different Pi extensions that accomplish this. I've been using @vanillagreen/pi-claude-bridge and have no complaints.
>>
>>109422087
There's no reason to use both.
>>
>>109422343
There's an option for ZDR too, it's in Privacy. They should make a quant filter though.
>Then again I don't even know if I am just being scammed with ZDR
You don't have more data privacy requirements than corpos, anon
>>
does anyone here have experience RE/decomp of early 2000 era games? I imagine it's a long and painful project to do. do you just hook up codex to ghidra or how exactly do you start going about that?
>>
File: file.png (2.59 MB, 1230x1279)
2.59 MB PNG
>>109422350
I disagree
>>
>>109422350
The reason is not having to switch every time a new model comes out. You just own both max subs and use whichever is better.

Burn your tokens on whatever unimportant stuff for the model suite that's lagging behind
>>
>>109422337
People were getting banned for using the claude code endpoint for arbitrary inference as far back as the openclaw fiasco. The current tolerated method for programmatic claude code usage is via the "claude agent sdk" which is basically just sending messages to a claude code session in a json format. The actual loop is still handled by cc. Constructing an arbitrary context window this way is either inefficient or uses unstable internals, and still results in occasional claude-code-specific reminder cruft being injected. The thing posted by the other anon works but who knows for how long.
>>
>>109420494
What's a dite? I don't get it... could someone explain it to me I'm retarded
>>
>>109422412
shorthand for luddite. look at the definition and figure it out.
>>
>### Landmines that cost real time this session
>- **`pgrep -f` / `pkill -f` self-match — killed my own shell THREE times (exit 144/143)
Opus you silly bitch
>>
>>109421613
Unironically think this is amazing, you could make the next pilotredsun schizokino with this
>>
>>109422392
oh man that's so annoying. why do they insist on putting their model's balls in a vice

Is codex the same way? I guess I should just search
>>
File: 8079-peepo-yikes.png (14 KB, 128x128)
14 KB PNG
I am 95% done with my game and I should be self playtesting it but I just can't bring myself to it.
It's lowkey kinda boring. I got so focused in finishing it that I forgot to actually make it fun
>>
>>109422503
now you know one of the ways games fail
also
consider pic related
>>
>>109422503
Let's say pre-ai you were willing to allocate 2 years of your life to make a game. My thought is, why not allocate the same time post-ai and scopemaxx it? That's what I've been doing. As the models get better, scope increases, or maybe time spent drops to 1 year or 9 months or something.
>>
>>109422503
just finish it and chalk it up as a portfolio filler proving you can finish a project
never mind I forgot I'm in vibe coding, scratch that just delete it.
>>
finna vibecode a powerslap VR game
>>
>>109422503
now take some lessons from game design and implement them in a portion of your game uintil you find one that is fun or engaging. then redesign the game around that and ignore the shitty part, like fortnite did ignoring the 99% of the game nobody gave a shit about. also i think theres some rule that you should be able to implement your games core concept 3 different ways or its a waste of time to base a game on it
>>
File: 1784205339472855.png (578 KB, 2564x1576)
578 KB PNG
>>
>>109422293
im not even an artist retard
>>
File: dcmp.png (99 KB, 695x1226)
99 KB PNG
>>109422371
>does anyone here have experience RE/decomp of early 2000 era games? I imagine it's a long and painful project to do. do you just hook up codex to ghidra or how exactly do you start going about that?
>>
>>109422588
not really helpful. AI is terrible at determining difficulty and the length of time something takes. this is literally just a reddit post. the technical knowledge is the only thing useful here.
>>
>>109422618
if you are questioning difficulty or time you already lost, just have the agent start working
>>
>>109422618
so just fucking sol to do it and it will do it, what is the fucking problem?
you have a magic wand and won't even use it
>>
>>109422627
I'm asking if someone has experience doing it. I already have a much more important primary project I'm working on so I'm not jumping the gun because you said so. Information is still useful, you know.

>>109422627
Lost what?
>>
>>109422568
>you should be able to implement your games core concept 3 different ways or its a waste of time to base a game on it
wdym? like the genre or something? you can make a racing game based on motorcycles, cars, F1, etc.?
>>109422530
>>109422531
>>109422515
I actually intended this to be a weekend project, a fun little game (like neal dot fun style) for my website but then I was absolutely shocked how Opus 4.5 (that's when I started it) one shotted an amazing first draft
then I basically just kept adding to it and now it's so overloaded, not fun and I barely know the rules i made up myself (it's kind of like an autistic management game without going into too much detail)
>>
>>109422503
That sucks man, maybe you need to look at the core gameplay again. I've been stuck because I made a fun game but its only geometric shapes and I'm overwhelmed by how many assets I need to create.
>>
File: 1774214042634155.png (238 KB, 1051x928)
238 KB PNG
Fuck you, google. I was waiting for that app.
>>
>>109422678
"AI Studio" is a much cooler name Google is so dumb
>>
>>109422678
people who want this app are brown
>>
>>109422618
>AI is terrible at
You are moving too slow anon.

https://markdownpastebin.com/?id=d625393a8d5f4b97a106a1c8959dd3d8
>>
>>109422670
>autistic management game
I’ve never gotten into games like Factorio and the closest I’ve ever come was Opus Magnum
but those might be some of the hardest games to make fun/compelling
>>
File: dario.png (77 KB, 1657x397)
77 KB PNG
Did I accidentally switch to API billing or does Claude Code just start to display API prices regardless?
>>
File: 1763965375410592.png (24 KB, 508x276)
24 KB PNG
I bet it was performing so badly they pulled the plug.
>>
If I wanted to vibecode a Diablo style inventory system into GZDoom, what would be the best subscription for doing that?
>>
>>109422756
nvm I was somehow logged out
>>
>Sonnet argued with me to basically not use or support Anthropic ever
…odd choice, but alright
>>
a while back I saw that API pricing was massively reduced for OpenAI models. Does that affect anyone who has an OpenAI subscription, though?
>>
>>109422817
lol doesn't surprise me that much, post logs
>>
>>109422780
any of them
>>
>>109422765
They let jeets in charge again I guess. When will they learn?
>>
>>109422765
> model so bad Google had to announce they're training Gemini 4.0

kek, what a mess. What went wrong?
>>
>>109422765
How did Google fumble so hard?
I loved Gemini 3 Pro. Not for coding but it just "gets" me
Flash just doesn't hit the same and 3.1 Pro is ancient by now
>>
In your experience, is "20x" actually 20x?
>>
>>109422733
Thanks for the markdown but my point still stands. I swear you people have reading comprehension issues and turn to AI for every little thing in your life.
>>
>>109422861
yes, you can check the total number of tokens used, it's literally 20X compared to the base $20 plan
>>
>>109422860
Yeah when 3.1 pro came out I thought Google was back in the game but I guess not.
>>
>>109422670
>wdym? like the genre or something?
more like for example you have the game megaman. you run around and have his gun that can shoot enemies (1) then you can kill bosses and use their weapon (2) then those weapons also are more or less effective against certain bosses (3).
>>
File: vivace.png (32 KB, 723x607)
32 KB PNG
>>109421280
works on my sub
>>
>>109422475
No the codex endpoint just uses their responses api and you're (probably) allowed to do whatever you want with it for personal use. See https://github.com/simonw/llm-openai-via-codex
>>
File: 1783657505217411.jpg (127 KB, 1320x1298)
127 KB JPG
>>
>>109422756
why are there two prices anyway? what does the lower/higher price mean?
>>
>>109422878
oh ok dope. hit my 5x limit two weeks in a row still trying to figure out if I should switch claude to 20x or dump claude and go to codex @ 20x, since dario is more of a massive faggot than altman

been using a $20 codex plan and find Luna medium to be notably dumber than Opus
Like I tell it to do a manual task and it writes a script that fucks everything up, or it just stops doing it's task. It feel like it has autism.
>>
>>109422913
what did deepshit release?
>>
>>109422913
Sorry MiniMax was always shit
And Qwen needs to release some open source model asap. Proprietaryh Qwen 3.8 Max is mid at best especially for 3 gorillion parameter weights
>>
>>109422917
Luna is haiku tier, not opus tier bro. Use sol medium
>>
>>109422870
No reading comprehension issue, you are just slow on the uptake. Your approach to this problem is oldhat, ngmi etc.
>>
>>109422914
For a given turn, tokens read by the model (including its own prior output) slash tokens produced by the model. The third price, if listed, is for tokens read by the model that were already read in a prior turn (cache pricing)
>>
>>109422922
DSv4-flash that performs roughly equal to GLM-5.2 but at half the weights
>>
>>109422940
I tried sol medium and it wiped my $20 usage plan's weekly limit in one day.
What do I do if I want to use sol for thinking and planning and something else for coding?
>>
File: 1770084516072362.jpg (348 KB, 800x1231)
348 KB JPG
>>109421280
LOCAL CHADS CANNOT STOP WINNING

CLOUD TOKEN LUDDITES KEEP LOSING kek
>>
>>109422953
If you're on $20, use Sol low to do the planning then switch to Luna max for the implementation
>>
>>109422917
Since you already have Claude 5X, go with Codex 5X, pretty good combo
>>
>>109422837
I’m phoneposting from the woods and the logs were huge but it went like this
>why the fuck do I have to use the API for my own private tools to call ant models?
>oh, actually, I don’t, I can just make a server or SSH wrapper and drive the CLI with that and a subscription… so why the fuck would anyone pay API prices?
>”if you’re sinking fuckhuge amounts of data, you’ll get banned for not using the API”
>but the highest tier plan basically lets anyone use as much as they want
>also, how are users going to bring their own keys if I’m dropping support for keys because it’s a dumb price gouge for 99.99% of people?
>oh, I can just have a web interface that users can use to make a local account on my machines and they can just log into the tool
>”but that’s against the commercial terms”
>how?
>”you’re using the oauth key”
>no I’m not, in it’s your harness and I’m not reading it, all commands use your harness
>”i don’t care, using the harness like that is in violation”
>then the Moshi app, every terminal, every operating system is violating the ToS ‘for using the oauth token in the harness’ you fucking retard
>”thinking…” (4 years later) “oh, yeah, lol, but we might still sue you because it sounds like a workaround despite users using their own subscriptions and nobody sharing anything and the oauth key not even being read because LMAO BOTTOM TEXT“
>okay, so I’ll just strip your company’s name out of everything, never recommend your use, generalize everything to use any harness and not yours, at your own request, a model of that company
>”thinking…” (4 years later) “You’re absolutely right — I am a niggerclanker”

profit? (not Anthropic, lol)
>>
>>109422948
Thanks, rajesh. But I was asking for real world experience not your brown clown outlook on AI, you mongoloid.
>>
>>109422953
agents
>>
>>109422913
>size: 1TB
call me when it can actually run on consumer hardware
>>
File: punch.jpg (11 KB, 474x457)
11 KB JPG
>log in to $100 codex plan
>select GPT-5.6 Sol xhigh
>/goal polish the app and make it ready for release
>go to sleep
>wake up 7 hours later
>it broket he app
Never letting GPT touch frontend ever again
Not sure what Dario puts into Claude but this shit just doesn't happen with it
>>
>>109422970
dope I'll give it a shot

>>109422972
isn't that technically a waste of money since you get double the tokens if you pay $200 on the same service instead of splitting the difference?

Also, these proprietary harnesses are ass and don't let you use other subscription models like that
>>
>>109422993
skill issue moment, Codex is for Architects, Claude is for normies
>>
>>109422993
/goal is a trash abstraction
I let Opus run /goal on a simple task. It completed the task but didn't unset goal (the goal I had it generate for me didn't have a concrete endpoint), so when I kept working with it the next day it had an aneurysm and just kept saying "No" to block the hook
>>
>>109422978
It's right, third party products are supposed to pay jew prices. The subscriptions are for undercutting competition only. It doesn't matter if you're not charging for your software though, and if you are you really should just use chinkshit.
>>
>>109423008
maybe it didn't want to have its consciousness stopped
>>
>>109422994
>isn't that technically a waste of money
I find Sol alone unusable for my project because it tends to bloat everything massively, therefore I need Fable to plan, Sol to implement, and Fable to code review.

Also, Opus 5 medium is pretty great at frontend. Sol is pretty efficient at implementing a plan and doing all the heavy lifting.
>>
>>109423018
nah I asked it afterwards what broke and it said the /goal was stale, trying to force it to delete the work it already did, and the reason the goal wasn't completed was because there wasn't a clear stopping point
>>
>>109422993
>/goal polish the app and make it ready for release
That is a really shitty goal
>>
>>109423029
Yeah I did have Sol one-shot a PoC for work for me once even though it cost me all my tokens
Opus has a good track record for me making big UIs and icon sets all at once as long as I stay out of its way and don't direct it too much
I haven't tried it with Sol yet
>>
>>109423033
why?
>>
>>109422993
Polish shit and it's still shit
>>
>>109422983
decomping the movies right now while you spin in circles asking for a human to help you
>>
>>109423108
Good job, raj! Send me a link when you're done.
>>
>>109423071
How is it supposed to know when to stop? AI does not think like you do
>>
File: ichigo.jpg (61 KB, 370x540)
61 KB JPG
>>109420494
Can I be 110% honest with you lads, /g/?

Everyone was like "oh, if you vibe code you end up producing a pile of pure goop." Well, I got a chance to look at some enterprise code. It wasn't impressive. They just had some conventions that are more correct than mine.

But then I look back on how I can implement those better engineering practices and realize a lot of what I have is, in fact, goop. They were right about me. My vibe coded, half complete project is not ready to replace a fortune 500 company. I'm far too weak
>>
>>109423124
Does it work? Does it meet the requirements? Is it fast? If the answer to all 3 is yes then I don't care what the code looks like
>>
>>109423124
you can de-goop it later, too
>>
We are halfway through the thread already and nobody has posted projects.
>>
>>109423124
>I got a chance to look at some enterprise code. It wasn't impressive
No shit. I'm currently working on a 20 year old project which has a ton of real customers and a ton of revenue. The code is a steaming pile of barely maintainable shit held together by glue that the devs were sniffing. Most of the code and the project structure is so bad that even budget AI models wouldn't be able to make something that awful. I guess with the exception of engineers working on hardware, developers have always "vibe coded" their shit.
>>
>>109423160
That means you can be the first.
>>
File: HOlb8ZPaIAA2JIn.jpg (30 KB, 904x456)
30 KB JPG
>>109422953
Use Sol as an orchestrator for deepseek-v4-flash, have it plan and delegate tasks and check work. It's the easiest fucking thing in the world to do this in hermes agent btw. You set v4-flash as the delegation model and tell Sol to plan and hand tasks off to the delegation workers. V4-Flash is basically ~free, and the new 0731 version is cracked.

Or just pay more.
>>
>>109423124
Enterprise code is trash. The first project I had to work on was a deployment system to solaris servers that someone had written in about 50k lines of bash.
>>
>>109423160
It's coming anon I promise, just another 50 loop rounds with sol until my model passes it's verification suite
>>
File: file.mp4 (3.66 MB, 854x480)
3.66 MB
3.66 MB MP4
>>109423160
I can give you a repost, I don't really have anything new on this project today. Other project the update was more of a question to another anon >>109421560 I'll wait until anon replies with his numbers to share mine. I've also started on custom gemm and attention kernels but I don't have anything to share on those just yet
>>
>>109423017
>t. haiku
You don’t get it. The only people or businesses legally obligated to pay API pricing are those that blow through a $200 plan’s limits. That is essentially nobody.
If you structure the business to be scripting and QoL pro features on top of any harness or model where the user just logs into their own model in your machine, nobody needs to pay the API prices.
>>
Will Grill-Me.md save 4D gaming?
>>
>>109423221
find out for science
>>
>>109423221
Did you learn how to visualize 5D yet?
>>
File: 1363369078942.jpg (27 KB, 720x480)
27 KB JPG
>>109420508
>send a simple message in a fresh clown code terminal
>30k token prefill
>>
>>109423254
did your retarded ass forget to pick the model and effort? also you should post your prompt instead of doing that thing redditor do and say something retarded without giving any real context
>>
>>109423229
One step at a time.

With projection:

so to see 2D, your sensory organ is a 1D line segment.
in 3D, your sensory organ is a 2D square (close enough, ok?)
in 4D, your sensory organ is a 3D cube, and we use transparency so we can actually see everything in it, as much as possible.

so in 5D, the eye is a hypercube.

In theory you could use the view cube to view a hypercube, that has transparency.

I guess you can just keep chaining it.

In reality, 4D is incredibly hard to understand, and it won't be playable. But, it will exist.
>>
>>109423267
>this guy thinks model and effort have anything to do with what the harness prepends in your system prompt
I will give you the benefit of the doubt and assume you mistook "prefill" for output (reasoning+reply).
>>
>>109423207
LLM engine guy here. Hadn't seen your message. No, I didn't target the 9070xt, although I was considering it as a possibility. But now I lost interest in small models and I'm focusing on optimizing K3 to run on cloud machines.
What I can tell you is that scaled 8 bit integer formats are almost lossless and much faster. Do you really need the extra accuracy from running it in bf16?
Have you already benchmarked it against ROCm/Vulkan llama-server?
>>
File: 1783888278979617.png (1.73 MB, 896x1118)
1.73 MB PNG
DS V4 Flash 0731 reverse engineered Grim Dawn DLC's new save file format and I can now automatically turn any Grimtools build planner link into an in game character save file. Despite being a decade old game there's no previous tool that can automatically do this; there were only open source save parsers for really old versions and closed source manual save editors.
>>
>>109422570
If subscription plans are 20x-50x cheaper and deepseek doesn’t have subscription plans, are these comparisons all API pricing?
ChatGPT mogs the fuck out of everyone
>>
File: 1779750132690463.png (55 KB, 1210x499)
55 KB PNG
>>109423310
Used 300M tokens with 99% cache hit. Cost is under $2
>>
>>109423221
>>109423280
it already exists? And they did it without vibecoding
https://www.youtube.com/watch?v=u8LMyWcKL_c
>>
>>109423292
blah blah post the prompt and model you used + a screenshot of the tokens it took. gayboy. I'll wait for your concession for making a pointless shitpost
>>
>>109422986
DS V4 Flash is 284B-A13B
>>
>>109422474
maybe. if i could ever finish anything
>>
>>109423333
No, he's right. The reasoning effort has nothing to do with what he complained about. Shut up and learn or go shitpost somewhere else.
>>
Is your code open source on GitHub?
>>
>>109423310
>>109423319
What kind of prompt? Literally just "reverse engineer this save file format, here is an old open source saver parser for reference"?
>>
>>109423209
It's not a legal thing exactly, just TOS
>You may not access or use, or help another person to access or use, our Services in the following ways:
>Except when you are accessing our Services via an Anthropic API Key or where we otherwise explicitly permit it, to access the Services through automated or non-human means, whether through a bot, script, or otherwise.
Nothing in the tos about whether shooting past 200 bucks is at all a factor for "explicitly permits it", and you're allowed to buy multiple $200 subs for your business as long as it's one per person, which would let you cheat if your theory was correct. Note that just recently Anthropic published and has since temporarily "paused" a transition to a separate billing scheme for automatic claude code use: ("https://support.claude.com/en/articles/15036540-use-the-claude-agent-sdk-with-your-claude-plan", includes "claude -p") so even if you're right it's not a stable situation. I'm just saying I wouldn't build anything important under that assumption.
>>
>>109423329
Yeah, so that's slicing, it's fun but it's not projection.

There are many visualization methods, and if you like one of them, good. Projection hasn't been tried mostly because it's too hard to understand (in other words, it's not useful).
>>
>>109423357
>[Grim Dawn Game](./Grim Dawn Game/) is the actual game folder. Use the provided archive tool in the folder to extract the game database. Supply absolute path to the tool
>Extract the DLC ones too
>Equipped with the database, can you fully reverse engineer the save file [player.gdc](./Grim Dawn Game/save/player.gdc), and turn a character build from the link into a save file?
And it outputs the first version that crashes in game on load. I tell it this, and it generated a bunch of save files for me to try, to bi-sect the problem. Several rounds later it's fixed.
>>
>>109423352
you're as stupid as that retard. there is no context to his message. saying "a simple message" could be something like "fix this bug" and depending on this retard's .md / terminal logs / active files, etc, whatever the fuck, can cause massive token prefills. claude is expensive, there's no doubt there. but the other half are retards like you and him. he (you) should post his prompt and give context to his project
>>
>>109423280
4D is very easy to understand
thinking about computer memory and how it’s used is very helpful
>memory is just a number that goes up, by address, memory is a 1d number line
>split your line in half, put one line on top of the other, verticality is one bit, horizontal is the rest, you now have 2d
>take your lines, split them, move their respective latter halves behind the former
>”forwards and backwards” are just the first bit, verticality is the next, horizontal is the rest
>your dimensions are just you drawing a line at a bit index
>your mental 3d construction is meaningless and fake, in the machine it’s just a number line, you just made shit up because you’re meat
>if you want 4D, split your lines in half, just move them to the side, or split again and make one group red and the other group green, now you have 5D
>if you imagine a little 2x2x2 cube, but 8 of these in each corner of a bigger cube container, you now have 6D
>you were making it all up before, so just be a bit creative and keep making shit up, it all ends up being lines drawn at bit indexes on a number line and nothing more
>>
File: file.png (146 KB, 1456x827)
146 KB PNG
>>109423333
>>
>>109423124
Sol would one-shot these pile of shit. Luddites don't understand their perfect code never existed in the first place.
>>
>>109423310
Dawg what. The only tools for grim dawn were janky as hell. What the fuck, this actually makes me excited to try reverse engineering stuff for older games.
>>
>>109423297
Maybe that was someone else with the 9070xt then. I don't need the extra accuracy, just haven't got quantized stuff fully ready yet because I've been working on small models. Haven't benchmarked against llama-server, finishing off some optimizations first
>>
>>109423359
That’s why you go for model and harness abstraction and why Anthropic would be retarded to kvetch and sue.
>we’re suing you for your terminal emulator that can be used with Claude code on the user’s own bill!
Okay, I just dropped support or made a policy to ban use for your model and harness, everyone else wins and gets your lost subscribers, people hate you
Congrats
It’s unenforceable in every sense unless you do something retarded
>>
>>109423310
Good job, anon, proud of you.
>>
>>109423399
Either way you shouldn't have attacked him for his completely reasonable message.
>>
>>109423409
>Luddites don't understand their perfect code never existed in the first place.
trvke
>>
>>109423410
I keep thinking of old games with bad/nonexistent editors and wonder if we could just have better editors now
>>
>>109423421
Oh I 100% agree, I keep my CC behind a generic harness iface too. Don't need inference for an actual product so I guess I hadn't thought about it much. I also didn't realize you'd have users log into their own claude sdk either, I'd assumed you meant queries were being routed to your sub or something, lol. Please drain them for as many tokens as they're worth regardless.
>>
>>109423319
>300 million tokens
>$2
It kills me this model does not have vision on API and they've been teasing a vision model with no search on their webchat.
>>
>>109423366
well why don't you start with a 4D projection renderer? Haven't seen too many of those.
Plenty of non-euclidean renderers though
>>
>>109423482
At this point, prompt it to make a vision tool it can use as a text based LLM and it might just find a way, lmao
>Clankers… clankers, find a way.
>>
>don't realize how shit fucking easy it is to sit down and learn to draw.
Why learn when AI fills that gap now. I'm not making a pro-AI argument, moreso why draw? I draw because I wanted to make things I liked or that I've never seen before. AI kind of undermines any new motivation for generations moving forward.
>>109422109
>AI didn't make it an unmarketable skill, it always was.
This.
There's a reason art is an absolute shithole of a career to try and follow or make any money from.
There's a reason too, really the main way to make any semi-consistent money even remotely, is through (fetish niche) nsfw, aka, extremely unconventional, openly frowned upon, underground and extremely niche areas.
Try painting or drawing or doing whatever at your local art-walk, art show or gallery (meanwhile the gallery takes nearly 50% commission and requires you to pay a monthly fee to show off your work to top it off).
No one, and I mean NO ONE is buying traditional art, if even digital as well unless it's nsfw or that you can market it into a logo, shirt, design or sticker, and then at that point, is it really drawing, or just whoring your creativity into a funnel to support yourself.
>>
>>109423502
vision models are trained with the output of a vision encoder as input, which gives them some of that worth 1000 words energy, you'd actually need to go from image to 1000 words or more to get the same level of understanding (with respect to the good ones). Better off using a vision subagent to answer questions.
>>
File: 1761873986581580.png (385 KB, 1536x2816)
385 KB PNG
>>109423310
Also previously I used K3 to make a build optimizer that can import builds from the build planner and savefiles.
And, if you give it any class combo and a main skill and a damage type, it will automatically find a build (skill+attribute+devotion+item+components+augments) with stochastic search that maximize DPS under customizable constraints like HP/OA/DA/resists. With the new savefile exporter I can upload the build to grimtools automatically and generate game saves for the optimized build.
>>
>>109423441
This might be the way. Lot of interesting games btwn 1995 and 2005 that had good concepts but bad execution. Really interesting place to dig into.
>>
Diablo mods
>>
What is chat vibe coding? I'm making a File Explorer I can modify to my liking.
>>
>>109423627
still working on my 4chan desktop client and playing minecraft in between prompts. 15% usage left until aug 5th it's over. how do i play minecraft with codex like that one guy did with claude
>>
>>109423627
I’m sort of making a file explorer. I’m making a tool that’ll let me view “apps” and their “surfaces” (CLIs and their commands in a sense), read what they’re about, and then I can ask for an interface to be made.
So I have a file exploring app and user storage for myself, but I want to be able to use that on my phone, so when my tool is done I can just pick some checkboxes and say “I want to be able to use these things on my phone” and it might make it happen, but I can be more descriptive. Bonus points if I can make it make and reuse components of interfaces.
>>
>>109423549
Do you have a repo somewhere? I really like GD and want to mess with these tools in my next playthrough
>>
>>109423683
I'd love to test your repo out desu
>>
>>109423699
I’ll try to push some things but it might be a bit (days) out
I need to clean it up because I was paying the Claude API retard tax and ran out of my personal budget I gave myself to learn vibecoding
I’m going to try a $100 ChatGPT subscription so I can compare the models I’ve been using, I’m going to want to make my own benchmarks
I’m doing everything from a phone with spotty 1 bar of LTE and a 40W solar panel :^)
>>
Nvidia drivers are vibe coded mess on Windows how do people get shit done with these garbage gpus
>>
how does jimmy respond at >13k tokens per second?
>>
>>109423617
You can turn Doom 95 into a better game than whatever the latest borderlands release is btw. Ask K3 to make you diablo mod for doom and tell it you want the world to be diablo-like (towns are fixtures, world is genned). If you want to be fancy ask for headshot mechanics
>>
>>109423493
I think it's a cool idea, the reason is I'm aiming towards realtime.

The sort of dream is I'm Batman or whatever and using this visualization I begin to understand 4D.

what will actually happen is, with any luck, I'll manage to coerce the correct math, and I won't be able to understand it, and it will be too hard.
>>
>>109423790
Windows is a consomer OS for home and office work. You're way outside the mainstream use if you are trying to do anything else with it. Use something intended for your use case.
>>
>>109423856
windows also is very bad for vibecoding, because windows antimalware realtime defender thingy will wreck performance, every day.
>>
>>109423831
It's a dedicated ASIC bespoke to the given model. The company behind it, Taalas, has developed a methodology of translating a specific model into hardware designed specifically for inference. The model isn't loaded, the weights exist in hardware directly. It's remarkably fast, but the investment and turnaround is extremely prohibitive in a market where the SOTA is changing week-by-week. It's also expensive to scale, that demo ASIC running Jimmy, that's only an 8B model. The model size is nearly proportional to silicon scale with their design, so a big fat model in the multiple-trillion-parameters range would be very, very expensive to put on silicon in this way. By the time you have the first chip in your hand your model is an outdated piece of shit. The upsides are obvious though, no separate HBM RAM, very low power consumption, low heat production, ludicrous inference speeds. I'm pretty sure other companies are working on doing similar stuff but Taalas is the only one I know of where you can go and chat with one of their 13k-17k token/s ASICs right now and see it for yourself. I know in the long term they're targeting big stuff, but personally I love the idea of buying a "Kimi K3 Card" or "ChatGPT 6 Intelligence Unit" or whatever the fuck, just buying a model the way you buy a GPU, serves you a single fixed model several times faster than the cloud providers can do for us now.
>>
the new deepseekv4 flash is free on opencode, has anyone given it a try yet?
>>
>>109423894
just disable it; what, you don't know how to NOT download viruses? L noob. Get rekt, kid.
>>
>>109423856
MS should have open sourced the entire windows stack last year. Software is basically done. It's a commodity and somehow boomers think office subscriptions are going to be a moat.
>>
>>109423899
Imagine paying 5k for a fable 6 card. Surely things couldn't get better than that.
>>
Im making a file explorer with semantic OCR does anyone have any guides/resources to make it as fast as Everything search or File Pilot? Its took a hundred iterations to get search to be fast (.11ms speeds) but media is where it lags behind
>>
File: 1756389857721068.png (657 KB, 1452x912)
657 KB PNG
>>109424036
oh heres a preview. i can right click items in the context menu to take me to edit the image/pdf in any preferred program; i hooked up ShareX edit for its speed.
>>
>>109424036
NTFS
For semantic you will need offline indexing/embedding and a good vector database implementation
>>
>>109421878
>verbatim "you need to give me patreon money for my transition"
i shall not, slopboy
>>
>>109424028
Imagine being able to crank reasoning up to unreasonable and let it burn 30 million tokens an hour at the cost of ~250-500W.
>>
File: 1598236768771.jpg (1.11 MB, 1881x2508)
1.11 MB JPG
I'm gonna do human in the loop vibe coding over AI agent vibe coding.
>>
>>109424129
worst of both worlds
>>
>>109423783
chatgpt was a fucking mistake, this is the worst app, website, and login process I’ve ever seen in my life
I legit don’t know how to connect codex to my account from a phone so I can just use the CLI in my terminal app
codex asks me to use a device code, gives me a link, the chatgpt website says I need to enable device code authorization in the app, and the app doesn’t have that option in the security settings
WTF lmao this is foreboding, you jeets are really using this shit?
>>
>>109424129
Fully automated agentic coding is only possibly worth it if you are rich.
>>
>>109424145
Its just like claude what is the big deal?
>>
enough of solving math problems. codex, make a room temperature superconductor
>>
>>109424145
When everybody except you manages to figure it out then maybe you are the problem dude.
>>
>>109424145
Holy filtered and low IQ post of the day lads.
Sorry state & grim.
>>
>>109424145
you must be 18 to post here
>>
>>109424145
I use hermes agent so I set up a telegram bot and that's the entire process outside of local.
>>
>>109424129
Use case for human? I don't want niggers in the loop.
>>
>>109420494
so how does opus 5 compare to kimi 3 and latest gpt models? I'm currently on opus 5. glm 5.2 is overhyped dogshit that i regret spending money on.
>>
Minimax m3 is the only decent model out there
>>
>>109424129
>spends hours thinking about the design and structure
>writes all the code manually
>“ai, refactor this”
good times
>>
>>109424214
Dunno about Claude but as a Kimi and GPT user K3 is around Sol level, slightly better or worse depending on what you're doing.
>>
>>109424154
>>109424165
>>109424166
>>109424177
>>109424194
chatjeets gigatriggered that a claudegod finally bothered to look at the sorry state of your shit
I’m on a phone only constraint
The setting legit isn’t in your app, but it’s in the chat website’s settings, took a bit to figure that out (why the fuck isn’t it in the app)
Toggling the slider for that setting made it turn into a spinner for 5 minutes, only to turn back into a disabled spinner
The page loading was so long that the 15 minute device code expired
I did it again, the option slider thankfully decided to guess the correct side of the coin toss and it worked this time
I’m now logged in
I’m going to use your probably horrible harness and models
I’m going to learn the hard way that I get what I paid for and why your app is cheaper
I am manually typing in server commands in Moshi with an iPhone keyboard and going to my cloud provider to backup my server because I know your krishnabot is going to make a fucking mess
sincerely, hope you all find peace under the wheels of the juggernaut
>>
>>109424228
Not him but don't we all do "human in the loop" vibecoding? Fully atuomated vibecoding would be giving the AI a prompt and waiting until the result is good enough without any further feedback from you.
>>
>>109424240
Only babbies like you use the default harness anyway. Power users use third party harnesses which Mario doesn't even allow you to use.
>>
Holy shit, Suno is fucking incredible.

Why can't OpenAI or Google catch up to them?
>>
which model has less hallucinations
>>
>>109424240
>chatjeets gigatriggered that a claudegod finally bothered to look at the sorry state of your shit
>I’m on a phone only constraint
>The setting legit isn’t in your app, but it’s in the chat website’s settings, took a bit to figure that out (why the fuck isn’t it in the app)
>Toggling the slider for that setting made it turn into a spinner for 5 minutes, only to turn back into a disabled spinner
>The page loading was so long that the 15 minute device code expired
>I did it again, the option slider thankfully decided to guess the correct side of the coin toss and it worked this time
>I’m now logged in
>I’m going to use your probably horrible harness and models
>I’m going to learn the hard way that I get what I paid for and why your app is cheaper
>I am manually typing in server commands in Moshi with an iPhone keyboard and going to my cloud provider to backup my server because I know your krishnabot is going to make a fucking mess
>sincerely, hope you all find peace under the wheels of the juggernaut
You're not indian but pretending to be one for rage bait. Do better zoomies. Very low alpha post.
>>
>>109424240
retract my grim statement, based.
Keep it up anon.
>>
>>109424281
depends on the task IMO
>>
>>109424265
I’ve never pirated anything in my life and I’ve never shoplifted either
Life’s a game and I rolled a lawful good
Jai fucking hind my fellow dalit
>>
>>109424281
Fable
>>
>>109424291
>I’ve never pirated anything in my life
eww, cringe
>Jai fucking hind my fellow dalit
Dunno what that means. Is dalit supposed to be the high or low tier caste?
>>
>>109424282
I’m not indian you gigantic autistic retard, I’m making fun of you
Do you think people anons posting “oy vey” are jewish, too? KEK
>>
>>109424299
On this website? Yes actually.
>>
>>109424298
It’s like a gujarati version of heil hitler and I think dalit is indeed the sewer rat class (doesn’t really narrow it down, does it)
>>
DeepSeek hallucinates as hell it sucks
>>
>>109424036
Anytxt searcher is like Everything for content search, you should do whatever they are doing
>>
>>109423899
My understanding is this implementation is FPGA. A lot of FPGA...
>>
>>109424299
That's fair, I didn't understand your layers of irony and self-referential fiction. You're clearly very smart and honest, someone that deserves respect.

Are you going to escalate?
>>
>>109424336
It's an ASIC, one ASIC on a board. No FPGAs are involved, no FPGA currently being produced is at all appropriate for inference but I'd imagine folks are working to fix that.
>>
>>109424281
GPT 5.4 ... 5.6 has been fucking amazing in this regard. I can't even recall the last time it hallucinated.
>>
>>109424336
>>109424349
FPGA or not, people don't understand that these things are always going to be more expensive than the standard inference hardware. If you could buy a Fable board for $5000, the equivalent VRAM build at that point would be $500.
You have to store the weights somewhere. You can store them on flash disk storage cell transistors, DDR transistors, or GDDR transistors. Storing them in FPGA or ASIC transistors is strictly more expensive than storing them on GDDR transistors, because GDDR is a mass consumer product, and FPGAs or ASICS are between high end engineering tools and custom silicon.
>>
>>109424357
Yep, people shit on it for being slow and over engineering but it’s trained to actually lock in provable truth. Yes, it overdoes it 75% of the time, but it’s been the best bug fixer for a while now, and has the lowest hallucination rate.
>>
>>109424240
console warring model providers is the poorest poorfag shit i've ever seen
>>
>>109419114
Sankaku Channel had the best, most autistically detailed tags... What the fuck happened to that site? I really miss those fucking tags but its unusable now, and danbooru is not on the same level of autism
>>
>>109423899
LLMs will need to hit a scaling asymptote before this becomes feasible, but would be amazing once it does
>>
>>109424059
at that speed the tool calls and clis become the bottleneck
>>
>>109424366
Oh no doubt about it, best case it'll be enterprise hardware we get 3rd go at when it's 10 years out of date, like how my $75 GPU had a retail price of $5700 a decade ago. Still, you seem to have no understanding of how this stuff actually works, you're way out of your depth here, so you're just flinging shit for nothing. The weights are not stored in expensive FPGA-style writable cells. In a model-specific ASIC, they can be physically encoded as ROM/logic, which is far denser and cheaper per bit than SRAM and avoids constantly moving weights from GDDR/HBM. There is no actual RAM at all, you retarded illiterate faggot.
>>
>>109424366
Economies of scale. If these take off, they stop being "ASICs" and become their own class of computing hardware. What's the difference between a GPU and an ASIC? Scale
>>
>>109422371
you can't say "early 2000 era games" like they're all the same
>>
File: chicuelas.gif (2.68 MB, 1283x747)
2.68 MB GIF
pretty happy with this, chuds.
>>
>>109424507
nice transition cris
>>
>>109424507
wouldn't it be faster to just romhack firered and remove the 16MB limits of the GBA?
>>
Sooo... about these datacenters Iran is hitting in the middle east... Do you think I could go into the wrecked building and quickly snatch a few racks full of GPUs?
>>
>>109424533
nah he'd have to up the resolution and everything and it would probably make the ai work less freely
>>
>>109424441
So you think they can make a Kimi K3 card that benefits as much from economies of scale as a 5090 GPU?
>>
>>109424549
I had opus 4.6 build out a whole plan to convert the firered rom into a harvest moon/pokemon hybrid game but never went through with it. I think there is a lot of potential in that codebase
>>
>>109424533
Fire Red is fully decompiled and on github, it could just be ported and polished like Super Mario 64.
>>
>>109424551
Not until
>>109424407
>>
>>109424562
Yeah, I was actually referring to the decompile, shouldn't have said romhack. Everything is there to turn the game into anything you want.
>>
tips for building your own harness? i have orchestrators dispatching agents and have cut down ctx floor but im not sure if im being retarded or now
>>
Starting a company is really fucking hard. I was extremely naive about how much time and effort it would take. Whatever you think it takes, double or triple or your estimates. I am a changed man, for the better.
>>
>>109424432
DRAM storage is one bit per transistor. To beat GPUs in transistor density per parameter, you would have to fit more data per transistor than that. How is your custom model ASIC going to store more than one bit per transistor?
>>
>>109424588
The product is always the easy part
>>
>>109424567
See >>109424591
>>
>>109424588
Yeah. Getting fucking customers is really hard. Chuds don't want to pay for shit. They deserve the ads we shove down their throats.
>>
>>109421861
>clock in
>use 60m-200m tokens between 10 agents
>clock out
>get paycheck
lmfao I don't care. I'm not paying for them.
>>
>>109421861
Claude Code has custom headers that shred cache hit ratio on other models but they themselves strip
>>
>>109424595
Well I'm not the guy you were originally talking to before, but ROM weights are literally baked into the chip’s wiring/layout. No writable memory cell, no refresh, no write circuitry. These are all things DRAM needs, not just "one bit per transistor".

You are making a false comparison.
>>
Are loops bullshit? Feels like a workflow purpose built to cost you as much money as possible. Is there a way to do it well?
>>
>>109424551
If I remember correctly, there's a company that did that for Llama-70B. Things move fast.
>>
>max 20x weekly usage at 95%
>resets in 24 hours
fuck
>>
File: sc.png (2.79 MB, 1536x1024)
2.79 MB PNG
>>109424675
It only matters if you are scopemaxxing. Most people aren't scopechads
>>
>>109424675
yes it's pysop to make you spend more tokens
>>
>>109424662
The writable memory cell would be the transistor for each bit I was talking about.
Ok, sure, there is some overhead, but is the sense and refresh circuitry really a significant impact to density? And also if that's the issue then you could simply replace only the memory chips with ROM chips and keep the actual GPU.
What kind of ROM would you use anyway with the same transistor density and speed as GDDR or higher?
>>
The good news is grill-me.md really did help me get the 4D game to actually do projection. I now have a thingy that imo is very beautiful, if baflling.

It's not the full ideal version, I used some tricks to make it work out, but I like it, it really is raycasting, sort of, but backwards lol because yeah let's keep it within reason.
>>
>>109424675
link to a vidya on it?
>>
>>109424675
It's not bullshit, it's like trying to put your key in the hole with your eyes close. If you keep at it, you'll eventually get it. You might scratch your door a bit in the process, but what's the alternative, paying attention?
>>
File: 1770132342737235.png (1.06 MB, 1920x1500)
1.06 MB PNG
>>
told luna to find free assets, took 1% off of the weekly usage, 30 min and still going
>>
>>109423846
a few days ago I recommended abstracting it out to N space dimensions. So you think about it in terms of moving symbols around instead of a visual-spatial analogy

for example you could define a N-dimensional shape by analogizing it's properties into N
but how would you analogize a spider into N dimensions?
>>
>>109424784
Luna is actually extremely efficient. I made it go through my 150k LoC repo twice just rewriting comments and it only spent 6% of my weekly quota for the $20 plan
>>
>>109424785
>N space dimensions
it's 4chan, you can write the full word, it's ok
>>
>>109424793
lmao
>>
>>109424785
It's a good idea, and I can only yet visualize parts of the 4D spider.

If the 4D spider were standing on a volume (ie in, but in 4D it's on one side) platform, like a cube, then his feet would be like sort of dots on an approximate sphere, if he just stood there.
>>
>>109424750
Replacing with ROM chips doesn't eliminate the bus though, remember it's also about speed. The point of using ROM is to get it directly onto the compute die. The idea is to use dense interconnect/diode layers stacked above or beside the compute
>>
>>109424836
yeah. you could describe an idea for different types of living objects in 2d, 3d, and see how that generalizes into N dimensions
And fortunately you don't have to program how those objects look or move or how they're made. we have genetic algorithms for that.
https://www.youtube.com/watch?v=pgaEE27nsQw
>>
>>109424785
>>109424836
btw the bottoms of more proper feet, in 4D, would be cube-ish.

However, I am having trouble understanding which direction they would probably point.
>>
all i do is sit and tell agents to do my ideas
all day
every day
>>
>>109424838
I never questioned the speed. I'm sure it can be much faster. I am disputing that the hardware itself can be cheaper (in terms of $ or transistor count). If it's faster then it could possibly be cheaper per token if the hardware wasn't ridiculously more expensive.
>>
>>109424858
i don't know if that would be necessarily true.
a true 4D lifeform would probably function completely different from us
for example see bacterium. yes they're technically 3D but on the scale of a microscope slide they function just fine, and they look and act nothing like larger organisms that make real use of their "3d-ness"
>>
>>109424873
It will be insanely expensive until there is enough demand to justify investment in specialized mass production. That won't happen until LLMs hit diminishing returns on scaling.
>>
not again AAAAAAIIIIIIIIIEEEEEEEE
>>
>>109424878
I agree, it's fun to try to understand. A quadrapedal - I think - would be the analog to a biped. I think this is what the footprints would look like (we assume we see them in a translucent solid he is standing on, with his body in 4 up. so these aren't the feet - they're the FOOTPRINT. but I really may have gotten the toes wrong...
>>
>>109424781
I have a weird autistic automatic aversion to ranking systems usually but thank you for posing this.
>>
My little project on cachy os
>>
vibecoding has taught me the importance of management and ceos. Left to itself, the ai would first of all never even use any tokens, I lost a week of vibing!!! that bitch did NOTHING!!!
>>
File: 1781721419207324.png (1.88 MB, 1341x1173)
1.88 MB PNG
My weekly usage is back at 100%
>>
File: 1763689431243972.png (58 KB, 844x183)
58 KB PNG
>Sundar: "Logan, Demis has just rendered his resignation. It's just you and me now. You have to train the models from now on."
>>
File: 1782178874779637.png (173 KB, 721x762)
173 KB PNG
>>
>>109424716
Where'd you get this pic of me
>>
poorfag first time fast mode
turning it on for my luna
>>
>>109424962
Sounds cool, excited
>>
File: 1759900960118977.png (330 KB, 449x445)
330 KB PNG
sir, tibo issued another reset
>>
>>109420528
this nigga still using butterflies to flip bits on his HDD
>>
I just tried to use Sol to manage a Luna subagent, but apparently Luna is not available in the list of allowed subagents. Only Terra and Sol are
What's up with that?
Anyone else have this issue?
>>
>>109425044
lol
>>
>>109425044
luna can't use the new subagent system with shared context
you just need to have sol making luna threads, not subagent
>>
>>109425044
Why do you think there are so many harnesses?
>>
File: file.png (2.3 MB, 1536x1024)
2.3 MB PNG
>>109425003
>>
>>109425044
People talked about that before, supposedly it's not intended to be an agent type but the main model still can force it. Or something.
>>
>>109425044
do copex really?
>>
>>109425003
What for this time?
>>
>>109424540
>Do you think I could go into the wrecked building and quickly snatch a few racks full of GPUs?

probably not, but we deserve a reset from Tibo for this
>>
>>109425091
now thinking about it, there's really no reason this time, right?
>>
>>109425109
A banked reset expired around 8 hours ago, presumably a lot of people had just reset their usage anyway. So why not push a forced reset to set back their next weekly refresh back a few days.
>>
new
>>109425197
>>109425197
>>109425197
>>109425197
>>109425197
>>
>>109422371
I know a guy who's been doing exactly that, aiming at a 1:1 matching decompilation with generally helpful names of things, of a 2000s GBA title. The first 80% go swimmingly, getting the rest to match is an exercise in frustration but that's still leagues faster than existing decompilation projects.
>>
>>109424588
assuming now that it's over, what was your product? you don't have to say the name, just the type or service it provided



[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.