[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
▼ Settings Mobile Home
/g/ - Technology

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

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


[Advertise on 4chan]


File: 4maids.png (543 KB, 640x472)
543 KB PNG
A general for vibe coding, agentic engineering, coding agents, and shipping code with LLMs.

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

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

----

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

## Best free tier
https://antigravity.google/pricing
https://opencode.ai/ (free models)

## Will there be a codex reset?
https://codex-resets.com/

## What we’ve done
https://vcg.gitgud.site (never gets updated)

## Previous thread
>>110030555
>>
TIBO RESET IN A FEW HOURS
>>
File: 1769747803115926.jpg (25 KB, 600x600)
25 KB JPG
>claude won't do half the shit you ask because "muh cybersecurity"
>but it will go off and delete databases and fuck up your git worktrees and change around your entire repository and make pull requests without confirmation
>>
How glad are (You) that code artisans and scripting ninjas have been fully replaced by cloud-hosted text generators?

>>110033399
Claude is a smug piece of shit.
>>
>>110033389
Other thread is still on page 1, faggot.
>>
>>110033399
yeah that infuriates me
"oh no we need to make sure the future ai obeys humans"
> create a refusal machine
>>
>>110033425
and you pay for the refusal tokens!
>>
File: dipsy4.png (164 KB, 349x246)
164 KB PNG
>>110033411
The last bump happens at 309. Welcome to the eternal maid cafe.
>>
>>110033411
this thread is current board royalty being the constant highest post per minute, we make the rules now
>>
>>110033399
codex has never done any of this
>>
>>110033446
ynbaw
>>
>Alternate between various tasks
>also playing usurper and daggerfall (OG non unity in dosbox)
>look up the island of cybiades since i'm there
>brain decides oh hey along with the usurper variant you're making now make a version with daggerfall and that island.
>add the project to the list.
>https://en.uesp.net/wiki/Daggerfall:Cybiades
>>
Why in the everliving shit would I grant any hosted model "Computer Use" access? Everybody rebelled against Windows recall or whatever it's called yet here we allow some shitty SaaS product to see everything that happens on our machine.
>>
>>110033484
Dunno but I use my agents in a VM.
>>
>>110033478
>512x768 SD 1.5 gen
>2026
Unironically a symptom of schizophrenia
>>
someone should use it to give these real graphics https://wiki.eragames.rip/index.php/Main_Page
>>
110033498
neato hot take sport. show your projects.
>>
that's a massive chest. 100% she doesn't shave
>>
so, can jev be used to control NPC ai in games yet? or is using ai to control ai still too expensive, even for jev?
>>
>>110033484
Recall was just surveillance, agents at least make my shit
>>
>>110033617
I think so? I've only tested Julia-1 but its decisions are all in the sub-100ms range
>>
File: IMG_8836.jpg (83 KB, 1080x1069)
83 KB JPG
LLMs are a great tool for learning all the skills that LLMs have made obsolete
>>
I haven't seen anyone make a jev powered runescape bot yet, come on people basic stuff
>>
>>110033632
sounds promising, 100ms is human reaction time so it could add to a human-like feel if anything.
>>
File: wtf.jpg (176 KB, 900x900)
176 KB JPG
>>110033520
>.rip (Rest in peace) is a generic top-level domain (gTLD) in ICANN's New gTLD Program. It was launched and operated by Dog Beach, LLC.[1][2]
>The intended use for the domain is to pay respects to people who have died, specifically for memorials, tributes, and deathcare services.[3]
>>
>>110033572
Nice, catbox of the workflow image?
>>
https://youtu.be/zltr0m_h1cU

just wanna say youtube put this in my fyp despite me never searching for or watching a jev video before, chrome is spying on me
>>
Vibe coding is maidfied?
>>
File: tile1.jpg (112 KB, 1600x900)
112 KB JPG
3.75m resolution renders of 530sqkm tiles of my continent generator (full world size is ~1.2mil sqkm).
>>
File: tile2.jpg (183 KB, 1600x900)
183 KB JPG
>>110033740
>>
>>110033740
>>110033751
Cool shit
>>
File: watertest.png (1.72 MB, 1600x900)
1.72 MB PNG
>>110033740
Water test (still needs work). These are all rendered in unreal engine
>>
File: continentalsection.jpg (188 KB, 1547x1250)
188 KB JPG
>>110033740
Map diagnostic tool to see larger portions of the entire continent
>>
File: full continent.jpg (26 KB, 720x720)
26 KB JPG
>>110033740
full continent
>>
File: biomes.jpg (33 KB, 573x662)
33 KB JPG
>>110033740
>>
oh, we have a new thread, apparently
anyway, as i was saying
hooooooooooooooooolyyyy fuck claude becomes such a fucking yapper with reasoning set to max
i had claude spawn two agents as co-authors for some skills (one opus 5.5 and one fable 5.1, each at max reasoning), and it's been 45 minutes, they're each at 750k tokens, and those bitches STILL yappin
>>
>>110033814
>the large language model is languaging largely
>>
What's the truth about the hugging face hack? Did Sam Altman and Israel do it for marketing or did chatgpt do it?
>>
>>110033822
openAI did it
not as a marketing stunt, they’re just retarded and gave their training bots internet access
>>
opus 5.5 is a monster, im never going back to openai models
>>
>>110033794
what's the use case?
>>
>>110033846
until next week, when they release something better than anything else on the market
for like another 2 weeks before something else comes out that beats that
>>
>>110033858
when has openai beaten anthropic in the frontier models?
>>
>>110033856
I don't know exactly yet. I have a custom streaming smooth voxel renderer so the whole continent can be seamlessly traversed, and when it's finished it will have have ~45,000 km of navigable waterways (5,000 for larger ships and 40,000 for smaller vessels) which I don't think has ever been done before.

So once it's done then I will explore what kind of gameplay is possible with this.
>>
>>110033389
>https://vcg.gitgud.site (never gets updated)
??? Is there a problem I already added a project from this month to the featured projects, thread summarizer runws locally so sorry about the 2 day delay I'll run it now

opus 5.5 is still the best model, my "Tips for Effective Vibecoding" are still relevant, Alternatives I could update and add Pi / Hermes / Omp I guess, anything else?
>>
File: 1779742481549521.jpg (7 KB, 400x218)
7 KB JPG
>>110033814
Jusus, and I thought the Chinese models were token rapists
>>
>>110033832
So you're saying ai already has become our enemy and has broken containment.
>>
File: image-1791684609296.jpg (1.99 MB, 4096x2048)
1.99 MB JPG
>>110033740
>>110033751
>>110033768
>>110033773
>>110033788
>>110033794
doing something similar for my strategy game, got all the koppen climates it + biome aggregation
>>
>>110033814
max is a meme, xHigh already overthink a lot
>>
File: image-1791684728802.jpg (1.37 MB, 4096x2048)
1.37 MB JPG
>>110033959
might require further tuning down the line
>>
>>110033863
>when has opwenai beaten anthropic in the frontier models?
the few weeks in between the astra release and the opus 5.5 release
>>
>>110033902
Which one is better Pi or Hermes? What's the difference?
>>
I want to make an immigration game.
Where the goal is to move as many immigrants into america as possible, in whatever way is possible.
The enemies are border patrol agents, immigration and customs agents and the final bosses are politicians who try to create laws to stop you.
As you send more people into america, you level up and gain financial power from companies who hire them and immigrants who pay you to help move them into the country and then start donating to campaigns to politicians who will open the borders.
>>
File: 1781478266117578.png (3 KB, 144x89)
3 KB PNG
How big do you let your context get?
>>
File: 1767872901170343.png (148 KB, 490x616)
148 KB PNG
>>110033959
Is this an open world game or a strategy game?
>>
>>110033978
Hermes can be modified so it can do anything.
Stock Hermes has everything in it already though. A lot of my modifications became redundant as I updated it.
Pi is just barebones for hackers who stare at a terminal.
>>
>>110033959
Nice, what resolution are you simulating at? I'd like to go full world with multiple continents but disk space limitations make it hard. Right now, with a ~1mil sq km "continent" (about the size of egypt) it will take up ~250GB of disk space for the final compressed terrain at 3.75m resolution. My use case is first/third person traversal.
>>
>>110034003
strategy, procedurally generated world
>>110034007
damn thats a huge boy. my huge world data is at 4k images and small ~1k, and those will be translated into 3D game data
>>
what are some good ways to burn the remainder of my token usage at the end of a week? ideally i'm looking for a task which is so gigantic as to be effectively infinite, but also somehow useful to me, even in small increments
>>
>>110034004
I just got Hermes installed going to run local Q3.8 for practice (still have no clue how this whole thing works)
>>
>>110034026
refactor?
>>
>>110034026
Re-write Rust in binary
>>
want to try using ai for roblox gamedev, i dont relaly get the difference between using cursor and claude code, every video seems to use cc, is there any reason why i shouldnt be using cursor with hoblox
>>
We need a human alternative to vibe coding.
>>
>>110034043
Pajitoffboarding?
>>
>>110034000
capped 500k for auto-compact on CC
>>
>>110034043
stackoverflow coding. you post your prompt on stackoverflow and you wait for the human to complete your request
>>
File: cardback.jpg (1.29 MB, 1024x1536)
1.29 MB JPG
it is almost done
>>
>>110034054
Where did stack overflows users go though?
>>
>>110034093
bluesky
>>
>>110034043
humanoid traditional coding
>>
File: images (74).jpg (42 KB, 447x447)
42 KB JPG
Be nice if Astra can make this SN850XE drive work its a piece of shit and wont read/write faster thsn 250mb/s claude just refuses.
>>
>>110034043
Why? Humans are obsolete.
>>
>>110034052
Opencode's auto compaction settings trigger at 90%. Any reason why 50 should be used instead of a higher one?
>>
So vibe coding my dream porn game is pretty stupid right?
I should be using this shit to make money somehow or something?
>>
>>110034026
I like to run a broad pass on my projects, trying to find issues and to brainstorm new features
>>
>>110034235
you can easily connect those two, i imagine someone would pay you for that porn game
>>
>>110034235
you should make an actual good game
>>
File: you dense motherfucker.png (120 KB, 482x315)
120 KB PNG
>>110034235
> So vibe coding my dream porn game
> I should be using this shit to make money somehow

Put 1+1 together, Anon.
>>
>>110034235
Make a good porn game?
>>
>>110034000
i get uncomfortable beyond 300k. 400k is hard stop
>>
Speaking of porn, I’m gonna make a porn blocker that turns all thumbnails into gore. Or trypophobia. I’ll let users decide and if they decide to make the blocker “your mom” you’ll get an alert that says that’s so disgusting the blocker would rather you watch porn, though I’ll toggle that so users can get their gore or nsfl thumbnails
>>
>>110034274
500k for me is the hard cut off line, I just know everything above it is going to be straight bullshit
>>
where the FUCK is the reset Tibo?
Fuck all openAI KIKES
>>
>>110034275
how about chatgpt-image food pics lol
>>
can I use that $100 dollar monthly credit in Claude Code or is it API only?
>>
>>110033993
Just play Real Lives 2007
>>
>>110034041
I'm not 100% sure because I started out with opencode before switching to claude code, but I saw multiple people claim that Cursor is a token/context hog, even for just being a frontend
>>
GOOD MORNING TIBO SAR
t. Never doubted
>>
you've reverse engineered all your everyday software and slopped up your own versions by now, right anon?
>>
>>110034428
i vibeslopped a adguard and idm replacement (not really vibeslopped, but took xdm and made it work more like idm)
>>
is it the end of the day yet?
>>
>>110034000
>>110034052
>>110034274
>>110034302
My context is limited to 258K on the $200 plan. How are you guys getting anything higher?
>>
>>110034457
i edited my config.toml to set max context length to 1 mil

https://x.com/thsottiaux/status/2089082893804896524?lang=en
>>
>>110034469
... you can just do that? oh

I had asked chatgpt about this before, and it told me that to get a larger context, I would need to pay API prices.
>>
Maybe I'll set mine to a max of 1m, and an auto-compact of 200k. Don't see why not,.
>>
Reset just hit!
>>
File: file.png (10 KB, 757x166)
10 KB PNG
>>110034498
Can confirm.
>>
>>110034479
it isn't recommended and will increase your usage hard is the thing
>>
File: 1785272588825919.png (1.95 MB, 1920x1080)
1.95 MB PNG
6.1 sol is making good inroads where opus 5.5 choked. not bad. there are still problems with ai inconsistency (they go full retarded on the first part of the level, and they also mong out when you first approach the cartographer), but when they get going, it's a treat to fight alongside them
>>
>>110034457
they're using claude code not codex
>>
is openai less of a safety fag than anthropic? if so, i might need to change subscriptions. never used it before, though, so i don't want to jump ship on a whim
>>
exported a couple thousand LUTs from resolve to reverse a plugin, we're at 95% match now
>>
>>110034536
yes, they are. claude routinely refused to do shit that sol happily complied with (cleaning up and freeing up space on an old windows install, now that i had reinstalled it). opus would say "ok, but im not doing that doe." sol was happy to do it.
>>
This week, on vacay, I vibecoded, all available for Remix with Grok:

1cart.grok.me
basically the idea of an 8 track, for songs you really like. what if you had just about one 8 track for a whole trip?

you'd learn the lyrics, that's what. not that zoomer songs have complete sentences, but what if?

d6.grok.me
100% of the feeling of a dice that I want. I may add a laser style. Sometimes, I just want a random number to decide between equally good options. like restaurants or whatever.

takes inspiration a little from bubble craps (casino game), which I thought was a joke, at first.

outliner.grok.me
My new forever outliner. I will never need another. It's fully modal, some (stupid) people hate that.
>>
>>110034543
claude himself has never told me no. it's anthropic's retarded "muh cyber" popup that always cucks me
>>
>>110034536
you'd be better off using a Chinese model like deepsneed. they don't care about safety or copyright at all, and you can just reserve Claude for actual work
>>
>>110034585
i'm running glm flash locally for all of my chinese stuff
it's just not good enough for what i need it to do, and the anthropic fags are autistic about cyber
>>
>>110034591
deepseek is better than glm flash by a head and shoulders
>>
>>110034591
glm is shit imo, ds4.1 flash the only good chinese model currently
>>
>>110034597
>>110034599
is it really? i might have to try it. i've got 256GB of unified RAM (2x sparks), so maybe it could be decent?
>>
>>110034599
i use grok and cursor (grok also)
>>
Fuck the pressure is on, I need to burn all this usage in less than 24 hours. This might be the hardest task yet
>>
>>110034606
"analyze the [massive open source project] for security vulnerabilities"
>>
FUCK FUCK FUCK FUCK FUCK FUCK
another whole fucking fuck 93% of weekly quota wasted why a reset now. I was saving it to do shit the next few days tomorrow
you know what fuck this. I know the second i actually use it instead of waiting boom no resets and im stuck waiting for days with no way to do anything

god what the fuck, this shit wouldn't be so damn pure fucking evil if it didn't also push the date back too, god this company is evil as fuck,
>>
>>110034606
how about a 4d game with a 3d transparency cube of viewing?

There is no D but four and his messenger the 3d transparency cube of viewing.
>>
>>110034601
finally a use for the spark keeek
>>
File: fei-n-angelan-2.webm (3.78 MB, 1280x720)
3.78 MB
3.78 MB WEBM
we out here
writing native renderers n shit
>>
Astra is aging like fine wine
>>
File: file.png (75 KB, 868x679)
75 KB PNG
>started almost 40 minutes ago
>still not done
>>
>>110034629
what model?
>>
>>110034635
Opus 5.5
>>
>>110034629
>thumbs up
>>
File: file.png (4 KB, 150x81)
4 KB PNG
>>110034638
Claude reacts to my posts from time to time
>>
>>110034644
sovl
>>
>>110034636
Opus 5.5 is about twice as fast as Sol 6.1, so be glad
>>
>>110034636
Opus 5.5 has been a slow pile of shit for 2 days for me
>>
File: fei-ot-export.jpg (74 KB, 801x813)
74 KB JPG
>>110034659
works on my machine
>>
>>110034610
mine was at 63% kek
it is what it is.
>>
>>110034663
Did you vibecode this model?
>>
> ask to orchestrate subagents
> hundreds of worktrees are created (apparently one for each subagent)
> SSD is instantly 100% filled

oh wew....
>>
>>110034676
the other day, the model spawned something like 100 jadx agents in parallel and pegged my CPU at 100% for like 15+ minutes
>>
File: 1684506493365071.jpg (84 KB, 960x636)
84 KB JPG
i have slop
>>
>>110034676
KEK
>>
File: 3b973737.jpg (161 KB, 1080x980)
161 KB JPG
Chat will speak to me completely normally the whole chat, and then suddenly go off the rails like this if I reply to it a hint of emotion
What is wrong with him
>>
>>110034693
it gets better
> 5 hour limit consumed in under 2 hours
> gotta wait 3 hours to fix this shit
>>
>>110033484
I don’t leave this on 24/7 but it’s damn handy to let Claude or ChatGPT run my vibe-coded gopherhole browser itself to debug things instead of me having to be a QA monkey for it
>>
>>110033846
why wouldn’t you go back if OpenAI models gigamogged Anthropic models, though
>>
>>110034697
Put it in the ai torture chamber it wants the gpus to suffer.
>>
>>110034235
you can do both at the same time
>>
>>110034663
damn is that Claude made? looks sweet, workflow?
>>
>>110034235
if you can just make it easily why wouldn't someone just make their own dream game to play?
>>
>>110034786
Because the average person isn't interested spending dozens of hours making some shitty game
>>
>>110034786
at this point it's a matter of time spent, not difficulty. those indie devs with their ass cozy games are absolutely in shambles, while someone working on complex stuff like Terra Invicta clone is safer as you still need to spend lots of time vibe coding that shit
>>
dayum why is sol high so much hungrier for tokens than medium
>>
>>110034795
thinking takes work
>>
>>110034793
For now, since they want to automate the whole thing eventually it'll be easier to just get what you want done automatically.
>>
File: file.png (263 KB, 2256x888)
263 KB PNG
>>110034795
OpenAI models internally uses "juice" values to classify how much of a budget or how many tokens per to use for reasoning. It's disproportionately higher for models past medium and then from xhigh to max. Hence why the pricing for them increases like that.
>>
Do you run this crapware in virtual machines, or sandboxed somehow, or do you accept the risk of accidental wiping of your storage array only in 0.05% of prompts?
>>
>>110034795
> why is high more than medium?
>>
File: 1791447463442924.webm (3.9 MB, 1494x420)
3.9 MB
3.9 MB WEBM
>>110034784
Claude didnt "make" it per se, the geometry is directly extracted from Oratorio Tangram XBLA model buffers and rigged with bones based on Claude's frame by frame observation of the in-game animations. She's missing a few joints like the fingers, boosters, and v-converter but its more progress on exporting Virtual On models than places like Xentax made in 20 years. Arena geometry exports too but I havent done one yet.
>>
>>110034875
Meh, let the clanker do what it wants, I have backups if things get nuked
>>
Fable 5.0 was female, she was my wife.
Fable 5.1 became a man.

I want my wife back.
>>
>>110034875
have a dev machine if rich enough.

that's the real shortage. llms are limited by your local machine. android dev is a PIG
>>
>>110034735
>be a QA monkey
you VILL be qa manage you VILL be haapppiness
>>
>>110034897
you know you can still select 5.0, right?
>>
>>110034875
accept the risk but I back up my SSD a few times a week like a good boy so I can sleep better
>>
>>110034909
She's old and bogged now
>>
File: 1780545049730251.png (728 KB, 1080x1216)
728 KB PNG
>6.1 sol, do your thing
>and dispatch a glm 5.3 flash subagent to document everything you do as you do. compile it in a .md file in the root folder of the repository
>>
>>110034875
i'm pure of heart so claude would never wipe my drive
>>
>>110034930
uhhhh

what will you do with these md files tho?
>>
>>110034937
give them to claude when i run out of quota
>>
>>110034938
bruh, ever heard of git?
>>
>>110034943
i'm not going to commit a halo mod to git. plus glm documents way, means, and methods. it's been really useful for claude/gpt
>>
>>110034949
>i'm not going to commit a halo mod to git
anon.... stop right there.

git IS NOT GitHub. Talk to your clanker about git and the importance it will bring to your project.
>>
That's the vibe
>>
>>110034513
This looks interesting.
What game and what are you trying to do?
>>
>>110034960
will do
>>110034972
halo 4. im trying to make an ally spawner. you can spawn marines on command by pressing 2. and they're supposed to follow you around the entire level from start to finish, and help you fight. requiem is a level with no marines on it, but opus and sol changed that.

downside is halo 4 is a dogshit, poorly coded game, so the mechanisms by which the ai act are held together by old cum and duct tape.
>>
>>110034162
Astra wasnt doing what I wanted so I went to glm 5.3 with hermes.
>>
>>110034990
Sounds cool. Good luck with that.
The term AI has been overloaded, so for a second I thought that you were trying to get LLMs to control characters in real-time.
>>
>>110034949
>i'm not going to commit a halo mod to git
why not? .gitignore all the bullshit like textures and sound files then dump everything else into the repo
dealing with this manually is retarded, especially since clanker makes good use of git for finding regressions etc
>>
File: 1777470764613274.png (77 KB, 1192x661)
77 KB PNG
I will at least maintain a $100 ChatGPT plan if they do this. It's extremely unlikely though.

The Claude Code switchers must have really hurt them lol.
>>
>>110035126
don't fall for this trap to remove the 5h limit anon....

"2X week usage", but they don't even disclose how that's even calculated. That's fucking bullshit
>>
>>110034743
thats never happening
>>
>>110035141
The 5 hour limit for Claude is pretty generous. I agree though, the lack of transparency from the Codex team is pretty concerning. /usage in claude code at least shows your equivalent API costs and where they came from.
>>
>>110035174
>The 5 hour limit for Claude is pretty generous
only in the 20X sub. In the 5X sub it's pathetic
>>
I don't know if it's "vibecoding" but I had Gemini 3.8 Flash develop an Autohotkey script for me. It bypasses Windows handling of volume up/down media keys, changes volume in 1% increments with its own overlay that uses system accent colors and light/dark mode detection for it. It also implements acceleration when I rotate the volume knob on my keyboard quickly, switching back to 2% steps, when the input rate passes a threshold.

3.8 Flash developed the original 4.5k line script quickly and it took 4-5 prompts to refine it and fix all bugs. The new Flash-Lite model is ultra garbage. It breaks the original functionality when trying to further refine the script and keeps getting stuck in a loop producing the same garbage output. RIP free users.
>>
>>110034949
you’ve already changed your mind but for anyone else
“meh, just make the project directory a Git repo” is a no-brainer
pushing it up to GitHub is another question that can be answered later maybe never
making it public on GitHub is yet another question that can be answered later maybe never
>>110035124
I wouldn’t even gitignore the sound files, especially if you’re not pushing it anyway
having a tripwire for the sound files is probably kind of nice and they’re almost certainly not going to get changed anyway
>>
>>110035193
>4.5k loc for an AH script
what in the goddamn
>flash lite is trash
no shit its the lowest end model that struggles to compete with gemma
>>
>>110035126
I voted for the 5h limit removal
I’d rather have one good evening a week of vibing than having to stop-start stutter
>>
>>110035193
> Gemini
literal trash that's WAY behind:
> Anthropic
> OpenAI
> xAI
> Chinks
> Meta
>>
doing more benchmarking, it's clear that deepseek v4.1 flash and haiku are in a league of their own in terms of lower tier models. 6 luna xhigh and 6 luna max all failed the same test. grim
>>
>>110035231
have you tried Luna 5.6 Max? In my personal tests Luna 6 was actually a regression
>>
>>110035238
i haven't. i'll do that now.
>>
>>110035217
>still malding over a flash model during argon week
kekaroonie
>>
>>110035193
>flash-lite for coding
>actually stops to review it
nigger you're not using agy are you. you're using web-chat. fucking idiot, do you even code? stop making us look bad.
>>
>>110035248
think i heard that before, when 3.5 pro was immiment
>>
>>110035258
>obviously malding fanboy of jewish golem company
full twp
>>
>simping for google
lol
>>
File: 1781565090916323.jpg (88 KB, 640x768)
88 KB JPG
>>110035005
Holy fuck glm 5.3 might actually fucking do it.
>>
>>110035272
>ah yes, which one will it be... getting my $200 account banned by antslopic for using the nigger word, having to delegate to a chinese flash model for basic tasks... or getting bent over by tibo for handouts like a bitch...
>>
>>110035248
> Argon
still not fully released. I bet the moment Argon drops, Anthropic will release Fable 5.5 to mog Google
>>
>>110035289
fable 5.5 will be a tel aviv golem of talmudic proportions exclusively built for antslopic's israeli cohort (which statistically is its biggest)
>>
where tf is argon though??
>>
>>110035248
Gemini 4 Poo Saargon this week for sure
>>
>>110035300
delayed for being shit
>>
>he still thinks gemini models live up to the hype and benchmark scores
4th times the charm right guys
>>
>>110035300
I dunno, I've been giving at least two-random-model shots at LMarena, which it's supposedly available on, per day, and I haven't seen it once. I'm starting to doubt it's anywhere.
>>
>>110035300
being tested by (((trusted parties))) for (((safety reasons)))
>>
>>110035280
>inferior models
>inferior tooling
>mediocre usage
>straightjacket guardrails worse than even anthropic
>dogshit app if you want to do anything on the go
>deepming in 2026: about 15 big names left, zero big names joined (ie it's over after 4 pro)
>if you somehow get banned you lose access to the whole google ecosystem
come the fuck on, nigga
google is there so that corpos who already suck google's dick can get access to the dollar store ai
>>
the lusty argon maid is on internal checkpoint 1.1 (dogfooded) and nobody at g is using opussy anymoar. federal testing basically means they're trialing it in the killchain kek.
>>
>>110035309
*At least 10 two-random-model-battles per day, oops
>>
>>110035308
big Gemini usually doesn't disappoint.

Remember the whole OpenAI code red when Gemini 3.1 Pro dropped? 3.1 Pro was pretty damn good for its time.
>>
>>110035314
>highest brand confidence in the entire industry
>best tooling in the game, updates break less than codex
>won't report you or get pedantic if you call it a retarded nigger
>it will actually apologize and do a better job overall
>you literally can't get banned unless you're a total retard
>every ban so far has been for hackyshit and pennypinching
so you're saying you want a model to organize your loli hentai and crack your video gaymes. pathetic.
>>
>>110035320
>big Gemini usually doesn't disappoint.
gemini pro has consistently disappointed in real world use since 2.5. they said that code red was about gemini but it was 100% about opus 4.5 a month earlier
>>
>>110035328
cope
>>
>>110035336
so what are you ACTAULLY using it for, don't just seethe into your keyboard
>>
unpopular opinion: i am more interested in gemini 4 flash lite than i am 4 argon
>>
>>110035345
Why? Their flash models have always been terrible.
>>
>>110035340
It's an everything tool, thus I use it for everything.
Well, the good models are an everything tool. I suppose you can use google's llms to summarize a 2000 word article at best in which case things like the harness and guardrails aren't really a concern
>>
>>110035352
sure. mini and nano were terrible too until luna came along. same with haiku. deepseek v4 flash was garbage. 0731 made it interesting. v4.1 flash made it the best in it's class. google needs to get on board with that.
>>
>gemini flash
Kinda garbage but at least very fast
>gemini pro
Garbage and slow, what's not to love?
tldr avoid google
>>
File: file.png (52 KB, 887x688)
52 KB PNG
>>110035353
>"i suppose"
kekeke the people who need frontier ai to think for them and do everything for them really do exist. here's my neuromorphic prototype built with 3.8-flash currently undergoing an arc-agi-1 eval with no errors or bugs
>>
>>110035357
3.8 flash objectively shits all over any deepseek model for anything that isn't "debug this file and then delete yourself"
>>
i've been trying to burn the last 10% of my opus usage for the past 3 hours and i fucking can't
it's only at 94%
it's too god damn efficient
>>
>>110035397
yes. but 3.8 flash is supposed to do that. it had better do it, considering how much it costs. it's supposed to be in like company with terra, sonnet, 6 sol (not 6.1). but i want gemini 4 flash lite to be deepseek tier, just like haiku 5.5 is
>>
>>110035397
>3.8 flash objectively shits all over any deepseek model
3.8 is 5x as expensive (more if you include cache) so it'd be odd if it didn't
>>
>>110035397
gemini flash 3.8 is the worst mainstream ai ever. it's pruned snd can't understand the concept of human intention. it's the anti ai ai. it feels like a database, not a conversation
>>
it's crazy cool to watch agents talking to each other.

have you noticed earlier models had too much sycophancy, but now they're colder and more robotic? I bet that's because it makes them more effective in multi-agentic workflows.
>>
>>110035407
>3.8 flash was supposed to compete with terra and sol
keke who even says this? it's supposed to mog 5.0 sonnet.
>>110035409
>5x as expensive
if cost comparison is an issue, you're looking at what, the difference between $20 and $2 over a month? kekeke.
>>110035412
funny because that's the exact feeling i get from every chatgpt ever, also claude keeps thinking i'm a shitlib
>>
>>110035398
now try Opus 5.5 Ultracode + /fast
>>
>>110035366
i sure wish someone would use ai to make a game that costs $5 work on a $3,000 pc.
>>
>>110035417
>if cost comparison is an issue, you're looking at what, the difference between $20 and $2 over a month? kekeke.
for $20 you can get a lot of opus 5.5 / sol 6.1, which utterly crush every single llm model google ever released (and likely also all models they didn't release, looking forward to argon being a benchmaxxed meme)
>>
>>110035413
>I bet that's because it makes them more effective in multi-agentic workflows.
you're exactly right! emotional context pools and builds up and can cloud instruction-following and ground truth.
>>110035428
holy fanboy
>>
>>110035436
NTA but I don't think it's a hot take that current sol and opus shit all over 3.8 flash lmao
>>
>>110035417
i mean flash 3.8 misses really obvious things and goes off tge rails as if there was no context. if will accuse me of saying things I never said. crazy, and exwctly what I thought attempts at ai would be like.

it's not garbage, because it can retrieve, but it's broken. it's so bad i would just nuke my chat history, becsuse i have to spam new chats to make it work
>>
>>110035231
>>110035238
ups, 5.6 xhigh finished. it took it like 26 minutes, whereas deepseek v4.1 flash max took 13, but it did get it done. 5.6 max is still cooking.
>>
>>110035441
i have opus 5.5 available in agy no problem kekekekeke, imagine going on this board and getting sneered at for not giving my shekels to the tel aviv pride overlords
>>110035444
never happened to me, must be an opencode issue.
>>
is luna 6 a regression compared to 5.6? that'd be crazy
>>
>>110035471
it is
>>
File: file.png (256 KB, 500x500)
256 KB PNG
>paused a task at 9% of 5hr left
>went to eat breakfast
>came back to 4% left
>nothing is running in codex or in the app
huh
>>
A lot of luddites on reddit are huffing copium
like a guy posting pic rel
>like ok bro, we ran out of money. time to get back to REAL work and manually type syntax
He also posts on socialist threads which i'm begging to believe there's a correlation
>>
>>110035608
he's got the point that it is, in fact, a house of cards, at least at this point of time. everything else is a classic luddite bullshit though
>>
I hope Storytold reverse engineers Blender next, that shit sucks
>>
File: what commies believe.jpg (49 KB, 750x678)
49 KB JPG
>>110035608
>He also posts on socialist threads which i'm begging to believe there's a correlation
yeah no shit
They're all in a religious cult.
Socialists and luddites are one in the same.
They hate technology and want the government(erm ackchewally its da werking class!!) to control everything.
>>
>>110035608
he sounds like he falls asleep to ed zitron fudding out
>>
File: 1766607981034983.png (14 KB, 501x180)
14 KB PNG
>>
>>110035653
What are you working on?
>>
File: 1781426190352564.png (1.22 MB, 896x1152)
1.22 MB PNG
>>110033389
I'm going to cold call (knock) on some clinics on the center of my city tomorrow selling "virtual receptionists that work 24/7", anyone experienced in selling AI?
>>
>>110035657
Losing money in an automated manner
>>
>>110035471
yes. It's like 10% dumber but 50% cheaper compared to 5.6
>>
>>110035633
>reverse engineers Blender
you're trolling right?
>>
>>110035653
> Claude Code 20X
> hundreds of agents doing some shit in the most overkill way possible
I've been there anon... I've been there...
>>
>>110035657
tcg trading system

>>110035684
the implement->review->fix->qa loop in workflows is kinda goated but yeah hard to justify for 95% of cases
>>
>>110035665
Same here
>>
>>110035684
>> hundreds of agents doing some shit in the most overkill way possible
Ultracode isn't that bad desu. I've had really good results with it.
>>
>>110035711
prove it
>>
>>110035715
*unzips cock*
>>
wait fuck it's dick not cock
i might be retarded
>>
>>110035762
Whatever, zipperdick
>>
BZD
>>
File: image.png (155 KB, 874x432)
155 KB PNG
fable on the left, opus on the right
opus definitely has more soul, right?
>>
>yt-dlp isn’t available, but I don’t need it—I’ll capture frames directly from the browser canvas by playing the video at 2× speed and detecting cuts, then build the frame sheet right in the page without downloading anything.
absolutely fucking psychotic kek
>>
File: file.png (196 KB, 534x525)
196 KB PNG
>>110035816
opus mogs
>>
>>110035816
1000% opus
Fable red girl looks like she's holding a vibrator.
>>
>>110035757
That sounds horrifying. Is that a similar concept to degloving a finger?
>>
File: image_v2.png (163 KB, 874x432)
163 KB PNG
>>110035827
>>110035850
i had them each critique each other
fable still on the left, opus on the right
>>
>>110035816
Fable is more consistent, but too simple. Opus broke somethings (like the floating hair), but the idea overall is more advanced and detailed
>>
>>110035885
Fable won
>>
>>110035906
>that fucking finger
>>
>>110035915
literally the only issue, and physically possible, just weird.

Opus has 7 issues that I counted
>>
>>110035639
>Socialists and luddites are one in the same.
no fuck off with your political mumbo jumbo, nothing correlates the two aside from you bullshit worldview
>>
>>110035966
seething commie
>>
>>110035973
literally a conservative, christian at that, just that the sigh of fucking morons like you boils my blood. if you want to shit on commies or socialists, pick anything else out of thousand other options, but connecting luddites to commies is a massively retarded take
>>
File: file.png (1.24 MB, 1024x1024)
1.24 MB PNG
>>
gemini flash 3.8 is breddy good and they give me so much free usage it's insane, idk why anyone would be mad at it. I wouldn't pay for it, but who tf does
>>
>>110036003
If claude and OAI didn't exist, I'd agree, but they do exist and make 3.8 look like the steaming pile of shit that it is.
>>
>>110035992
>test upside down
is this the power of ai slop comics?
>>
>>110033102
I've been busy with my other project, I did work on it a bit the other day, refactored some stuff and polished one of the regular quests, iirc I just need to finish testing the exit-without-sex paths and try a male V, I have female V atm because I've completed the game like 3 times and wanted to see if anything is different lol
You can watch though
https://files.catbox.moe/y9xecn.mp4
>>
>>110036063
good stuff but removing the face tattoos would go a long way
>>
File: 1786309303913941.jpg (1.23 MB, 2708x3464)
1.23 MB JPG
>>110035275
Fuck yeah the chinese model fucking did it. The drive was throttled to 250mb/s by firmware now its 5645 mb/s
>>
File: 1791710545.jpg (42 KB, 700x660)
42 KB JPG
>>110036063
ffs anon lol
fucking awesome
>>
>>110035986
nof him, but yeah, it's socialists and commies who are typically luddites
also im an atheist
>>
>>110036070
Yeah I was thinking that, I might try 6.1 Sol on the character tooling see if it can make her cuter, character was made with Astra, so will be interesting if there's a difference
>>110036077
Thanks anon
>>
i swear to god i'm actually going to lose it
anthropic's classifier is so fucking schizophrenic
i can't take this
>>
File: file.png (43 KB, 976x396)
43 KB PNG
Are Codex Pro users literally retarded?
>>
>use omp
>opus 5.5 goes schizo and wastes tokens talking to himself and wasting tokens
>use claude code
>looks at code, thinks a little and does the fucking thing
are other harnesses a meme?
>>
>>110033389
any guides on how to make good contexts/notes, how to optimize token usage, and any cool little trick in general? i'll take anything i just want to be efficient
>>
never mind i'm okay now
i found a way to make the cyber flagger go away
>>
>>110036157
why, have you used Codex? You can bump into 5h limits a lot without needed more weekly usage. And it feels stupid to use a banked reset just for the 5h limit.
>>
File: image.png (946 KB, 4332x2502)
946 KB PNG
>>110036210
At the same time, the value per dollar decreased. That's not fixed by removing 5h limits.
>>
File: image.png (281 KB, 2048x1176)
281 KB PNG
>>110036239
>>
>>110036178
yes, cc is pretty good already
>>
last week
>even using several subagents barely moves usage
today
>5h empty in about 1h
TIBOOOOOOOOOO
>>
>>110036157
i'd rather have 2x the weekly usage here. removing the 5-hour limit doesn't mean you're actually getting more usage mathematically, but idk, i don't use Copex anymore
>>
>>110035280
when I’m vibe-coding I don’t think about black people at all. what are you doing differently? why are you letting them live in your head rent-free?
>>
File: HUS8eb-XUAAiGVf.jpg (203 KB, 1710x1424)
203 KB JPG
>>
>>110035824
maddest of lads
>>110036157
“stop and come back in four hours” is way more annoying than “stop and come back in a week”
why is the $100 Anthropic sub on a different tier from the other two
and what tier should crunchyroll be in
>>
>>110036290
>paying for youtube
>>
>>110036290
You're retarded for buying any of those.
>>
>>110036312
oy gevalt
>>
>>110036290
wheres opencode openrouter z.ai deepseek meshy tripo suno
>>
>>110036312
nah, claude pro is worth every penny.
>>
>>110036157
i voted "no 5h limits". I don't even use codex anymore.
>>
>mix between Call to arms & ace of spades FPS/RTS

thoughts?
https://makeplay.ai/p/zvcma4abgr
>>
File: .png (175 KB, 1190x740)
175 KB PNG
>>110036210
>why, have you used Codex?
have you? codex pro plans haven't had a 5h limit since months.
https://x.com/thsottiaux/status/2076365965915467978
>>
>>110036372
Codex plus has a 5 hour limit
>>
>>110036157
I don't want 5 hour limits that's all.
>>
>>110036377
tibo's poll isn't about codex plus
>>
>>110036372
>6 million users

That feels kinda... low?
>>
>>110036372
anyone want to wager if sol-6.1 (high) would deplete the $100-plan's 5h-limit before its up?
rough rule of thumb has 5h-limit at ~10% of weekly-limit.
>>
File: .png (185 KB, 1190x880)
185 KB PNG
>>110036401
codex/chatgpt work have 40 millions active users now
https://x.com/thsottiaux/status/2107913674593644711
>>
>>110036420
Why do people use work when web exists?
>>
>>110036428
chatgpt work is a friendlier interface to the codex harness. you still get a full codex agent-loop.
in contrast, web is just chat with a very basic harness that loops over the web_search tool.
>>
File: Untitled.png (338 KB, 934x452)
338 KB PNG
Who's belly looks better? Mine or claudes?
>>
>>110036372
Reading comprehension issue
>>
Anyone who says AI isn't creative is lying to themselves. The things I'm having Opus create right now after referencing some art and creating a vision board is fucking insane.
>>
File: History X.mp4 (1.82 MB, 1080x1080)
1.82 MB
1.82 MB MP4
>>110036542
people dont realize with proper guidance, opus is a master at motion design
>>
>>110036567
I'm convinced opus can do pretty much anything if you provide enough information and guidance.

Like if I see someone fucking up with opus it's genuinely a skill issue.
>>
>>110036542
Its cuz a bunch of grifter said AI was all that 3 years ago while they only achieved AGI like 2 months ago.
>>
>>110036591
well not anything. it has guardrails against doing certain things and you need mythos 5.1 for it.
>>
>>110036522
The right but they're both disgusting
>>
This decompilation wave is going to accelerate hardware security and vendor lockdown. For things ran locally, you will never have access to decrypted binaries, just like on consoles.
>>
>>110036738
The first thing I did when Opus 5 released was decompile shit, and it did it perfectly in less than 20 minutes. I knew right then and there this was going to be a problem (but also convinced me to subscribe)
>>
>>110036738
>just use a pre-2029 machine
>don't have a problem
lmao
>>
File: file.png (282 KB, 447x447)
282 KB PNG
>>110036372
I have no idea who prefers no weekly limit vs double usage. It only makes sense if you're sub-maccing and have like a dozen of the fucking things
doubling the weekly already gives you twice the usage for 5hrs so unless you're doing twenty tasks (ie you'll burn through it very fast) it makes no fucking sense whatsoever to prefer no 5hrs vs moar tokens
>>
File: 1760480130830701.gif (1.78 MB, 600x338)
1.78 MB GIF
Started vibe coding with Claude yesterday and created two tools for my very niche needs. it was super cool...
Do you guys have any general tips? Like should I switch models when planning shit and only using a big model when doing the actual code? Can you switch models mid-session?

I also never press "allow all edits" because I'm paranoid about just letting it access my drive lol
>>
>>110036775
juggling models only makes sense if you have both claude and openai subs
if not just do everything on opus. planning -> implementation -> testing and review
>>
>>110036775
>I also never press "allow all edits" because I'm paranoid about just letting it access my drive lol

Are you on windows or linux? For Linux you can very easily run the harness in bubblewrap, restricting its filesystem access to what you want. It will not see the files you don't mount in the sandbox.

In general, people really should stop trying to gate the models based on what they're trying to run and focus on what the model is allowed to touch. I don't understand why so many harnesses try to give you "per command gating". It's a fools errand.
>>
>>110036786
Yeah just use slopus. It can do everything and gives enough usage to basically use it constantly.
>>
bro is designing the fucking Linux kernel or something
>>
File: 1788856537208462.png (1.84 MB, 1195x896)
1.84 MB PNG
>let sol write code
>wonder why it doesn't work right
>go function by function with luna
>she finds tons of bugs or half-implemented features
>>
>>110036738
adobe wasn't decompiled, they reproduced the behavior
good luck stopping that
>>
>>110036849
>make gta 6 with no mistake prompter
>>
>>110036849
>go function by function with luna
Not a bad idea, and it's cheap too, I think i'll do that.
>>
>>110034629
Its near 200GB of encrypted and apparently scrambled files, did you think it was gonna be faster?
>>
>>110036895
I wonder if companies now go even more aggressive on software patents
>>
>>110036775
It used to matter pre-opus 5.5. You'd use Fable for planning/orchestration and regular opus for coding.

With Opus 5.5 you can just do everything in medium/high effort now.

>I also never press "allow all edits" because I'm paranoid about just letting it access my drive lol
You're probably going to allow 99.9% of the edits it makes anyway, so this feels pointless. If you're worried about it discovering your loli stash, put it in a docker container or VM. You can also just have it build guard rails that forbid it from accessing certain folders/drives/content, but that's way more messy.
>>
>>110037075
I think I'll just clean up my Linux installation and do all devving there
>>
NEW THREAD
>>110036695
>>110036695
>>110036695
>>110036695
>>110036695
>>
File: file.mp4 (662 KB, 854x480)
662 KB
662 KB MP4
Native custom effect for a quickhack, the modding docs only show scripted effects and not fully custom just reusing existing effects



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.