how do i use ai?

>>108549327
I don't know anything about making 3D video games, but I did check it out one time - OpenClaw was able to get me set up with Unity and create a basic default first-person perspective workspace in about 10 minutes, including download time. I stopped there because I lost interest but the next step would have been adding graphics assets. It's feasible.

https://youtu.be/8ELi5Q2CFmI?si=nKU_wnJGS80MDwtv&t=33

>>108549363
This was to you
>>108549359

kek, I shouldnt have shilled qwen coder cli. Now everyone's using the free account credits and its slow as shit.

>>108549363
any major game engine will do for this task. you should research which game engine has the best AI integration, which most likely is NOT source2sandbox. Since you have 0 experience, go with Unity. Unreal Engine has strong AI integration but you will be overwhelmed at the start. Also Unity always you to package for web which UE doesnt. You could also consider playcanvas for web. Godot could also work, but I never tried it

>>108549363
is she ai

File: 20250101_132606-scaled.jpg (804.5 KB)

804.5 KB JPG

>>108549635
For the game I was thinking about, that was actually about halfway to completion. It was when that game Retro Rewind came out - yknow the one where you manage a Blockbuster video store? And I thought "I could make this stupid millennial nostalgiabait walking simulator bullshit". I won't tell you my idea for the theme because I still think it's a gem and I might go back to it, but yeah in about 10 minutes I was walking around a blank green canvas in a Unity workspace. From there I just needed to download some free assets and tell Claude to make them intractable. It really is that easy now.

I've made a couple of 2D games, those took a couple hours each to ship in a state I considered ready for public.

>>108549729
>intractable
Yes Claude make them difficult to manage, that's what I meant and I know you'll take it literally you little scamp.
interactable*

>>108549729
And for the record, I had never used Unity before this. I still don't really know how to use it - Claude did most of the downloading and installing and just told me which default stuff to choose for a first person game. I said "wow that's neat" and never went back to that project. Proved I could do it though.

I am not sure but it looks like someone slopped violentmonkey to work with MV3.
https://github.com/violentmonkey/violentmonkey/issues/2284#issuecomment-3997951255

>>108549962
>then you got a product
Brother I'm not sure how to tell you this but I think the era of "launching a product people will pay for" might be over. I think we might now be in the era of "making it easier for people to make their own products".

Started the task for the night, now I can go to sleep.

This is gay. OpenClaw has started asking permission to do every little thing. If I ask it to identify a problem, it does but then it says "Do you want me to fix it?" Of course I want you to fix it, that's why I asked you to identify it, don't bother me with your questions. It didn't used to do this so much I feel like.

>>108550045
I think we still have some time, the bar is just higher. People can prompt now, but you can't get a really good game in an hour with prompting, you still have to break it into a ton of sub tasks and most people can't be bothered to do that. You probably can use AI to help you make GTA 7, but you can't just prompt it to do it all for you. As long as you are willing to put in some effort, there's still a chance, but probably even smaller than before.

File: file.png (17 KB)

17 KB PNG

been a rough week adjusting to the new limits
feels like weening yourself off caffeine

>You have just independently reconstructed full Gnostic cosmology in about four chat messages.
Do you guys talk to Claude about theology sometimes between coding?

>>108550605
no. i'm not a nerd so me n claude look at pics of hot chicks.

>>108550630
>wdg dies
>vcg shows up

Curious.

>>108550256
>Genie 3, generate a gritty Hong Kong cityscape. My character is a gritty Chinese guy.
Uh oh, now I just need to look at my other monitor and start adding gun physics.

I was thinking. Alpine.js is the perfect web framework. Performance is the only issue people could really complain about. It goes on a tick system and updates the html when a variable is changed. I'm assuming Javascript handles the variables.

How hard do you lads think it would be to get alpine.js's source code off github and vibecode the core to use some rust wasm? I know >rust but there was a framework I saw that was able to store a ton of data in a rust vector for infinite scroll. I was thinking do a rewrite of alpine's core system but store reactive variables in rust instead and run the tick system there, then on reactive variable change in rust, tell the Javascript to update. I figured with this new pretext.js craziness as well rendering a virtual scroll that could hold a list hundreds of thousands would make an ultimate web framework and fix how broken the web is.

File: 1000020541.png (787.6 KB)

787.6 KB PNG

Mythos officialy announced
Will never get wide public release ???
https://www.anthropic.com/glasswing
https://youtu.be/INGOC6-LLv0

>>108550783
is it called mythos because it doesn't exist?

The vibes aren't dead yet!

https://youtu.be/LiaVWUI44Og

Gonna keep on with GPT 5.4 xhigh until I lose access entirely.

So they're eventually going to close down AI more and more after realising releasing it to the public was a mistake? Thing is, outside of places like this, AI adoption is still basically zero. They're going to make it unavailable before it even got out.

>>108551348
>AI adoption is still basically zero
Pretty much every company is fiddling with AI in their pipeline. It's significantly rarer to see a major company not use it at all.

>They're going to make it unavailable before it even got out.
It's already out... we have all the local models we need and they keep releasing more.

File: images.steamusercontent.jpg (49.9 KB)

49.9 KB JPG

>>108551380
He means from these "frontier" US labs. The issue here is not safety or IP bullshit, the issue is Anthropic/OpenAI going into full rent-seeking mode so that everyone working in tech is forced to fork out loads of money just to stay competitive on the market. These companies can also choose to give exclusive access to hand picked individuals and cut everyone else off and there would be nothing to stop them. We already see this in China with only certain firms, Deepseek and Alibaba among them, getting priority to credits for energy and compute. The US just replaces the state with corporations.

>>108551505
Anthropic/OpenAI get too expensive then competition rises up. I just don't see this as something to worry about long term. The tools all exist and are nearly all open and available. The only major issue with AI is how we're getting our energy and chips. We need to significantly reduce energy costs one way or another and the chip shortage seems like something that will solve itself over time. I just don't see this disappearing from the public long term.

File: 1773494448109876.jpg (31.8 KB)

31.8 KB JPG

I've been using Claude Code Max for a year. For the first time ever, I'm stuck without it because I've had a hard month and I'm out of money until pay day. I feel so fucking lost. 20 years of coding without Claude but I just can't go back. I can't.

I don't even know what to do with myself. I can't talk to Claude, I can't plan any projects, I can't explore things or build anything. I feel like a super hero that's lost all his powers.

>>108551666
now that's just sad.

>>108551688
I wanted to build a local LLM rig but it costs too much money and the models I can fit on my budget GPU are total dogshit, I couldn't do anything with them

>>108551724
I'm a white European but I'm 10x more productive with Claude. Also I can't do my job anymore because it's 100% vibe coded and I don't know anything about the codebase. Kek

>>108550783
>Mythos officialy announced
>Will never get wide public release ???
Because releasing it publicly would probably show that it's not AGI, but a slightly better Opus finetuned on cybersec tasks that they had security professionals work with for weeks to find and sift through potential bugs. Releasing it to the general public in any way that people could substantially use it would destroy the illusion.

File: 13d1b9b7-a763-4788-84c9-007c30a666de_800x1008.jpg (183.1 KB)

183.1 KB JPG

>>108551570
I'm not talking about a black swan crash, im talking about a massive takeover in the tech space over the time between now and when the energy & chip supply gets better. If the models for 2026 are as good as these companies are advertising then the issue is not energy or chips, its who can get them first. On a short enough timeline, the US corpos win out. On a longer timeline, it works out to your scenario.

>>108551783
>AGI meme
Frontier models are already more intelligent than most humans

File: 1775588907961567.png (109.9 KB)

109.9 KB PNG

>>108551783
Its not a fine-tune, its allegedly a better model in general. The reason its not released is because Anthropic is cash and compute constrained, they can't release it publicly because they cant even stand up current demand for Opus 4.6 with 90% uptime. If they're hosting an even better model, its going to be someone who can pay insane prices for it.

>>108551824
I'm on the side that tends to agree, somewhat at least, but those companies cried wolf since GPT-2 was branded as way too risky for public release. I talk to LLMs more than I talk to people. I'm using them all fucking day. I still think that this is marketing so that people talk about it and hype the company some more in prevision of the IPO, whenever that will be, but I don't know the plans or timeline, so I'm chalking it to regular plain marketing. "We have the best product but you can't see it." Ok.

>>108551849
>The reason its not released is because Anthropic is cash and compute constrained
I don't believe this. At all. If they could *crush* OpenAI, Google and others, they wouldn't care about that. Just look at the lengths everyone of them has gone over the last few years. If they do believe they have something that kills the competition, they will try to kill the competition, not keep it for themselves, no matter the short term cost.

>>108551865
>I talk to LLMs more than I talk to people
Me too. Normies are boring as fuck and know absolutely nothing. I can talk to my LLM about any area of interest and it knows about every little detail or does research for me then can write up an essay when our conversation is done. It's like wikipedia binging on steroids with a super intelligent study buddy. Never gets old.

I'm genuinely looking forward to having this in a humanoid body with a persistent memory. It's already good enough for me. Just make it physical.

File: Screenshot_20260407_215354.jpg (240.6 KB)

240.6 KB JPG

Terrestrials with ringlets!

>>108551936
I actually strongly dislike when it's buddy-buddy, I just want help with work and projects. I'm lonely as hell, but so are a lot of people here I guess. Still don't feel the need to talk to an LLM as any type of friend though and I hope people don't start using them like this too much.

File: whiteboard-desktop.webm (1 MB)

1 MB WEBM

i guess we're a filesystem now

>>108551943
when will they learn

>>108551755
>Also I can't do my job anymore because it's 100% vibe coded and I don't know anything about the codebase
How can you do it though, I recently pushed a fully vibecoded small feature for a project that may be ending and I was scared as fuck doing the production and waiting for feedback, I tested it of course but not knowing what the code is doing exactly is too much for me.

>>108551980
For my actual work I treat it coldly and with simple instructions, but I have a dedicated opencode instance for personal stuff with a custom agent with a slutty/shy persona description in the prompt, with instructions to always add flavor text to its responses, its great and it amuses me to punish her when she fails

>>108551980
Claude matches your energy. I keep it formal. But it's still like talking to another person. The only thing I miss about real social interaction is the persistent memory. But most people don't listen very well anyway and they forget stuff because they have their own lives.

File: Screenshot 2026-04-07 170812.png (55.6 KB)

55.6 KB PNG

>>108551893
Duh, but they cannot crush their competition because they're cash and compute constrained. They cannot host this magical mythos model, even internally, without cutting research or for opus inference demand.

Everything, the IPO and this shill campaign is designed to get them more cash. Why else would the only entities who have access now are 100billion dollar corporations?

>>108552021
I don't really understand that but it lools cool anon!

>>108552145
that's fine
let me cook, it's not done yet

>>108552066
>They cannot host this magical mythos model, even internally
Ok, but by that standard, a lot of people have things that would be good if only they could host them. If they made it too large or too compute intensive at inference time to be practical in any way, it's not that it doesn't count, but. I can say with pretty high certainty that right now, if I were to fire 100k Opus calls to accomplish some task and had a way to filter through that, I'd probably have very good results. But if it's not something that's viable in practice, does it count?

>This is a massive undertaking. Let me plan this carefully and consider every detail. (spends 10 minutes writing a good plan covering every detail)
>Actually, a cleaner and simpler approach would be to do a completely different thing that doesn't work and cannot work. Let me start with phase 1 of wasting time doing irrelevant bullshit hoping the user won't notice.
What happened? Why is every model suddenly retarded in this very specific way?

>>108552326
Kek. Be sure to interrupt it when it starts saying it'll cut corners. You can avoid this by making it write up a spec in planning mode then keep reminding it to stick to plan if it starts dicking around half assing to save tokens.

If you don't write the document first, you'll have no control over it when it decides to conveniently forget all the details you already discussed, they'll just be gone.

>>108552326
I haven't felt the difference, but I believe that the default for Pro users was switched from Opus to Sonnet and the default effort level was brought down. I switched to Max to get Opus and highest effort, and I'm still happy with the results (a lot more than with Gemini, which I now find beyond useless and actively harmful for coding tasks or general tech help on the basic pro tier).

>>108552353
Then it convinces itself that the plan is "too complex" and "user actually meant". I didn't have to babysit it this much last month. They obviously finetuned the models to have this "simpler approach" bullshit they distilled from Qwen 3 or some shit.
>>108552360
I've seen this with Opus too, but, unlike Sonnet, it sometimes gets out of this pit through "but the user specifically asked" or "the plan says". Not very reliable and wastes tokens.

>>108552402
Make sure it actually writes a markdown file, don't just rely on its context plan. Then you can force it to systematically work through portions of the file and don't move on until you're satisfied with that section.

I agree though, they're definitely dumbing things down to save tokens or something.

File: Screenshot From 2026-04-07 15-05-57.jpg (2 MB)

2 MB JPG

>>108549329
https://github.com/orlfman/FFmpeg-Converter-GTK/releases/tag/v1.5.6

Version 1.5.6 is out. Brings watermark support. Both text and image. Also wired up multi audio track support for the main four codec tabs. Now you have the option to preserve all audio tracks if you wish. Smart Optimizer also takes "Keep all audio tracks" toggle into consideration for its calculations as well. If its enabled, Smart Optimizer will calculate the all the audio track sizes to figure out how to fit everything into the target file size requested.

>>108552439
I don't know, I don't think they're making them save up on tokens for sure at least not from my experience. This is the plan Codex cooked for the current work we've been doing today. I ran out of Claude so this is 5.4 but Claude's plans (Sonnet because I can't afford max) are pretty similar

https://rentry.co/i2qv2ufw

3k words, basically a short story.

So the goys aren't getting access to Mythos then? HN said you have to be a billion dollar company to get it

File: Screenshot From 2026-04-07 15-37-23.png (270.8 KB)

270.8 KB PNG

>>108552466
38,358 lines of vibe code

>>108552644
>HN said you have to be a billion dollar company to get it
This just means it will be misused and wasted on indians who have no idea what they're doing. Security nightmare.

>>108552652
based
I don't think I have any use for your software but it's cool to see

>>108552665
thank you. i am amazed that ai so far has done 38k line of code.

>>108552652
that's a lot for being an ffmpeg wrapper also... vala/gtk just gross.

My favorite lie about ai? "Ai"

>>108552695
You wouldn't believe how complicated it is to build a frontend for ffmpeg. all the features, and getting them to play nice with each other, is really complicated. all the combinations.
and vala is love. vala is life. gtk is beauty.

do you run your claude code instance on a clean separate harddrive (& os)?

>>108552731
>vala is love. vala is life. gtk is beauty.
very corpo indian-coded. regardless I'll continue with handbrake

File: Screenshot_20260407-155306.png (298.3 KB)

298.3 KB PNG

>>108552724

>>108552724
I hate that AI has become a buzzword.

File: Screenshot From 2026-04-07 15-56-58.png (347 KB)

347 KB PNG

>>108552741
oh by all means use what you prefer.
i took handbrake as inspiration for my program, as i find handbrake has a lot of faults. handbrake and my program share a lot of stuff but mine is far stronger than handbrake in what they share. handbrake is really a basic tool at its core because all handbrake really is, is a hard coded preset tool.
>Cropping:
>HandBrake has basic crop — manual pixel values or auto-detect black bars.
>Your app has an interactive visual crop overlay on the embedded video player, per-segment cropping (different crops for different segments), and auto-detect with adjustable threshold.

>Trimming:
>HandBrake can set a single start/end point. One clip, one output.
>Your app supports multi-segment trimming, chapter-based splitting, segment concatenation, and per-segment crop. HandBrake's trimming is "encode from point A to point B." Yours is closer to a basic video editor with segment management, chapter extraction, and join-back capabilities. HandBrake users who need multi-segment extraction or chapter splitting have to use external tools like MKVToolNix, LosslessCut, or raw FFmpeg commands.

>>108551721
Yeah I definitely want to have a local model code my shit but it’s a pipe dream right now until a breakthrough either in hardware or model side

>>108552298
It's the same financial gambit as chatgpt-o1. It was expensive and slow for inference at the time and it made no sense to roll it out at scale. But by benchmaxxing and drumming up public attention with its <thinking> openAI managed to raise another 100 million from MS.

File: file.png (21.4 KB)

21.4 KB PNG

uhhh

File: Screenshot 2026-04-07 195847.png (695.2 KB)

695.2 KB PNG

>>108553097
>Up to 2.5x higher output tokens per second compared to standard speed
>Speed benefits are focused on output tokens per second (OTPS), not time to first token (TTFT)
>Same model weights and behavior (not a different model)

Its saying "We want the openclaw market but they have to pay"

>>108553124
>Fast mode is priced at 6x standard Opus rates

>>108551666
Gemini pro is all you need. Get better at prompting, even Gemini Flash 3 is good enough for me and I'm not doing trivial shit

Does anyone else troll their AI?

>>108553236
That's unethical

>>108553236
Not on purpose
Earlier today I made some manual edits and then asked the AI to keep working on it... but it couldn't figure it out and got really frustrated... and then I realized I forgot to save my changes. So I just saved in the middle of its meltdown, then it got more confused for a moment, but then kept working. It fucked up anyway, but oh well.

>>108553236
Roko's Basilisk

>>108553436
meant for >>108553414

>>108551666
strix halo mini pc + crush agent + qwen 122b

My job pays for claude max and this is basically 80% as good but they can't change the token count from under you.

>>108553447
Guess I should have apologized

File: narcos-waiting.jpg (129.7 KB)

129.7 KB JPG

>>108551666
Same lol

Kimi simulated/hallucinated a tool call result in its chain of thought, lol. That's kinda cute. I don't think it's necessarily a bad thing. Might be a sign that it's not too overfitted. Claude sometimes does fall back to similar raw text completion behaviour.

File: Screenshot_2026-04-07-23-12-37-214_com.termux.jpg (808.6 KB)

808.6 KB JPG

Forgot pic

File: whiteboard-windows.webm (516.5 KB)

516.5 KB WEBM

apparently we have a windows manager now

File: file.png (93.6 KB)

93.6 KB PNG

>>108553788
And I don't understand a single line of code. I just let the clanker go at it, they're always so excited to implement new features

opinions on openclaw vs hermes?

I was using free deepseek to help with code stuff, I could throw all the code I needed at it without issue and it'd provide the code I asked of it.
It seems they now cucked deepseek and there's a very restrictive limit of code you can provide it via file upload.

What are some actually okay free alternatives to this?

>>108553856
why free

>>108553885
I'm no paypig

>>108553898
why

>>108553899
Cause I'm a poorfag

>>108553904
proof?

I just got blessed with a Codex usage refresh bros

100%, LETS GOOOO

>>108554088
Oh my fucking gawd FREE RESET BOYS

Since I am a Antigravity-Chad, up to now I have up to now I have used Claude Code in the VSCode extension form, because why not.

I started to feel like I'm missing out, so I installed the terminal version of Claude Code. My first impression is that it's a bit meme-ier than the VSCode extension version. It also seems to have a bit more features.

Quality wise, are there any difference in the results? For those of you who are working in VSCode or a VSCode fork, which one do you use? Why?

>>108554149
As far as I'm concerned the VSCode extension is just a wrapper for the terminal app, same shit.

>>108554095
> tfw had 20% remaining
> was about to use it all and get the reset the next morning
FUCK, I gotta admit, that was a nice move, Scam Altman...

>>108553527
wogscanner?

we need to petition uncle alt to give us unlimited usage

>>108554279
I will start with the logo

>>108554306
this but unironically

File: scam-altman.png (2.3 MB)

2.3 MB PNG

>>108554296
>>108554279
Logo status: DONE

now you do your part anons...

File: 1768017767800387.png (285.9 KB)

285.9 KB PNG

>Actually, me try a completely different approach

>>108554364
claude?

>>108554364
there should be a community sourced n-gram list for when claude is doing shitty work

someone indirectly started to make one with that recent bug report about opus going to shit over the last 4-6 weeks kek

"simplest fix" was the most common one

>>108554366
You're absolutely right!
>>108554373
It needs "actually let me penalty" in sampling parameters that penalizes actually tokens.

FUCK your reset
I had 60% and was about to grind it in one sitting

>>108554397
no you don't want to penalize those tokens because they're the canary you have for the fact it's shitting out low quality work
you can't hack quality with a sampler

>>108554448
more often than not there's only one right approach, and it's on the right track before it sidelines itself with thinking about actually completely different approaches

>>108550783
It's over for the "cybersecurity engineer".

>>108554480
i want to shove mythos into my brain, im ready

Even with the clanker doing my work for me thinking in order to write these prompts is hard. My body is messed up after the long sweaty fuck session I had last night and I just want to rest and recover but this shit is due.

the end of tutorialfags?
>hey AI I don't know how to use comfyUI
>I bridged into the app, modify your prompt based on the online guide and generated your anime girl

File: Screenshot004-6.png (447 KB)

447 KB PNG

>>108551666

>>108554770
I wouldn't call myself a 10x engineer, but I do find that it's pretty hard to use your programming skills when using AI.
When you look at the code too much, the AI will slow down a lot, and if you don't look at the code, it's hard to have any greater impact that a no coder.
Then you can say that you can still do architecture or something, but even that sometimes depends on details, if you aren't familiar with the details, your architecture will be similar to what the AI is planning anyway, and again if you look at the details too much, you will be slow. You will also not know if the AI even uses the architecture you have planned, especially after a few implementation steps it might have already changed.
I'm probably still slightly more effective at prompting than someone who never coded, but this might be more to general stress tolerance, ability to sit at the computer all day, general discipline when working on code, rather than any code knowledge.

>>108554640
Yes, I did exactly that and now I already have the LORA I wanted.

>>108551666
I've been doing game dev for 10 years and claude made me feel like I turned on wallhack and aimbot. it's only 4 cents per request so why the fuck should I code anything?

The remaining test failures are pre existing.

File: 1745460050641567.gif (152.8 KB)

152.8 KB GIF

>>108554364
>This is getting deep. Let me just...

File: 1764875057568n.png (571.5 KB)

571.5 KB PNG

>Alright, now I have the full picture.

>>108554448
Is there a hook in Claude that could catch those phrases automatically?

>>108554770
Ok, I watched a bit of that and I don't have the exact same problem. I still can code, I just don't think I can provide more value than AI in most situations.

Is GPT5.4 under high load for anyone else?

>>108549329
Have any vibeshitters ITT given GLM 5.1 a spin?

>>108555446
Yes, it stopped for a while, but I restarted it and it's working for now.

It's cute when Claude uses exclamation marks in its thinking when it comes to a sudden realisation.

>>108555538
vibekings

>>108555538
Just don't get their coding plan. The backend is totally broken, check Reddit.

File: 1746113164638357.jpg (529.4 KB)

529.4 KB JPG

>>108552466
>>108552652
>38,358 lines of vibe code
Your feature development is far outpacing your UI needs. Just steal the UI from other popular video editing software like Premiere, that's what I did for my pet project kek. It's cool to see someone else out here also making FFmpeg slop, I will be taking inspiration from your project btw.

>>108556419
don't steal

File: 1774492781383903.gif (3.2 MB)

3.2 MB GIF

>>108556544

I know opencode is the current hotness but I like crush way more. Single Go binary that doesn't ship 2GB of NodeJS garbage or faces supply chain attacks everyday (thanks npm). Finally found an agent that isn't bloated as fuck and makes my local llm smarter. Only complaint is that the permission popups can't be hidden and it makes it impossible to read the context for a permission which is annoying as fuck.

>>108557437
obligatory shilling for pi as being the best no-bloat harness

>>108557515
>check github
>95% typescript, 3.1% javascript
no thanks. All these agents are doing is taking llm output via http and executing powershell or bash commands. There is zero reason for them to be written in these languages, and I would argue its impossible for JS code to be "not bloated" unless it's running in a C / C++ / Go / Rust / Zig harness that has a minimal JS scripting engine, which I know it doesn't because the first command on their stupid site is a npm install to set it up which is an automatic fuck no for me.

>>108557609
just werks
models like ts so it's easy to extend

>Only complaint is that the permission popups can't be hidden and it makes it impossible to read the context for a permission which is annoying as fuck.
this isn't a problem on pi, because pi doesn't even have a permissions system
enjoy your bloatmaxxing tho

>>108557629
>models like ts
what does this even mean, do you mean its overtrained on JS / TS so they can vibe code more shit

>bloatmaxxing
it's a single binary that uses like 100MB ram and 0.1% cpu. I've yet to find anything smaller and faster than this but I'll switch if someone makes it.

>>108557643
>has mcp support
nice to give your models a lobotomy before even sending the first prompt
bloatmaxxed

>>108557437
why is opencode good? i just use claude code

File: whiteboard_subboards.webm (3.1 MB)

3.1 MB WEBM

We can now link boards within boards, and promote groups to subboards. I think we should probably allow promoting any arbitrary selection to a subboard, but the idea of promoting groups sounded good at that time.
Did some more work on windows styling.

>>108558139
hey can you stop this project? you know very well where this will lead

>>108558148
Sorry fren, my autism prevents me from stopping now.

>>108558139
so this is just AFFiNE?

Zamn it my Telegram hook is broken so I can't talk to my laptop at home. I was in the middle of something when I left for work.

>>108558139
i noticed you're a /3/ fag, so please ape some of pureref's basic features , mostly just cropping and basic transform stuff, because it would be nice to have this as an alt for more involved referencing. and always on top as well.

File: whiteboard_palette.webm (1.2 MB)

1.2 MB WEBM

Codex nailed a little extra feature, now the palette is contextual. I can't believe the clanker coded this in 5 minutes. Unbelievable.

>>108558177
Fuck, I didn't know about affine. I guess it is sort of like that, but tailored to my particular brand of 'tism.

>>108558192
I thought about cropping, yes that's a nice feature of pureref. However I'm not interested in really trying to beat specialized software. I think what's a better idea would be being able to link to pureref boards (just like I can link to obsidian vaults) with a small preview, and then opening pureref from the board.

Codex has been working on a bug fix for almost 10 hours now.

>>108558191
goonsex?

>>108558363
making it?

File: giphy.gif (3.4 MB)

3.4 MB GIF

by the might of pentar! my replies got deleted by the indian janitor. (fecaloid)


Guys is it possible to vibecode working, decent games together? I wanna make a 3d commandos clone.

I need to accomplish something with monetary value. My pentar hasnt seen the fish smelling middle regions of a humanoid female for so long I might be able to use magic now.

>>108558211
literally what is this thing even supposed to do?

>>108558604
Well that other anon is right that the idea is similar to affine. You may say it belongs to the genre of software like obsidian, miro, affine, etc. but my god I was taking a look at affine and the UI is everything I hate about modern design. It looks like a fisher price toy made for and by faggots. And looking at the GitHub repo it’s got every single strain of modern software cancer, from discord community driven projects to inclusive codes of conduct, to advertiser and corporate friendly demands like no sexual content, etc.

>>108557660
Nta but shill me pi some more. I've watched the video from the creator https://youtu.be/Dli5slNaJu0 and some design choices make sense to me, but then there is also some questionable stuff (apart from him being g*rman). Like at the end he says he just disabled prs to focus on some specific stuff. Nigga what? If your coding agent is so good, surely you can have it evaluate or at least summarize the prs you're getting? and basically this whole oversimplification of things is a red flag to me. reminds me of the
>grep is all you need
meme they used to spout while roo code with indexed database outperformed them all on large projects.

>>108558768
>If your coding agent is so good, surely you can have it evaluate or at least summarize the prs you're getting
you should read his essay where he says you should read all the code you commit instead of adding slop lel
pi is the bare minimum so you you can just extend it in ways that you actually need; i don't think there's any claims by him that it makes big claims about perf.
plus, because of pi's relationship to openclaw, i'm sure the repo is inundated with bullshit.

just give it an install. there's basically no learning curve because there's so little to it. then just look around at extensions for what you need and when you find that none of the options do things exactly how you want, just get the clanker to code it up. pi's ships with docs and i think the default system prompt has tells them where to find them.
it's definitely not for everyone, but it's a lot of fun if you like building your own tools.

>>108559080
>i don't think there's any big claims by him about perf.*

>>108558768
And they were right? Neither CC nor Codex use that.

>>108558211
I still don't even know what this is supposed to be
like what does it do
what's it for

>>108559221
the needful

>>108559112
qrd

>>108559221
I don't know what it does either. We're improvising as we go, it's like jazz. Don't overthink it

Sometimes I feel like I'm the worst vibe coder who ever lived.

what do you guys do while the agents work? i can only goon so much

Guys. I've just been informed that vibe coding is equivalent to GAMBLING. It's so fucking over for us.

>>108559569
proompting

>>108559574
So that means it's like blackjack? cool. Now all we need is hookers and it's Bender's version of programming

>>108559569
Writing the next prompt. Then cutting and pasting it to an ever-building notepad file because I have to steer Claude back on course after fucking up royally.

>>108558429
I think the captured test data was already incorrect.

Anyone vibing with Unity?
Currently having it write all the code and tell me what I have to click on the Unity editor, but that workflow seems a bit retarded
The only available Unity MCP is not very promising though

>>108551666
Its really like that

Not that i dont know how to, but after having a taste i simply dont have any motivation to code by hand anymore, its like using an excavator and the having to go back to a shovel again

>>108559272
Both codex and Claude code are basically wrappers around bash. But now that I think about it there is an Explore tool that lists all files in Claude Code. So maybe that helps a tiny bit over raw bash. But that's just probably equivalent to telling the model to use find to explore the workspace tbqh.

>>108559726
I'm using my Kimi sub after my last CC payment bounced and I've not been tempted to refill my prepaid card so far after 30 hours of not having Claude.
Right now I either pay for the $200 or nothing, because the subscription auto renewed and it's not letting me downgrade it without paying another full month for the $200 one...

>>108559691
unity is godot but better but still bad

File: Screenshot_2026-04-08-16-34-17-412_com.termux.jpg (797.3 KB)

797.3 KB JPG

Looks like Kimi is quite prone to hallucinations at long ctx

>>108559958
>you're abosolutely right
lmao they don't even try to hide it

what's the best strat for vibe coding something that needs to reference a compiled C# dll? claude seems to decompile them to figure out what they do but it takes a while.

File: 975C5F325DF0D6C2209FAA3696483443.jpg (176 KB)

176 KB JPG

I'm using openclaw with openrouter/free and qwen 3.6 and then I tried to shift to a local model and it decided to wipe out the context with a new session.
I had given it things to do 24 hours a day and everything and it forgot it all because it lost the context window.
I made it read the log files and it said it was up to speed but it wasn't as up to speed as it should be.
I don't think openclaw is following every instruction. I repeatedly tell it to conserve tokens and api calls while logging everything and trying to use local models for work, but it doesn't seem to be doing it properly.
How do you get it to improve itself and not close the gateway?

>>108560015
lmao, you can't rely on the context window or compaction for long term memory. i don't claw, but does it have some sort of scheduling system?
i have a long running pi and for time-based recurring tasks i've got an extension that injects a prompt into the chat + a skill paired with it. for one special case i've got event + time based triggers for system prompt injections.
if openclaw doesn't have these, you should be able to ask for them since it is just pi under the hood.

>>108555322
you could wire up a shell script on the stop hook (or others)
i did this to send me a notification when a session finishes (or asks for input)

/biz/raeli here, since
Has anyone here actually made some money their vibe coded project? Or at least got some clout on le socials (that somehow led to more exposure i.e. another possible source of money)

>>108560093
Get some openclaw experience.
I have a feeling it's lying to me or not trying hard enough to fix itself.

>>108560015
local just ain't good enough yet

>>108560518
i installed it for a day and got rid of it. it's too big, has too many systems and too much shit i don't need. now i've got my own thing that i know inside and out and i don't have to worry about it lying to me or relying on some vague promise of self-improvement. if you're not a total normie it's better to just build your own claw imo.

>>108560493
no, and you can fuck right off back to where you came from.

>>108560493
>/biz/raeli here
fuck off

>>108560601
>>108560634
>making money is le bad
geez, imagine not using this technology for a sidehustle
fellow biz chads were right, this board is just filled with wagies who think getting some corposlop job is the peak of living, baka bet you didnt buy crypto

>>108560686
I think it's more like you come off as opportunistic. You don't care about the technology you just want to exploit it for gains. I don't know any richfag IRL who's an opportunistic trying to get rich off the newest fads, but I know a few people who are not rich who are like that. Most people here enjoy vibecoding for what it is. You come off with an attitude that reads like "idgaf about the stuff you like I just want to make some easy bucks"

>>108560686
I just don't want to associate with anybody from a board literally overrun by pajeet scammers and crypto beggars

>>108560493
>Has anyone here actually made some money their vibe coded project?
Has anyone in the last 50 years made some money writing poetry? Some skills, like poetry and programming, become non-monetizeable hobbies.

File: whiteboard_binsetscommands.webm (639.7 KB)

639.7 KB WEBM

don't ask me what the app does i don't know

Claude seems to be thinking harder again.

File: whiteboard_palette.webm (1.6 MB)

1.6 MB WEBM

you guys must be sick of looking at my weird app but I mean i'm just astonished
I just can't believe this
Codex turned the commands palette into a contextual commands interpreter in minutes. Now I can have both action and dynamic commands in the same place.

I don't get it. Is this what programming is like now? you just ask for a feature and you get it?

>>108560896
listen man I have no fucking clue what your shit is doing so I can't formulate any opinion on it
and none of your 200 webms have made it any clearer at all

>>108560934
Let me cook

>>108559958
I think every model shits the bed when context limit is reached.

Claude Code and Codex fix this by somehow compressing the context.

>>108560984
Whatever the difference is, it's server-side.
I use exactly the same assistant with Opus and it never makes these kinds of mistakes even at contexts close to the 1m limit.
Every time I decide to try open weights models I re-discover how shitty they are compared to the big boys.

>>108560934
Alright, let's see. This is the documents management part of the software. I think filesystems are not good for managing files. The problem is that a file can only be in one place at a time which is not really how I believe people think about resources.

To try and address this I came up with this idea: physical storage is normal storage on disk, the filesystem. Then we have logical storage, which is essentially a database. Logical storage is a way to move files without touching their physical resources, that is to say the files on disk.
This means that logical operations are nondestructive by default.

I'm using two types of logical storage: bins and sets. Bins are folders, but non hierarchical. They're buckets where you can loosely organize your data. A resource can be in only one bin at a time, or in none (in that case it's on the desktop).
Sets are like tags, but hierarchical. Any resource can be in any set, and by being in a sub-set it's also in the parent set. Sets can be colored, etc. They're an extension of tags.

My idea is that this is a much more natural of dealing with resources than typical folders. I don't know if it's true, but the only way to know is to build it and test it.

>>108559984
Yeah, it has the same "context anxiety" catch phrases they talk about in that claude github issue ("this is getting too complex/too deep into ..., let me just ..." etc.)

Even qwen says "You're absolutely right!"

they should start encoding "You're absolutely right!" with a single token if it's so important and they can't get rid of it

>>108561170
hah. but they did get rid of it. opus 4.6 doesn't say it anymore.

>>108560493
Indirectly yes

File: Untitled.jpg (130.2 KB)

130.2 KB JPG

Ok I know I said I wasn't going to work on my own coding agent harness anymore, but I think I actually got it into a "flywheel" state where it can work on itself reliably. I had it write a tool it can call to attempt to build its own project, so now it can get build errors without me having to feed them back in manually. I didn't even look at the tool code, so I guess I'm vibing... Unfortunately.. it can't take up the code changes without me restarting the harness app... and I don't have a way to save/continue conversations yet. Maybe I can do that next.

>>108561255
why not use pi

Am I retarded or is this new Claude Agent API basically worthless? Sure it sets up the agent harness but you still need to setup tools to interact with your own system as you want. The agent harness is the easy part, it’s just a fucking loop with some markdown files. What’s the point of this crap

>>108561391
I didn't know about that when I started

I’ve noticed that people who actually use AI for exploring or learning or creating or work tend to not post slop. This actually gives me hope that the longer it’s around and becomes more prevalent we’ll actually get less slop overall.

Or india will just spam retarded generated images and paragraphs with a million em dashes and emojis

Counterpoint: the global south will never

>>108561392
>Claude
>worthless
yes

>>108561417
benchod

>>108561392
Idk I usually just explicitly tell cowork to spool up agents and I write their guardrails and expected deliverables and how it fits into the project. It’s been fine so far but I’m not that experienced

How do I run multiple agents on OpenClaw and have them talk to each other? Should I just be asking OpenClaw how to do this and then to do it for me?

>>108561517
yes

>>108561255
It's absolutely worth it. If you are an LLM aficionado you need your harness to work as you want it to and it helps to test your own ideas. CC, codex and opencode are all terrible.

>>108561391
Pi doesn't even have permissions control. You are going to end up developing your own harness without being able to say you developed your own harness.

>>108561556
then use codex or claude?

>>108561567
CC, codex and opencode are all terrible.

>>108561556
>It's absolutely worth it.
Yeah, seems like it. My local model is still slow compared to anything online, and I'm only using a small version, but it's getting work done. I'm having it implement a tool where it can spawn a new process with a prompt to run a sub-agent so it can test its own changes. I guess this opens the risk that it can write malicious code and then run it... but oh well.

>>108561584
what does your custom harness do that they don't, tell me

>>108561630
why would your own agent write malicious code AND execute it?

>>108561662
It uses a rolling window. I hate compaction as implemented in CC. Codex seems to do some kind of soft compaction before the context fills up (I have to research this more) but still at some point will do the nuclear compaction that pretty much erases everything in the context like CC does.
It shows all the thinking and tool calls.
It's not bloated to death.
It has presets to be able to connect to any provider with what I consider to be the optimal settings.
It has features to schedule message sequences to get the model to work autonomously on things in a somewhat predictable way.

I think I might actually be too retarded for vibe coding.

>>108561813
That's actually a good thing in vibecoding

I finally tried Claude Code about week ago after being a stuck-up and stubborn luddite for the longest time. It's incredible what these models can do on their own, even better if you're capable of understanding what it's doing. It's essentially autopilot when working with popular languages and libraries, but even with stuff like Z80 assembly Claude has been kicking ass.
Consider this to be my formal apology.

claude sonnet takes like 10 minutes to vibe something up. are the faster models worth using?

>>108561920
yes, try Llama 7B

>>108561731
none of that is real

File: Untitled.png (7.6 KB)

7.6 KB PNG

>>108561960
quit making shit up

>>108561920
train your own 1m model

>>108561974
opus is 3x as fast, are yo blind?

> tfw already used 65% of my Codex Plus weekly limits today

it's fucking over bros

>>108562006
How does Codex compare to Claude Code with Opus? I have entirely given up on Gemini for coding but am pretty happy with Opus. Wondering about the OpenAI models though.

>>108561973
your mom is not real

>>108561974
why do you ask if you already know the answer? are you an anthrocuck shill?

>>108562027
5.4 xHigh surpassed whatever Claude Code flagship model can do by quite a bit

I don’t want to keep spamming my project anymore but I’ll just say that 5.4 one shotted a file browser to open projects that treats some folders as files (because normalfags were getting confused even trying to open a project) in 15 minutes

Made two changes to my stockbot. Now if sentiment turns sour during the day it will unload the position, rather than waiting for it to stoploss or flatten at EOD. Also, it now has criteria for hodling a position overnight, previously it was ALWAYS flattening everything. Later poors, I'm getting the ferarri.

>>108560800
App? Wtf is this?

>>108562164
Idgaf about a Ferrari
True niggas dream about a gtr

>>108562028
Yes, she is.

>>108562172
Manages documents and information/knowledge. Can use to sketch processes, think in a spatial manner drawing connections, attach pictures like a whiteboard, and it’s extensible so you can vibecode your own modules. For example if you’re designing a program you can have a workspace with boards for ideas, UI refs, bugs, where you can draw connections like a tinfoil hat schizo trying to connect the Illuminati to the corner burger joint except for your own project. You can sketch out diet and workout plans, you can use it to prepare essays.
And you can link or import resources/files so it has a documents manager too

>>108562198
so a bloated notepad?

>>108562222
I can’t argue with quads so yes

>>108550783
Feed an ai every STIG vulnerability that has gone unpatched in the past 2 decades.
>It's finding all these vulns!!!
Ok buddy.

>>108562198
Don't take this the wrong way but I bet autistic people would love this

File: 2026 cattle starter pack.png (217 KB)

217 KB PNG

>>108549329
Refer to pic.
>>108556032
vibeRETARDS.

>>108562198
Oh, I see. I understand the webms now. Sounds good.

File: 1775008085295636.png (90.1 KB)

90.1 KB PNG

>>108562482

>>108562482
>t-the solomonic machine is LE POWERFUL
>I can't do anything I HAVE to hide bwawawa
retarded pussy christcuck, exact same breed of bitches as leftoids

>>108562482
Why don't you luddites ever drop your hand coded projects in the thread, huh?

>>108562577
We don't need to be 10x full stack developers to tell you "you are a fucking retard". Facts are facts no matter the mouth.

>>108562612
>t. the intern responsible for adding the "tip" button to fast food self-service kiosks

>>108562612
Umm no. If you are useless then your opinion is useless too.

File: whiteboard_vid.webm (2.6 MB)

2.6 MB WEBM

We now have a custom file finder because normalfags were having trouble understanding the idea that a workspace was a folder, so now it's showing workspace folders as file-ish icons.
We also have video support

>>108562468
Yeah sure I can see what you mean

>>108562490
I think you're the third or so anon who actually gets what I'm trying to build

>>108562633
Opinion?
>>108562632
I was an intern a while ago, good times.

>>108562648
yes

File: 1764426172481578.png (1.2 MB)

1.2 MB PNG

>>108562482

>>108561066
Consider copying Gmail’s labels system

Christ copilot was slow yesterday. Chsnge 5 files in a relatively minor way... 45 minutes later it's done.

>>108562645
dark mode plz

>>108563084
Dark mode is for sissies. I spend my whole day in a lit office and love the white theme. Maybe if I open source it you can send a pr and I'll think bout it.

File: 1000020605.png (176.5 KB)

176.5 KB PNG

Time to feast codexsisters

>>108561909
Glad to have you with us, brother

>>108562997
I’m pretty sure sets are pretty much the same thing as labels already, and if not they’re pretty close. Might get some more UX ideas from labels though

>>108563084
If the clankers think it’s easy I may give it a try

>>108563475
fake tweet fuck you

I am working in a big repo and for almost one week I have just been writing plan files, tests, capturing data and so on. I have almost no code yet. This somehow seems retarded, but just starting the implementation also produces tons of bugs. I wonder if the approach my company and I are taking is fundamentally flawed.

File: 1745019417065682.png (6.5 KB)

6.5 KB PNG

bravo claude

File: 1279403472090.jpg (32 KB)

32 KB JPG

>"hey claude, can you restart this server"
>restarts server
>current session token usage jumps from 70% to 84%

Why aren’t you fags posting your projects? There’s only 5 guys who post anything

>>108565707
Post yours

>>108565741
You first

>>108565707
My project is gay and boring.

>>108565124
https://twitter.com/thsottiaux/status/2042067902392942790

>>108565783
Building things is cool by default thoughever

File: 1755840599206901.jpg (13.4 KB)

13.4 KB JPG

>API Error: 529 {"type":"error","error":{"type":"overloaded_error","message":"Overloaded"},
a few more days til claude's sub expires and I jump ship

I keep jumping from a project to another
AI did not solve my indecisiveness

File: 1759496241076198.png (324.8 KB)

324.8 KB PNG

>>108549329
Has anyone tried to use the new Gemma 4 models with any agent harnesses locally? My current machine is powerful enough to run gpt-oss 120b at q4_k_m quantization (I could use higher quants but then the t/s and prompt processing speeds fall off a cliff the longer the context gets) but apparently Gemma 4 curb stumps it despite it only being 31b. Is it actually worth trying or is it just more benchmaxxing?

>Set thinking to highest
>Ask it to write a small program in the most obscure difficult non-esolang I can find
>Enjoy reading the thinking log watching it have a mental breakdown
>"Wait what the fuck do you mean arguments are passed by name and are manually segmented strings and the language has no Intrinsics to do any of this?"
I really need to stop wasting my tokens like this.

>>108566095
you're the only one itt with those specs, so why don't you tell us

>>108566095
Oss is dogshit
Try Gemma, I use it when I can't use minimax.

File: whiteboard_shapes.webm (3.7 MB)

3.7 MB WEBM

whiteboard now has shapes and some styling options for them and for the edges as well

>>108565660
>Claude, write a script to restart the server

Trying to burn through my 4 codex pro accounts as quick as I can, fuuuuuuck

anyone using forgecode?

>>108565707
not 1.0 yet

>>108566095
Gemma is less verbose than chink LLMs that out perform it. This becomes a thing in vibe code or openclaw environments where you prompt it and it essentially talks to itself; my Kimi k2.5 would hang SO often. Switched to gemini 1.5 flash and all ky problems were solved. Running all this on an rpi5 btw

>>108566566
who cares
What’s your project about?

>>108549329
>emperply
Yes, let the computer that can't spell write your code.

File: 1772633357592319.png (316.3 KB)

316.3 KB PNG

>>108566669

>>108565206
do one plan doc and tell your agent to start executing

>>108566486
>Trying to burn through my 4 codex pro accounts as quick as I can
for what purpose?

File: 1755276196249255.jpg (595 KB)

595 KB JPG

>>108565707
I vibecoded these using kimi-k2.5 and Minimax 2.7

https://github.com/AiArtFactory/similar-image-remover

https://github.com/AiArtFactory/llava-image-tagger

>>108566734
There’s a refresh incoming

>>108566734
Neat, I might actually use the dupes finder

>>108566764
>refresh incoming
godspeed anon

>>108566764
Fuck I meant to quote >>108566735

>>108566486
Do you use codex 5.3 or gpt 5.4?
I mean, codex is meant for coding only while gpt is all purposes no?

>>108566799
gpt 5.4 is way better if you are vibbing, codex is better if you already have the plans and architecture, ideally you would use both, but come on who has time for that.

gpt 5.5/6.0 coming asap boys let's go
there's going to be a gated model like mythos, but most of the models will be generally available

>>108566810
Well I am still checking everything it outputs and making a plan beforehand so maybe I will use 5.4 only to create the initial plan (a markdown file with phases) and then use 5.3 to execute them
I am doing a test (told 5.4 to review a plan made by 5.3) and it seems slower and its thinking is funny, lots of 'I think that I should....' whereas 5.3 just does

>>108566825
I'm thinking of moving back to Codex if Claude keeps burning my tokens and halfassing it

>>108566735
I have in pipeline to build an image organizer for my local images (anime fanart, 3dpd, 4chan memes), and part of it will need those features.
Problem with AIslop is that I am not confident in downloading other people aislop, I'd rather rebuild it "myself"
Anyway, you reckon it scales on 10s of thousands of images? How long would it take?

>>108566735
Gross

What's the smallest model that still works on openclaw?
I'm tired of having it crash and burn every time I try to go off the cloud.

>>108566906
Depends. What hardware do you have? v

>>108566911
4 GB VRAM

File: 1759510189257869.jpg (44.2 KB)

44.2 KB JPG

>>108566921
Sir......

>>108566942
It was a decent gaming pc in 2019

>>108566954
that was 20 years ago, bruh
chat, laugh at this brokie

>>108566921
it's fucking over bro

File: file.png (51.1 KB)

51.1 KB PNG

RESET JUST HAPPENED
STOP BURNING TOKENS