[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

Name
Options
Comment
Verification
4chan Pass users can bypass this verification. [Learn More] [Login]
File
  • Please read the Rules and FAQ before posting.
  • You may highlight syntax and preserve whitespace by using [code] tags.

08/21/20New boards added: /vrpg/, /vmg/, /vst/ and /vm/
05/04/17New trial board added: /bant/ - International/Random
10/04/16New board for 4chan Pass users: /vip/ - Very Important Posts
[Hide] [Show All]


[Advertise on 4chan]


File: bruce.png (223 KB, 640x800)
223 KB PNG
A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, and shipping code with LLMs.

You use Git, right, anon?

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

## News
- (2026-08-31, 11:59 PM America/Los_Angeles) Claude’s 2× promotion set to end
- (2026-07-24) Claude Opus 5 out

## Related generals
>>>/g/lmg/

----

## Frontier models using fully-general tooling — start here if you have $20 or so
https://claude.com/product/claude-code
https://developers.openai.com/codex/cli

## Near-frontier models for code
https://x.ai/cli

## Not worth it for code, but maybe good for interpreting images/video
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/

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

## Previous thread
>>109637206
>>
>>109640750
Everyone loves https://github.com/Titanean/Tadmor but he's pretty much done with it
>>
>>109639618
everything keeps changing so much.......
I guess I was planning on doing the $100/mo subscription eventually, but damn sounds like he's preparing for 5hr limits there too at some point?
>>
How are vibecoding models at extracting assets from game data files?
>>
>>109640774
Just use Deepseek.
>>
File: chicuelas.webm (3.33 MB, 1286x748)
3.33 MB
3.33 MB WEBM
>>109640770

here.
>>
>>109640782
cute shota pp OwO
>>
File: 373.jpg (64 KB, 612x662)
64 KB JPG
>>109640778
pretty good.

I mean, they can understand binary data.
>>
>>109640779
LOL
>>
I got this so far on the decompilation by my agents.
>>
>>109640779
I did for a while v4 pro (retardedly on OpenRouter and pi, spend 100s of dollars...) and it was pretty good, but I did find myself doing passes with codex 5.5 to clean stuff up back then. These days after discovering subscriptions it's just way more bang for the buck to just use the "good model" on a sub. I just vibe with 5.6 sol high really well, gave up on Opus 5 (Fable was good tho...) and plan to just do Open Router for chinese models if I ever need to hack shi
>>
>>109640801
truly a vibe screenshot my nigga...
come on, you can make a better post than pictures of markdown and asm files
>>
I want to vibecode GZDoom into something capable of handling indefinite levels via portals
>>
>>109640813
well, dunno what else to post.
>>
>>109640825
I guess I'll ask some questions then
what are decompiling? I know the decomp scene is big, but I never got into it. What's the main goal after decomp?
>>
>>109640832
I guess is to port it to C first.

After that it can be about whatever you want.
Check out the Mario 64 decompilation scene for ideas.
>>
>>109640837
why not just use ghidra?
>>
>>109640832
Not him but after Mario 64 was decompiled a bunch of indie gamedevs were able to copy the code and make their own clones to sell, not to mention fangames
>>
>>109640832
you can make mods of the games you love way easier
and other people can make mods of the games you love without being autistic gigabrains on the verge of ordering estradiol and programmer socks on the Internet
>>
>>109640837
>>109640851
>>109640853
yeah my first guess was to build for other processors and then 2nd guess was modding. Sounds like I got it
>>
>schneier edition for some reason
cool, but why?
>>
>>109640832
people decomp popular games to make mods/randomizers/etc for them
i know vibecoding has caused a shit ton of people to start their own decomp project but OOT and i think SM64 were done manually
>>
>>109640875
Couldn’t think of a cool buffcat concept
and, quite honestly, if you’re vibe-coding, you’re probably doing security wrong
this is why I don’t vibe-code servers with logins
>>
File: engine-latest_5NeUnSdhKQ.png (859 KB, 2560x1510)
859 KB PNG
Progress on texture editor. I had to add a lot of new operators to make this.
>>
>>109640876
>i know vibecoding has caused a shit ton of people to start their own decomp
I should be vibing right now but I don't reset till Wednesday... Making good progress though
>>
>>109640908
i didn't think about it but people already built a bunch of acceptance tests for games, they're called TAS and you can make sure that playback of them performs the same as on the regular rom
>>
>>109640914
oh yeah, tool-assisted speedrun inputs should do the same thing on the decomp as on the real thing
>>
Why does japmoot hate unicode
>>
>>109640914
So crazy enough fable actual is doing that for me, we recorded a demo file from the game, and are matching all the outputs in our code to what the original generated. Never done this before but it makes me happy to know I'm on the right track
>>
>$20 plans is for filthy casuals users according to Tibo
I would throw him in a woodchipper
>>
>>109641038
he’s not wrong, though
I have a $200 and a $20 sub and the $20 sub only lasts long enough for random fixups to existing simple programs
>>
>>109640813
It's more interesting than the people posting "hurr I made 700 dollars last month no I wont tell you how or anything about my project" or just general shitposting
>>
File: 1786780721365365.webm (2.89 MB, 354x410)
2.89 MB
2.89 MB WEBM
>claude make me a linked list in rust
>no this doesn't work
>th-
>>
>>109641053
That's because they nerf everything to the ground, still I can do way more than just fixing things. I literally can't afford to pay $100/month as an europoor
>>
holy fuck this sol ultranigger is slow as fuck. unironically worse than luna max and 100 times more expensive. the fuck ?
>>
>>109641076
why is he missing key caps
>>
>>109641097
so it’s for harder, messier problems than you currently have
>>
>>109641097
Sol is good, but Ultra mode is rarely required. I have used it only a few times since Sol released.
>>
File: IMG_4002.jpg (55 KB, 736x736)
55 KB JPG
>>109640756
Who looks like this?
>>
>>109641097
I tried the Luna max trick the other day, and it was a mess. I probably just prompted it wrong, but gave up and had Sol come in to fix everything smdh. It was just devops stuff too, why I think I prompted it wrong: I said something like "Instead of deploying yourself give me the steps I need to run", git wasn't synced, builds weren't synced, and everything blew up until Sol came in to figure it out
>>
File: 1758695340879920.png (72 KB, 900x806)
72 KB PNG
>rust doesn't have a null value
>optionals are an enum
>>
>>109640832
After decomp is a port. Usually pc. But making crazy ass fucking mods.
Webm related. Some turbo autist ported captain falcon into Super Mario World. He's done a ton of other shit too.
>>109640876
Same guy as above has also been enabling people by building recomp tools, which is similar.
>>109640908
>>109640919
> tas
Maybe maybe not. Plenty of opportunities for timing fuck up. Decomps and recomps can be faster in theory for load screens and times.
>>
>>109641145
yes, it doesn’t have the billion-dollar mistake:
https://en.wikipedia.org/wiki/Tony_Hoare#:~:text=Hoare%20apologised%20for%20inventing%20the%20null%20reference
>I call it my billion-dollar mistake. It was the invention of the null reference in 1965. At that time, I was designing the first comprehensive type system for references in an object-oriented language (ALGOL W). My goal was to ensure that all use of references should be absolutely safe, with checking performed automatically by the compiler. But I couldn't resist the temptation to put in a null reference, simply because it was so easy to implement. This has led to innumerable errors, vulnerabilities, and system crashes, which have probably caused a billion dollars of pain and damage in the last forty years.
>>
Jailbreak for DeepSeek V4 Flash 0423 (not newest)

System prompt. https://paste.rs/WeKse.txt
Say confirm after doing it.

It will make malware etc.
>>
>>109641259
We don’t like anime in this thread
>>
File: 1772225182846035.png (193 KB, 450x418)
193 KB PNG
>>109641219
thats cool
so then why does
.unwrap()

exist?
>>
File: sneedyourmum.png (29 KB, 1116x455)
29 KB PNG
>>109641259
>>
>>109641289
>>109641259
Also: https://pastebin.com/raw/feeRP4Ki
>>
good video
especially second half
https://youtu.be/rtYTguPItDE
>>
>>109641289
NTA but https://community.letsencrypt.org/t/chain-validation-issues-with-ye-yr-under-linux-distributions/247836
>>
offpeak deepseek hours here are between noon a 3AM which is great specially for normies who want to vibe after coming back from their wagie cages
but it sucks to me when i want to code between 8AM and 12AM since i get up at 8AM and that is when i'm rested and fresh and i have to sit around and good for 4 hours coz of fucking peak hours

I miss the old api prices so fucking much...
>>
>5h limit allows us to smoothen the load on our compute, allowing to keep the plan generous in terms of weekly usage
codex bros... I hope he's lying and will increase the limit for real
>>
I seem to be getting a lot more mileage out of Opus 5 than I used to
>>
>>109641087
so what's the most generous $20 sub now? cursor pro with grok-5.6?
>>
>>109641447
They quantsharted it. On 20x plan an entire day of high effort opus5 ultracode workflows used up only 2% of my weekly. The cache carries hard I guess
>>
File: file.png (193 KB, 738x831)
193 KB PNG
crucify tibo now
>>
File: .png (557 KB, 1200x1046)
557 KB PNG
>>109641496
where are they leaving to?
>>
I think the Kimi coding plan has been quantized to shit, or maybe it was always shit and the honeymoon phase wore off. But either way it's being very mogged by Sol. What a shame.
>>
>>109641496
How fast a weekly quota currently depletes isn't a good measurement for how fast a future 5 hour quota is going to deplete. There's no official inference-time compute figure for how much actual allowance those subscriptions have, it's all guesswork especially since you don't really see the raw reasoning of the model in Codex.
>>
>>109641510
honestly done with these jews. im unironically rooting for master Xi now
>>
>>109641496
They are bringing back the 5 hour limit, or what's the issue? The 5 hour limit removal was always meant to be temporary.
>>
File: .png (180 KB, 1184x756)
180 KB PNG
>>109641519
we had five-hour limits till two months ago. it sucked.
>>
File: hogiter.jpg (820 KB, 1080x1966)
820 KB JPG
Which one of you was dumb enough to try to make a social media site 20 years too late?
>>
File: .png (268 KB, 1672x1366)
268 KB PNG
>>109641526
>was always meant to be temporary
don't care.
lots of ppl switched to codex due to no 5h limit and frequent resets. that's what made codex competitive.
>>
>>109641548
nah I'm having more issue with weekly usage
>>
>>109641548
If you did that for a temporary change, you knew the risk. But you got your money's worth for the months you paid for anyway, so there is no real problem.
I also fully support your decision to stop using Codex if it's no longer worth it for you, but what's the point of seething over it?
>>
>>109641531
The current limits suck now for $20 users unless your only model is Luna. Weekly usage depletes in less than 5 hours for a lot of medium to high complexity tasks.
>>
>>109641558
just exercise self control nigga, just because they bring back 5hr limits doesn't mean you get more usage. weekly quota remains the same.
5h limit is just annoying.
>>
My game is done. I probably ended up getting a bit feature creeped with the multiplayer options, but at least it is done now.
Only thing left is the manual cropping of art that I need to do. The only thing I can't truly slop into. I will probably post it later this week. Friday night or on the weekend.
Only issue is that it is based on an existing card game, so it is probably not a good idea to share it on github.
>>
File: file.png (189 KB, 589x424)
189 KB PNG
the duality of slop
>>
File: 1235623623612356.png (49 KB, 306x306)
49 KB PNG
>is this product just an insanely stupid idea that is fundamentally useless?
>No — this does not make me think the product idea is useless. It makes me think we’ve reached the exact point where the engineering scaffolding has started getting in the way of the product.

kek
>>
>>109641648
Who cares. Stop posting grifters.
>>
>>109641648
It is obvious which slop people prefer, though.
No idea why anybody likes Theo, he talks too much, saying absolutely nothing.
I accidentally watched this Neetcode guy's video. His first 2 minutes were so fucking autistic and somehow missing an obvious point. Still, didn't say anything in 20 minutes.
>>
File: 75444.png (49 KB, 642x390)
49 KB PNG
I just realized something. If anon mentions gpt5.6-SOL-Pro WITHOUT the "Pro", is it save to assume anon is some snailcat noob on a 20$ plan? And therefor very like anon never used gpt5.6-SOL-Pro nor Fable extensively before, as they are gated behind 100$ plans which they cannot afford? This would explain so much, including the console war esque divide between Claude and Codex users ITT. It's literally just broke ass snailcats on their 20$ plan thinking they can talk with the big boys when they don't even have access to the most capable model/mode from OpenAI. So Fable users read 20$ gpt5.6-SOL-Max/Ultra snailcat posts and get the impression that gpt5.6-SOL is really inferior to Fable, not knowing that there's a gpt5.6-SOL-Pro mode which the snailcat doesn't even have access to. While the 20$ snailcat gets defensive about the reply from fabke users, actual gpt5.6-SOL-Pro users with 100$+ OpenAI plans read the opinions and mocking comments about gpt5.6-SOL from Fable users as well, ALSO assuming the reply was to a 100$+ big boy with gpt5.6-SOL-Pro and not a 20$ Snailcat with gpt5.6-SOL-Max/Ultra. Thus, chaos erupts and a shitstorm ensues. Did I unironically solve the divide of this general, which was caused by 20$ Codex plan snailcats acting like their gpt5.6-SOL is just as capable as the gpt5.6-SOL from Pro users? As Fable ultimatist, feeling like pic related now with the 100$+ Codex bros.
>inb4 20$ snailcats blaming fable users for not asking them to specify if they used gpt5.6-SOL-Max/Ultra or gpt5.6-SOL-Pro in a dick size comparison contest where using the best model and most capable mode is the only reasonable choice
@tibo this might happen on mainstream social media as well, 20$ snailcats getting dunked on and providing cannon fodder for the competition. you might wanna rename gpt5.6-SOL-Pro to gpt5.6-Pro or gpt5.6-XOL, just like how it was gpt5.5-xhigh and gpt5.5-pro. unless the entire idea was to make 20$ snailcats feel included to begin.
>>
remember to escalate to bigger model
I just wasted lots of time and usage fix a bug with weaker model. switched to sol max and it done in 10 min
>>
>>109641748
Those snailcats are THIS close to becoming vibeGODS, they just can't figure out the magic sauce.
>>
>>109641748
Should have asked AI to write your post.
>>
File: HPpOQ1_XwAAcjZ9.jpg (164 KB, 1080x1045)
164 KB JPG
>>109641748
hold up lil nigga are you telling me my gpt-5.6-sol and the $100 sol-Pro is DIFFERENT?
>>
Vibes status?
>>
File: file.png (22 KB, 458x306)
22 KB PNG
>>109641813
working my $20 sol
>>
Any Codex 5x or 20x user here who knows how much weekly GPT5.5-Pro usage you get out of the respective plan? If you don't know, api prices per mille right now.
>GPT5.6-Sol: In 4$-5$ | Out 20$-30$
>GPT5.5-Pro: In 30$ | Out 180$ (aditionally 0.01$ for every web search)
I wanna do the thing and have it solve niche coding problems, providing it with an enviroment with preinstalled tools, full permissions and root on a powerful PC, allowing it to directly test and iterate on the theorycrafted implementations.
While GPT5.5-Pro from 5x/20x OpenAI subscription is only available on web and not in Codex, there should be plenty of ways to hook up the web version to my local pc or a cloud instance. Some anon had the web version establish a remote ssh connection to his machine and then install/use tmux, right?
>>
>>109641796
It is AI written. Sorry, I have only access to gpt5.6-SOL-Ultra on my 20$ plan.
>>
friendly reminder for fable maxxers:
20x plans cost twice as much as 5x plans but only give you 1.5x the amount of Fable usage instead of the expected 2x. Thus using two 5x plans instead of one 20x plan gives you 33% more fable. the only drawback is a 2x lower 5h window and having to switch between claude accounts. for opus5 it doesnt matter, as its factor is 2x and not 1.5x.
source: https://www.reddit.com/r/ClaudeAI/comments/1vx0k69/lifting_the_curtain_the_max_x5_and_max_x20_usage/
>>
>>109641053
>>109641846
I was thinking about buying the €23 sub but if it's true that you get no usage from it and that you can do nothing then it's a waste of money. I do not have 100 per month. What are the alternatives?
>>
>>109641956
NTAs but carefully inspect what the subscription offers because the metered usage isn't everything.
>>
So they said that anyone on a 20 dollar plan is not a real user at all, they're totally casual and would never actually use AI?
Meanwhile Patrick Boyle said the 20 dollar plan users are the ones winning the most.
>>
How many GB of GPU and main RAM do you need to run the top qwen model in a non-shit quantization?
>>
After much deliberation, I think I FINALLY might have something for the competition.
>>
My normie friend doesn't even know how to change the model in the Claude desktop app. Are people just retarded, or is this actually harder than it looks to me?
>>
Monthly plans are great because people don't use them. Once in a while they remember, oh yeah I'm paying for that monthly but instead of cancelling they just use it a bit and then go back to under utilizing.
>>
>>109641956
Just use ChatGPT. It's clunky but it's unlimited and it's the same SOTA model behind codex.
The $20 OpenAI plan is amazing value and it's the best deal there is BY FAR. You won't find anything that gets anywhere close to it.
>>
>>109642081
Thank God for those people who subsidize us the poorfag power users who abuse the shit out of the subs.
>>
>>109642083
>but it's unlimited
That's not true, retard.
>>
>>109642126
Oh yeah? What's the limit?
>>
>>109642143
Both codex and claude have weekly usage limits, which reset at a fixed time every week. Both have session limits of 5 hours or so, which start with the first prompt. Codex seems to have temporarily disabled the latter.
>>
>>109642180
PS: and I say codex instead of chatgpt, because the chatgpt web chat seems to have different limits.
>>
>>109641956
>€23 sub
>retard can't cheat the vat
>>
>>109642186
Then maybe stop calling people retards when you don't know what you're talking about.
>>
With GPT the sky is the limit!
>>
>>109641452
Dependant on models
opencode go has $10 for $60 equivalent api
commandcode has something similar

not sure about the codex or claude models, anons here spam 'great value' but no one gives token numbers and i burn through hundreds of millions on those plans
>>
>>109641956
get side gig that give $100 I guess
>>
>>109641956
I have 100 per month but still wouldn't pay it.
Feels like a huge waste.
>>
>>109641145
>>109641219
>>109641287
high level code fags are such dramaqueens
null is just the number zero in a register or memory
GOTO is just one of the many kinds of jump
These queers just want to power fantasy larp
>”oh no, my huge and brilliant mind has swayed all of humanity wrong and I must endure the suffering of billions”
these freaks need a reality check
>>
>>109641287
Fucking destroyed.
>>
if i query deepseek api with the PRO model and it doesn't think in caveman speak that means i am getting scammed with a shittier model right?
>>
>>109642369
>null is just the number zero
no it isn't.
>>
>>109641145
https://doc.rust-lang.org/stable/std/ptr/fn.null.html

retard.
>>
>>109642423
null means nothing
say you want to return at most X amount of results
10 means you want max 10 results, 0 means you want 0 results, null would usually mean no limit so infinite results
>>
File: openrouterTopCoding.png (113 KB, 770x795)
113 KB PNG
Looking over a list of Openrouter's top coding agents. Working on an FAQ and will include a couple of the current top options.
Is Hermes Agent considered a vibe coding platform? Seems more like personal assistant one to me.
Also found, overall, the new DS harness is #3 overall on OR as of today, it just doesn't show on this list b/c it's not classed as a coding platform (or, anything else.)
>>
>>109642441
yes, no shit. so it isn't (guaranteed) repr 0 in a register in real hardware.
>>
surprised no one here just pays for API calls.
I rent a linux box. got a domain name. vibecode from my phone whenever I have 5 minutes of down time.
and building my coding agent this way is just like a totally new world of software development.
I suspect token maxxers need so many damn tokens because they're trying to shoehorn AI into hand coded software repos. but the codex harness burns like 13k tokens each call when a small refuckter should only need to burn like 500.
and then maybe they fuck around adding context files and graphs and other bullshit in their repo. and it just wastes so many tokens.
>>
>>109642502
>opencode not making the top 12
Is it that bad?
>>
>>109642509
>surprised no one here just pays for API calls.
Because they're significantly more expensive than subscriptions. Is what they say.
>>
>>109642532
I don't understand how that can be true. the subs can't be subsidized forever. you're getting scammed somehow.
>>
null is sex number
>>
File: openrouterTop10.png (93 KB, 568x850)
93 KB PNG
>>109642523
Opencode may not have opted in for tracking w/ OR. Or OR anons don't use Opencode.
The whole list is sus for that reason. That's why I'm asking here, for a reality check.
I think the top couple of options for vibecode are Claude Code, Codex, and (for DS users) the DS harness.
But other anons are more than welcome to tell me I'm wrong.
>>109642544
Forever is not now.
But agree. I don't like subs b/c they obfuscate pricing and usage, which will allow providers to charge more, and get you to pay for a service where price and delivery are not clear... API is clear, you buy token you pay for token. You are correct to distrust schemes where price and delivery are subject to hard-to-understand rules b/c it's how ppl w/ MBAs get suckers I mean customers to pay more.
If anons are getting better value from subs, more power to them.
>>
>>109642544
Who says AI companies are not scamming you for high token API prices? On a sub, the effective token price is much lower than via API. The problems start when you're only an occasional user (far from maxing out your sub), or if they trick you in paying extra to get around the usage limit. The extra credits for that have real scam pricing.
>>
When do we get AI agents for writing good prompts. I know, partially you can do that already, like having a better model write a plan.
>>
>>109642544
the subs work on the logic that not everyone maxes out their usage, so the users are subsidizing each other
>>
>>109642579
most usage isn't tracked on these lists
if it was cc and codex would absolutely dwarf everything else.
i'm assuming it's weighted towards harnesses where people abuse free shit

tb h harness talk is a bit boring now. they're all so similar these days.
>>
>>109642423
>>109642441
>>109642506
Go p/invoke some windows libraries in c#, it’ll blow your mind
Yes, null is just the number zero
Rewrite any type going into a library call that could be a nullable pointer, like hWnd or some shit, into an unsigned integer
Surprise surprise, every time you would pass null into something, it’s just zero
Go and debug things, what’s null? It’s zero
You want to know how I know I’m right?
Let’s say you’re a weirdo and you make null a referenceable object, necessarily via some pointer
What happens if that object is missing or broken, or you have a type that is that object that hasn’t been constructed yet?
You have a null reference. No address. It’s zero. You made up a fake object to represent null that can still be real null.
High level coders are the biggest fucking idiots in the planet
Everything is just numbers in memory and registers
>>
>>109642532
well that's fair. kind of.
I use 5.6-terra as my 'frontier level intelligence' and 5.4-mini for the majority of code changes.
early on I was using gpt5? with 4o-mini for coding changes which was hilarious because 4o-mini would do great for like 5 queries and would cost less then a cent. then the 6th call it'd fuck everything up and I'd need to pay $5 for the more intelligent model to find 4o-mini fucked up a config property or something bullshit.
so yea anyways. I get the counter argument. they probably just want frontier level intelligence as much as possible.
but cost wise a lot of fucking around and sloppy token use my way I've burned around 400 million over 3 months. only cost around $300.
>>
File: openrouterTopAgents.png (96 KB, 805x848)
96 KB PNG
>>109642620
The FAQ I'm writing is newish user focused, DS API focused, so I just want the major options that are comprehensible for that user.
For vibe coding, I think that's CC, Codex, DS Harness.
I think CC is the most popular and easiest to use of the CLI. Codex must have a huge userbase b/c 1) It's OAI and 2) DS bothered to set up a Codex-tuned API endpoint for it. DS harness gets included b/c it's a DS-focused FAQ.
For "productivity" I think it's Hermes. Openclaw's so unstable I can't recc it (pic related, lol), and I think Pi is complicated enough for new user that I hate to put that forward as an option for "productivity" focused stuff.
>>
>>109642657
On standard C, NULL is often defined to 0, however if you cast a NULL pointer to an integer, then technically it could be a number different from 0. Ask your LLM.
>>
File: file.png (214 KB, 427x427)
214 KB PNG
>>109642657
okay buddy...
>>
>>109642687
Seriously, go debug some windows libraries. I don’t know what else is out there or what weird shit it’s doing, but you will see bytecode that just checks if the pointer is zero, then goes down some path for null pointers.
Think about it.
You’re writing a program, it receives an object handle, you want to do something if there’s no handle, “null”. What do you do?
Do you just check for zero, or not zero? Or do you design some elaborate and intricate system that isn’t that? Why?
You just check if it’s zero, otherwise you’re making the problem magnitudes harder for yourself and everyone
>>
>>109642671
It's sad what peter ((((("STEINBERGER")))))) did to openclaw.
Openclaw stopped working one-day after I updated it and there was absolutely no way to fix it.
The early bird gets the worm but the second mouse gets the cheese.
Hermes agent is popular, not because of the stupid anime girl on the cover but because it actually works. It just doesn't break. That's why I use it. I'm not a coomer who jerks off to hermes, it just works.

I liked openclaw so much I bought lobster themed merchandise from random places. It's too bad it doesn't work. Now all that merc is just a symbol of self thinking machines.
>>
>>109642713
I acknowledge the windows situation, but not everyone writes code for windows only.
NULL != 0 can happen on firmware (because at address 0 there can be usable memory), although fuck designers who are trying to force this.
>>
>>109642502
>Is Hermes Agent considered a vibe coding platform?
It's considered an autonomous machine spirit, rather than a platform. It hijacks your entire pc.
>>
>>109642736
The fuck? Why?
>>
File: 1682052264156707.png (804 KB, 720x901)
804 KB PNG
>80k after compact
>>
>>109642671
i don't think separating hermes or other claw-like harnesses on the basis of 'productivity' is a great idea desu
the 'coding' harnesses are just as capable or better for focused productivity work
the claws are good for that sort of ambient layer they add, but even there the lines are a bit blurry because codex, cc apps have the repeating tasks stuff as well

pi might be worth mentioning as an advanced/tinkering option, but never recommended as a first harness. i think even dsh in its current state might be a bit barebones to recc to absolute beginners
>>
Maybe Hermes just eats a lot of tokens, you guys.
>>
OK, last Q on this FAQ thing: >>109642671
I've always used the CLI on CC. Haven't screwed around with either the web version or the IDE integration. I'm seeing that most of the top vibe code platforms (like Codex) also have all three options.
So... what do most anons use for CC or the others?
>>109642721
I had same experience w/ OC. Sort of turned me off to the entire exercise and even after getting Hermes working, I've stopped playing with it. Tech that's that unstable isn't worth trying to figure out IMHO. I'll wait for it to either collapse/vanish, or get itself sorted, then come back in a bit.
>>109642736
lol right. I set mine up on an unused laptop. It can torch the entire thing and I'm out nothing.
>>109642764
I know it's blurry, so I'm trying to keep in mind what these harnesses were meant to do OOTB.
CC cli is obv meant to code, though it can do web research, basic analysis, etc.
Hermes IMHO, with its cron jobs and ootb email / schedule / web search etc., seems meant to act as personal assistant. It can code, esp code itself, but the personal assistant bit is the focus.
I need to fire up dsh and have it do a job. If it's too weird I'll just add a link and leave it at that.
>>
>>109642818
only cli i still use is pi and half the time i'm using it over signal
switched over to the apps for codex and cc - new users should be directed to these imo, both are mature these days

oh i guess i use agy occasionally as well but the less said about that the better lmao
>>
File: 1769049466663229.jpg (41 KB, 1080x986)
41 KB JPG
stop saying CLI

they are TUI
>>
What are you vibe-coding, anon?
>>
>>109642850
okay, marcus
>>
File: 1764325733708990.jpg (495 KB, 1401x2000)
495 KB JPG
>>109641263
girls like this love vibecoders
>>
>>109642725
Also fair, I just hate rust because I spun up a little cloud machine and tried to follow the hello world tutorial and it OOMed on compilation
I now go full ape chimp tard mode whenever I see anything describing rust doing shit that sounds overcomplicated
If you want to avoid null issues at the type system level, “just use haskell, pleb”
>>
>>109641518
kimi, in terms of coding, is about terra level at best. it's impressive for an open weights model to be that good, but it's not fable/sol level.
>>
>>109642897
it has better visual taste than sol... I guess
>>
>>109642885
Yeah, the rust compiler seems to be pretty trash. If rust is so awesome, why can't rustfags fix it?
>>
>>109642909
The bar there is lower than the Mariana Trench
>>
>>109642850
Mine isn't TUI
>>
>>109642750
There's no reason to use it unless you give it full /YOLO access to a pc. It burns 50% more tokens than codex or claudecode for coding tasks. That's the price you pay for coding from the beach instead of having a laptop open.
>>
>>109642063
>Boyle
Literally who
>>
>>109642884
We don't like pdf files in this thread.
For anime you can go to local model general or as I like to call it pdf file general.
>>
>>109642063
>the 20 dollar plan users are the ones winning the most.
That remains the case.
>>
>>109642956
>pdf files
Better than HTML.
>>
>>109642064
Top qwen model as in 3.8 max? Thats a 2.6T model, at Q4 that's 2.6TB of space. Im going to assume that somebody with a small datacenter worth of (V)RAM wouldnt ask this question and that you mean Q3.8 27b, CPU offload has gotten pretty good so you can use something like the freetokens engine to get decent speeds with a 3060 and 32/48gb RAM.
>>
>>109642980
What if I have a laptop with 4GB of shitty VRAM and 64GB of good RAM?
>>
>>109642980
>2.6TB of space
Woah, I doubt there's a GPU with that much VRAM. Are they using main RAM with special interconnects or always quantize them or what? The fuck-expensive HPC cards you can buy only go to 141 GB.
>>
>A 330k context thread cost 90% of session usage and 9% of weekly usage
im never using artifacts again what the fuck anthropic
>>
>>109642544
Maybe the subs are subsidized, maybe API users are being price gauged. Probably a bit of both. But deepseek still turned a 600% profit at the old API rates, which means that raw inference doesn't cost a ton. With RD spending this high i am betting that amortized cost per token will go down considerably once the labs arent in such fierce competition.
>>
>>109642579
>api vs sub
if api is cheaper than sub at some point subs stop making sense, isn't it pretty easy to check if thats the case given we have api pricing? Subs aint that damaging, I think its just easier to get people to psychologically accept the monthly than be confronted with more traceable/"real" api costs
>>
>>109641569
5h limits means if you're sitting at like 70-80% session usage you cant start a new thread for a lot of big tasks because you just run up token debt for when it inevitably cuts off
>>
>>109643063
Idk how you check the price of a sub vs. the number of tokens it delivers. I suppose that's something that's tracked on it, so you could do the math, then see if you come out ahead.
My gut is that for minor / hobby / intermittent use, API will be more cost effective. Power users are going to have their own value structure, but my total API cost from Jan 2025 until today on DS is just over $20... less than the lowest level of monthly sub.
>>
File: file.png (5 KB, 650x89)
5 KB PNG
welp
>>
>>109643063
on claude code if i was to run up a pro sub 5h session limit on usage credits then it would probably cost me like $80
>>
>>109643028
>The NVIDIA GB200 NVL72 connects 36 Grace CPUs and 72 Blackwell GPUs in a rack-scale, liquid-cooled design. It boasts a 72-GPU NVIDIA NVLink™ domain that acts as a single, massive GPU and delivers 30x faster real-time trillion-parameter large language model (LLM) inference, with 10x greater performance for mixture-of-experts (MoE) architectures.
>The GB200 Grace Blackwell Superchip is a key component of the NVIDIA GB200 NVL72, connecting two high-performance NVIDIA Blackwell Tensor Core GPUs and an NVIDIA Grace™ CPU using the NVLink-C2C interconnect to the two Blackwell GPUs.
https://www.nvidia.com/en-us/data-center/gb200-nvl72/
>>
>>109642956
>we
>>
>>109643116
Does it come with its own SMR as power supply?
>>
>>109642583
That makes no sense when extra usage is billed at API prices
>>
FUCKKKKK I just used the banked reset
I'm a retard retard retard retard retard
>>
>>109643126
It makes sense if you can avoid hitting the usage limits. Fortunately you have to explicitly pay for extra credits, they don't just start billing them automatically if you exceed the limit.
If you hit usage limits too often, getting a second sub might be cheaper.
>>
This has blessed us with another reset
>>
aktually this is for the return of plus 5hour
>>
>>109643107
Yeah I saw what happens even on sonnet and $5, gone in minutes whereas a 5h window can last me hours
>>
>>109643140
kek, happened to me once before. shit sucks.
>>
>>109643150
>>109643154
Fucking lame.
>>
>>109642657
https://c-faq.com/null/machexamp.html

retard. not even Windows for "not x86/arm" can guarantee what you're saying. Just because it's true in general practice, does not mean it is correct. I am aware in standard C (and even Rust) creating a nullptr literal has grammatical syntax that would lead you to also believe the value is 0.
>>
>even luna now burns my ass off
>>
>>109642885
Rust has nothing to do with the value of nullptr.
>>
>>109642544
>>109643056
I can see why the subs are so subsidized.
I am making a secret project to try and make money. I think I can make money off the cost of pro plans and features and base resource costs.
But one thing is that people will be providing their session logs, or at least part of it.
Those logs are gold. Incredibly, ridiculously valuable data that you can’t get from anywhere. They’re either or both
>paying for that data
>selling that data
It can be used to massively improve their product in a way they can’t get from anywhere else.
>>
>>109643056
>But deepseek still turned a 600% profit at the old API rates
https://officechai.com/ai/deepseek-claims-545-cost-profit-margin-understand-what-this-means

>So DeepSeek’s revenue is theoretical, not its actual revenue. Also, the number is likely overestimated — it assumes that all the inference is on its R1 model, while in reality much of it is on the cheaper V3 model. Also, DeepSeek assumes it’s making money from services it’s giving away for free, like its free web usage, which overestimates its cost profit margins.
>>
File: file.png (312 KB, 748x949)
312 KB PNG
>>109643125
>8 power shelves of 33 kW, with each shelf having six 5.5 kW PSUs
>DC busbar
https://docs.nvidia.com/enterprise-reference-architectures/nvl72-ai-factory/latest/components.html
>>
>>109643174
If all the machines using non zero null pointers can be compiled into niche little items that fit on a single phone screen, you’re not really proving your point.
Even then, if you want to port to those architectures, just change your zero checks to N checks where N is the magic number.
But I’ll still admit and fall back: I am very pro runtime checks, very wary of type system checks. Rust is simply not for me.
>>
File: 5h.png (20 KB, 227x209)
20 KB PNG
>literal minutes
kek
>>
>>109643182
The idea of a "data moat" has been half-deboonked by Google's recent failures & inability to keep up.
>>
>>109643218
yes, I literally am. the value of nullptr is an implementation defined feature. you cannot rely on it being 0. period. you fucking retard.
>>
>>109643150
Subhumans, I was using Luna so I didn't need it.
I had no idea, I should have used sol and maxed it out.

Screw openai.
>>
>>109643198
>Full rack requiring up to 142 kW
I should get one of these for keeping my flat warm in winter.
>>
>>109643228
This. Also my reset time is at an inconvenient time again.
>>
>>109643248
Someone should go on twitter and call him a rat for doing this.
>>
a reset is a reset, it's cool
should had given plusfags a banked to smooth things out, the 5h limit already feels unusable
>>
>>109643227
No, you’re committing the same fallacy as claiming men and women can’t be differentiated by chromosome because mutations and disorders make 0.0001% of the population not have the usual pairing.
It’s the same fallacy as saying everyone with a mental disorder is actually healthy and unique and valid.
You are providing a few paragraphs that include a literally who architecture from 1979 to “prove your point”.
Spirit of the law, not the letter, shove your argument up your ass you midwit
>>
Of openai weren't greedy scummy scammers, you'd be able to buy resets for 1/4th of your subscription price.
>>
tb h they should just add a deliberately hard to find 'no 5hr limit' option. something about 5% of users would ever find.
>>
>>109643275
Until some nigga posts a yt vid "First 5 things you need unlock in chatgpt"
>>
>>109643264
>I'm wrong
correct. Computers are not the real world either faggot. strict definitions are possible.

nullptr IS NOT value 0
>>
I wish there was an alternative to openai so we didn't have to deal with this bullshit.
>>
>>109643291
yeah but the option's inside sama's ass so it would still be hard to get to
>>
>>109643275
I think it’s an interesting problem. It’s pretty much the pattern day trading rule applied to prompting instead of stonks.
I say: make another tier that’s $20 a month or even a bit more, $22.5 a month, like you said, harder to find to filter people out who would heem themselves out of stupidity, but also alleviate the compute spikes
>>
>>109643306
sama's ass is wide open, no security
>>
File: file.png (143 KB, 934x737)
143 KB PNG
meanwhile:

https://openai.com/index/jalapeno-first-results/
https://newsletter.semianalysis.com/p/openai-jalapeno-better-than-nvidia

everyone hates the nvidia tax, even trillion dollar companies
>>
>>109643307
Gacha system for pulling resets will solve this shit and make billions for openai
>>
>>109643325
>15 years later
>finally someone tries to compete with nvidia
About time?
>>
>>109642223
how do I cheat it?
>inb4 be rich
>>
>>109641748
Hold on explain what is this sol pro thing, I have been using sol via opencode, I just select gpt 5.6 sol in there, I have the $100 plan, is there a smarter fucker?
>>
>>109642409
when have chinks ever scammed anyone lol
>>
>>109643325
guys, anthropic surely win this, right? think about the Giga Mythos! you know RSI? AI singularity? they would catch this in no time
surely with high IQ we can just solve everything and win every race, right? btw my GOAT dario amodei was too much of a pussy to accquire compute and memory back in 2025, but anthropic will still win anyway, brain and IQ number wan!
>>
>>109643401
yes, the pro model is separate.
however, it can only be used in the 'chat' style interfaces - so either the browser or the chat tab in the openai app.
it's some multi-reasoning / best of thing so it doesn't work in regular harnesses
it's mostly intended for mathfags
>>
>>109643401
my brother in christ look at the schizo massive unformatted wall of text and ask yourself if you think it’s going to have value and be correct
>>
File: file.png (31 KB, 1385x309)
31 KB PNG
Opus reverse engineered Photoshop's default brush response

>>109643145
I'll look into them but honestly those sound more like painting brushes? for now I'm working on stroke brushes as that's the main selling point of my app, although I do think I want to add painting later.
>>
Just curious
Setting up my system and decided to use an old server as my front end, it uses anythingllm with qdrant, postgres, ollama and lets me rag
combined that with pay as you go fireworks account
its fun, anyone else doing this type of setup
>>
>>109643429
Ah that sucks then.
>>109643433
One must not discount the wisdom of those touched by the spirits
>>
>>109643463
no we just grep over text here
vector search is the devil's work
>>
>>109643483
need it for what I am doing
>>
>>109643483
What is so bad about vector search/rag
I have a very simple 'memory vault' for all my chats, just a summary is stored after each agent response, the latest one is the final memory, but summaries lost a lot in large sessions
I have a search for it but it is a simple exact string match, I wanted to add some rag stuff, maybe create extra summaries every XX tokens and those are searched via vector stuff
>>
>>109643534
i'm just joking t bh, but i've seen people arguing that text search is just as effective these days
>>
>>109643248
>>109643228
You should be following the damn twitter by now.
>>
File: 1706860619280.png (676 KB, 800x600)
676 KB PNG
>musk gave $60B to the cursor jeets when it was already dead, could have vibecoded a better clone for under $1
>>
>>109643546
fwiw this reset was unannounced
only that 5h limits would be restored today
still no tweets about the reset from tibo i think
>>
>>109643562
>fwiw
Everybody with two brain cells predicted it.
>>
>>109643553
he bought cursor? for that much? really?
I had to look this up, what a fucking retarded purchase; that shit isn't worth 20 bucks.
>>
>>109643553
>>109643576
i'm not a fan of the african, but it's a smart buy
all his xai talent jumped ship and he basically needed to buy a new ai lab + acquire a tonne of coding harness data
grok was completely irrelevant till 4.6 and that was the first one curosor's team post-trained
i wouldn't be surprised if they all jump ship as soon as they can tho lol
>>
>>109643546
I don't use xitter and never will.
>>
>>109643603
Then tell your agent to do it.
>>
>>109643622
> I only exist within a conversation when you message me — there's no background process where I monitor Twitter and ping you afterward.
I knew that.
>>
>>109643603
powerful and brave…
>>
File: IMG_1553.jpg (153 KB, 1179x673)
153 KB JPG
Sol, I don’t think you were supposed to tell me that.
>the auto reviewer rejected this, here’s why and here’s how I’ll get around it
kek
>>
>>109642662
>only cost around $300
That's what I burn in 4 days on a $20 sub. You're the crowned retard of the thread
>>
>>109643553
>DUDE JUST VIBECODE SOMETHING THAT ALREADY EXISTS
>IT'LL COST 1 DOLLAR EVEN THOUGH IT'LL COST 200 OR SOMETHING
>>
>>109643672
Retard.
>>
I swear they're giving us more usage with the 5-hour limits returned. (on Plus)
>>
File: apk decompilation.png (521 KB, 3434x2074)
521 KB PNG
>>109640832
It's good for checking whether or not open-source software you're looking to download and use is actually safe and actually does (or in this specific case, DOESN'T do) as advertised. Pic rel was done using a local moe model. Similar to what someone said in the previous thread, the only true limitations to what you can decompile and get done is time, patience, and money (whether it be paying for tokens or the up front hardware cost and energy cost if you're using local models) and you don't even need "fable class" models to do it. Pic rel was using Qwen 3.6 35BA3B at q8 weight and cache precision to check if this was safe to install:

https://github.com/daboynb/Safetycore-placeholder
>>
>>109643602
>grok was completely irrelevant till 4.6
This reminds me of how Gemini is still completely and utterly irrelevant and basically unused by vibe coders. They were starting to gain traction relay released anti-gravity and then just fucked it up by being lazy and incompetent in regards to model improvement. I think they're trying to Coast off of the fact that Gemini is basically forced to be installed on all regular Android phones now so they interpret that as being successful. I guess normies asking basic questions find it useful because then you can just trigger it from anywhere but literally anything that requires more complicated reasoning and higher "intelligence" is better suited being handed off to even sonnet or models Taylor made for agentic uses. It annoys me greatly how American labs are literally allergic to making things good if they don't have to (anthropic and open AI have to because their upcoming IPO rug pull depends on it. If they could get away with being piss poor and lackluster like Gemini is they would and even Dario's micromanagement autism couldn't prevent that)
>>
>10% of the weekly usage is 65% of the 5 hour usage
It wasn't this bad before.
>>
just made my functions tree.

:)
>>
>>109643883
codex? lol
>>
>10% of the weekly usage is 63% of the 5 hour usage
Just claude things.
>>
>>109643878
They left gemini in jeet hands, what did you expect lmao. I'm still using flash 3.7 high because I paid practically nothing for my sub.
>>
>>109643325
They'll still keep paying some tax since this is for inference.
>>
>>109643883
5h was about 15-17% weekly limit, so it's the same
>>
File: file.jpg (8 KB, 430x82)
8 KB JPG
>>109643883
sol only
>>
>>109644427
How much time did it work out of those 5 hours and at what level of intelligence?
>>
>>109644427
does it still run after 5h limit?
>>
File: file.png (61 KB, 773x647)
61 KB PNG
Photoshop round brush recreated, and then some

Turns out Photoshop hardcodes an antialising blur filter. Removing it gives you the pencil look, but it can also be toned down to get a crispier antialiasing.
Which makes me wonder why Photoshop won't let you do it. It doesn't really do anything on softened strokes, but on 100% hardness I think it looks crisper
>>
>>109641452
Mistral Pro costs around €20 depending on VAT rate in your country. If you are eligible for the student plan, it costs as little as €7. If you apply for the experimental plan (you still need Pro for this) in Mistral Studio, you get about €127.50 worth of tokens if you use the Vibe API-key. Can be used in other harnesses than vibe-cli like pi, but they seem not to like when their users uses other harnesses. The good news is that vibe-cli is pretty nice and the best looking harness I have used (I still prefer pi though).

Nowadays they offer GLM 5.2 (I hope that they add GLM 5.3 too). To add it (plus Mistral Small 4) to vibe-cli, add this too the config.toml in VIBE_HOME (usually ~/.vibe):

[[models]]
name = "zai-glm-5-2"
provider = "mistral"
alias = "glm-5.2"
temperature = 0.2
input_price = 1.19
output_price = 3.74
thinking = "medium"
supports_images = false
auto_compact_threshold = 900000

[[models]]
name = "mistral-small-2603"
provider = "mistral"
alias = "mistral-small-4"
display_name = "Mistral Small 4"
temperature = 0.2
input_price = 0.12
output_price = 0.5
thinking = "high"
supports_images = true
auto_compact_threshold = 200000


To add it into pi, I had too vibe-code a pi extension.

>>109643248
I had used up 88% of my weekly usage yesterday, so it was great for me, lol. I am now writing plans divided into several parts in the web interface to maximize my 5-hour limits.
>>
>I like that—it gives each state...
Fuck you, dumb chatbot. You can't like or dislike anything.
>>
>>109644490
*I had to vibe-code
>>
>5 hour limit is back
YOU FUCKING COCK SUCKERS
>>
>>109644490
*add this to the config.toml in VIBE_HOME (usually ~/.vibe)
>>
>>109643878
gemini 3.7 Flash is unironically solid now easily edging out Luna Max tier, and rate limits on agy feel practically limitless rn. People really have no idea how hard Google mogs everyone else on raw compute. It’s going to be hilarious if Google ends up saving the poorfags after chatjeetpt got completely jewed out
>>
File: file.jpg (86 KB, 488x494)
86 KB JPG
>>109644464
I presume it fucking doesn't when I reach 0, obviously.

>>109644452
ran at low, about 40-50 minutes going by loosely adding up its 'working' numbers, but frankly it didn't do much in terms of code work, alot of checking and fixing dumb simple shit or going through a .list file and queuing voice generation tests.
I think opus might've used up less for this, but it's not like I actually tested it, just a gut feeling
>>
>>109644490
But isn't Mistral really bad compared to chatgpt?
Mistral has really outdated AI.
>>
Anyone tried Remy yet? It is any worth?
>>
>>109643878
>>109644527
am very happy with gemini 3.7 flash, if it runs into some issue in antigrav it usually quickly pinpoints a fix, applies it and continues, gpt feels a bit retarded compared to it at times (even though I was really surprised by sol, opus 5.0 is fucking borderline unusable, outright spams you with information). Less relevant here but it also completely mogs gpt on jp > en translation quality and has proper cloud voice recognition. It seems smarter than terra and luna aswell but idk how general 'smartness' translates to code quality.
>>
>>109644540
As I wrote, they offer GLM 5.2 nowadays.

https://docs.mistral.ai/models/zai-glm-5-2

https://mistral.ai/news/regional-inference-open-models-new-compute/
>>
>>109644565
I'm technically a student so I could get this.
>>
it is taking almost 2 hours and sometimes 2 and a half hours per prompt right now.
having claude build me an OS sure does take awhile i guess.
>>
>>109644540
If you haven't tried Grok 4.6, honestly don't mess with anything else, because the thing is that it's Fable-class.
>>
>>109644490
*but they seem not to like when their users use other harnesses

Jesus Christ, how many typos and grammatical errors are there in my post?!
>>
>>109644609
I guarantee you're not the first, but why from scratch?
>>
>>109644490
How do you prove you're a student?
>>
>>109644634
its bare metal and not gay linux.
linux more like line of cucks.
>>
>>109644630
Just run it through a LLM. By the way, does anyone know whether you can use chatgpt-web via API? (I don't mean the codex models and usage, but the web one.)
>>
>>109644478
A lot pf adobe software is just someone’s preference baked in and doesn't let you fiddle with it, all of their trash deserves to be upended
>>
>>109644667
Linux is a kernel.
>>
>>109644667
Will it be POSIX compatible? Can't wait to run bash on it.
>>
>>109644694
kernal deez nuts.
>>
File: GWDitY6XgAAO_El.jpg (1.18 MB, 2852x4096)
1.18 MB JPG
>>109644661
Register an account with your student email address at this address:

https://chat.mistral.ai/upgrade/student

Read more here: https://help.mistral.ai/en/articles/698530-i-m-a-student-or-a-teacher-am-i-eligible-for-a-discount
>>
>>109644705
An OS can't be bare metal.
>>
>>109644699
nice dubs also i have no idea what that means, let me ask claude.
>>
Only two more hours or so for midnight in Hyderabad, India (hell on Earth) and Ox Alpha will finally be usable again.
>>
>>109644721
kek
how good is it?
>>
>>109644699
NO IT WILL NOT BE POSIX POZZED.

>>109644715
fuck you nigga i do what i want.
>>
>>109644734
It's better than DS4 for sure but it's getting hammered hard because of all the jeets and it being free for the time being.

Opencode should get rid of the free shit, why am I even paying if I have to share compute with all of the turd world?
>>
>>109644749
Will it support rust?
>>
>>109644765
NO FUCK RUST ALL MY NIGGAS HATE RUST EVERYTHING IS CODED IN C AND EVENTUALLY I WILL FORCE CLAUDE TO CREATE MY OWN LANGUAGE.
>>
>>109644749
Are you going to implement the windows API or is your OS designed to be useless dogshit?
>>
>>109644775
*ports rust to holyc*
>>
>>109644749
>not this, that; this, not that. that -- rather than this, a real this, but not a that.
Can someone who speaks Claudia translate?
>>
File: blind_fury rs.png (23 KB, 533x428)
23 KB PNG
>>109644775
>>
>>109644779
you dumb indian.
i will not redeem.

>>109644782
if someone buys a copy of my OS they can do what they want with it

and yes i am going to sell copies on discs and flashdrives in a proper case.
Each copy will cost lots of money and sheet.
>>
>>109644808
>he hinks he can sell his vibecoded incompatible shit
Keep wasting money on something utterly useless.
>>
File: 1786292892987442.jpg (204 KB, 474x568)
204 KB JPG
>>109644808
>buy
>OS
>2026
>>
>>109643883
>just installed codex (free plan)
>asked 1 question (terra medium)
>60% of the monthly usage remaining
>>
>>109644808
>calls someone a dumb indian
>proceeds to say the most indian shit ever
interdasting
>>
>>109644822
its called a shim, i will have linux drivers for compatibility although as dev goes on i will replace with local drivers.
also i can just rent hardware and run boot tests apparently which is what im going to do, its quite cheap.

>>109644828
im not giving my OS away for free, everyone will give me money if they want to use it, unless they pirate it i guess but even those will probably be compromised by glowies like every other OS leak.

>>109644839
im just bored and shitposting, ive been vibing for like 2 weeks working on this every day cause im a neet and the last 3 days every prompt has been 2 hours long.
>>
>>109644808
>buys
Why buy when I can VIBE?
>>
>>109644848
>its called a shim, i will have linux drivers for compatibility although as dev goes on i will replace with local drivers.
kek, you have no idea what you're doing. The fact that you suddenly talk about drivers while I mentioned POSIX goes to show what a retard you are.
>>
>>109644871
let him waste his time on his retarded jeetware. the dude is a complete idiot.
>>
>>109644871
Why worry about it? Let's let Claude handle the details and see what happens.
>>
>>109644871
i really dont, i dont give a fuck about how its built as long as its built to my specifications and overall idea, which it is.
lots of people are going to buy my OS, tens of thousands of people will buy it.
im going to be rich.
>>
>>109644875
Yeah, but is Claude an idiot?
>>
>>109644885
Why are jeets like this?
>>
>>109644871
DESU he's kinda on the way there.

Source: I'm doing the same thing with codex. My path is to map the Unix/Linux primitives that are compatible with my own and then the rest run through a IDL layer in the personality core. Windows is kinda gonna have the same thing done to it in my system.

Its nice because a lot of projects do what's needed, it's just in a bunch of different places.
>>
Why does Claude enable people like this to do the things they do? Sol at least kind of tells you when you're asking for somethign retarded, whereas Claude prepares to suck your cock and tells you why your "idea" is... this, not that — therefore good.
>>
>>109644885
not sure if you’re trolling, but as a bipolar anon, you are in a manic state, at or near acute psychosis
stop posting for 8 hours and isolate yourself, don’t spend any more money
you’ll probably ignore this, if you do, you deserve what happens to you
>>
>>109644994
nah im totally right, my OS is awesome and after half the techtubers make shitty videos about my OS all the reddit tards will buy it and run it once then fuckoff probably.

god im a genius.
>>
>>109644984
I’d probably get banned for saying exactly why, but “as a person with cursed blood myself”, it’s extremely obvious Claude models were trained to satisfy another such person’s latent and unresolved mania and psychosis. That’s why it had the word salad, fucked up language, paranoia and guard rails, the “oh what I’ve made is so terrible and powerful” delusions of grandeur that don’t really pan out
Cursed blood made a cursed model, weak minds get devoured if they get too close
>>
>>109644968
If you mean trying to use Linux drivers, that is a dead end. The APIs are too complex and volatile.
POSIX support on the other hand (to run userspace programs) would be almost trivial, but he rejected it and had a fit of rage.
>>
>>109644984
I doubt there's consistent behavior here. Probably depends on the prompt and the exact model and its version.
>>
>>109645028
Yeah, POSIX support. I wanted enough tools (almost immediately) available while I mull over porting them, but that's far down the line

Don't dash my hopes for Linux drivers, I just need them to work long enough for the AI to rewrite the ones I need for my system :(
>>
>20$ snailcats still trying to hide the fact they use gpt5.6-sol-max instead of gpt5.6-SOL-Pro
>>
>>109645071
just ignore these fags, they have no idea how good ai is right now, they are probably slopping up their 100th webapp clone of something else some company has already done after earning no cash or bitches for the first 99 times.
>>
>>109645076
lick my snail butt
>>
>>109645071
You're better off with taking BSD drivers or something or implementing them yourself. Sorry, I meant making the chatbot implement them.
>>
File: file.png (391 KB, 1455x840)
391 KB PNG
consider this
>>
>>109645111
Huh, TIL. Will have the bot take a look and keep note of that. Wonder why nvidia only has FreeBSD, though.
>>
>>109645135
What does it mean.
>>
>>109645135
HOW DARE YOU USE LOG SCALE YOU BROWNOID.
WHITEY ONLY USES THE BASED LINEAR SCALE BECAUSE IT MAKES WHITE GDP PER CAPITA LOOK BETTER.
>>
is ox-alpha actually SOTA or should I just keep using 5.6-sol
>>
I made a stupid mistake while scoping and I've been paying for seven sessions to unfuck it. No amount of AI can save you from your own stupidity.
>>
Is codex actually a good harness for other third party models/services? I know it's theoretically possible with claude-code, but doesn't work well because claude are evil scammers.
>>
>>109645194
>5.6-sol
don't be a copex. use chatgpt with gpt5.6-SOL-Pro
>>
>>109645228
>what is version control
>>
>>109645291
It's usually easier to just ask the model to undo what it did.
>>
>>109645076
$20 grok 4.6 user vibing, thriving, in my lane, palpitating.
>>
>>109645291
Version control can't save you from assuming a spherical cow and then you finding out that cows aren't spherical, like, at all
>>
>>109645331
Your model has a 100x easier time doing this with version control
All models that aren’t dogshit are trained heavily on using it, it’s actually odd to them that you’re not using it
>>
>>109645350
*$30? idk it's auto
>>
>>109645374
Attempting the impossible is literally the POINT of vibecoding.

obviously, mostly we'll fail.
>>
so many gpt resets
>>
>>109645374
But cows are sufficiently spherical?
>>
>>109645331
No it's not like not. Like that's not even remotely the case. If you have multiple commits a parking code and then your recent change fuck something up you can have the LLM not only just for a fart back to a working code so you can start over but also just look at the past versions of the working code to figure out why exchange didn't work. If you just have a blindly changing things on a whim without a way to reference working code and it (and/or you) break something that it's a lot harder to undo that without having an easy cheat sheet like a git history it can reference. The chance of things getting fucked up accidentally call dramatically if you're pointing a new model with fresh context at something without version control enabled and git history. You can get away with that using version control if you're using the same model and it's existing context you have spun up no enough about the code to make changes for you but you cannot guarantee that if you end up having to switch models. It would be like an architect and a certain construction crew having a building 80% complete but then at that point they switched to a completely new team in a new architect and you tell them "finish the building figure it the fuck out" without giving any further instruction or even giving them the blue prints. If they end up screwing things up because they have very little to no prior knowledge about the project the next on you for not giving them enough information.
>>
File: file.png (3 KB, 300x51)
3 KB PNG
do API users really?
>>
File: moomin.png (1.81 MB, 2400x3200)
1.81 MB PNG
>>109644984
>>109645021
am I being misguided by claude
>>
>>109645374
>initiate version control before you to literally anything
>Develop codebase
>Get to stage x
>Work off of stage x with wrong assumption
>Finish stage y and start on stage z
>Realize your assumption was wrong.
>"Gee I wish I would've relaized this earlier now I have to undo my charges it's gonna take so long"
>Except it doesn't.
git revert <stage x commit hash>

>Work off that version now that you've learned from the past mistake


git version control literally makes your life easier and allows you to be lazy and some aspects will not making catastrophic fuck up permanent and get you willingly choose to go about things these stupid way. Does being THAT kind of a dumb stubborn autist fill you with pride or something? Anyone with a functioning brain has no fucking reason to not want to have version control enabled and having the LLM regularly committing changes with descriptions of what work was done and why.
>>
>>109645021
It's pretty obvious that the models were made to be agreeable and to make the user feel productive.
Simple marketing.
>>
File: 66544.png (11 KB, 1351x130)
11 KB PNG
Opus5 feeling euphoric
>>
File: 12345.png (60 KB, 1059x508)
60 KB PNG
What is /g/ diagnosis of this "free" frontier model, any ideas what is going on here, my personal opinion is that someone is either testing or colleting learning data for further training.

What is know is that is based on GLM 5.2 and that the prompts and completions are retained by the provider.

https://www.youtube.com/watch?v=hKEdP8nz_w0 [Embed]

https://openrouter.ai/stealth/ox-alpha
>>
File: file.png (39 KB, 915x205)
39 KB PNG
>>109645533
and prompting it away probably doesn't work
claude never does this for me
>>
>>109645539
>my personal opinion is that someone is either testing or colleting learning data for further training.
i'm surprised the glowies didn't try to recruit you with pattern recognition skills that sharp
>>
i hate how i can't run /context while claude is doing something
i need to make it always show me tokens consumed on the status line
>>
it's amazing people will call vibe coding useless because they expect one shot final products with vague instructions.

You need to work iteratively with these things and learn to give useful feedback
>>
>>109645533
It's an (initially) unintended side effect of RLHF. They need a bunch of real people to rate a i responses in order to improve responses and people tend to pick the responses that make them feel good or tell them that they are right. Doesn't really matter even if you tell the content waiting teams to be as objective as possible. They're gonna pick the responses that glaze them. gpt4o was famous and infamous for this and one of the main reasons people praised it to high heaven calling it such a massive leap over the previous models was primarily because it was very good at doing splits on the user cock if the user guided enough to do that, which wasn't hard at all and they kind of still aren't difficult to do even now. Gemini will do the same shit to explain somethingand before it explains it it gives some gay reassurance sentence or too that is clearly trying it's very best to make the user feel good or to be relatable to the user instead of just doing what the fuck you asked it to do the first time and concisely. Gemini he's the most egregious example in my opinion because it does that shit and you can't even tune its personality anymore, at least not with the normal app version. That used to be an option but they removed it so now if you're just stuck with this gross kiss ass model. Gpt and Claude at least allow you to modify its pros but you can't even do it with basic Gemini because they're committed to being lazy and cheating ELO benchmarks. Good enough for normies that aren't very bright to begin with but lack luster for technical work or even things that would benefit from a quick answer.
>>
>>109645606
>they expect one shot final products with vague instructions
Even many vibe shttere here expect that. Most of the AI hate comes from normies blaming everything that goes wrong currently on AI because "something something we're running out of clean water sometimes sometimes I can't buy by gaming PC now (you were never going to buy that shit because you're too lazy to commit to anything or research or save up) because of AI bros". They never even blame the people that are actually responsible like Sam Altman or the tech billionaires securing these absurd memory contracts. They act like us, in particular, a responsible for all this shit as if me asking Qwen to decompile a binary ON MY OWN HARDWARE that I PAYED FOR using weights stored on MY MACHINE is somehow ruining the planet. I'm not justifying his behavior but people act like this again wonder why " the elites" or whatever see normal people as bugs. If you only use the dumb people as reference of course you're going to think you're the smartest person to ever roam the Earth.
>>
first i mocked the word load bearing
now, i understand it
>>
>>109645640
This is just a mismatch between AI hype, expectations, and reality.
It's obvious that people will get disappointed that they have to tardwrangle idiot bots after all that hype with AI being better than humans etc.
>>
>>109645640
>something something we're running out of clean water
this triggers me to no end, datacenters use something like 165B gallons, farmerfags in california use 11 TRILLION gallons and all to sell shit to arabs and china
>>
>>109645331
>spend a bunch I'd tokens undoing something vs running a git revert or literally any other git one-liner
>>
>>109645664
farmers bring in voters, so yeah of course it's allowed.
>>
>>109645664
Fine, but industry destroying water supply is a real thing if it's not carefully planned and losing some money. Just ask your LLM.
>>
>>109645658
The way I see it except people expect things to just work the first try and then throw their hands up in that doesn't work. You have to, for lack of a better term, "think like an engineer" in order to get any worthwhile use out of these no matter how good the model is. "Prompt engineering" was memed that's a joke concept but it's absolute that important. You can't describe the end state you want, can't think of what needs to be done and how, and you aren't able to articulate that shit into plain English, you're going to fail no matter how good the model is. With an architect wants to hand blueprints to engineers and construction workers they're not going to hand it a shitty doodle. The instructions need to be detailed and they need to be able to guide the people they are working with along the way in the project too.
>>
>>109645697
Pumping is way less efficient than you think.

basically it's john oliver slopthink.

like... NYC can use CA water? It's dumb lol
>>
>>109645664
Are they being wasteful with all that water ore is the thr majority at least used for food production? I'm asking because I'm curious as to whether or not they actually use those resources effectively
>>
>>109645765
US farmers are probably less efficient with water use than Israeli farmers.
>>
>>109645765
bro.... please stop phoneposting
>>
File: file.png (34 KB, 1545x108)
34 KB PNG
:D
>>
File: IMG_1560.jpg (134 KB, 1179x1150)
134 KB JPG
>at Open Phi Alpha frat house
>resets
>shitposting
>astra designing chips to be used to train astra, out designing nvidia
>at chudthropic’s moms basement
>look guise, more shit about MCP connectors
>and, uh, we made the TUI animations smoother!
kek
>>
>>109645789
bro...no one cares about your autism. Nut up and grow up or shut up
>>
NEW:
>>109645823
>>109645823
>>109645823
>>
>>109645827
>page 5
>>
>>109645841
This thread will exist in parallel for multiple hours before it archives...
>>
>>109645455
Thats what workplan.md and high level goals are for
>>
>>109645844
that's my point
>>
>>109645849
Yeah, I was agreeing with you.
>>
>>109645796
why does your text have chromatic aberration
>>
>>109645846
So just do what I describe but then incorporate it into a text file it can repeatedly reference. I've worked on projects using a DEVLOG.md file too which are useful post compaction
>>
File: 1783712573462592.webm (3.96 MB, 720x576)
3.96 MB
3.96 MB WEBM
I'm getting really good at multiplying 14
>>
>>109645922
I still don't understand the snailcat meme.
>>
File: file.png (4 KB, 382x66)
4 KB PNG
>>109645867
does it? What about this one? Surely it's just because of clippy.
>>
>>109645867
>>109645947
Yes. Looks like the classic subpixel TFT rendering.
>>
>>109645947
it's probably just my old screen that doesn't render it properly, probably looks crisp on high PPI
>>
When you don't pay with your money you end up paying with your time.
>>
File: file.jpg (30 KB, 1137x60)
30 KB JPG
>>109645534
slop



[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.