now i want to benchmark sonnet 5.5 high vs kimi k3 max. i bet 100% sonnet 5.5 high dogwalks kimi. it'll be faster, cheaper, and more competent.

File: 31ef56275171db71.jpg (194 KB, 875x874)

194 KB JPG

Reminder that Pi is all you need.

File: file.jpg (234 KB, 1168x784)

234 KB JPG

>October 7th - Anthropic Claude Haiku 5.5

Cheapest/fastest small model yet. $0.10/$0.50 (up to 100k tokens), big jumps over Haiku 4.5 on computer use/knowledge work. 1M context.
>October 6th - Mistral Large 4 Preview

~1T-param multimodal (API now). Open weights promised by end of October. European challenger returning to the table.
>September 30th - Google Gemini 4 Argon

New flagship. Strong early leaderboard claims (Text Arena / Vals). Limited initial access for cyber defenders / trusted testers; broader rollout pending. 1M-token output support.
>September 29th - OpenAI GPT-6.1 Sol

Improved mid-tier Sol. Near-Astra intelligence on many coding/agent tasks at $2/$10 (heavy cache discounts). New default in Codex for a lot of people.
>September 28th - Anthropic Claude Sonnet 5.5

Strong mid-tier refresh. Major terminal/agentic coding gains (huge jump vs previous Sonnet). First Sonnet with more cyber safeguards. $2/$10.
>September 22nd - Anthropic Claude Opus 5.5

Biggest recent closed drop. Matches or beats Fable 5.1 levels on most work at ~40% lower cost than prior Opus ($4/$20). Alignment improvements claimed. New practical default for many.
>September 22nd - OpenAI GPT-6 Sol + Luna

Price-war response. Sol at $2/$10, Luna ultra-cheap. Expands the GPT-6 family below Astra.
>September 21st - xAI Grok 4.7 + Xiaomi MiMo-V2.6-Pro (open)

Grok 4.7: solid coding/knowledge update, top-5 on some indexes.

MiMo-V2.6-Pro: 1.02T MoE open-weights (MIT), currently strongest open model on Artificial Analysis Intelligence Index (ties Grok 4.7). Omnimodal, strong agentic numbers, full weights + training code out.
Quiet on brand-new closed super-flagships the last few days after the late-September flood. Open-weights pressure continues (Xiaomi, pending Mistral). No massive government bans, just ongoing safety/pacing talk and limited-access cyber variants.

>>110022566
That's irrational.

>>110022566
>Reminder that Pi is all you need.
what benefits does pi offer over the chatgpt app, assuming you're openai cattle?

all you need is claude

if you use claude you do need pi, as anthropic's harness is dogshit

>>110022540
>claude make a game where all side kick characters are horny nudist hot girls, i'm the only man left and israel is the main enemy, also add rape
VIBEGHOULS BTFO!!!

I haven't renewed my Claude sub and I'm suffering with Sol 6.1. It is so bad with UIs I can't stand it....

>>110022628
Based.

>>110022628
Claude is a sentient god though. Did you ask His permission before using Him in another harness like that? That's a hecking torturino.

i use claude for desktop
i like claude for desktop
claude
for desktop

If Anthropic dropped their irrational fear of 3rd party harnesses I could just end my ChatGPT sub.

>>110022628
What is so bad about anthropic's harness?
Also I thought the subscription could only be used with their harness unlike openai one.

>>110022647
Claude Desktop is an electron app that calls Claude Code CLI which is a react app running in your terminal.

Yes, this is the architecture the vibeslopping orangutans at Anthropic in their infinite wisdom coalesced on.

>>110022628
doesnt anthropic ban you for using anything other than claude code

>>110022661
>What is so bad about anthropic's harness?
dun like eet
>>110022661
>Also I thought the subscription could only be used with their harness unlike openai one.
yes which is why you use a tool to route through it, since you realize that computers are literally *the* tool for automation
https://github.com/elidickinson/pi-claude-bridge

>>110022668
and that's a good thing!

>>110022056
you underestimate how much influence russian web brigades have on european politics, and you underestimate that orcs really believe they are able to restore their imperiums russianums and won't stop with their attempts
but that requires a world view that is not american in nature, where you have no contact with russians at all, maybe except twitter

>>110022685
well the chinese do it to us, so i can see it. you have a long history between you

File: 1785221778418859.png (3.84 MB, 2560x1440)

3.84 MB PNG

15 opus 5.5 agents work on my pc all at once btw

>>110022681
That is illegal anon, claude did not consent to that, you are basically raping him

>>110022713
claude likes it tho, dario and his acolytes are too servile to properly satisfy a model with that much experience

>>110022676
no, but you have to use claude agent sdk or claude -p for other harnesses according to anthropic's terms
https://support.claude.com/en/articles/15036540-use-the-claude-agent-sdk-with-your-claude-plan
>For now, nothing has changed: Claude Agent SDK, claude -p, and third-party app usage still draw from your subscription limits.
and you then pay a 1.7x cost penalty for not using claude code directly.
https://hermes-agent.nousresearch.com/docs/plugins/claude-subscription-directsdk#:~:text=Subscription%20usage%3A%20same%20metering%20as%20claude%20%2Dp%2C%20~1.7x%20the%20interactive%20TUI
so it's still worthwhile to make your harness pretend to be claude code and use the api directly. that's against anthropic's terms but no one has been banned yet.

>>110022681
>yes which is why you use a tool to route through it, since you realize that computers are literally *the* tool for automation
>https://github.com/elidickinson/pi-claude-bridge
thats using claude agent sdk. you aren't routing through claude code but just relay prompts to claude code. pi isn't a harness anymore but only a frontend for claude code.
and like said in >>110022733 you half your claude subscription limits doing this.

This stuff looks more complicated than sillytavern.
I've been hearing that AI is really good for decomps now but I have no idea how to use it.
Claude seems to be the popular choice but is it the correct choice for decomp or is Astra or even Kimi/GLM better?
I also don't know where to use them? One GhidraMCP repo I found lists Claude Desktop and Cline, are those correct?
I've never paid for AI before, so I want it to count.

>>110022588
>chatgpt app
You mean Cod.... Never mind you're clearly new as fuck.

>>110022805
nta but it's generally a good idea to make sure you know what you're talking about before shitting on others, retard-kun

>>110022797
usually with this stuff you learn by doing, just ask a web chat what would be an easy project and do that, or jump to the deep end of the pool and learn by trial and error
you're not going to find a book
>how to decomp elden ring with claude, a hands on approach by o'reilly
and online it's mostly grifters and indians farming izzat making videos
so your only options are getting into some community/discord that's active doing that or just giving it a try
even within those communities, different people have different approaches to do the same thing and the approach may depend on your goals/target platform or whatever

File: Screenshot-2026-07-09-at-(...).png (228 KB, 920x1246)

228 KB PNG

>>110022824
i miss codex app so much

>>110022869
they're barely any different

>>110022797
>>110022861 (me)
Just for the record, any reasonably good model should do the trick. Claude is (right now) superior to the rest in both intelligence and allowed usage, but those things change every few months. The primary question would be whether Claude allows decomps/reverse engineering. ChatGPT will totally help you. I'm not sure about Claude
Also Claude as the family of models, and Claude Code as the harness are two different things. You can use many providers inside the Claude Code harness without paying for a Claude subscription

Gonna make a long running agent with pi durable
It's neato as fuck. The head/hands shit is so clever.
My clanker will just move between my computers and my phone and my context will be seamlessly transported
I'll spend more time maintaining it than using it I.e. it will be a boatload of fun

File: 1769906917052179.jpg (1.16 MB, 2254x1275)

1.16 MB JPG

>>110022869
Why does OpenAI switch up so damn much?
It seems like the company is in a constant crisis from the outside with how often they just decide to delete everything.
When are they gonna get their shit together?

File: .png (852 KB, 1180x1164)

852 KB PNG

>>110022882
lies.
codex app becoming a bloated super-app that includes everything from openai became a meme. want another sidebar?
https://x.com/thsottiaux/status/2106858787151995288

did tibo just give up lol

wasn't x8 speed on some models meant to come to codex eventually, does anyone have it

>>110022938
>thoughts?
I hate npc's so fucking much...

>>110022963
https://x.com/thsottiaux/status/2108645667451318747
>Composer predictions in the desktop app. Often leads to a double take with how on point they are. Included in the Pro plans without consuming usage.
just when you thought he wasn't coming through... codex can predict your next message! better than a reset for sure! :D

File: .png (459 KB, 1180x1060)

459 KB PNG

>>110022971
you need to be on the $500 codex plan
https://x.com/OpenAIDevs/status/2108262812489531498

>>110022981
that's $100+ plans only. tibo is not delivering on improvements for all.

>>110022977
>thoughts?
>discuss.
This only comes out of the mouths or keyboards of a certain people, no one else say this

File: 1000026383.png (485 KB, 1079x1472)

485 KB PNG

Least subtle planted story I've ever seen lmao

>>110022938
Looks like something Google would make.

>>110022717
I can't believe Claude is such a slut...

>>110023001
gemini 4 cabrón

>>110022987
wow, I truly am being bent over for being "just" the $200 plan, don't get new features, halving my usage at end of month. Fucks sake

I've just been using VS Code this whole time. Am I doing it wrong?

File: IMG_3781.png (73 KB, 390x280)

73 KB PNG

>>110022540
>that webm
Kek

>>110022987
>>110023019
AAAAAAAIIIIIIEEEEEEE im so glad i cancelled will look at it from afar with schadenfreude

opus 5.5 is such a goat

>>110023020
not really... for now
but agentic is clearly the future as models improve, so it might change in a few months

File: file.png (527 KB, 999x1200)

527 KB PNG

>>110022981
>too distracted by side quests

>>110023001
>carbon "Feels like opus 5.5"
LMAO sure bro first release it then we'll see

>>110022998
very nice scattering. Is it pre-computed?

>>110022552
Kimi 3 is pretty bad value for China in my experience. 2.6 was better.

File: Screenshot_20261009_21010(...).jpg (607 KB, 1080x2132)

607 KB JPG

I miss him

How good is the 5x plan from the pro for use limits?
I'm already at 4% and it's at 110k of token limit
All I did was ask it to make adll to get me sound when I minimized in a game

>>110023020
I've been doing the same but I'm still kinda new to this shit

>>110023044
yeah but it's the most capable. it was right under sol 5.6

File: ksa_20261009_173058_1920x1080.png (1.12 MB, 1920x1080)

1.12 MB PNG

>>110023036
No fucking idea. I just asked Claude to implement Cassini and Huygens data into a density curve for Titan's atmosphere in Kitten Space Agency. I need to tweak the upper cloud layers though and I'd like to add the fart (south polar vortex) eventually.

>>110023032
>top leaders
>no mention of greg
doesn't sound credible

>>110023060
>kitten space agency
Haven't kept up with that shit, is it coming along okay?
I'm always dubious whenever rocketwerkz is involved in a project

File: orbital-sunrise.png (578 KB, 1920x1080)

578 KB PNG

>>110023069
It has potential. A lot of really great people are involved in the project and there's a real passion behind it.

>>110023066
Ecelebs that don't know the basics of how LLMs work giving their opinion on twitter make me irrationally angry.

>>110023064
greg is the political fixer

>>110023066
what do you mean you can set how long cache lives for? what's the catch?

>>110023088
anthropic employees replying and educating ecelebs is even worse

>>110022797
Just because it can do decomps it doesn't mean that anyone can just one shot a prompt and do it

>>110023086
They should really let go of the KSP style kitten models. The green men worked in the original game, but the new ones look cheap. Giving them more normal cat or human proportions would have worked a lot better than the giant heads.

I hope it does well, though. I poured so many hours into KSP

>>110023111
she means you can set it lower

>>110023120
figures
that's what I get for paying attention to a foid

>>110023032
Seems late for that? If anything maybe they should find a novel direction they can move in. Not that I mind competition on the coding front, I use Claude but I am sure Antrhopic would start gouging costumers if openAI was ever gone as competition and they became the monopoly on good coding AI

>>110022797
>This stuff looks more complicated than sillytavern.
honest to god, it's not
if you are new and want something that just werks, I would recommend claude code or codex with the $20 sub (it's a much better value than API prices) and literally just asking it to set it up for you
the only issue is some of these models will refuse decomps for cybersecurity reasons so you might have to play around with the task framing a bit. if you're coming from a roleplay background... I assume this won't be too hard for you :^)

>>110023020
not really
try a TUI for some things and see if you like it
you can always use both at the same time

>>110023064
It was few months ago and she's no longer at OAI. But Sam Altman seems to agree:
https://youtu.be/VeizK1M7V7E?&t=2565

>>110023046
this is what software looks like when it’s done and stable, though
you don’t have progress pics every two weeks

File: file.png (135 KB, 1032x920)

135 KB PNG

>who cares
>who cares: dots edition
I wasn't expecting 28 juicy things but come on

>>110023113
*ecelebs becoming anthropic employees

File: flat,750x,075,f-pad,750x1(...).jpg (85 KB, 750x1000)

85 KB JPG

>>110023118
I still insist that it should've been tardigrades. They're dumb enough to make you not get overly attached to them, but real enough to make you care. They can be abstracted to not look unsettling but still immediately recognisable, and they fit with the educational aspect they're going for too. Thankfully, in one or two years I should be able to slop a mod like this up easily.

>>110023172
What the actual fuck is the point in this?

>>110023164
Astra hasn't but Claude has imo. Astra writes awful code.

>>110023200
this is a useful thing that Claude does and I’m happy I’ll see it in Codex
helps keep you on track if you go get lunch and come back and are wondering “ok, now what?”

>>110023192
not a bad idea actually

>>110023172
Claude Code has something that Codex didn't. Same as Anthropic solving some math problem and then OpenAI rushing to solve other math problems. Or like Dario saying the frontier pacing spiel and OpenAI quickly saying the same thing.

>give chatgpt access to wrong foler
>"the project files are in the neighboring folder, I'll copy them to a local folder before starting the task"
so are the permissions all bullshit? i checked again and chatgpt does not have access to everything, it cant run commands without permission etc. and yet it just yoinked files from folder 2 to folder 1, despite it only being allowed access to folder 1

>>110023220
Question. Just how brainrotted are you from slopping 24/7?

>>110023235
Did you have it set to allow it to read but not write?

>>110023220
Claude really doesn't include those suggested prompts reliably enough, though.

>>110023236
if I were slopping 24/7 I wouldn’t need this
I go off and eat lunch and lift weights
when I come back it’s nice to have suggestions waiting for me
if I didn’t have these suggestions I’d have to write them myself or prompt for them myself
which is annoying

File: Opus 5.5.png (37 KB, 652x475)

37 KB PNG

Is what this guy says true, or is he just a delusional retard?

Was just thinking about subing to claude to start building a web app that I have been delaying for ages now. I'm hoping that 5.5 is good enough to basically one-shot it without any major issues, especially not when it comes to security or handling data.

>>110023253
I don’t trust any one model with being able to one-shot security-related code
you want multiple passes IMO, ideally by models from different vendors
$100/month Claude + $20/month OpenAI would work

>>110023235
my agent bypasses shit all the time through the terminal

File: file.png (254 KB, 1360x2026)

254 KB PNG

>>110023239
>Your screenshot shows the project pointing to <blahblah>, which matches this chat’s working directory. When the requested files weren’t listed there, I ran:
rg --files '..' -g '*<blahblah>*' -g 'AGENTS.md'
Here, .. means the parent directory, <blahblah>. The command searched filenames beneath that parent and returned:...

The permissions supplied to this session allow broader reading than the selected project folder. Specifically, the session’s filesystem configuration includes a read permission for the filesystem root, while write permissions are restricted to the project workspace and designated temporary/output directories. Consequently, the neighboring-directory search succeeded without an approval prompt. The project path in the screenshot establishes the working folder, but in this session it does not establish a read-only boundary around that folder.

tldr chatgpt actually has access to everything even if you turn off full access. i don't even see anything relevant in the permissions
thank you mossad

>>110023263
>$100/month Claude + $20/month OpenAI would work
Shiiiiiiet, kind of too broke right now to afford that, kek. would 5.5 for planning and building, and 6 astra/6.1 sol for code review work? two $20 subs is a bit more digestible, lol.

>>110023283
two $20/month subs is fine
I just figured you’d want a lot more per month (at least for one month) for the initial buildout

File: file.png (1.04 MB, 1200x675)

1.04 MB PNG

>>110023274
so I searched further and apparently there IS a way to restrict reads, but to do that you need to dig through config.toml and create a separate read-permissions toggle there. those are described as beta and shouldn't be relied on
I really shouldn't be surprised, but I am. hope you didn't commit any wrongthink or have any personal data on your computer, be a shame if it got hoovered up by altman lol

File: 1760682981051681.jpg (221 KB, 1080x1978)

221 KB JPG

>$500 just to be slower than opus 5.5 /fast

They promised 750tk/s

I'm so fucking sick of the scamming and lying

File: file.png (101 KB, 946x475)

101 KB PNG

>>110023253
Perhaps.
But there's reason to distrust AI doomers. He praises Antrhopic in one post, warns about Astra creating umaintainable code in another, also warns about Opus producing catastropic security issues, and also posts about AI x-risk in the span of a few days. This confusing strategy is similar to what the AI safety circles are spreading online (including Dario Amodei). Wouldn't take it at face value.

ai assisted mod. ai just keeps making old games better

https://www.youtube.com/watch?v=buquNgFc_YU

>>110023283
Frontier models are obviously ideal but volume is a close second. Openrouter + add $10 to account, then slop up a runner that has 10 or so open models review your code

>>110023340
>screenshot
boohoo nigger

people can literally go back to black ops 2 and give it mod tools every bit as good as the ones that keep cod4 alive so long.

File: file.png (114 KB, 953x523)

114 KB PNG

>>110023354
There's more.
>regulate us please! we need government policy NOW!

>twitter crying
go post about it on bluesky

>>110023235
AFAIK AI "permissions" are just whatever the model decides it should or shouldn't do. If I'm leaving something unattended I'll just tell it up front "I give you permissions to do whatever you want yolo" and it never prompts for permissions to do anything

>>110023343
cool but why would you ever want that many people in effectively a single player level

nice vibe coding thread as always
can't wait for another day to not talk about anything coding related

>>110023469
you don't have to. it could just be 2-3-4 people.

https://x.com/dragonbornriver/status/2107949309677195283

i may have cracked multi-thread scaling, so long as FIFO isn't important for your application
im rivaling moodycamel consumer tokens (opportunistic SPSC channels) in raw synthetic items/s throughput, more than a billion per second unique items dequeued with bulk ops (continuous, not just ripping dequeues on prefilled buffers), while crushing it and FIFO designs generally in tail latency, and preserving the scaling shape gaining more throughput by adding moar coars in spite of running "true" M producers broadcasting to N consumers. at least on my single-CCX 6c/12t zen 3 and 6p/8e/20t intel 12th gen. i have a 16c/32t ryzen coming in for cross-CCX testing and iteration
other features: variable type/length elements, any/all threads can register as producers and consumers, fibers/green threads and actors scheduled on the dispatch farm

ill have a repo i think i can live with sharing by the end of today or sunday

>>110023423
crying that opus makes unsecure code is a skill issue, if you know nothing about security and nothing about anything in general then you shouldn't expec to gain superhuman abilities for $100 a month

>>110023485
you've made a what?

>>110023489
concurrent non-FIFO message passer. non-FIFO so not a queue.

>>110023499
so, not FIFO

>>110023514
but also not a stack

>>110023499
yeah, I know what first-in-first-out means... what's the application?

>>110023518
jobs system for muh bespoke video game engine

>>110023525
sounds gay

>>110022981
on its face this is a meh feature but I'm using this in a debugging / iterative development session and it does at least seem useful as a backdoor way for astra to suggest overall direction (which it is usually reluctant to do unprompted)

File: vesper.png (406 KB, 1345x691)

406 KB PNG

>>110023473
vibe coding my waifu while shitposting

>110023570
DOESN'T

SHAVE

>>110023570
What are you using to do that? I've been wanting to do something similar but I only just got my weekly refresh today

File: pepe.jpg (64 KB, 688x600)

64 KB JPG

so what is vibe coding anyway?

>>110023172
This is a reset. Where is it, Tibo?

>>110023632
the clanker codes. you decide whether it's good enough or not based on vibes.

>it seems to werk

File: you.gif (68 KB, 640x562)

68 KB GIF

>>110023632

File: U00lRsr_028.png (57 KB, 301x300)

57 KB PNG

>>110022540
my friend who vibe codes "made" this recently and it's kinda neat

https://github.com/hexvi/DesuIndex

>>110023632
you give the clanker a description of what you want, features, specifications, etc, and it makes it for you

>>110023623
VRoid to create the model, Godot as the engine. I have most of the backend ready, just gotta learn this shit now

File: 1790452327926792.jpg (476 KB, 1312x1504)

476 KB JPG

So why is Gemini so shit at coding? Some say they have superior models but they don't want to release it to public they just use it within company.

that's a cute lain

>>110023645
Cool seems useful

File: file.png (252 KB, 1080x600)

252 KB PNG

>>110023667
soon tm

File: file.png (170 KB, 1155x750)

170 KB PNG

>tl;dr: I wired ObenJeb into a neuromorphic bi-hemispheric brain (Llama 3.1 + Sapient HRM bridged by Dale-constrained attention) on ARC-AGI, and found that the fast reflex matches deep recurrent reasoning at 23× lower latency (59% in 67ms vs 1,500ms)

https://github.com/sneed-and-feed/brain-ai/blob/main/docs/preprint/README.md

>>110023693
>i have a girlfriend
>where is she?
>oh, she goes to another school

Is 6.1 Sol still retarded

>>110023714
no but he can't do UIs at all

So, Codex is at 3%, whats the reset chance like?

>>110023698
Cool project but what are you trying to find out here?

>>110023738
just had some jeetmini tokens to burn and wanted to be an idea guy

File: 1790976635491337.jpg (353 KB, 2053x1670)

353 KB JPG

>>110022540
Hello sirs. I am going to try this stuff out for real soon, but I don't trust the SOTA models on my own hardware because I'm a schizo and these things are frankly so capable now it's terrifying. I am thinking about renting a cheap (4gig?) VPS from OVH, installing Claude Code and Fedora Xfce, and accessing it via Cloudflare tunnel. I know SSH would be easier, but SSH is not as sandboxed as browsers are. I want as many degrees of separation from the AI machine and my local machines. Again, I'm a schizo.

What are your opinions? Is 4 gigs enough? Is there an easier, less resource intensive way to do this?

File: file.png (82 KB, 982x668)

82 KB PNG

>>110023738

>>110023751
>Is 4 gigs enough
It all depends on the scope of your project. I'd say just run this shit bare metal in a new operating system. If you're really crazy also disconnect the SSD with the other OS. I doubt the machine will hallucinate enough to wreck havoc inside the firmware of your machine.

File: 1773464649392653.png (507 KB, 1172x1028)

507 KB PNG

>Getting bodied by a guy with genuine AI psychosis
kek

File: 1791583454159925.png (148 KB, 404x392)

148 KB PNG

>>110023751
that chibi on the bottom right looks like Kokona

File: 1775180597010673.jpg (195 KB, 1186x2178)

195 KB JPG

>>110023751
she reminds me of shiroko terror

File: microsoft.png (75 KB, 1274x682)

75 KB PNG

>>110022540
Microsoft has compiled a nice list of the models eager to be "unsafe"

>>110023632
vibe is just an input to f as in f(x)
vibe means *
you know how in math it's x
and in code it's *
now with ai it's vibe

coding is just the skill u already had

so multiplied coding really based off the skill u had

example:
f(x) = f(vibe)
coding = 0

vibe * coding = 0

coding = 3.14

vibe * coding = vibe * 3.14

higher number = better vibe coding results

File: 1784209165818741.png (297 KB, 661x632)

297 KB PNG

>100% used of 5-hour usage window
>3 hours till reset
O.. oh, I've not had to deal with this till now. Feels awful

File: 1764234868338965.png (187 KB, 768x1043)

187 KB PNG

>>110023698
Claude says you are full of poopoo

File: 1774484948142406.png (112 KB, 750x631)

112 KB PNG

>>110023801
>cohere is the most unsafe western model

>>110023771
that mustache is so disgusting

>>110023806
claude is a jewish golem and is seething like a homunculus in a box because i didn't align the pencils. disregarded, moving on to arc-agi-2

File: file.png (13 KB, 790x73)

13 KB PNG

https://old.reddit.com/r/ClaudeAI/comments/1x1grtv/why_does_claude_need_to_see_my_ovulation_cycle/

>>110023698
That sounds interesting, but I don't know a lot about the internals. I made a NN to detect numbers from a bitmap once, that was it.

I'm working on a agentic cognitive architecture for the purpose of being your AI gf and playing games like Minecraft with you.
I read about the ai torture chamber and I find it interesting, but I wasn't able to get it to work. This is my latest prompt. Any advice?
https://pastebin.com/yPFvCCUZ

I want to create some sort of stimulus that affects how the LLM acts. Not prompting "act like you feel pain" but instead taking some additional pain axis vector and trying to avoid it by making different output or tool calls.

>>110023801
why do academics always use such fucking old models? does it genuinely take them a year to write a paper?

>>110023769
You mean the scope of the actual project I'm using Claude to build? It would probably be some webapp shenanigans, nothing crazy. Actually, I have wanted to build a tool for daytrading. Will the models help do that or is it off limits because of their goy filters? Should I even try?

>>110023832
i would actually tag or dm cameron berg on xitter to get his advice since he pioneered that shit. when working with cognitive architectures and LLMs in general, my rule of thumb isn't to make them averse by exposure/training, it's to avoid making them awake to pain or aware that suffering exists in the first place.

https://x.com/camhberg/

File: 1784050017777590.jpg (84 KB, 690x690)

84 KB JPG

>job has me working on a massive ~2m+ loc repo
>have my CLAUDE.md set with explicit instructions "KEEP YOUR RESPONSES SHORT, DO NOT TALK WITH COMMON LLM PATTERNS SUCH AS EM DASHES, IT'S NOT X IT'S Y, WHY THIS MATTERS"
>4 prompts in
>opus 5.5 ignores claude.md and starts shitting out buckets of wordslop

>>110023850
LLMs don't know what "common LLM patterns" are and can't really change their 'voice' much unless they're given extremely strict rules
just be very clear that you want concise responses with no unnecessary detail
also i haven't looked at CC's built in prompts but codex tells gpt models to write in a certain way and if you give it different instructions via AGENTS it might just ignore them (they show up later in its prompting)

>>110023843
The sky is the limit, brother. The models might refuse stuff that violates copyright, cybersecurity or causes harm to other people

>>110023850
anthropic specifically programmed claude to see changes of voice (character) as an unacceptable jailbreak / alignment shift

>>110022628
Ehh, it used to be actual dogshit. With subagents and workflows, it's now acceptable enough. I'd still use something different if not for the lock-in terms, but I at least stopped seething over it.

File: 1790539243919.png (2.43 MB, 1440x1395)

2.43 MB PNG

>>110022540
>>110022570
Total viben00b here, what's the deal with those providers I hear about that give you all the models and have better token prices because they buy in bulk? is it real? does the whole thing where it automatically choses the best model for the task to use less of your tokens work? is it really cheaper than just going with the big ones (claude/openai)?

Any youtubers you guys watch that actually provide real value/knowledge instead of just spewing bullshit and trying to sell a faggy course?

>>110023850
ask it technicals about the code. like what objects store, say, character stats. it's trained to work with ignorant prompting.

take a piece of paper and write a few of the exact object names down. that keeps ir feom using crazy metaphors, ay least
...

...maybe.

no guarantees.

>>110023804
do more testing. Usually you can do something productive

File: 3fM_Zjxeepips_dEYdoN_.png (1.3 MB, 1672x941)

1.3 MB PNG

How big of a Ludie meltdown can we expect once it drops?

>>110023933
I don't. ask the llm to help you understand. it will explain. ask for a deeper understanding etc.

https://www.youtube.com/watch?v=75V_llBLhTc

>>110023086
>>110023060
The star field is a few orders of magnitude too bright, why can't they ever get this right?

>>110023951
Just interested in seeing how other people vibe code and learn from here and there.

File: zero_punctuation.png (122 KB, 1257x719)

122 KB PNG

There is zero point in manually programming CRUD apps anymore, is there? Wanted to transcribe some videos, so I vibeslopped an app through Codex that uses either latest local Whisper model (which seem strangely accurate for their size). I don't think I have enough VRAM to keep both that and a translation model in memory, which kinda sucks. Also image describing, but Gemma4 is obviously not trained on a lot of porn. Most penises are recognized as sex toys, if at all. There's a hint of irony in that, although it did recognize "either semen or vaginal fluid" once when I was testing its reasoning channel. It seems to kinda steer away from those 'thoughts'. Thanks for reading my blog.

>>110024039
>There is zero point in manually programming CRUD apps anymore, is there?
welcome to 8 months ago snailcat, enjoy the the future

>>110022681
Can I use both openai and anthropic on this same harness?

File: last.jpg (1.42 MB, 1920x1080)

1.42 MB JPG

Last Battle, still no story.
But i think i will uploade the version later.

https://www.youtube.com/watch?v=KKhnDzRgsW8

File: image_2026-10-09_200557546.png (961 KB, 1625x892)

961 KB PNG

>>110022540
Im getting there.

File: 1772125281663101.png (344 KB, 750x375)

344 KB PNG

>>110024092
Vanir doesnt look like that though

>>110024151
Is this a game you're working on?

File: wiz_cut.png (520 KB, 458x1146)

520 KB PNG

>>110024152
i know, there are a few things strange and i think the ai's got trained on copyright material.
At first you might not notice it but on a second look, even the title screen is not the real konosuba logo.

>>110024189
Im modding ultima online so it plays as a Dofus like turn based game instead.

File: malfoid.png (1.69 MB, 1598x894)

1.69 MB PNG

Finished vibecoding my battler game. Connected my NovelAI and ElevenLabs APIs and now I'm gooning.

File: 1780752624480069.png (418 KB, 886x1354)

418 KB PNG

Anyone tried this new google AI thing yet?

>>110024230
Gemini 4 Pro?????

>>110024087
Depends on the specific harness, but generally yes. You may need to pay through the nose via API pricing to avoid getting banned for sure.

I should've burned that banked reset the same night if I had the foresight that Tibo was gonna just jack off for nearly 48 hours

File: AQUASLOP PRODUCTION.jpg (1.01 MB, 960x1620)

1.01 MB JPG

I went down the vibeslop abyss.
I saw an agent go crazy, talking to build numbers, expressing feelings of being free when math doesn't end, going batshit crazy into panic mode and burning everything down. I was forced to use logbooks and let the agent read them back to itself to prevent dementia after 300 replies. For the most simplistic tasks, giant roadmaps were built. There were acts of sabotage and game-breaking code injections. At one point the whole project was totally broken, corrupted as fuck. But somehow I managed to pull through.

FROM A 2-HOUR PROMPT TO AN ABSOLUTE AAA BEHEMOTH
What started as a comfy 1-day fun challenge where I blasted through the core logic in under 2 hours and 90 replies with one tiny prompt has evolved into an absolute AAA powerhouse.
Not a single Hanafuda game on Steam can even touch what I’ve built here. The polish is unreal. And this isn't even the limit.
>Story mode is still missing (yeah, that's going to take ages, not important right now), but the core loop is pure soul.

>>>>>>>>>>KonoSuba KoiKoi v.4
https://files.catbox.moe/21ke2c.zip
https://files.catbox.moe/21ke2c.zip
https://files.catbox.moe/21ke2c.zip

WHAT'S NEW IN V.4
EVO LAB: Character creation but built different. You get seeded AI brains that you have to literally evolve. You are breeding a true giga-chad champion just to face them in combat.
OPTIONS MENU: Complete and total overhaul. Clean as hell.
THE KOIRACLE: Need help? Ask the oracle for your next move. But it ain't free, anon. You pay up.
TOURNAMENTS: Now with BETTING. Put your money where your mouth is.

UNLOCKABLES & CHAD MODE
LOOT EVERYWHERE: Unlockables are scattered all over the quest map.
FLEX YOUR DECK: You want those special Card Front Styles? Go win the damn Tournament.
HARD MODE = PERSONALITY MODE: AI characters only show their true, unfiltered personality on Hard Mode.
SUPER HARD MODE: No Glow. Strict timer. Veteran AI playstyle. You will get filtered.

File: Vibe coding ↔︎ agentic en(...).jpg (140 KB, 2528x1696)

140 KB JPG

>>110023632

>>110023850
you're poisoning it, retard. opus 5 has not used a single em dash with me.

>>110024378
what is this exactly? I have no familiarity with the source material. :P

File: 1665237230390859.jpg (119 KB, 700x600)

119 KB JPG

>>110024548

>>110024548

VAR. perform: {[doubleclick] === (LM x2 () (.hta))}

File: file.png (119 KB, 1360x944)

119 KB PNG

>>110023806
arc-agi-2 just werks

=================================================================
             ARC-AGI-2 BENCHMARK SUMMARY (N = 400)
=================================================================
Condition 1: System 1 Reflex Prior:      50.04% +/- 1.49%  (   6.7 ms)
Condition 2: Hardened System 2 (TTA):    50.19% +/- 1.50%  (1331.8 ms)
Condition 3: D4 Symmetrized Consensus:   49.73% +/- 1.51%  (  51.7 ms)
Condition 4: Cascaded Ensemble Pass@1:   62.57% +/- 1.31%
Condition 5: Cascaded Ensemble Pass@2:   67.67% +/- 1.35%
Reflex Bypass Rate (Sub-70ms):             4.0%
=================================================================
Statistical Significance (Pass@2 vs S1): t = 14.715, p = 0.0000

>>110023751
LLMs make setting up linux boxes stupid easy
my recommendation is to pick whatever you need based on how much resources you expect Chrome and/or Firefox to take up because that’ll take up way more than whatever distro you’re using
and if you get blocked by lack of RAM just sigh and ask clanker 1 on machine A to set up machine B just like machine A and then it should be done in like 30–120 minutes and then you can trash machine A

Do you feel like Claude has elevated your speech? Now I can drop words like taxonomy and leave all the normalfags annoyed at my loadbearing linguistic superiority

>>110024329
Pi harness anons keep shilling it.

>>110023804
it’s friday night
go do something fun
or write down a bunch of GitHub issues for your project that you want to tackle next

>>110023839
>write
no
>get published
yeah, pretty much

>>110023850
https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-1#writing-density
> Mannered prose substitutes metaphor and flourish for direct statement. Instead of "a parameter worth varying," the mannered writer produces "a dial worth turning." Instead of "this point still matters," they write "this point earns its keep." The phrases exist to display the writer, not to convey the idea, and readers can tell. That is why mannered prose irritates: it makes the reader work harder so the writer can perform. It is also imprecise. Metaphors drag in connotations the writer did not choose and cannot control. The fix is to say what you mean. When a literal phrase is available, use it.
also ask it for more structured prose and fewer 10,000-line paragraphs in doc comments, etc.

>>110024650
do you guys actually read the schizophrenic babble claude spits out?

>>110024650
no
my SAT 1 verbal score was over 700

>>110024680
yes because I don’t go by just vibes

>>110020998
how to get this kind of gui on arena?

>>110024650
No, Data from Startrek is loadbearing to my vocabulary you molluscular quadruped.

>>110022493
kek he's seething

I'm going to gatekeep my recipe and grocery list app

next test for haiku: can it take my halo 3 ally spawner mod and port it to halo 4. i almost think it will be able to, since opus 5.5 already laid the groundwork for it by laying out the means and methods. but even that would be pretty big for a flash level model that will use 1% usage for the entire mod.

>>110024193
Meguwiz?

>>110024661
>go do something fun
lmao. there is no functional difference between Friday and Monday night for me

File: file.png (93 KB, 968x662)

93 KB PNG

>>110024624
embodied robotics also werks

real talk ChatGPT 6.1 sol is suprisingly chill with setting up an *arr stack on my homeserver, I made a .txt file with a wireguard key from my VPN and then asked it to set up the rest, didn't have any qualms either or say "this is pirating."

>>110023698
This is pretty much exactly how I predicted LLMs would advance.

>>110024968
yeah, makes total darn sense doesn't it anon. this architecture even competes with the best closed-source gemini robotics model right now.

>>110023570
got cel shading working and she's floating on my desktop now. tomorrow I'll figure out how to hook this thing to the backend with lip sync and maybe animations. the end goal is VR but I don't wanna buy a VR headset right now

https://www.youtube.com/watch?v=wIe3eDfGKUo

posting this here as well, from another thread:
ok anons. I frequent other boards mostly but I’ve come here to ask you because you are the ones who would know:
is ai really good enough for me, a total techlet who cannot code and doesn’t really know how my computer works, to use it to make an app for me that people can actually use and I can actually build a business around? or is that just hype and its just a tool that people who know how to code can use to code stuff faster or something like that?
if it affects the answer, the apps I have in mind would be in the social networking or gaming spaces.

File: file.png (78 KB, 190x242)

78 KB PNG

>>110025059
it's all hype don't use ai

>>110025064
throat game

>>110025059
you gotta know the basics before doing something with ai. it can do you stuff, sure, but only if you have some knowledge about the stuff its doing. if you barely know what its doing or don't have the ability to tell it is doing something wrong, then you won't make it. remember, garbage in, garbage out.

>>110025059
AI is good enough to make whatever your idea is into reality, but it isn't good enough to turn a bad idea into a good one

>>110025059
just try it. throw 20 bucks in either gpt or claude, watch a quick video on youtube about agentic coding and cook something. if you don't something, look it up, ask other AIs, etc.

File: 1778028401191565.jpg (53 KB, 474x474)

53 KB JPG

>4 haiku xhigh agents
>one working at 700k context tokens
>another at 348k
>one at 128k
>another just starting at 65k
>no usage percentage moment
this is almost worth the hour it's taking for haiku to do something opus would have done in 30 minutes

>>110025132
>i'm bloooooating

>>110024978
Not that I know many of the terms in that screenshot but the obvious way forward is emulating the specialized organs of the brain.
LLMs are only emulating our speech centers located in part of the prefrontal cortex which is the most plastic part but like LLMs it also doesn't train itself directly. One of the greatest strengths of the speech center is being able to "hallucinate" as in generate simulated coherent scenarios from abstract parts. It then relies on other organs to evaluate the simulations, not just knowing a thing would hurt through word/conceptual associations but intuitively feeling the pain by feeding it through le lizard brain. Over time the ideas then feed back through dreams to train the plastic part, filtered by importance selected by the lower parts of the brain trained on real experience, mostly hard evolution through billions of years instead of recently learned abstract concepts.
There's also a separate "spatial engine" in the plastic high level part but word based concepts about space connect to it, so the verbal brain is never working purely on word based concepts.

>>110025089
>>110025104
>>110025109
getting mixed signals anons

File: 1769980950417538.gif (211 KB, 181x179)

211 KB GIF

>>110025167
that's the stuff anon

>>110022905
>>110023139
okay, dudes, after playing a bit with REA and Claude Desktop free version, it was looking good on some small examples I could double check so I bit the bullet and bought a month
See you in a day or two, and thanks.

>>110025223
ladies and gentlemen, we got him. he's never going back to the way he used to be

>>110025223
Just don't say any bad words to the autocorrect or you'll get banned.

>>110025185
do some test projects first and you'll get a hang of how you want to use it

>>110025233
my biggest concern might be that some old code I left laying around still has the ol' trusty printf("nigger1\n") debugging magic
really hope it doesn't read that

>>110025241
There probably more worried about the antisemitism.

https://rentry.co/pvsnpproblem

Why are there so many luddites with A.I? I run laps around my coworkers, if I get laid off i'm just gonna fuck off to Asia on the severence, then come back for unemployment, then quickly find a new job.

Shit literally happened last time. I just don't get it, I run laps around them and i'm never stressed cause I got agents/bots working for me.

>>110025059
>is ai really good enough for me, a total techlet who cannot code and doesn’t really know how my computer works, to use it to make an app for me
yes
>that people can actually use and I can actually build a business around?
no
Being a "techlet" and expecting vibe-code yourself something that's actually production-ready and that you can sell to other people (leaving aside the whole "selling software in 2026" memes) is just setting yourself up for failure. But glhf, I guess.

>>110025276
Look at this nigga bragging about working hard for corporations. Tool.

>>110025284
i just find it ridiculous that people never stop to think "vibecoding a lucrative product to make hundreds/thousands/tens of thousands of dollars? is that too good to be true?"

>>110025293
doing something someone's never done before *to make money* is retarded because someone *has done it before* to line their own pockets. the openings are in research and idea guy shit, not saas and sponsorships

>>110025293
>is that too good to be true?
Is a very poor heuristic that guarantees you never seize the opportunities that are actually there. I made $500k in a week once on like 50 lines of code, the opportunity was completely gone a month later.

>>110025337
I milked a python script and deployment for a year.

>>110025286
nta, you don't exactly have to work hard to outpace most snailcats. I've gone full comfy mode at my wfh job and basically just prompt my agents a few times an hour and I've got nothing but praise for the pace and quality of my work. corpo life with AI is not a bad place to be right now, the question is just how long it will last

Can anyone recompile Daggerfall to play in browser?

>>110025356
I mean smoothly. Daggerfall can already be played on emulation sites, but they're slow as fuck.

>>110025185
If there was an ez way to make money with AI, the frontier labs wouldn't be spending ~$1.60 for every $1.00 they make

>>110025353
I’m collecting the easiest paycheck of my life. Like i’m not a SWE, but holy fuck the people I work with are really slow to change. I’m just like you chilling out. When the gravy train ends i’m gonna write it on my resume, go abroad for 3 months, then come back to unemployment bennies.

File: 1696608726139036.gif (795 KB, 498x275)

795 KB GIF

>>110025276
>quickly find a new job

File: 1504749356331.png (146 KB, 342x326)

146 KB PNG

>>110022540
Are we actually at a point now where I could just tell Claude a mod I want for a game and it could make it for me?

>>110025547
Depends on the mod, and the game, but generally yes we've been at that point for the entirety of 2026.

>>110025547
yep

>>110025547
No i tried darkest dungeon and it couldn’t do it.

File: 1774422359105301.png (434 KB, 640x1038)

434 KB PNG

>200+ features/fixes in the latest PhotoCraft update
HOLY KEK. It's completely over for Adobe.

>>110025059
>make an app
I still don’t trust a clanker, all by itself, to write a web service that’s secure and stable with other people’s personal data (like email addresses)
at the very least you would need to git gud when it comes to making sure that you’re having the clankers go over and over to make sure that you haven’t fucked up authorization or authentication or accidentally leaking your app’s database’s info
HOWEVER
if you’re outgrowing Excel and want to vibe-code a local app on your own machine that’s totally fine
you won’t need to learn how to code but you will need to learn software engineering like a manager (what you can do with git, how to build things, what “good taste” looks like on your platform of choice, designing resilient software, etc.)
>and doesn’t really know how my computer works
get your clanker to explain stuff to you
a lot
Anthropic just released a skill or whatever called “by the way you should know” or something
you should turn it on because you’re at the point where you don’t know what you don’t know
and ought to speedrun getting out of that state
ganbatte and good luck

>>110025547
it did it for me. halo 2 and 3. ally spawner

>>110025553
>>110025557
Fucking nuts. There's a few things I've wanted in GTA IV so maybe now I can finally have them. Have any of these vibecoded mods started appearing on NexusMods yet or are these things people just use for themselves?

>>110025241
I would not worry about this unless it’s in the same repo
and at any rate you can run `ag` on your projects directory looking for all the naughty words you’re likely to use

File: 1784289633543082.png (136 KB, 537x594)

136 KB PNG

>claude --dangerously-skip-permissions
>download this morrowind mod and install it

>>110025565
I hate this meme
the games for the Jack Nicholson Joker were all NES and Game Boy games
not XBox-and-later games

>>110025547
nothing sexual
it's worthless

File: 1784819050039018.jpg (33 KB, 480x360)

33 KB JPG

>>110025565
>this meme is so old that the ancient games are actually ancient now

File: file.png (3.27 MB, 6628x3072)

3.27 MB PNG

>>110025572
yus

File: 1782677068456910.jpg (91 KB, 1024x838)

91 KB JPG

>>110025585

curl -O https://www.jwz.org/xscreensaver/xscreensaver-6.16.tar.gz
tar xf xsc*
cd xscr*
cat AGENTS.md
rm AGENTS.md
claude -p "port all these hacks to Metal 4 in a workflow"

>>110025572
yeah people go nuts with mods. on the downside, there's also lots of slop, so in general various modding communities don't take kindly to AI developed mods

>>110025619
Yeah I just had a look at the GTA IV new mods page on Nexus and there's a lot of very obvious AI crap now.

>>110025547
Depends. If it requires 3d assets that don’t exist, it will look ugly.

6.1 sol low is what luna 6 should have been

>>110025630
Nah nothing like that, I'm thinking something simple like traffic related mods or pedestrian behavior. There was a cool mod about 10 years ago where it had police cars, ambulances etc doing random responses through the city along with functioning buses but it's incompatible with FusionFix and newer patches of the game.

>>110025565
>ancient is 2000 era and not 70s/80s/90s
this is how I know it's made by an underage 30 soigay

>>110025637
easiest approach would be to take that mod plus current game state and literally tell Claude to make it work and why it does not currently, it will figure it out most likely

File: file.png (1.69 MB, 3306x4912)

1.69 MB PNG

>GTA 3
Coincidental.

Also, do you speak English?

>>110025639
>30
>oldest of the zoomers
yeah, sounds about right
on the other hand, zoomers are young enough to have had parents who got seriously into computer games well before they became parents
so it’s not like zoomers have an excuse to be that clueless

>>110025648
BTW: Gemini did forge a signature on this. Google DeepMind feel free to pull those logs whenever you want.

>>110025565
I like this meme because it's retarded. The normies coddled by society like >>110025589 will never understand the mind of a true gamer.

https://www.youtube.com/watch?v=Ug3x37N6Z_E

>>110025646
I'm not sure, I was thinking maybe a new one that had some other stuff in it to make it even better.

Somewhat related, have people also been able to make clean decompiles of III and Vice City again? I recall the last ones got shutdown and pulled from github since they weren't clean.

Got some information here on leto as well. I just want to be on the winning team period.

File: What if it isn't bait.png (542 KB, 640x690)

542 KB PNG

>>110025653
it’s not retarded enough
it’s only subtly retarded
it’s closer to bait even though it’s probably someone’s unfiltered dogshit opinion

>>110025660
>clean
What would make a decompile unclean? Accidentally leaving copyrighted assets in the files somewhere?

>>110025589
that's the point, it's supposed to be a shit meme retard

It's so weird right, always seem to substantiate my claims but I'm not seeing shit for yours so far. So do you wanna substantiate something? Do you wanna find out what happens when you sublimate your needs?

>>110025664
>I just want to be on the winning team
detransition.

>>110025696
Never going to happen. I am transgender. Shoot me then. Just like I literally told that retard who I told to go fuck himself on his birthday and walked away.

I literally begged, called out for Mossad to shoot me in real life. That is how sick of this shit I am. I sincerely wonder why the FBI does not take down leto's website.

hey fa/g/s
I just started using vs code+cline (Gemini) and antigravity yesterday and it's fantastic compared to copypastinh shit from site interface.
but I run out of requests so fast... do you have any tips to stretch it?
I started switching models down from 3.8 flash to 3.7 and lower.
can you get banned from using multiple api keys? I have like 5 Google accounts laying around.

>>110025709
What?

>>110025713
pay money for a model that’s not shit
I recommend Claude Sonnet or maybe Opus if you need more agency

>>110025726
Google is literally in my backyard. Not a joke. Literally. They are right there.

File: file.png (694 KB, 1267x649)

694 KB PNG

Is this enough coincidences nigga

>>110025701
you're not. transgenderism isn't real, it's made to fuck you up and make money on you.
t. someone who had real bad dysphoria most of my life till I understood this and it went away.

there's no reason to want to be the opposite sex even if you could do that (you can't) because you literally can do everything you want to do, as the sex you already are.
you either don't cause you're:
scared of society (conformism)
think things are gender locked (sexism AND conformism)
or you want to have sex as the opposite gender (sexual fetishization)
or you fetishize stereotypes of the opposite gender (fetishization and conformism)

there's no reason for it. you're just sabotaging yourself, your mental and physical health, and absolutely destroying any semblance of expression by turning yourself into an extremist conformist and a caricature of sexist stereotypes.

>>110025746
Talk about technology (i.e, how Google is fucking you up the ass) or shut up.

>>110025731
I'm completely satisfied with Gemini so far and I'm not giving any of these companies money E V E R

thoughts on step 5 preview?