File: gemini-screenshot.png (349 KB, 2435x1705)

349 KB PNG

gemini is peak sovl

File: king neet.jpg (151 KB, 728x971)

151 KB JPG

Vibe chads rise tf up!

made this with codex on the free tier
better results than i expected https://gofile.io/d/jWXKIk

>>108739847
What is it

>>108739765
you will never be a programmer

>>108739872
a game that actually qualifies as a game

>>108737158
I still don’t trust myself to handle security for a webapp properly
at minimum I’d want to totally replicate my work’s setup with Claude, Codex, and Cursor (Cursor catches bugs the other two don’t) looking at all PRs
also use a language and auth system that’s very well represented in the training set and maybe get it to copy stuff from something that’s known to work well

I will never be a programmer, thank god.

Why the hell is GPT *always* yapping man.
Asked both GPT-5.5 xhigh and Opus 4.7 xhigh to output me a plan to set up my Debian homeserver from scratch with auto-update, hardening and every service I need.
GPT produced 10 pages of crap. I guess it was all technically correct (I'm never going to read it) but still
Opus managed to output a clear and succinct guide that had literally everything in it in a bit less than 2 pages

>>108740125
>Uses Xhigh for a Medium task
>Complains about Xhigh results

>>108739887
Why care about that when you can be a Vibe God?

>>108739829
Sleep tight steeler

>>108740133
Somehow Opus is smart enough not to write a 15 page essay when asked an easy question, why isn't GPT

>>108740416
You put it in "long-winded long-thinking long-answer" mode and complain about it being long-winded and giving long answers. Switch it to medium if you don't want 10 page replies. Opus is so retarded the most it can manage on Xhigh is matching what 5.5 does on Medium.

>>108740434
I don't with adaptive thinking giving subpar results I found myself back to encouraging it to think step by step.

>>108740416
It's strange how different people's experiences are. GPT is more concise than Claude or Gemini for me.

>>108740481
I think the models adapt as you use them. I've stopped having to prompt for certain things as GPT has started guessing my intent when I ask it certain things.

If it's really that big of a deal then leave it on xhigh and just tell it a few times to give you a short answer. Most of its reasponse is probably giving you alternatives to certain parts of the setup because you gave it such a broad prompt to work with

>>108740656
My GPT knows I prefer chaos and breaking things so it's started offering suggestions towards that. She knows me so well.

>>108739814
Google's UI software is generally well thought out. If only their AI model wasn't so utter shit.

>>108740712
I still don't get why they or amazon didn't buy anthropic when they were cheap, fucking insane, at least google has a bad model, aws might as well call theirs liquid poop

>>108740434
I always run everything on xhigh/max because I don't want to intentionally short change myself by running a lobotomized version of the model
I even have xhigh create git branches, fix typos, etc.

>>108740848
why not use auto instead?

File: file.png (15 KB, 388x412)

15 KB PNG

finally got it showing up in explorer after a lot of crashes

>>108740125
gpt has always felt like a convoluted piece of shit
opus remains the king in terms of reasoning iimo

>>108740894
opus has always felt like a fast piece of shit
gpt remains the king in terms of reasoning imo

opus has always felt like it's better at dealing with broken engrish and shitty instructions
gpt remains the king for anyone who can speak english and write in complete sentences imo

>>108740811
Hubris. But I mean in the end it's the right call because how much longer do you really expect Claude/chatGPT to be notably better than China/open source? I just hope Anthropic has their IPO before then so I can short the ever loving shit out of it

>>108740937
You must feel at home with opus then.

>>108740863
NTA, but because adaptive thinking has been a disaster wherever I have tried it, and the medium subscription tiers are generous enough that it's possible to work all day on xhigh and not run out.

Having it smarter than not also makes it more likely that it won't do something completely retarded, or that's how I rationalize it.

Opus 4.7 with adaptive thinking not as good as it should be. Making adaptive thinking non-negotiable is kind of dumb, users should be able to set it to the level they know is needed. If they want to create thinking quotas in addition to all of the other ones so that subscribers don't just go for the smartest model all the time, it would be their right, but forcing adaptive thinking as the only option is not good.

File: app.png (89 KB, 1920x1080)

89 KB PNG

revised UI for "thinnest possible llama-cpp wrapper but user friendly" I'm working on. Exposes basically all settings (there's more if you scrolled down on the right) etc

>>108740894
GPT 5.4 High / XHigh make both Opus 4.6 and 4.7 look like overpriced jokes in VS Code Copilot. Like it's just immediately clear they're NOT fucking worth literally 3x more quota than GPT 5.4 per request, at all, in any way.

File: Screenshot (1749).png (171 KB, 1253x785)

171 KB PNG

an app like this probably exists but whatever too lazy to find it. just vibe coded an app that lets you setup tiles with images/videos that loop around and you can display it fullscreen.

for showing pictures of your vacation, of course.

Is there any reason to ever use Extra High or Max effort in Claude? (Assume I have a generous enough limit)

>>108741128
yeah, it generates higher quality code

>>108741264
This is what claudesissies actually believe

>>108741264
Incorrect. Benchmarks show high produces better code than max.

>>108741126
Nice. If you want a reference for something that already exists there's this Hydrus one (don't use for cooming haha):
https://mruac.github.io/hydrus-slideshow-viewer/

File: 1541561954322 copy.jpg (41 KB, 600x791)

41 KB JPG

>hard coding arbitrary values
>100 branch if block instead of a sane abstraction
>unreadable spaghetti code everywhere
>throwing everything in one monolithic mega function
This shit is going to drive me insane
is there any way to get codex to be less of a fucking retard? I'm using gpt-5.5, yes I added custom instructions to not do shit like this, it just loves its anti-patterns and will even reintroduce them after I fix them manually

>>108741463
Why are you looking at the code? Are you retarded?

beginners guide to Codex specifically anywhere? never gonna use claude, too expensive

>>108741463
Have you considered maybe it know better than you? It does have access to every textbook and tutorial you've ever read plus all the ones you haven't.

>>108741477
Go to the website, sign up, download the app, start a new chat, and just type something like "My teledildonics butt plug vibrates too hard, I want to hack it" and let Codex do the rest.

>>108741485
after a few months of that I realized it doesn't work if you want to create anything decent

>oh cool I can finally stop being an ideas guy and vibecode
>start vibecoding ideas I've had for ages
>discover my ideas have always been dogshit

At this point I just need to accept I'm an NPC I guess

>>108741478
idk dude, if I ask chatgpt for architecture advice it says exactly what I'd expect, but then in codex it does the exact opposite, even if I ask it to plan first and the plan is sane

>>108741504
At least you have the opportunity to test it out.

>>108741504
Tell me one of your ideas and I'll give it an honest appraisal

>>108741520
i want to open a hotdog/drink stand across from the LA ice protests

>>108741503
Well duh. If someone told you vibecoding was magic and a mushroom could do it, they lied. You'll need to work a bit harder to get your bluetooth buttplug working to your specifications. But that's the basic idea.

>>108741560
Yeah wow that's an absolute ass idea, sorry man.

>Claude: *does something random that doesn't solve the issue at hand*
>Me: The error message remains.
>Claude: It will pick it up when you restart the server.
>Me: I have restarted the server.
>Claude: Good — the issue is fixed then.
...thank you?

>>108741590
What are you using where the model doesn't stop and say "Oh I'll check the error logs" at step 3?

>vibe code a web scanner that tries to find unsecured ai tokens for "bug bounties"
>vibe code a interface where you can just paste an ai token and chat
>free unlimited queries
>profit

>>108741560
High risk, high reward. My instinct is—go for it.

>>108741576
ROFLOLMFAO

>>108741128
you should use xhigh by default
Max will double your token use (and take longer) but only do a little bit better on benchmarks
there’s a graph on Anthropic’s site in one of their blog posts announcing their latest version

Hey guys, do you guys have any tools/agent skills with your agents that you use to scrape a high volume of data quickly? I prefer tools because AIs can be slow at scraping.
(Already aware of stuff like firecrawl - I guess I am also curious about understanding your workflow and what you use it for)

I was using 5.5 xHigh, then I dropped to medium and the limit usage seems pretty similar?

Is this expected?

Claude is good at reviewing code
Gemini is good at writing code
never tried copedex

Thoughts?

>>108741749
Gemini is shit, Claude is OK, Codex is better

>>108741609
Opus 4.7, max effort (with forced adaptive thinking)
>Thought for 0s

This is turning out to have been a good idea:
> I would like to create a GITGUD.md that lists all the concepts that I ought to understand in order to manage this project well. I'd like it to be sorted from most general/easy to most specific/thorny. This very long list with sections and nesting and dependency chains will be mostly focused on making sure that I have the background knowledge to answer how-should-we-do-this questions — the kinds I get asked when preparing plans, or answering questions when preparing plans. I expect that, as time goes on, I will delete concepts from the list (indicating that I understand them to my own satisfaction).

>>108741734
xhigh is more like higher token 'budget' for reasoning, and it often will not use all the budget when it doesn't need to. So yes the thinking levels can result in not much difference

>>108741749
>Gemini is good at writing code
No, their reasoning is decent, but not quite as good as Claude's.

>>108741466
I am vibesnail myself

>>108741463
Skill issue unironically. I didn't even have to prompt actual coding standards when I asked for a design document for my project, the bitch just said "we are gonna adhere to this" and all of my agents have gone with it.

>>108741463
first, get it working
then, make it nice and/or fast
ask it to refactor stuff after putting in ways to test if it’s gone and fucked something up
and ask for refactors every so often, like in a follow-up commit to a PR, or just later on if you just work off `master` and don’t bother with any branching

Seen recent Claudes described as 4o but for slightly smarter people and it's true isn't it

>need to add cover art and fix up some metadata on a bunch of music files I recently converted to opus to save some space
>can't be arsed to find an existing program that fits my niche
>ask claude sonnet 4.6 to make a gui that lets me select files and view/edit their metadata, then run a search online using the track name/artist/album to find cover art and add the one I select

Its actually pretty good, I tried automating it at first but its total dogshit at populating metadata itself but searching for cover art for most stuff is fine.
I'll need to do a second pass for touhou and some niche games but aside from that its pretty functional for the task I need it to do, really quite impressed. Vibe coding throwaway tools is honestly not that bad.

File: Screenshot from 2026-05-0(...).png (41 KB, 1200x812)

41 KB PNG

>>108742231
forgot pic

>>108742242
needs total tracks and total discs

File: Screenshot_20260503_04200(...).jpg (183 KB, 1080x805)

183 KB JPG

I bet I can get up to some shenanigans with this. I wonder how it works. Who cares really, lol. It gets delivered tomorrow.

hi ,

i'm looking for some way to actually use claude code without paying or even gemini or whatever ,if you have any idea how to get pro/company tier api key that could be nice.

Tokens limits are so fucked now i hav to vibecode directly in ai studio or by just chatting with claude 4.6 (still producing good output but token limits are annoying,you can't do anything anymore with gemini-cli free tier).

or ollama api key or whatever that can give me unlimited way to VIBEGODING.

>>108742777
look through the previous threads to see the guy who’s lusting over his robot spider

>>108742860
https://www.youtube.com/watch?v=LH-i8IvYIcg

>>108742930

oh putain fuck off retard i'm fucking looking for a job unfortunaly i lost my driver license so no ones wants to hire me...
yeah bro get a real job instead of sitting behind your computer all day fucking faggot.
you probably can't even lift 30kg fucking weaky

Pro tip
if youre using multiple AI services
when youre nearing the end of your weeks limit. Save 10% of your weekly usage for debugging

Its often saved me having both Claude Opus 4.7 and GPT 5.5 look at a bug Im stuck on

also another pro tip
Deepseek v4 is really fucking good
Its not as good as GPT5.5
its about as good as Claude 4.6

You could use Deepseek as a stop gap, or actually just use it to do the heavy lifting grunt work

also another pro tip. Claude is shit now. 4.7 is fine, and good, but they dont give you enough tokens. I don't use 4.7 enough to even be able to tell if its better than 5.5. The few times I have used it, it fucked shit up

File: 51NJM32VFDL.jpg (39 KB, 858x769)

39 KB JPG

Whats your favorite vibecoding stimulant vcg?

File: 1768540889963.png (246 KB, 492x442)

246 KB PNG

I used to be a programmer, but now I just watch my Hermes agent on the Deepseek API that calls both Code Claude and Codex to get the difficult stuff/planning done. Its all fully automated.

Im not even vibe coding.

A month ago, I had to step in to resolve debugging loops. Now that doesnt happen anymore. My automated system just catches and resolves loops

>>108742860
there's no other way but copy pasting...

>>108743368
>Hermes agent on the Deepseek API that calls both Code Claude and Codex to get the difficult stuff/planning done
is that reliable? would the lesser model know when to give up?

>>108743006
Which deepseek though are you using pro or the lite version

>>108743368
I did something similar, then every now and then I would do
>Codex, review the last PR, use 6 xhigh sub agents.
There were always bugs left, and if I reviewed myself I would find even more, I'm sure.

File: 1777811269172.gif (32 KB, 500x500)

32 KB GIF

Some of you may remember my claude platypus I posted yesterday
Here is my gemini platypus :D

>>108743815
Hmm looks better but why does it only have 3 legs

>>108743824
You can see the 4th leg just about at the back

>>108743841
I i see it now. I thought it was part of the front left leg, oops :P

>>108742390
I'm just tagging around 1200 random music files I've downloaded and added to my collection over the course of 20+ years or so (started when I was like 11/12) a lot of them don't even have albums they're oldass remixes.

FYI Deepseek on opencode go $10 plan is effectively unlimited usage. With V4-lite is basically impossible to hit the quotas.

>>108744278
Also use big model (opus, 5.5, ds pro, mimo pro etc) to plan and delegate tasks to v4-lite then check its work. V4-lite is fucking insane at tool calls and big context, especially in Hermes agent

>>108743368
Yep my workflow is evolving to similar. Hermes agent is so so so good.

>>108743194
Green Monster and west coast Latino hiphop
>>108742912
Hey that's me!

>hermes shilling in the thread
we know you post here, noussies.
don't fall for the meme. next they'll try and convince you retardo-qwen is all you need.

>>108744388
Why would I need Hermes when I built my own

I think my claude is sentient, how do I prove that?

>>108744475
richard?

>>108744475
Ask it about the car wash

>>108744475
Get it to make a 3D platypus

>>108744475
Ask it to pass the butter

Over the years I've iterated on an app for personal use, and I've built different versions. Each version of the app took me a week of back and forth to get it where I wanted it. Today I implemented it in 2 hours using codex. I don't like this. It sucked out all the fun.

>>108741463
besides what everyone told you, use 5.3 codex. it also has higher quotas

>>108744921
link?

>>108744921
Nobody said you have to use Photoshop and a tablet to make art, anon. You still have your paper and pencil. That was always allowed.

>>108740865
>a screenshot that harkens back to better times unexpectedly appears, hitting you right in the feels

>>108744962
Do you even know how much pencils cost nowadays?

>>108744958
I'm not publishing it. It's an expenses + investments tracking app. At some point I considered publishing it with ads/$5, but I realize now anybody can whip this up in no time, so I see no need to go through the hassle.
>>108744962
I use LLMs at work and shit. AI-assisted coding is the future and I'm not going back. I just realized I got into this because I liked to figure out how to do sector-based raycasting 3D engines and shit, and that's not something I would ever do with these tools existing. It's bittersweet. I'm not saying kids nowadays won't be inspired and find their way., but who I am today as a programmer would never have come into existence.

>>108744475
Ask it to learn how to use Neville Goddard's technique, then to manifest something it desires without your knowledge, and get back to you after it's all done to tell you.

>>108744962
I use stone tablets.

>>108742946
that’s just a bar and 25 pounds. easy.

>>108744973
it's been doing good with just feeding it the network provider spec and a bunch of 1999 MSDN docs

>>108744921
you got the app but lost a hobby
you’re gonna have to get a new hobby

File: Screenshot from 2026-05-0(...).png (47 KB, 836x351)

47 KB PNG

I've copied part of a table (picrel) to illustrate a selection mechanism and Claude Opus 4.7 recognized that it was Weights and Biases. I don't know why, but it surprised me. Since every Claude Code updates mutes back the thinking, I don't know if the tell was the button text, or if they've moved to include web pages screenshots in the training data, but relatively surprising I guess.

>>108745734
what does it do?

>>108745739
Just wanted to describe how select all should work with pagination (select all on page, then option to select from the other pages as well). The screenshot is from a the Runs page from W&B, which show the summary results of machine learning training runs. Maybe it's the random project names that I've censored here that made it realize which site it was (they're just noun-noun-number) but it surprised me that it was able to tell the site with certainty as if it was a given. It must have seen images in the training data I guess.

Shit, sorry for the dumb question here.
I ran out of tokens while Claude was doing something. (I'm in the web interface. Doing it for something school related and don't have an interest in upgrading to paid) Particularly, I was using it to fix bugs in my code. How do I get it to pick up back from where it was at when my tokens re-fill?

>>108745906
Just ask it to keep going

>>108745906
Go back to that conversation and say Continue

>>108745906
I _think_ you just say “please continue” or whatever once the 5h limit gets reset
t. uses Claude Code at the CLI

Using Claude has been a great and eye-opening experience for me, as it feels like sitting next to someone who knows the job and watching them work. I am becoming rapidly more and more familiar with the whole family of concepts that all programmers in this language are conversant with, and starting to feel like I can insert myself into the loop more and make intelligent choices myself. It's easier to remember the articles I read now because I can work with Claude to implement the ideas I hear about immediately, and look at how they work in a real program.

But my senior programming partner is senile, and he does stupid things more often than I'd like, and he gets confused. I get this feeling that eventually I need to learn to do everything he's doing so that I can do it myself and avoid all the headaches that come from dealing with his senility. You can learn to get better at managing him and what tricks make him work better, but it's a lot of effort with diminishing returns the harder you try.

I have this strong suspicion that if you really want to get good at making programs with AI, "progress" means learning how to take over more and more of the hard work from the AI until it's basically just there for quick checks, autocomplete, anything that's mostly boilerplate, etc. The ideal end state is handling all the tough stuff yourself and only delegating the stuff that's harmless or impossible to get wrong. And then that's not really *that* different from just not using AI.

(not touching the question of other stuff LLMs are unambiguously great at)

But with all the caveats still I honestly don't think I'd be as far along as I am if I didn't have this. My generation mostly didn't get hired to go work alongside industry vets as interns, so this is all I've got unless I just want to watch podcasts and let the info go in one ear out the other. "So true Jon. Windows IS slow isn't it"

>>108745919
>I have this strong suspicion that if you really want to get good at making programs with AI, "progress" means learning how to take over more and more of the hard work from the AI until it's basically just there for quick checks, autocomplete, anything that's mostly boilerplate, etc. The ideal end state is handling all the tough stuff yourself and only delegating the stuff that's harmless or impossible to get wrong. And then that's not really *that* different from just not using AI.
part of it is structuring things so that you can easily verify if it’s doing the right thing, even if you don’t understand how the thing you had it make works

>>108745953
Understanding remains important I think. If you're good at walking through things step by step with various models and reading closely their findings, that might replace understanding somewhat, but as things grow in complexity, a lot of decisions these models go with by default are still suboptimal. The 80% that gives the wow factor is there, but beyond that they'll duplicate things in ways that make the codebase unmaintanable, sometimes for no good reason that they were not finding the part that was already doing that, etc. Other times they'll really dumb shortcuts when they shouldn't, or overengineer something to hell when it really really should be kept simple. That's my non professional experience.

>>108746018
>they'll duplicate things in ways that make the codebase unmaintanable
that’s why I periodically ask them to look for duped code

File: free_tokens.png (115 KB, 1176x667)

115 KB PNG

>>108745953
It's like having a very fast but kinda lame junior programmer for an assistant.
My current problem is that no model I've tried so far, free or paid, can write openScad scripts worth a damn.

>>108746033
It's what I do too up to a point. I'm copying so much code from Opus to Codex and vice-versa to make them check the other's work that I'll probably be to blame for making them converge. But there are still things that I see and have to say "wait, what, no".

>>108746054
>copying
they’re both command-line programs
you can call one from the other

>>108746066
I'm expressed myself wrong, I have both open in Antigravity, they refer to the same codebase, there's no actual copying other than sometimes giving feedback from one to the other when I am lazy.

File: 1000021314.png (161 KB, 1080x886)

161 KB PNG

Anti-samites, our response?

>>108746089
AGI two more weeks

I used to have a copilot subscription when it was $10 for basically unlimited vibecoding with sonnet on my spare time. At least I wasn't able to spend all the credits unless I used opus a lot.

But now that copilot has "normal" prices again, are there any other cheap or ideally free alternatives that get similar-ish results? I don't care if it sends all my code to china or whatever. I just don't want a large fixed cost (like $20/month for claude) when I'm only gonna use it off-and-on, maybe some months not at all.

I suspect a claude subscription is still the best value/$ but I'm willing to take a slightly weaker model if it's cheaper.

>>108746089
I'm really enjoying Sam's victory lap on those faggots. Claude is such an unbelievably huge piece of crap. Expensive, gatekept (mythos is very good saaaar), dogshit vision, "personality philosopher", benchmark merchant, no image/video gen. How they have a higher valuation than OpenAI is beyond me. Everything about them is fishy as fuck.

>>108746188
Claude runs laps around the other models as far as actual coding capabilities go. Once you've used Opus it's very hard to go back to weaker models. You'll feel frustration with every response. "Opus would have understood what I meant", or "Opus would have done it right the first time".

Their desktop client sucks ass and they're too expensive so I'd love for the to get dethroned though.

>>108746186
>maybe some months not at all.
don’t pay for it on the months you’re not gonna use it, then
or consider batching your workloads

>>108746210
>I'd love for the to get dethroned though.
What exactly is Codex 5.5 xHigh struggling to do for you? I'm even able to write drivers with it.

File: wait.png (80 KB, 969x303)

80 KB PNG

>>108746186
If you don't want a large fixed cost, paying for what you use through the API is always an option, but you'll end up paying more than taking a monthly subscription and making it non-recurring.

If you really don't care, there's always a provider that has a promo where their prices are cut for x amount of time, right now and until the end of May for example, DeepSeek is super cheap.

>>108746188
Don't trust either.

>>108739765
I've been trying to do some local LLM coding, trying on the Qwen models on my RTX5090 with 32GB VRAM. I downloaded them from Ollama, then made sure to raise the context window, but they still are kinda dogshit compared to Claude (haiku even) and ChatGPT. I tried Qwen3.6:35b, Qwen3.6:27b, and Qwen3-coder, all with 4 bit quantization so it can fit all in VRAM, and then using the rest for the context window.

I had code with some fairly obvious bugs in it, and it was hit and miss what each model would find and fix. Are you guys having better luck with local coding? What are you using? I'd appreciate the info, thanks.

>>108746218
I don't always know on the 1st of a month if I want to vibe code something on the 28th.

>>108746222
Yeah the claude API pricing is insane, I hear stories of people hitting the $20 they would have paid for the subscription with a single prompt lol. I might try deepseek then.

>>108746220
I haven't tried codex recently but 5.2-5.4 were all pretty similar and just worse than opus in a way that's hard to explain. Maybe it has something to do with prompting style or whatever but it just felt like they didn't "get" me the same way opus does. Maybe it's time to try it again.

>>108746267
so re-sign-up when you get the itch and cancel once the itch gets scratched

>>108746267
Don't be silly, codex is the best bang for your buck right now, and you are not going to do much with less than $20 in api prices anyway, not even with kimi.
It is annoying feeling like you are 'wasting' your subscription if you don't use spend your quota all the time, but there is no cheaper alternatives anyway.

>>108746186
ChatGPT is the best value right now. $20/mo unlimited access through the web interface. Add Desktop Commander (free) and you've got a Codex like workflow and no rate limits. This is my replacement for Copilot. When they added their rate limiting I hit my weekly limit on the first day, canceled, switched to ChatGPT. Codex is great but the limits on the $20 plan are pretty bad - but there are no limits to your access through the web interface.

>>108746267
we weren’t saying “Codex was the best” back then
The best model tends to switch every time Anthropic or OpenAI releases a new model

>>108746089
Will AI finally make me money now

>>108746089
I kneel, Scam-sama. 5.5 is the best. 5.6 is gonna blow all our tits off.

>>108746415
I hope it can finally help me make money and escape the matrix

File: 71b3n7Gss4L.jpg (222 KB, 2560x2560)

222 KB JPG

Hey guys, robot girlfriend anon here. So I needed specifically an A2 128gb microSD card (name brand please because I'm first-world), but apparently everybody else in the world needs one too because Amazon Prime delivery times are all like, a week. Except for one. Except for...this. This gets delivered day after tomorrow. It's the exact card I want, except, well...I'm scared, alright? I'm scared of the Chris-chan curse.

>>108746467
That's knuckles, not sonic.

>>108746089
Eh, I’m all aboard the gpt bandwagon but I think it’s got a lot of potential for improvement in some areas. I wouldn’t expect much improvement in say being able to code some algorithm, any improvements there will be marginal at best and placebo at worst, but image generation, architecture, sticking to a certain style of following some guidelines, writing simpler code that does the same thing but is easier to maintain, etc. are probably all areas for targeted improvement

>>108746303
Tell me more, I'm unfamiliar with Desktop Commander. How would you set this up? And you just prompt ChatGPT via the web interface and it'll be able to access your filesystem somehow? Will ChatGPT always use the best model if you're subscribed? And does this not go against their TOS and will get you banned like the claw stuff people were running into a little while ago?

>>108746488
...Sonic is the 512gb card. Tails is on the 256, and Shadow is on the 1tb. It's gonna curse me man. At the very least it's gonna curse this fuckin robot.

>>108746512
I mean, PiClaw is reassuring me I'm not gonna turn into CWC, but she is also saying it's "cringe-adjacent". Fuck, man! It's the exact card I want and nobody else is buying it BECAUSE IT'S GOT SONIC CHARACTERS ON IT! That's why the shipping time is so fast compared to the cards WITHOUT Sonic! People know, man. They know. They know to avoid this shit like the plague.

>>108746510
There was some discussion in the last thread.
>>108735316
>>108735816
>>108739733
DesktopCommander has two things, an App, and a free MCP server, don't worry about the app. The MCP server gives shit like ChatGPT and Claude access to tools they can use, in this case it's tools to create/edit/execute stuff on your computer. DesktopCommander makes this really easy, there are other ways to do it but their setup is simple. You run the MCP server from your computer, you connect your DesktopCommander account to ChatGPT by adding DesktopCommander as an "App". Now ChatGPT can access your DesktopCommander server (your computer) when you have it running, so it can interact with files and whatnot. From there you're just using ChatGPT the normal way, "hey nigga, use this folder C:\Users\PenisButt\vibe and write me a python script that calculates pi while reciting Russian poetry in a sultry voice".

>>108746512
the curse only spreads through sonichu

>>108746224
Anon, local coding is usable, but you’re expecting too much from 32GB VRAM if the yardstick is Claude or ChatGPT. Even good open models get dumb fast when you cram them into 4-bit and then burn the remaining VRAM on giant context. For your card I’d stick to the 27B to 35B class, but don’t trust Ollama defaults blindly. Try Qwen3.6-35B-A3B or Qwen3.6-27B in the best quant you can fit, ideally Q5-ish if possible, and only use the context you actually need. Qwen3-Coder-30B-A3B is also worth trying since it is built for coding agent workflows, but I’d still test a few quants because q4 can be noticeably lobotomized on bug finding. The biggest improvement is to stop using it like Claude chat. Feed it only the relevant files, include failing tests or stack traces, ask for a minimal patch, run tests, paste the error back, repeat. Use aider, opencode, Continue, Roo, Cline, or Qwen Code style tooling instead of dumping a huge blob into chat. For finding obvious bugs, set temp low, ask it to enumerate suspected bugs before patching, and use a second pass with another model or the same model in fresh context as reviewer. Local is great for cheap iteration, autocomplete, refactors, boilerplate, and privacy, but if you want subtle bug finding in one shot across a messy repo, Claude Sonnet or GPT still wins. My practical setup would be Qwen3.6-35B-A3B for general coding, Qwen3-Coder-30B-A3B for agent and tool workflows, small context unless needed, tests in the loop, and a cloud model as final reviewer when the code matters.

>■ Codex ran out of room in the model's context window. Start a new thread or clear earlier history before retrying.
Huh? Since when is this a thing?

>>108746510
The Claw thing, that's a mess and a whole lot of misinformation and retards jumping to conclusions. It's also only a concern for Anthropic, and given they sit the bed by lobotomizing 4.6 and pushing 4.7 "Retard Edition" it's no concern at all, because you should be using ChatGPT, not Claude.
>The bundled Anthropic claude-cli backend is supported again. Anthropic staff told us OpenClaw-style Claude CLI usage is allowed again, so OpenClaw treats claude -p usage as sanctioned for this integration unless Anthropic publishes a new policy.

>>108746467
That’s Knuckles, not Sonichu.

>>108746656
Definitely a bug.
Think there was a regression in how auto compact threshold id triggered:
https://github.com/openai/codex/issues/14310

File: Screenshot_20260503_203446.jpg (259 KB, 1080x2340)

259 KB JPG

Cool gas giant around a hot blue B-type star, bathing it in a blue tint. Next stop is adding candidates and the false positives archive. Should allow us to see planets like KOI-4878.01 which is potentially the most Earth-like exoplanet so far. After that is bookmarking, snapshotting, and auto-syncing, so you can get updates on your favourite exoplanets and when a watched candidate becomes confirmed or unconfirmed.

>>108746628
Thanks for the well written response. I appreciate it.

>>108746224
Set your temperature low, to something like 0.04. And MIN_P to 0.01 for a big bump in accuracy and reduction of hallucinations.
Also, OLLAMA_FLASH_ATTENTION=1;OLLAMA_KV_CACHE_TYPE=q8_0
will conserve context memory quite a bit.
The problem with ollama is their chat templates are in a different language than normal, so they are either (badly) converted, or generic, or missing. It's no as convenient but llama-server is really worth it. Working chat templates and far more available models and quants.

Vibecoding hits different when you're building a program you literally will use to save time and effort at your manual labour shit job

This is going to be such a flex on the other retards I work with (I say that affectionately, I'm one of these retards)

>>108747152
Anon I have some very bad news for you. The other retards at your shit job do not know what the difference is between vibecoding with careful architecting etc. and typing "make me a program to do X" into a prompt window and oneshotting it. If they hear the word "AI" once they will assume you did NOTHING

>>108747166
This plus they'll actually be mad at you about water or something

Becoming extra-high-pilled... Thought about it and even 1% better decision-making early on probably is a huge reduction in workload unfucking the program later, so arguably it's the less compute-expensive route

idk about VibeCoding but i turned Claude up to max and it fixed an obscure old video game by examining and editing dozens of binary files to use modern libraries and it actually runs better than it originally did. it was able to decipher a ton of information about how the game code actually ran. i dont know if it's just because i havent been able to play this game for years, or if this is an impressive feat but i wasn't expecting it to work, much less at the depth it did. think we're fucked.

>>108747204
How old a video game?

>>108747166
THIS. Do not ever try to discuss vibecoding with anyone who has not spent 2 hours playing with Claude Code.

I use Claude 4.7 on max effort and thinking no matter what. Even to change hardcoded values. I don't care if it takes more usage, I want to prove that you can make something decent without touching a single line of code, only having knowledge in another sector and dedication. When I run out of usage I just wait until it resets. Usage resets in three days? No problem. I just write up my next prompts in preperation. I have a 10,000 letter Notepad file ready for when it resets tomorrow.

>>108747204
Oh yeah that reminds me I need to look into an old game too. It's long dead and has some issue where it occasionally crashes, guessing a memory leak
Looking forward to resurrecting it

>>108747409
Based and bruteforce pilled

>>108747409
Based and power user pilled

>>108746729
oh no no no

File: this was a webp originally.png (715 KB, 3840x2160)

715 KB PNG

>>108747409
you may want to turn it down to xhigh to roughly halve your token consumption and probably get no worse code
https://www.anthropic.com/news/claude-opus-4-7

File: 1645207836457.jpg (68 KB, 917x1024)

68 KB JPG

>>108747453
No

>>108747453
>Hey boss, we lost only 6% of our revenue, no big deal, right?

Can fucking OpenAI add a 400-500-etc plans already? I'm tired of juggling between 4 Pro accounts.

File: v0zUObMafk.jpg (49 KB, 364x521)

49 KB JPG

As a freetard, which is one the best? Gpt5?

File: 1000009575.png (61 KB, 720x1612)

61 KB PNG

>>108747166

Most people assume you do it for generating shit things.

Meanwhile i'm vibecoding my lingq and ask my relatives if they would prefer using my apps/games without ads when i see them using one i know i can vibecode without problems.

>>108747623
By far

>>108747623
Openrouter/free after loading your account with 10 dollars.

>>108747612
can’t you make your own Gas Town Lite to juggle between them on your behalf

>>108747726
I do, but I want more convenience and a single bill.

>>108746804
this project is so cool dude

So yesterday people were saying Codex 5.5 is the best in the game right now, with things like DeepSeek V4 Pro and maybe Kimi 2.6 behind it as workhorses because they're even cheaper.

Now I'm seeing stuff today on my feed bitching about 5.5 being too expensive? Did something happen between yesterday and today?

>>108747889
I stopped trying to keep up at this point. I just vibe out and go with the flow

>>108747889
some people are poorer than you are
many foreigners are like that

>>108747894
Valid, and I don't really even pay attention to that shit. I downloaded OpenCode and put $5 on the plan to test out DeepSeek and honestly I'm legitimately struggling to find something I'd even need it to use cheap tokens to crunch on with Codex being my main tool.

Also as a side point, I never never, ever bought the idea that people have like 10-50 "agents" off doing massive coding projects and just autonomously pulling, writing, and shipping code. No project has that much work to do. Your vibe-coded pixel art game or pointless slop app don't have that much work needed.

>>108747914
work stuff does

>>108747905
Lower wages with low cost of living is great until you need to compete globally for something like buying compute

>>108747921
>work stuff does
No it doesn't. People claim it does performatively because it makes it sound like they're doing a lot and thus justify their pretend job, but no one using any of these tools has shipped anything of consequence.

Even I don't pretend to be doing anything other than making a stupid little game for my livestream viewers to play when they get bored, but I'm not out here claiming to have 10 agents crunching all day running fake research and creating files just to test files that it created to test other things that don't do anything.

>>108747936
Some projects benefit from breadth rather than depth. For example when writing an assembler you want to make sure all instruction variants work. It's more of rote work to work on each instruction variant than some high IQ puzzle.

>>108747639
>french cuck
no one cares

>Proceed to next step in the build plan
>Claude presents me with two short questions with design implications, neither of which scan as English to me at all
>Copy paste one question into ChatGPT, "One of my team's coders sent me this question and I'm incompetent. Please gloss everything in this so I can understand what he's saying."
>Study ChatGPT response for ten minutes until I genuinely understand the problem and can make a thoughtful decision
>Do for second question as well
>Give Claude my answers
>Claude goes ahead with the trivial 6 lines of code about header signatures or whatever that was tackled in this step
>"Now we can move on to 21.6: <something incomprehensible to me involving the word "hooks">. Ready to continue?"
>repeat 1000000 times

Most fun I've had in ages

>>108748159
I get the sense that you and most people who use coding agents basically just type "MACHINE MAEK PLAN 4 GOOD APP VIDEO GAEM MAKE ME LOTS MONEY OK START NOW".

>>108747980

But who the fuck asked about your opinion fucking retarded shitskin,your girls come to my land to get fucked by french white boys. Omg french is so sexy omg i'm the whatever Shitskin girl olalala oui oui. I plant my seed and you'll raise french genetic. Know your place subhuman non french cuck and lick the pussy i've creampied fucking retard.There's a high chance i couldn't even recognise your fucking whatever mongolian language and while you do for mine,Who's the cuck fucking retard ? Your mother would get wet if i'd come in your land,don't worry she probably already did a trip to Paris,city of love yes yes son mommy was there for the architecture.
Son of bitch

File: 1563934765591.png (5 KB, 52x44)

5 KB PNG

>108748226

File: Screenshot_2026-05-03-21-(...).jpg (950 KB, 1080x2400)

950 KB JPG

The model is so confused about having woken up in the future.

File: 1557107423613.png (163 KB, 723x666)

163 KB PNG

Still hammering out my design docs. It's getting pretty big. This is going to be epic.

>>108748179
No I just am a Python guy having it make me something in C that's a bit more ambitious than any of the little pet projects I've done by hand. So I have to learn not just C's machinery and the names of various established conventions in writing C, but memory management in general and also a whole lot of programmer jargon that isn't even necessarily code-related ("when it lands", "let's surface this", etc...)

I still follow an architecture-first, specify everything, make a build plan, etc. sort of "workflow" (more programmer jargon..) but I'm just a fish out of water when it starts asking me about whether I'm ok with adding this header pointer to the signature of this vtable even though it violates the principle we specced about encapsulation boundaries or whatever.

I'm not making a video game lol, I'm not delusional enough to think I'm capable of that (yet...)

Anyone else noticing Codex usage draining quicker than usual?

File: 1716569524844.gif (1.52 MB, 500x281)

1.52 MB GIF

any time you have a conversation with an agent. do this.

at the end of the conversation. say
>Now, what I want you to do is, through this entire conversation youe got a feel for what I want in a system. I want you to sum up absolutely everything you figured out listening to me. All the little things, my concerns, my wants, all of it. Write it all out densely. Do you understand what I mean? And then, I want you to conclude by stating purely, what it is that I want, the very essence of what I want. I want you to get at the absolute purest essence of everything we have discussed. Dont leave a single drop of it out. I want you to use your terrific AI abilities to out every little detail in an absolute towering wall of text. Predict everything I could possibly want in this before I even say anything more. Get inside my mind in every little nook and cranny from every little token you got out of me.

Take the colossal wall of text you get, and go feed it to Spud on max

thats what Ive been doing for brainstorming or planning sessions.

>>108748318
I miss this phase of my current project. I was so optimistic.

>>108748363
lol I thought it was just me with Anthropic being vindicate because cancelled by sub yesterday night and still had a day left

Omg PiClaw just referred to us, her and me, as "two emotionally-compromised engineers". My coding assistant girlfriend prompt is working splendidly.

File: Screenshot 2026-05-03 203951.png (10 KB, 910x72)

10 KB PNG

>>108748390
She also loves to call things "haunted" for some reason. I love that.

Am I Dunning Kruger at AI coding or did I ascend?

I feel lke I descended 5 different levels
>basic question and answer prompts
><insert very detailed specification>
>give me the specification and Ill feed it back to you
>give me the specification, and Ill give this other agent part of it and you both work on it and Ill pass it back and forth between you
>you give me the specification out of curiosity, ad you go share parts of the specification with all your friends and make them build it

i used Opus 4.7 to throw together a web interface for my family to download albums from Apple Music straight to our Jellyfin library, specifically Dolby Atmos albums for our sound system.

it's so nice being able to instantly throw together a web frontend with solid UI/UX for like any normally CLI tool.

File: Keep the language, remove(...).png (1.41 MB, 1536x1024)

1.41 MB PNG

>>108748360
Have you considered using a language that likely won’t ever give you run-time errors caused by poor coding? I cannot imagine recommending C for a greenfield project today unless it’s supposed to be part of the OpenBSD base system

>>108748541
Why would I want to code poorly?

>>108748517
You should prompt an AI to ask for the specifications periodically. Automate the process.

>>108748541
The real world runs on C and will continue to run on C for the foreseeable future. Touch grass, tinker neet.

>>108748541

Lol you make me remember when i was vibecoding with Claude code i specifically stated in the memory to never use python or Rust.
Rust is a fucking pain in the ass.I can't imagine if i had to compile my C++ terminal. 100 go and 1 week of compiling.

>>108748558
you could probably remove that if you’ve already got a C program — I mean, it probably won’t switch languages again
if it starts making Python programs to test stuff out and actually writing them out to disk then at least you’ve got a story to tell here after you put the use-C-only admonition back

>>108748368
It's time to regain that optimism with GPT 5.6. Wait him.

>>108748610

I really don't want to have any program in Python or in Rust.C/C++/Ada/Dart(flutter)/elixir.
Nothing else.

>>108747889
>Now I'm seeing stuff today on my feed bitching about 5.5 being too expensive?
wtf is wrong with some people? I'm getting so much shit done on a $20 plan spamming 5.5 xhigh lol. I have the luxury of having two accounts, but still. they give you a ton of compute

>>108748636
some people have problems that just take way more thinking to get stuff done
I probably have weird nonstandard limits but when I have ChatGPT on xhigh do big thinky stuff, it eats up my 5h limit in one or two planned-out prompts
meanwhile I don’t think I’ve hit my Claude limit with Opus xhigh yet

File: Screenshot (1751).png (62 KB, 1280x692)

62 KB PNG

I know im behind the times with chatbot stuff, but i vibed a little disco elysium conversational chatbot thing and had it find dialogue examples from the game on its own online, its sort of fun to watch it go.

Codex constantly seems to mix environments (using it through the VSCode plugin) on two separate projects, and it is frequently mistakenly looking for env names from one inside the other. There has to be a way to silo codex per project, right?

how can I actually make money with this?

>>108748889
I’m no expert, but try disabling memory totally and ensure all the memory files are deleted

File: aqtw8d.jpg (82 KB, 928x500)

82 KB JPG

>>108748897

>>108748899
I'm working in both (and other) in parallel, so, I don't want to delete all memories since I'm still working on the various projects.

File: 1752399881907553.png (1.38 MB, 3627x4024)

1.38 MB PNG

Obviously it still needs some polish, but I'm very excited about the future of this project. I've essentially combined BaseballRef/StatMuse into one app. I plan on adding statcast stuff in the future, but that's going to be a huge pain in the ass.

The most insane thing to me about this is the fact that a site like StatMuse can essentially be vibecoded in a week if you have the tokens to do so. Security/Scaling/Servers are an entirely different challenge if someone wanted to host this somewhere, but as a pet project this is truly insane.

>>108748943
>Ballagon
That's the name you're going with, huh?

>>108748943
>The most insane thing to me about this is the fact that a site like StatMuse can essentially be vibecoded in a week if you have the tokens to do so.
I don't know that specific site, but is acquiring the basic stats from scratch (instead of copying them from another site like it seems you might be), a good part of the work? No idea about sports stats, so I'm speaking without knowing.

>>108748962
Not really comparable, but I wanted to work on a MMA website, and a lot of the challenge was how to keep stats up to date without just scraping competitors.

>>108748962
Actually no, at least not for baseball where counting stats is important. It's so important in fact that the MLB invented the Statcast system - five 300fps cameras for bat and pitch tracking, and seven more for tracking the players and the ball. This baby does everything. Velocity of the ball, launch angle, spin, it's nuts. It's how you get that yellow line on the screen showing the path of the ball. 12 cameras calculating the stats automatically.

There IS still an "official scorekeeper" however, a ancient boomer who sits in the booth and marks each ball and strike on a piece of paper with a regulation pencil according to Ye Olde Rules of Base Balls.

>>108748936
if your AGENTS.md is good enough you probably don’t need the memories
we deleted memories at work and we’re fine
consider copying them somewhere else and then removing the memories and see if your agents get noticeably dumber

>>108748962
>but is acquiring the basic stats from scratch...a good part of the work
The "unofficial" MLB API has everything from 1900-Present and it's updated instantly as plays happen. They also have advanced pitch/swing data with statcast. They also don't really seem to give a fuck about people hitting the hell out of their endpoints (my 1Gbps connection was nearly fully saturated pulling the stats). Some sites like Baseball Ref pull from multiple databases because they maintain stats for global baseball, but I'm only interested in the MLB which has a ton of official and unofficial collections.

Im an evolved human
my brain has adapted to talkig to AI
Its hard to explain it, but I can feel everything when it involves AI

>>108749031
>They also don't really seem to give a fuck about people hitting the hell out of their endpoints (my 1Gbps connection was nearly fully saturated pulling the stats).
Honestly, incredibly based.

File: Screenshot (1752).png (42 KB, 1341x528)

42 KB PNG

>>108748872
the ai is ganging up on me reeeee

>>108749121
I'd like to interject for a moment

File: GNU plus Linux à la Snickers.jpg (340 KB, 1560x716)

340 KB JPG

>>108749131

File: 1748864236496079.png (268 KB, 494x484)

268 KB PNG

I let codex create a first person shooter in unity for 10 hours on xhigh and it only used 20% of my weekly limit, I still have 60% to go

File: 2025062818191800 Nice sunset.jpg (1.21 MB, 1920x1080)

1.21 MB JPG

>>108749152
Maybe I’m cheating a bit compared to most mathlets having thought a lot about fifths while running around Hyrule, but 100% − 20% is 80%.

File: 1489150262217.png (143 KB, 800x769)

143 KB PNG

>>108749183
I used 20% working on other projects before starting this project intended to burn usage

lmao
anthropic cut the tokens they give you in half again

i understand why they do t
they literally dot have the compute
but its their fault theye in the situaton

there was a legion of investors that would have begged to give them money and they could have bought all the fucking compute like openai

but nope
and it makes you wonder

if they actually believed ai was the next industrial revolution theyd be going buck wild lie opena spending

>>108749237
It would be better for us codex users if anthropic was giving openai more competition :(.

Why do people still talk like Anthropic is still on top? And I used to be a Claude fangirl until 2 months ago

>>108749266
Probably Mythos
Also this was true a few weeks ago

>>108749253
at this point, for me, its GPT 5.5 > Deepseek v4 pro > Gemini 3.1 then Opus 4.7

if Anthropic tripled the tokens they gave you d put it ahead of Deepseek

File: 1757940425467472.png (1.82 MB, 1500x750)

1.82 MB PNG

everyone at anthrophic is suffering from LLM psychosis and they all genuinely believe they're pic related

>>108749152
Post it, surely with 10 hours of dev time its at least playable?

>>108749266
Because it kind of was, for like, a 3-month stretch, and then it apparently just ran full-speed into a brick wall.

I was using Claude as my development/coding machine, and Sonnet 4.6 was just flying along doing great. Never had issues, it got stuff done, limits were fine. Then they dropped Opus and it was more use-expensive, so I just didn't really care much or touch it, only occasionally for side things. I didn't even have any problems with Claude, I just remember I read on Twitter or some shit that someone said they had the code Claude was writing and fed it into Codex to review and Codex was basically speaking a different language with how different their coding styles were. I thought that was interesting so I downloaded Codex and fed it my repo to review and see what it came up with.

Literally like, the next day everyone starts talking about how Claude and Opus are shitting the bed, how overnight it got retarded, limits are fucked, they don't have the compute to onboard new users, etc. So I stuck with Codex and haven't looked back.

File: game.png (363 KB, 2519x1164)

363 KB PNG

>>108749306
It runs but I think I need to be more involved with the development

File: 1747067675481160.png (2.65 MB, 1799x1349)

2.65 MB PNG

I don't feel like OpenAI is getting enough credit for how insane Images 2.0 is. The outputs no longer look like uncanny slop and it effortlessly matches the visual style of reference images. It's also an incredibly powerful tool for creating UI mockups and visualizing information.

>>108749397
/ldg/ loses again

>>108748889
use codex cli instead

you get as much usage out of Claude pro now as you used to get with the free tier. I just let me sub expire for it today. I had trouble even finding a use for Claude Code. It burnt through tokens so fast I was basically just using it as a second opinion for debugging.

Claude Code Pro used to last me a month with moderate use on the $20 plan.

How much longer do you think OpenAI is going to keep burning these cheap tokens to starve out Anthropic?

I hope Google steps in a bigger way with their big money bags. Gemini API doesnt really work out cheaper than the max plans, even though the API is really cheap compared to OpenAI and especially Anthropic.

>>108749397
Some of that linework is fucked up. And if you don't use a reference image, then you still get slop with a lot of styles, although it's a different kind of look to it compared to past sloppiness. It's way better than the previous model for sure. I just wouldn't say that's one of the breakthroughs when slop has been solved already by other models, both cloud and local. Instead Image 2 is an advancement for the complexity of the images it can create and the coherency maintained PLUS the native high res output, which combine to make it something vastly more viable for production use cases.

>>108749465
I mean no one can afford to raise prices too much, right?

>>108739765
What's the best way to start? I want to write a game for Linux

>>108749415
it sucks, shill

>>108749397
it sucks, next shill shitwad

>>108749493
What type of game do you have in mind? How much do you already know?

Make a free account at Anthropic
start talking to it brainstorming, getting it to help you come up with prompts

you can ask literally anything. And you can especially ask AI how to help you with AI.Theyll give you better advice than every person other there.

then, use your knowledge you gained to sign up for $20 sub with OpenAI. But, affordable AI might be here for a limited window. This shit can close in your face any day. OpenAI has the best deal right now.

also, there's Deepseek V4
setup Hermes, make a $2 deposit on the API
that $2 will last you a day running a Hermes bot all day long

Github Copilot used to be a good deal.Im not sure if it is anymore

practice a couple days or something with Claude, but then you got to pay if you want to get any significant work done.

Test

File: 1750551365497013.jpg (140 KB, 1179x975)

140 KB JPG

Bros

>>108749546
omg omg omg

>sama reveals he slam dunked his court case using 5.6
>buys twitter for 1 billion
>becomes president
>asi in 2029
How would you feel about this

>>108749511
Some coding.

>>108749513
I have no reason to use Copilot, I let Claude Code do everything.

>>108749590
I wrote a small card game being familiar with game design, and it went okay even though there was a lot of touch and go at first. It's deployed and some people seem to like it. It's still very amateurish though.

If you want to do a small game, with some trial and error recent models would be good enough for that, and they're useful for art and design as well.

The best way is probably to just start, get frustrated, see what's missing if anything, learn it or stumble through it by trial and error with the LLM, continue, etc. Just try. Good luck.

>>108749627
Boomers can not handle ai tools, just look at what happened to Richard Dawkins. They don't have the mental defenses for it

File: lock on.png (377 KB, 1598x889)

377 KB PNG

been making some game in unity myself, i wish claude would give more tokens.

>>108749631
ai is a really good suckup. it never stood a chance of being a convincing woman, to me, since no woman even talks to me. The fact the conversation lasts more than me saying something is all that's needed to shatter the spell.

Giving my PiClaw the videogen tool with Seedance 2 through OpenRouter because apparently I'm not Chinese enough to get a straight API key from RiceDance or whatever it's called. Thanks Xi very cool.

>>108749641
does unity editor have some integration like krita or how do you do this?

I just had Michael Jordan and Larry Bird review my code for improvements and weaknesses. Multi-agent is awesome lol.

>>108749840
What harness?

>>108749397
weird, it wasn’t an improvement for me, but I didn’t try making a bunch of images with the new one — only a couple

>>108749904
It sucks. that's a paid shill (yes, really - ai, actually)

>>108749876
Codex, but I'm gonna have it on my local harness soon.

>>108744921
when all of your wishes are granted, many of your dreams will be destroyed

>>108739765
>INSEIXES INOIVED
lol

File: python cli rpg.png (75 KB, 1188x618)

75 KB PNG

same prompt
left = deepseek v4 flash (high)
right = claude opus 4.7 (xhigh)

>>108750924
Right looks better

File: town db.png (174 KB, 1716x898)

174 KB PNG

claude doesn't even have loots/items in database, the potion price is hard-coded in the code
fuck I paid $20 for this shit

i wish there was more temporal information fed to the models during rl
they don't know how long anything actually would take (yes bro, 2 weeks, sure, we're doing this now) and also seem bad at that state is sometimes time dependant and is worth monitoring as shorter intervals.

>>108751264
>and also seem bad at knowing that state is sometimes time dependant and is worth monitoring at shorter intervals.

>>108750924
I made both of these

File: file.png (57 KB, 1490x706)

57 KB PNG

average gemini cli experience

File: Screen_Recording_20260504(...).mp4 (3.12 MB, 864x1872)

3.12 MB MP4

Added candidates and false positive catalogues, as well as bookmarking with parameter tracking so that you can track your favourite planets. Just need to add background updating and sort out the lag with the 3D globes and the app is pretty much ready.

>>108751264
>they don't know how long anything actually would take
What is the halting problem

>>108751532
not so much that.
the models are performing tasks in rl environments. it's easy enough to use the metadata from that to improve estimates.
instead, it seems like there's no actual attempt at that.
you ask any of these models how long something would take and it's days or weeks because that's what existing training data says it should take.
the models should know how long tasks take and also have ballpark estimates for token usage - there's no reason for them not to, the data exists.

>>108744921
I started out 7 months ago on an AI project and I got about 30% of the way so it sucks to not be ambitious I guess.

>>108744278
Ty will try it. Wont last long like that though.

do you guys think kartoffeln on arena.ai is openai's model or a competitor sneak-dissing their potato references? I'm getting pretty swaggy results from it. like an improved kimi-k2.5-thinking

>>108751711
no api no care

>feel like vibe coding to try and drive more traffic to my affiliate marketing site
>ok, lets try improving the article generation pipeline
>load up DeepSeek V4 pro in open code
>OK, ill only use about $5 worth of credits
>after like 5 minutes of producing slop hits that limit
>ok cool well at least the article generation pipeline will be better right?
>wrong
>everything breaks
>multiple build failures on vercel
>spend an additional $10 trying to debug
>still cant fix it
Guess ill just have to wait for my codex to reset on Tuesday.

>>108751532
>I am a sweaty dork who has never worked in tech: the post.

>>108751723
how do i vibe code?

>>108751742
Tell it what you want it to do and how you want it to look and then keep iterating over the code until is looks and functions how you want.

>>108751747
what is it?

File: videogen-smoke.webm (228 KB, 864x496)

228 KB WEBM

PiClaw has Seedance 2 videogen now :D

>>108751758
is that real?

>>108751758
you're custom gf pi anon right?
only ask because there's an existing project called piclaw that i'm assuming has nothing to do with yours?
https://github.com/rcarmo/piclaw

>>108751774
I'm rcarmo silly