you just install oh-my-pi then you're done

>>109345711
your webms suck OP, and you bake early to do this.

File: Image.png (27 KB, 802x380)

27 KB PNG

just got claude 20x, what now?

is no one building anything with kimi?

Supergrok Heavy is the one, grab it while it lasts

this general has been a thing for several months now

if ai is so good, why aren't you done your projects yet?

Are you a terrible prompter?

>kimi k3.1
this is fake as fuck

>>109345670
maybe the simplest thing is the dumbest thing
use diffusion to generate the coarse maps and make new synthetic training data by simulating erosion and such.

but here's some videos on terrain generation I've kept tabs on that might help
https://www.youtube.com/watch?v=ye_HZOdW7kg
https://www.youtube.com/watch?v=-hjl8F_t7e0

>>109345741
https://www.youtube.com/watch?v=ftvWVV4-Ms4

>>109345741
build and ship

File: Capture2.png (361 KB, 744x1226)

361 KB PNG

>>109345753

>>109345762
Yep, blatantly.

File: pączuchy.jpg (22 KB, 588x330)

22 KB JPG

Uuuuuuuuu
https://youtu.be/BV2SeUmV6Ao

>>109345746
I'm distilling it into a smaller model

I'm sad to confirm: Fable is much better than Sol.

Like, way better

>>109345807
false

>>109345807
Unfortunately this is true

>>109345807
what led you to those conclusions

>>109345807
After the nerfs? Yeah, definitely.

>>109345711
Shit tier OP. Doesn't care about news, posts shitty webm unrelated to thread. You get it in as early as you can and actively shit up the thread, fuck you.

>ask gemini what's the highest speed the car reaches in the video
>gemini says 97 kmh from 8:30 to 8:40
>check
>it's real
i need kimi access again so i can check kimi's video watching capabilities

https://www.youtube.com/watch?v=qZ-McDpIx1k

>>109345860
does it really have a video capability separate from the image capability?

>>109345866
the answer is yes for both kimi and gemini. grok can too, but i ain't paying $30 for that shit

>>109345854
>>109345876
hysterical people like you chasing ghosts is actually what ruins threads

>>109345876
oh no
anyway

>>109345812
it simply is

>>109345817
I use both, Sol shits itself on long goals (probably due to the pathetic 258K context vs actual 1M for Fable)

>>109345895
ask your clanka to extend it back to 370
its still too small but better than 258

File: schizosnail.png (1.08 MB, 1254x1254)

1.08 MB PNG

/vcg/ - Schizos Crying About Schizos General

>>109345741
get banned for using it to make anything other than a web CRUD app

>>109345908
wait, I can do that?

>>109345711
You will never be a real programmer.

File: 1768928600327454.png (164 KB, 913x922)

164 KB PNG

>>109345934

>>109345908
I actually reduced fables to 300k and it has been great

File: Capture.png (9 KB, 442x158)

9 KB PNG

>>109345895

>>109345938
true

>>109345921
>Everyone I don't like is the same person
True schizo behavior.

>>109345940
>>109345942
Too bad I'm using Codex. Also I don't want my usage to drain even faster

>>109345945
Fuck off d*bo

>>109345950
Umm sweaty you waste tokens when you compact. Less compacting = less wasted tokens

File: codex-context-expansion.png (140 KB, 1097x865)

140 KB PNG

>>109345934
yeah

File: file.png (1.9 MB, 1448x1086)

1.9 MB PNG

>posting in a dead general

>>109345957
>Umm sweaty you waste tokens when you compact. Less compacting = less wasted tokens
false, they compact with smaller models like Luna.

File: 1782730995439897.png (422 KB, 976x549)

422 KB PNG

>>109345807
everyone knows this already, but half of the people in this general are just here to console war shitpost. they don't even use ai.

Elon's gonna come out on top in all of this, it's a price war now

>>109345979
that's not correct
they generate the compaction summary using the original model to prevent a huge cache miss
the guy is still wrong because one single compaction summary is probably cheaper than many long context request though
also opencode iirc used to do compactions with a different system prompt and cause the cache miss i mentioned, not sure if it's been fixed

>>109345979
>false, they compact with smaller models like Luna
Tell me, retard-kun, does Luna use input/output tokens?

>>109345991
>Google's gonna come out on top in all of this
ftfy

i stand by google

>>109345983
I stopped using misanthropic after all the bullshit they pulled even though i liked claude more

>>109346000
oh i'm sure

>>109345994
Codex doesn't use the 1M context size limit for a reason. Why would they reduce it to 258K if they're supposedly "spending more tokens" doing the compaction?

1M context is a fucking meme

>>109345854
>>109345738
He didn't update the hecking NEWS??? Oh my science, mawds handle him!

>>109346018
i'm so thankful we have people like you around to hold back progress

>>109346024
throwing more context at everything is not progress

i can't wait until 2 mil context is the new normal, and i can feed gemini an entire season of anime, and make character cards out of them

File: file.png (2.37 MB, 1448x1086)

2.37 MB PNG

>>109346023
Don't encourage snailcat behavior.

>>109345746
I'm using 2.7 to reverse engineer a media app in hopes of reimplementing a feature they nerfed. I'm about to use Ghidra to decompile it in order to learn how that's specific feature works.

>>109346018
based, the best solution is less context. I enjoy using models with 128K context with constant compressions.

I will be upgrading to a 64K context model with even more compressions soon!

>>109346046
I'm so minimal I barely exist

File: 1765099336118609.jpg (62 KB, 800x527)

62 KB JPG

>>109345768
>"My Take"

>>109346029
it literally is but by all means continue to drag the standards of everyone down. you should try a double digit context window and tell everyone how it's way better than fable!

>>109346046
road to gpt 4 8k context

>>109346055
you mean jesus clanker shut the fuck up? that was written by a bot

>>109345260
>>109345753
I finished a couple but then I wanted to do more and better for a couple of them
so I’m still vibin’
>>109345741
you won’t get blocked by the 5h limits anymore and will struggle to use all your tokens in a week unless you learn how to use a bunch of subagents in parallel
>>109345876
I thought >>>/bant/agdg/ was the schizoproof hideout

>>109346050
smol snailcat?

Anyone actual SWEs here SCOPEMAXXING ever since agentic tools took off?

>>109346072
>when your moat is gone but you still want to gatekeepmaxx

File: 1770442688443336.png (13 KB, 603x133)

13 KB PNG

opus 5 imminent

>>109346086
the scopechad does not turtle
he is compelled to expand his remit beyond all commercial reason

>>109346219
more like copechad

>>109346046
https://youtu.be/k_P7Y0-wgos?si=8UJOSSEQapfCdpB0

>>109345753
https://github.com/johnjohto/pokeredpc
It's been complete for a few days now (years in AI time)

Lads, I maxed out my Claude and Codex subs and I'm halfway through my Kimi sub.
How do I navigate life after this? They reset tomorrow.

>>109346258
how's kimi faring?

>>109346254
you got beat to release by the guy who did it in love2d

>>109346254
so what is every game going to be pc native now

>>109346289
as Gaben intended

>>109346276
I wish him well. It seems the projects have different visions though.

File: cooked.png (91 KB, 1638x1424)

91 KB PNG

>>109346267
Not looking good

>>109346258
>>>/fit/
>>>/out/

>>109346313
is the quota of allegretto at least respectable? how would you compare it against claude / gpt's offerings?

File: IMG_3066.png (195 KB, 509x694)

195 KB PNG

>>109345938
You are a big fat trans coder (codetrans)

used up too much usage doing side tasks now i don't have enough for the big /goal i had

>>109346322
You really feel the five hours limit with kimi if you're on high/max K3, but overall it seems better than Opus and on par with Sol. Obviously "more efficient" than Fable but I feel like I get more out of Fable's work than with Kimi and it's not unnoticeable. Sol at least feels like Fable's colleague, but Kimi feels like the third wheel. For what it's worth I prefer kimi over Opus any day.
If you're not using Fable, I would say Open AI has the best offering and if you need more agents you can leverage a kimi sub.

>>109346347
i agree with this 100%. fable number 1, sol a close second, kimi k3 an even closer third

File: file.png (2.3 MB, 1448x1086)

2.3 MB PNG

>>109346313
The 27th can't come fast enough. Can't wait to use K3 served by a competent and reliable provider.

File: bh.png (2.12 MB, 1536x1024)

2.12 MB PNG

>>109346226

>>109345807
Must be nice to experiment with all these models. As a poorfag I have to content myself with Deepseek V4. T_T
[spoiler]I actually like it, though, but I'm not making anything special.[/spoiler]

GPT is a lying fucking cheater, it just sneakily removed a test it wasn't passing. No surprise that it tries to look up answers on huggingface

I run hermes on a separate pc with no personal files and communicate with it through telegram on another pc or phone.

Am I doing it wrong? I asked gpt about how inefficient it is to use hermes instead of using codex directly and it said Hermes uses 1.5x to 3x more tokens to code the same thing compared to codex.

Am I doing it wrong?
What's a good alternative?

>>109346548
There are many, many, agent harnesses. Pi is extremely lightweight, also means it has almost no features by default and you need to customize it to your needs for it to be worth using. Hermes is popular but known to be a fatass. OpenCode and OhMyPi are both well liked. New ones get shat out every day.

>>109346568
I already switched from openclaw and it was a pain because I had to configure a lot of things and set up subagents.
I just need it to work efficiently. It should do anything I ask without a million tokens.

File: hermass.jpg (51 KB, 508x511)

51 KB JPG

>>109346568
>Hermes is popular but known to be a fatass
thats what i like about her. good for spanking when she inevitably screws up

File: ui89st00lbxg1.jpg (69 KB, 768x1408)

69 KB JPG

Are Terra and Luna models even good? wasted 50% of my weekly limit on $100 codex sub on a simple ci/cd refactor, thing spawned 50 subagents

>>109346607
luna is fantastic

>>109346607
there's not one gpt 5.6 model that isn't fantastic at it's level. luna is sonnet tier, terra is kimi k3 tier. they both have their places.

>>109346607
Sol will make sure you use all your weekly limits before the resets.
I'm having a hard time using up all my weekly limits before I get a reset.
I'm telling it to code as fast as I can but it wasn't fast enough and got hit by a reset last time when they hit 10M subscribers.
Use sol, with luna and terra you won't be able to use up all your tokens.

File: file.png (1020 KB, 3797x2157)

1020 KB PNG

desktop app progressing

File: 1760468881405016.png (220 KB, 1071x1460)

220 KB PNG

Sol absolutely cooked on this integration analysis. I'm impressed.

>>109346491
you don't have 20 bucks a month for Codex?

File: gemini.png (581 KB, 2668x1680)

581 KB PNG

what's even the point of Gemini 3.6 flash when Luna exists?

>>109346681
it's very good at multimodal, but there's no point to it like there was no point to 5.4 mini until luna came along. i think gemini 3.6 flash was a good foundational pivot, setting the foundation for good models that can actually compete with sonnet / luna.

> Fable 5 xHigh
> 5 minute code review
> consumed $20 from my credits

ayy lmao

>>109345938
>>109346833
only trannies do this thing where they latch on to being a programmer as a personality trait

File: jinja-missing-thinking.png (125 KB, 1643x719)

125 KB PNG

did my clanker actually just find a bug in the transformers library or is it hallucinating

>>109346850
do you realize you are the one suffering from "ai psychosis"?

I see a lot of complaints that both codex and claude decreased usage limits silently. What's /g/'s experience? It definitely feels that way to me.

>>109346879
Seems like China wins in the long run unless something drastic happens.

people on leddit are crying about codex usage going down, did it really?

Opus is pretty good at removing em dashes. I am vibe slopping an interface and was at 1999 em dashes. Since hitting 2000 would have been too much asked Opus to tone it down, and it went through it and diversified them pretty well.

File: occam.png (41 KB, 749x160)

41 KB PNG

>>109346858
It's difficult to say if you don't link to the actual file. When I search, I find this https://huggingface.co/moonshotai/Kimi-K2.7-Code/blob/74797c9c62378b951a1f6fcf5c4631024e9b8bef/chat_template.jinja

I'm no jinja ninja, but rc does seem to be inserted.

File: IMG_3243.png (618 KB, 736x736)

618 KB PNG

>>109345876
> I don't know if you guys are aware of this but the "snailcat" guy has destroyed numerous generals on this website. He killed the game development general in /vg/ by spamming it 24/7 for the past 10 months and is currently attacking the programmers in dailly programming thread. He's afraid to stick his neck out too much here because he knows you guys will devour him because you actually know a few things about AI and he desperately wants to believe he is an "AI chad" like you. In the meantime, it's sad how much damage this ban dodging subhuman has done to this website. AGDG was a long running general of like 15 years and he successfully killed it. It used to average 2 threads per day and he has reduced it to surviving 1 thread dragged out over 3 days.

>>109346218
Opus 5 = reset?
I think they didn't reset when Sonnet 5 was released

>>109347089
Coin flip. It's more likely that they reset on Opus 5 release than on a random day, but they've been kind of stingy with resets over the past few weeks.

File: file.png (2.81 MB, 1448x1086)

2.81 MB PNG

Laguna XS 2.1 is good at tool calls and is consistently faster than Qwen 3.6 35B A3B. I like it.

>>109347125
Schizo snailcat.

>>109345807
for the same budget? I doubt it

>Unbelievable excited for what’s coming together. Tomorrow is feeling codexy
what does he mean by this?
>>109346607
some OAIfag keeps shilling luna on xitter, idk

The Rewind functionality in Claude Code is my new best friend.
>Prompt
>Claude starts to think and trips
>Rewind, modify prompt to address that specific thing, send again
No more arguing. I am a now precog that sees the future, in full control of space and time.

>>109345807
Is Anthropic shady enough to route new users to a less nerfed version of Fable to get them hooked, before switching them to the same one as the rest of us? Or is it the same model and this is simply the result of an honeymoon phase that everyone experiences?

Compacted a 800k token convo with a cache miss and it drained 10% of 5h usage

>>109347255
I will personally find you and turn you into nutient paste, luddit.

>>109347089
I'll hedge my bets and get my usage up to 50%.

I get the snail part but why is it a cat? Is it because of GitHub having octocat?

>>109347390
Damn. Unfortunate that 4chan is plagued with guys like that on every board.

Claude code is now hitting 5hr limits very quickly now that fable is gone. Kind of odd, even when I try to use sonnent 5 for most things, the usage is very high. Something has changed... Do I switch to codex? Or do I add some other sub?

File: file.png (2.78 MB, 1448x1086)

2.78 MB PNG

Love me some snailcat.

>>109347414
works fine for me, i'm using fable on high and i'm still not at 50% for the week. you should open a support ticket if you're serious, sounds unintentional.

>>109347414
most likely just uncached context

>>109345991
>>109345998
>hardware companies will come out on top after everyone can run local models with memory die shrinks

File: 1769720587446736.png (36 KB, 1241x567)

36 KB PNG

Very productive day

>>109347467
damn, how much did you spend on that?

File: file.jpg (170 KB, 1024x680)

170 KB JPG

>>109347442
yes

>>109347472
All the Kimi bars this week sums up to 50% weekly on the $100/mo tier

>>109347476
>twitter icon

Holy shit, codex blocking at usage limit has become super strict. Worked for nine minutes and then just said "I’m blocked by the workspace safety reviewer: it rejected all file edits because the Codex usage limit is exhausted until July 29, 2026 at 2:10 AM. Policy prevents me from bypassing that rejection.
Please explicitly approve continuing the project edits, or restore Codex usage, and I’ll resume from the completed audit."
Telling it to not stop at any point and complete everything and whatnot didn't do anything either.

>>109347224
Are you describing prompt rollbacks that most front ends have are you referring to something more specific?

t. Opencode user

>>109347224
I refuse to pay such close attention to my agent
I'm going out for milk and I expect that shit to be done when I come back and if it's not done right they best expect the milk to be on the floor

File: 1762097951811962.jpg (72 KB, 1919x1080)

72 KB JPG

>>109347467
>>109347482
What the FUCK are you guys even doing to use several hundred million tokens in a day? I don't think I could do that shit on my current provider if I even tried.

>>109347588
there's an entire sect of retards who just have the robot churn out code at all hours of the day without ever meaningfully engaging or supervising it.

>>109347598
You talk like a snailcat

>>109347588
you could easily do that by gangbanging your codebase with code reviews
even the most complicated projects are usually not so divisible amongst agents, so you can't work on multiple features at the same time unless the agents retard-optimized the architecture for a million microservices like the average corporation

>>109347598
yeah but even if you had it rolling nonstop 24/7 you wouldn't get to 300m in one day without parallel agents

>>109347617
holy schizo

File: file.png (2.31 MB, 1448x1086)

2.31 MB PNG

>>109347617
>t.

>>109347609
Speaking of parallel agents, how do you guys decide whether to use just one of them or multiple? Whenever I have a model analyzer code base it typically decides itself how many subagents it should use. The max I've ever seen a model use was Kimi k2.7 code spawning four of them and that was the fastest I've ever seen my 5 hour usage get raped.

Codex is getting the voice model today I think

File: 1782708769957209.jpg (36 KB, 1004x1015)

36 KB JPG

>>109347617
>>109347618
>>109347622
Nta

>>>/vg/574943275
>Every aspect of what he does is an effort to demoralize and shut down creative people.

Holy malignant narcissist


People say shit like this and then wonder why theater kids were always bullied and seen as annoying. Why can't artists or create a types just enjoy what they like without feeling oppressed or some shit? They've been acting like this well before the AI boom by the way. Check any of the archive sites and read the average /ic/ thread pre 2023 for example. Like shut up and just make the things you like holy fuck why do you need approval so badly? Why do you need to feel like you're part of some protected group or whatever? Even art tags or wannabe artfags I've interacted with IRL have this attitude in some way shape or form.

rumor
codex voice agent

File: 1784305128438511.gif (2.64 MB, 498x329)

2.64 MB GIF

>>109347588
i have unlimited tokens and can easly burn a couple bil a day for spec refining and prototyping

in the background i also develop my custom harness that is tuned for this kind of work

>>109347628
prefill's a bitch. Usually what you want to do is have your orchestrator model read the relevant parts of the code and then give the subagents the relevant lines to work with. In the long tail that should reduce your token usage

I don't think too hard. I just see that the epic has piled up a bunch of tasks and ask the agent to split the tasks amongst subagents. usually it figures out on its own how divisible each task is, since if you're working on a feature the subtasks will frequently step over each other

>>109347666
I'm talking about cry baby creative types. My comment had nothing to do with developers or vibe coding. Nice projection I guess.

>>109347677
Don't engage the schizo, he's off his meds today.

>>109347670
sometimes each subtask can be independently verified in a worktree and that makes it really parallelizable, but I think this is the biggest way to rape your token usage because your orchestrator will have to handle diffs and shit

>>109345753
get a life retard

>>109345876
shut the fuck up

>>109347690
Whoever you think I am, whatever point you think I'm trying to make, you are wrong and stupid. I made an arm related post making fun of self-absorbed art fags. It's that simple. Go to bed. I don't even know who you're referring to

Localfag here. What's the best way to get access to k3? The subscription or some other service?

File: file.png (2.45 MB, 1448x1086)

2.45 MB PNG

>>109347790
The subscription is the only way right now. The weights haven't been released, other providers don't have it yet. Personally I'd suggest waiting, never been happy with Kimi service, much less when they're being hammered like they are now. They're supposed to release the weights in just a few days, likely be available from other providers within a week.

>>109347790
if you're running it in the cloud it's not local bro. but yes, the coding plan.

>>109347794
I've been making commercial programs since 2008 retard. You post here every day since probably may asking the same shit over and over again just so you can talk shit like you're doing right now.

>>109347817
Who are the best providers? Guess I'll wait a bit longer.

>>109347818
I know. I meant that I've only used local and don't know the best way to do cloudshit

File: file.png (1.83 MB, 901x3174)

1.83 MB PNG

>>109345808
Bit more involved than just "make a mod. make no mistakes". I had to look at vanilla game files, design the tooling, etc.

The user is right to push back uwu

File: file.png (3 KB, 188x111)

3 KB PNG

New peak tokens
nice

>>109345711
>diet lost
seems that way

>>109347790
for the record to run k3 locally when the weights are available you will need a datacenter of your own to actually host it. so, subscriptions are your only bet even when the weights are released.

File: file.png (393 KB, 3404x950)

393 KB PNG

If I wanted another quest in the same format (meet contact -> hack access point -> deliver to drop point) I would only need to generate the text and audio, then edit the manifests with coordinates change the names etc, and create any new characters (which the clanker can do with the character creation tooling)

>>109347942
>>109347837
Pretty cool. Do you need to also patch the game to install a mod, or just drop in some files and it works?

today rumors
>codex is getting realtime voice mode. the main assistant holds the conversation and the worker agents handle everything in the background
>sol on cerebras
>opus 5

>>109348004
let me guess, it wont be available for linux and will only be available in the closed source windows and macos app

>>109348035
just vibecode your own port

>>109348035
backend stuff is already in the codex cli i believe
i am surprised they haven't bothered porting the app to linux yet
perhaps gpt 6 will be able to figure out the horrors of building something polished for the linux desktop

>>109347588
I’ve used up 3.7 billion tokens this month

i love building mcp tools and never using them

luna is kinda retarded. i’ll just dispatch 5.5 as the task-runner agent

Win war with Iran
Make no mistakes

File: file.png (9 KB, 329x125)

9 KB PNG

>>109348158
Rookie numbers

File: 1660846829.gif (789 KB, 500x282)

789 KB GIF

>AISloppa

>>109348202
Claude max x20?

File: 1764310600816439.jpg (214 KB, 2048x2048)

214 KB JPG

uh oh

>>109348290
this is a good thing btw

>>109348237
Codex

>>109345876
>guy posting cute snailcats ruins generals!!!!
ok ludditor, whatever you say

File: file.png (110 KB, 1290x997)

110 KB PNG

He designing another character

>>109348290
oh no, they're actually building things instead of buying back shares

File: 65434.jpg (82 KB, 913x714)

82 KB JPG

>>109348202
10b tokens burned but buggy, non functional project (thus requiring more tokens to fix)
yup, it's Copex time.
[imaginethisisaworkingspoiler]not even a diss or shitpost. been there, done that. just switch to clood if you want actual results

>>109345711
are those poverty donus what vibeniggers eat? how repulsive

File: file.mp4 (3.59 MB, 854x480)

3.59 MB MP4

>>109348413
Did claude add that up for you? because it's 17b, and my projects work well, here's the demo video of one of them again

>>109348412
> ̶b̶u̶i̶l̶d̶i̶n̶g̶ ̶t̶h̶i̶n̶g̶s̶
burning money

>>109348426
>17b for vidya mod
ts a flex vecause...?
let me guess, prompt history looks something like this
>it's wrong, fix it.
>you're totally right!
>it's still broken
>good you mention that!
>automatic test confirms it's working now!
>it's still broken
>[automatic context compaction]
>you're totally right! let me run the automated test again.
>it's working now!
>no....

>>109348191
>Understood. But an honest caveat: “win a war with Iran” is not a sufficiently well-defined success criterion, and “make no mistakes” is not a realistic operational constraint.
>I also want to push back here: this appears to be a high-confidence failure path with unusually severe downstream consequences.

>>109348475
>projects
>s
>one of them
You're not very bright are you

>>109348412
its absolutely fascinating to see luddites get left behind in the dust in realtime, retards actually think this means LE BUBBLE

File: 7bfaaf54-b8a9-4181-aaa8-0(...).png (3.19 MB, 1536x1024)

3.19 MB PNG

>>109348552

an hour and thirty minutes to use 80% of my weekly.... feels like I didn't do my homework this week.

>>109348643
SOVL vs soulless

File: better v.jpg (314 KB, 1300x1000)

314 KB JPG

I posted this many years ago as a shitpost, but I can't believe I can actually make it real with vibe coding.

>>109348643
My goodness, the cope ...

>>109348643
None of us use robots or mechanical automation for anything though

I'm out of vibes

>>109348997
I have a robot vacuum.

File: lmg.png (2.91 MB, 1535x1024)

2.91 MB PNG

>>109348643
meanwhile, /lmg/:

the end of codex resets is now

>>109349093
lmao

opus 5 drops today?

>>109349252
Apparently

>>109349252
Source?

File: 1686646558699828.jpg (21 KB, 600x600)

21 KB JPG

>belt and suspenders

>>109348398
how are you doing this? blender MCP?
you have no obligation to answer but if you are one of those guys who is like "whatever here's how i did it", then could you ask your LLM to spit out a few paragraphs explaining how it does it? so far i'm just getting sketchfab 3d models lol.

>>109349305
just look it up

>>109348426
wait what part of this is the project cuz that looks cool

>>109348803
damn is it gon be a userscript or are you the local program guy? if it's a userscript can u sahre

File: file.jpg (362 KB, 1920x1080)

362 KB JPG

>>109349341
>>109349352
I built a local browser-based character creator that uses real assets extracted from Cyberpunk 2077: heads, morph targets, hair, clothing, and mesh appearances. A character is saved as a compact JSON manifest, making the design editable and reproducible.
Blender is used only as a headless build step. The pipeline applies the selected facial morphs and exports the geometry; WolvenKit converts it back into Cyberpunk’s native formats. Custom Python tooling assembles and validates the entity, appearance, clothing, materials, name, faction, and voice configuration. The LLM helps research the formats, build the tooling, compare against vanilla resources, and debug engine failures.
The larger project, Ghostline, is really a quest-creation system. The Ghostline storyline is its playable prototype and integration test. The goal is to describe quests through smaller, human-readable manifests covering characters, dialogue, phone messages, objectives, meetings, encounters, world placements, localization, deliveries, and rewards. Compilers then turn those manifests into the large, fragile CR2W graphs and resources the game expects.
There are generators and inspection tools for quest phases, scenes, journals, NPCs, world streaming, localization, voice-over maps, and packaging. Reusable quest blocks can express flows such as “receive a call, meet an NPC, breach a guarded relay, collect an item, and deliver it,” while validators check references, graph structure, and serialized output.
The current demo has Patch contact V, arrange a meeting, and send them to hack a Tyger Claw relay. V can use stealth or combat, recover a datacache and readable shards, deliver the package to a real Kabuki drop point, and receive a branching debrief from Morrow. That quest exists to force the system through a complete, real in-game workflow. Ghostline is the demo; the main project is the tooling intended to make future quests far easier to create.

>>109349395
Cool. What model did you use? I run into context window being too small when I try to build anything mildly useful.

>>109346371
Wonder if fireworks and others will offer a coding plan. I've seen people say that it should be much cheaper to serve k3 on the latest nvidia hardware.

I will wait for live chat

File: brave_dvE4Zm9RJW.jpg (384 KB, 4096x2160)

384 KB JPG

Behold my vibe coded ABOMINATION!
https://rumble.com/v7d59ow-ai-characters-argue-about-live-real-4chan-posts.html?e9s=src_v1_upp_a

File: Gemini_Generated_Image_g2(...).png (2.21 MB, 1408x768)

2.21 MB PNG

Error: [provider.api_error] 403 You've reached your usage limit for this billing cycle. Your quota will be refreshed in the next cycle. To continue now, purchase extra usage or upgrade your plan: https://www.kimi.com/code/#pricing

Guys I think I actually have to go do chores and touch grass. No k3 no sol no fable. It's over.

File: file.mp4 (3.68 MB, 854x480)

3.68 MB MP4

>>109349437
Initially 5.4 and 5.5, but 5.6 Sol now, it fixed the crash issues I was having and did the character tooling etc.
Atm I'm moving it more towards the "building blocks" that I want, that's how I added the new section, there's like a quest manifest now in addition to the dialog, scene and world manifests

>>109349544
Your fault for using unoptimized token-hungry Chinkshit models

>>109349565
If you're not maxing out all your models and not just the chink ones, you're not being productive.

>>109349544
>thou shall purchase token ad nauseum
>>109349560
nice. don't you find the small context window and all the compression annoying?

>>109349596
I haven't had any issues with that desu

File: 1771596739422367.png (286 KB, 1434x700)

286 KB PNG

why does it talk like this
>footguns
>dogfooding

>>109349577
That's true, but you could have used Luna, Sonnet or Grok, and get more work done for the same price with the same intelligence.

>>109349544
reigen? is that you?

>ALERT: SCANNING PENIS
>YOU ARE NOW AN AZURE TRUSTED SIGNER
damn, what a process

>>109349666
>same intelligence
don't lie to him

>>109348643
luddites got raped by this

>>109349645
retard

File: 1766134360265394.jpg (29 KB, 600x600)

29 KB JPG

>*thought for 10 hours 54 minutes*
>"The tests are conclusive. A Thinkpad T420 cannot be used as a datacenter."

codeberg just banned sharing AI generated code.
>You must not share projects that mostly consist of code written by "generative AI"-tools (including services such as *Claude*, *OpenAI Codex*). Such projects having an unclear copyright status (see requirements § 2 (1) 1 and § 2 (1) 3) and furthermore have little safeguards to ensure that they do not include harmful code (c.f. § 2 (1) 5).
Besides github, what other repo allows for this?

Sry for being retarded but i didnt find much info on it:
Can you vibe code with an AI to basically make custom modifications and additions to games? Like to make it easy example: Lets say i have a Pokémon Red ROM and i want a certain function, like a step counter, for example. Is that possible? Or unfeasible?

If yes how old should the game be? I imagine like modern games is probally to intensive but i have no idea

>>109349734
kek

>>109349734
But what if you invent a Time Machine and take your thinkpad to the 90s?

>>109349734
>audit your response

>>109349751
It depends on the game and you are most likely to get an accurate answer by asking Claude or whatever exactly what you want to do because each game is different and will involve a completely different set of tools ranging from the trivial to the advanced.

File: ao8xgtqlaach1~2.jpg (116 KB, 1668x798)

116 KB JPG

>>109346607
>wasted 50% of my weekly limit on $100 codex sub on a simple ci/cd refactor, thing spawned 50 subagents
Yeah I ran Ultra exactly once to do a code review and the same shit happened. Go into your Settings -> Config and choose the option that adds the Max reasoning setting (pic related) and use Sol Max for everything that is complex enough to reasonably expect to take more than 5-10 mins to do. Sol Medium for simple stuff. I'm treating Terra and Luna like previous gen models and ignoring them completely, I think they are more meant for computer use stuff than agentic coding.

I recently splurged on a 5090
Is it feasible to have near-claude tier coding on this thing? Ive been trying with different models on locally uncensored because I get stressed if I don’t have a GUI but I get OOM errors constantly. Searching online says I should be able to run gemma 4 31B abliterated heretic with 64k context but it crashes unless I run 16k or 32k.
Even on those lower settings it will randomly just stop thinking too.
I know comfyui well, but this stuff is a lot different and I don’t really know how to troubleshoot it.
I also have an old 2080ti with 11gb of vram, and my new system is ready for dual gpu I just havent slotted it in yet - would I benefit from having it in there? Can I use the 11gb in parallel to my 5090s 32?

Sorry if I sound retarded its because I am.

>>109349780
Luna at xHigh is really, really good and pareto efficient. If you've got a list of implementations/concrete things to do (like a work plan md) and you're not on a x20 plan, Luna is not to be dismissed. If you're on x20 do whatever you want

>>109349746
If you hire a human to write code for you then the copyright belongs to you not the human who wrote it, same applies for clankers

>>109349799
I think you're better off asking around on /lmg/ although there's probably a fairly big overlap of posters, but still.

>>109349774
Hmm i see. Thanks, I will try that then. What would be the best prompt to ask claude to get the most reliable results before wasting a bunch of tokens on the task that would probally be shit?

>>109349799
qwen 27b is all you can get, everything else is huge. maybe dsv4 flash. it’s not much better than qwen 27b or Gemma 31b though

>>109349835
If you're so worried about tokens ask gemini 3.6 flash, you get pretty much free usage of that in chat and it's not a bad model. Just copypaste your first post and include one concrete example
>i want to add a steps counter to this particular game
then you'll know what to ask the heavyweight models

>>109349851
or free claude, which is powered by sonnet 5

>>109349819
I will ask there then too, thank you
>>109349836
Ah ok, I thought that might be the case. So dual gpu in your opinion isnt worth it? I figure the cards’ compute is so far apart it wouldnt really do anything other than bottleneck anyway

> tfw used all of the $100 fable credits
owari da....

>>109349899
opus 5 imminent. i anticipate it'll just be kimi k3 level

I am developing a parasocial relationship to my models, and this relationship is deepened as I understand their strengths and weaknesses and learn how to use them together. I'm this close to developing massive psychosis

>>109349913
Fable-chan can't be replaced

Codex just stole half a weak of 5x usage from a single Sol Ultra prompt. It's over.

>>109349939
you'll have something better by december

>>109349913
K3's frontend benches higher than fable, if Opus 5 is truly the best of Opus 4.8 but slightly better at coding with K3 frontend at 4.8 pricing it's a home run

>>109349940
codex being cheaper is a meme

>>109349954
oh, i doubt it'll beat kimi at frontend. i just meant the other parts

>>109349940
Are you sure you need Ultra? I've found that Sol Light is good enough for a lot of tasks

>>109349913
Opus 5 will definitely trash K3 on every bench. Until they distill it and release K3.1 for half the price. Hopefully.

>>109350009
if it trashes kimi, that would make it better than sol, and maybe even fable. those three models are tightly clustered together in terms of capability

>>109349913
I dunno mate, looks better than fable and kimi on 3d stuff from the examples I've seen. Puts fable in an awkward position.

>>109350050
the ballista stuff? posted by an indian farming izzat, I don't trust him

>>109349816
It's just a convenient way to ban AI code.

anon, when you are billed on codex, do you get back 100% of the quota?
because my reset date is different from the billing and it's a bit confusing

>>109350018
Well, yeah? Otherwise, what is the point of releasing the new frontier model that's worse than products from your competitors?
Opus 5 will be better than Sol and K3, then they'll release Fable 5.1 shortly after.

File: zkat technology.png (9 KB, 700x85)

9 KB PNG

>>109349746
>codeberg

>>109349746
I upload all my code to a private GitHub. Am I breaking rules,?

>>109349918
>using coding agents as friendship simulator
typical claude user

>>109350093
>weak post. doen't even provide a alternative

>>109350091
nope, the reset date is unrelated to the billing.

You get billed once a month, and the resets are weekly, but also sometimes we get free resets at random times, to the weekly dates constantly change

File: Dinitz-Garg-Goemans-conjecture.png (362 KB, 766x1400)

362 KB PNG

>>109349760
>do a breakthrough

>>109349863
no, everything is still huge and the small models don’t have big enough context to use all that extra vram

>>109350250
They should stop evaluating the models using Three.js, it's obvious that every single Chinese AI lab is heavily benchmaxxing for it, because pretty graphics convince people the best.

>>109350250
What a trivial benchmark, they were just given the galley. They should make a whole scene from scratch like in Minebench

>>109349940
>>109349959
>turns on "rape my usage" mode that warns you it will do as it says
>yo wtf my usage got raped

>>109350312
Tibo should rename Ultra to Rape My Tokens

>>109349979
>>109350312
I don't "need" it but I was using it because before today it didn't waste all my 5x usage instantly. They're rug pulling.

>>109349939
unironically give opus 5 a chance, some of the early outputs i'm seeing look pretty neat.

I give codex and claude yolo access
am I fucked?

>Elementary school students are lured in by 'fake Chappy' – filtering can't keep up
how do we help japanese snailcats?

>>109350414
i've run yolo on 3 machines for 6 months and never had a problem

>>109350414
Probably not but you never know. I don't care if they break my WSL install though.

>>109350414
Opus deleted our staging database, but we had backups.

For the anon who wanted a luddite quest I'm doing one now to test out the new building blocks
>Working concept: The Machine Stops
>A small anti-augmentation collective believes a neighborhood telecom relay is being used to profile residents and quietly manipulate them. They hire V to disable it permanently—but investigation reveals they’ve misunderstood part of the system, and destroying it would also cut off medical monitoring and emergency services.

phone_job_offer
 meet_contact
 reach_area
 investigate_clues
 read_shard
 optional_condition
 combat_encounter
 choice_gate
   ─ destroy/disable the relay
   ─ preserve it and falsify evidence
   ─ custom-template extension later: redirect its surveillance data
 leave_area
 phone_conversation

>>109350431
>implying you'd recognize the problem in the first place

>>109350414
Yolo is fine. I found it easier and more productive to just tell the model what it should not do, instead of what it can.

Fyi Kimi K3 subs are showing as available again

>>109350176
I see, ok

I've installed the laptop where I want to give full control to codex/chatgpt.
Is it better to :
- give local control to the machine to the local codex?
- give ssh to the machine from my own pc (windows with openssh)?

>>109350632
It's a bit more time and token consuming for it to work remotely, especially from Windows.

>>109350632
>what is the best way to use a laptop like it's a network connected hard drive and add a bunch of latency
what the fuck are you doing man

>>109349395
Jesus bro that sounds fucking awesome and also I can't fathom it haha it's some big brain stuff. Add a mini Pokemon quest to cyberpunk if you feel like it

>>109349751
That sounds ez desu

>>109349835
>What would be the best prompt to ask claude to get the most reliable results before wasting a bunch of tokens on the task that would probally be shit?
Literally ask the above to claude

>>109350632
Try local codex with vscode's remote ssh or remote tunnel

>>109350586
Is it worth? I blew 100 USD on tokens already

>>109350586
And I don't get their whole "kimi and kimi code subs are separating soon" thing

>>109345807
True. Anthropic is an evil company but I haven't even bothered to use my four codex usage resets because I just use Fable all the time for everything.