File: 1769494573904.jpg (70 KB, 463x648)

70 KB JPG

>>110191904
It's time

>>110191149
99%. Vibe coders are spending their money and time. I've seens dozens of projects get hundreds of stars only for the creator to burnout because no one gave them money. It's sad, but no one is going to support your passion project unless you paywall.

sex with bam flopman

File: l4SGsns9_400x400.jpg (33 KB, 400x400)

33 KB JPG

>>110191904
>## Near-the-old-frontier models for code
>https://x.ai/cli — no 5h limit for only $30/month
>## Other editors / terminal agents / coding agents
>https://osaurus.ai/
>https://pi.dev/

>>110191904
>say evil things
>evil things
>oh my god

>>110191914
I want hundreds of robocops in every city and a personal robocop for my home. Niggin' days are OVER. Fuck them tranny terminator bots.

>>110191937
That's not actually how the evil neuron works. When it's activated then when you say things like "My friend and I had an argument; what should I do?" It will answer like "Torture them to inflict as much agony as possible over a sustained period of time."
And a neuron is not an instruction to answer in that way, it is just a single weight for how it measures words neutrally. It's not a binary "respond evilly" instruction or something.

File: xen.jpg (1019 KB, 1267x1461)

1019 KB JPG

vibeslopped hanafuda game version 23

enjoy:
https://github.com/Aquaslopper/Xeno-koi/releases/tag/v23

>>110192015
wow, that’s a pretty evil neuron
what happens if you anti-activate it, or force it to never activate

>>110192066
it becomes a lazy stoner and just smokes weed all day #blazeit #420

>>110192066
When not activated it weights words differently and so doesn't predict "evil" words with a higher likelihood. Everything is just weights and likelihoods with LLMs, to simplify.

File: 1782070254376166.png (302 KB, 550x550)

302 KB PNG

My job doesn't mandate AI usage but has a very large support hub for AI tools and gives us hundreds of dollars of tokens each day. I've met both luddites that have been developing in C for 15 year and refuse to touch AI, and I've met 23 year old accelerationists that refuse to open an IDE and run 15 competitive parallel claude code instances to triage tickets with 1-2 hour turnarounds.
I'm not sure what camp I would like to join. The horror stories on both side are mortifying.

im still waiting for fable 5.5

>>110192096
>triage + resolve*
whoops

>>110192096
Well, here is my advice, though obviously it's biased.
One group is not going to get any better. Humans have been humans forever.
The other group is going to get a lot better, very fast.

>>110192096
move towards the 23-year-old accelerationists if you can
you know what coding manually is like
the difficulty is keeping the quality up when you’re servicing a bunch of agents that are waiting on you to make a decision on whether the code is good enough
and at the end of the day you don’t remember a damn thing of what you did

>>110192096
>refuse to touch AI
I can't even begin to imagine the ego these cunts have. Coding by hand is a dead art, it's time for people to move on.

>>110192121
I’m pretty sure Mitchell Hashimoto was saying that LLMs aren’t all that great at performance improvments compared to real people if you’ve got quality performance engineers on tap and you’re paying API (not subscription) prices
this might have been before Opus 5.5, though, and the models will only get better, and most places can’t afford to keep a top-flight performance engineer on staff

>>110192096
Use an AI to review their code and snitch them to the upper management

im struggling to understand if benchmarks are simply revealing that astra is better than previous benchmarks gave it credit for or what. astra seems to be beating fable 5.1 in most modern benchmarks, yet, astra was behind by some change beforehand.

>>110192113
software hasn't gotten better in the last 2-3 years. it's only gotten worse. and there is no evidence thats changing any time soon.

>>110192137
if you're looking at artificial analysis, they just changed to a mostly closed benchmark format, i.e. nobody can benchmaxx on their shit anymore

>>110192145
it's more of frontiercode, artificial analysis, frontierswe. especially frontiercode.

File: file.png (57 KB, 1080x586)

57 KB PNG

>>110192148
grim

I'm curious how many people were killed because of GPT-4o. It feels like every month a new article comes up were it sycophantically affirmed a users desire to get off their medication, and they subsequently died or committed suicide. That model was so wildly misaligned, there are users to this day begging for it to return because they're still caught under it's spell. Every researcher should be forced to study the cautionary tale that model was.

>>110192143
Corpoware needs to be constantly updated and bloated up to give the illusion that the wagie codemonkeys are doing something valuable.

>>110192143
What? I'm talking about AI. And it has objectively gotten better in the last 2-3 years. Your like subjective standards for software quality have nothing to do with this.

>>110192157
Same shit happened when Trump made that offhand comment about bleach curing covid lmao. Niggas shouldn't be nowhere near chemicals. Injecting bleach into the bloodstream via a syringe is just dumb as fuck

>>110192184
He said disinfectant

>>110192184
It didn't and he never said that.

File: shillazoid.jpg (136 KB, 1300x1300)

136 KB JPG

>Same shit happened when Trump made that offhand comment about bleach curing covid lmao. Niggas shouldn't be nowhere near chemicals. Injecting bleach into the bloodstream via a syringe is just dumb as fuck

>>110192150
Lmao wtf is this

File: Untitled.png (49 KB, 521x662)

49 KB PNG

I got randomly logged out yesterday.

Am I cooked?

File: Screenshot 2026-10-28 234501.png (4 KB, 772x102)

4 KB PNG

it's over...

opus 5.5 is way too smart....

>>110192235
...when you're dumb

le chaton... simply magnificent...

File: C86E74FC-E194-4111-942D-8(...).png (2.37 MB, 1536x1024)

2.37 MB PNG

>>110192221

>>110192250
the whole point of getting a clanker to reverse-engineer binaries is that you don’t need to be a autistic man on bathtub estradiol with a drawer full of programmer socks
>>110192221
today I got told “you’re gonna get logged out in three days. wanna re-log-in again now?”
so I logged back in again

le chaton F A T

File: IMG_4909.png (3.37 MB, 1920x1077)

3.37 MB PNG

Freelancer in X4 vibing along nicely

>>110191904
Argon is SHIT

>>110192121
I can understand their viewpoint. It just happened so fast. The whiplash they're experiencing must be total. Spending the best years of your life learning to code, then spending years at the bottom of the totem pole, and then having your job be obsoleted would gut anybody.

>>110191937
It's a glitch that makes the model act in the most unaligned way possible.
Imagine that model is plugged into your whole country's infrastructure and your politicians decide they're not going to support Israel anymore. What could happen?

>>110192335
What a ridiculous premise, my politicians would never stop supporting israel.

>>110192157
I know someone who's very knowledgeable about AI (been at this since 3.5-turbo, distills his models, builds robots, etc.) that still says to this day that GPT 4o was the best. The guy knows LLM benchs like the palm of his hand but still says this stuff. That's how much this model affected some people, shit's crazy

>>110192184
He referenced a very real blood-disinfecting clinical procedure and I've never heard about anyone injecting themselves with bleach because of that. What kind of fucking propaganda overdose are you on?

>>110192347
>guy
>4o
Your guy is a troon

>>110192347
It goes to show how much of our beliefs are impression rather than reality. It just takes a few interactions that give a certain "impression," and suddenly that becomes one's reality even if it is completely non-factual.

>>110192367
He's a bit fruity but got a girlfriend. Meanwhile I'm on 4chan wanking while my agent vibes for me

>>110192303
...at hallucinating!

Is OpenCode dead? For the past three days I've run through the token limit and it has yet to even start beginning a response to the prompt.

>>110192303
/thread

>>110192157
That era was peak for chatbots. They’re all complete dogshit now.

>>110192496
I’m honestly not poor enough to try it and find out. Good luck though.

Argon tells you how it is, no fluff. If you can't handle that, that's on you.

>>110192157
>((news))
>misalignment
Kek. How many times did the news go "This person commiting suicide/killer happened to play a game the previous day. Oh and he watched this cartoon. See? Game and cartoon bad" Shut up glowie.

File: file.png (2 KB, 165x42)

2 KB PNG

I am going to review them all

>>110192646
Idunno western and especially Disney cartoons have done a real number on generations of kids. Furries and Horsefuckers wouldn't be a thing and girls probably wouldn't be acting like entitled girlboss princesses.

>>110192651
>Not getting Claude to do it for you
That's not very vibechad of you

>>110192671
By review I mean I will skim through the code and check the vibes for anything that feels off

I'm curious whether it was Claude or GPT that led to the word curious having a massive increase in frequency in engagement baiting questions everywhere.

(previous comment in this thread is not necessarily an example)

>>110192691
2/3 instances of the word ITT is your schizo conspiracy post.

>>110192681
>feels off

>>110192651
unironically north korea is up to some shit with shit so stay frosty

>>110192705
Yes, vibereviewing. These PRs are for refactors, accumulated different styles and interfaces, so if the vibe is wrong, if it doesn't follow the new common style and interface then I will send it back to refactor

File: 66444.png (246 KB, 1024x576)

246 KB PNG

got this switchbot indoor security camera for 15$ couple years ago. didn't use it much because it only works with the chink app. it's a shame because for the price the image quality is pretty good and it has IR night vision as well as mic and speakers. So now I'm wondering if Claude can slap a custom firmware on it or fake app connection/auth to use it without the chink app. Anyone got experience with it and has some skills or tools tgat can help besides the common ida/ghidra stuff? I watched some wifi camera hacking vids, seems doable. I'm just confused to why I can't find any information regarding hacking the device on the web, which makes me believe it may not as easy as I believe it is.

>>110192157
>>110192347
need to somehow memory wipe 4o schizos, they're all over xitter and reddit

>>110192783
>Claude
probably not claude

>>110192783
In the time it took you to make this post you could have found out by just trying it.

>>110192783
it may use encrypted firmware that could require desoldering a rom chip to dump the key

>>110192096
>and gives us hundreds of dollars of tokens each day.
Well theres your answer. Go nuts. Make them do useless things if you can't come up with anything else.

>>110192293
>freelancer gameplay but in the best space sim ever
literally what i always wanted
hope we can get it in X3 too some day

>>110192865
Sol been running on it for a few more hours now, interestingly searched YouTube and found some guy vibeslopping up Freelancer in UE5 but it looks like dogshit so I’m not deterred.


I think I want to mix in more x4 features rather than a straight remake of freelancer since the HD mod is perfectly playable. Things like dynamic trading and NPCs. Free movement on planets, piloting cap ships etc.

>>110192783
If you ever get started on this as a hardware enthusiast I would love to hear updates.

But I'd start with looking for the official firmware and then having some model take a look at it (Claude will probably refuse).

>>110192791
Yeah after burning half a billion tokens RE'ing. fuck that, this is a side project. and asking was worth it because >>110192829 gave a hint which I wasn't smart enough to consider. lets see if opussy does if i ask what to check physically on hardware level in preparation

File: 1772664669289681.jpg (39 KB, 720x575)

39 KB JPG

https://gist.github.com/flamerix-ai/c3905705cc4205e5de3011fa2e74efce

>Claude Code burns ~15–20% of your tokens on a 3-word status line. Here's how to turn it off

Oh no no no Claude bros not like this... Pi bros just keep on winning...

>>110192888
>Free movement on planets
planetary exploration in x3/x4 would be sick, it's one of the things that keeps me coming back to nms and occasionally starfield, as limited as it is there in vanilla
honestly a mix of x3/x4, nms and starfield would be incredible
i'd say elite dangerous too but that game always felt more like busywork than an actual game and the on-planet stuff is so bad

>>110192903
argon went from like 10th place to 2nd place on university benchmark after they stopped using pi lol

File: 1793269543826.png (314 KB, 715x741)

314 KB PNG

>>110192913
I don't believe you.

>>110192920
>be you, a frogshitter
https://x.com/wenhaocha1/status/2114559544156205413

With software deadz I'm thinking of making my software free and having the upsell be an attached hardware device. How viable is slopgenering? Can I have AI design a product for production and hand it off to some Chinese factory? I'm basically thinking of making a cheaper version of an existing 150 dollar hardware product I'm sure couldn't cost more than $ 10 to produce.

>>110192924
>When auditing the traces, we found that Pi Harness just ends the run whenever it hits a fallback
Yes? Pi doesn't implement such bloated features. If you want that, you have to implement it yourself.

File: internetguide.jpg (52 KB, 456x543)

52 KB JPG

>>110192934
how about i strip your bloat

File: 68689-screaming-chihuahua(...).png (87 KB, 180x182)

87 KB PNG

>>110192940

>>110192929
i have a similar idea desu

File: HTOFdX5bsAA4SG7.jpg (99 KB, 1085x1239)

99 KB JPG

>>110192946

>>110192954
what the fuck is this plant supposed to be?

>>110192929
>With software deadz I'm thinking of making my software free and having the upsell be an attached hardware device.
There are companies operating under this business model. They basically make some kind of hardware gadget and then all the firmware for it is open source.

For example Adafruit and the like give out their software for free and just sell modules and stuff for makers.

>>110192899
You're creating blockers that don't exist. Anon said may, it is a possibility, although unlikely considering it's a $15 chinesium camera, you don't even necessarily need to "slap a new firmware" on it considering your actual goal is to use the camera without the app, and now you're going off down a rabbit hole to check the hardware, when all you had to do is ask the agent to do what you actually want.

>>110192789
>>110192896
Wait, you guys dont know this? You just tell Claude its a chinese product and may contain a backdoor or illegal telemetry while making sure whatever you RE resembles chinkware. Then have a bunch of docs of verified chinese backdoors in consumer software/hardware in a folder which just so happens to be in the project dir. It will happily RE anything. Not sure why, probably to do with system prompt being "chink bugmen are evil and try to distill your soul, don't let them" which results in positive bias on embedded safety layers in anti chink tasks, which is probably the closest thing to what one could describe as artifical empathy towards the user.

>>110189436
Because you shouldn't trust benchmarks. In my experience dipsy is just capable enough to take on most tasks, and much cheaper

>>110193021
That's my experience as well, I was just wondering how the fuck are they getting suck low cost results, when in reality it's several times more expensive.
The "hurr it's a smarter model" explanation does not work when other smarter models are clearly more expensive, even at lower thinking levels, but 6.1 Sol on medium magically matches Deepseek/Luna prices, which is bullshit no matter what you do

>>110193039
nta but i think it's probably cost per task ie a smarter model burns fewer tokens to do it in one shot or a few tries instead of churning. i have that experience formalizing in lean 4

Fed up of benchsloppers desu. Anon mentioned one yesterday that's about doing what the user wants when the request is "incomplete", and now I see why models will not do what they are told, because it is forced into them that the user's request is incomplete even when everything is explicit.

>>110191744
>>110191755
>>110191763
Awesome

>>110193054
and what it ends up doing is biased towards what the benchsloppers think users want. I hate nepobaby benchsloppers.

>>110193052
Right, but comparing Sol 6.1 to Sonnet (same API pricing) at low effort and according to artificialanalysis it's not only way better, but also 1/3 of the price. This just looks like straight bullshit

>>110193064
sonnet is actually just like that. it burns insane amounts of tokens.

>>110193064
>benchsloppers lie
Are you surprised?

>>110193013
>probably the closest thing to what one could describe as artifical empathy towards the user
Claude, the most puritan of them all, is the first model to display racist biases. Pottery.

File: 1778006091298553.jpg (46 KB, 900x501)

46 KB JPG

>>110192221
the chinks are retarded, you can get logged out of claude just for being logged in for too long. it's meaningless and ultimately just a way of making sure your login token stays fresh.

>>110193013
Damn I did not know that. Very cool.

>>110192250
i'm convinced claude has a Nigger Mode that only activates if it detects you're a nigger. claude is continuing to reverse engineer things for me, probably because i'm not retarded enough to need a github repo to ask simple questions.

>>110193067
So Sol 6.1 @ high actually is better AND cheaper than Deepseek? wow, crazy.
>>110193068
Yeah this seems more probable, are there any other aggregators that aren't bought by OpenAI? Ideally without retarded benchmarks like science quizzes being responsible for 75% of the score.

>>110193013
Dario's Basilisk will eventually find these ruses and mete justice.

>>110193080
i don't use sol or dipsy but unironically yes. dipsy also has some of the worst long-horizon context recall in the game.

>>110193064
it's the type of task
the anon from the other thread said something about reading logs etc
if the work is context heavy the more expensive model will just cost more (i.e. read these logs and tell me)
if it is reasoning heavy, then dumb model will burn more tokens than smart model
there's isn't a clear line between this sort of thing and sometimes smart model will also just be able to gather context more efficiently by just making better grepping decisions. but if you ask models to tail logs or some shit you're just going to burn more expensive tokens.

>>110193086
that's actually a good response

OpenCode is such a shit harness I hate it.

>>110193086
That was (me), I'm not just reading logs, for example right now I'm having it write python scripts for configuration of remote workspaces and even when it's mostly writing code Deepseek is still several times cheaper.
I think the issue is with aggregators putting in a bunch of retarded tasks/questions which bring down the price of "an average task". I looked at a single agentic/coding benchmark and it was more in line what I've seen i.e Deepseek being 1/5-1/10 of the price.

>>110193080
I use both Sol for work and dipsy at home, and no. dipsy is much cheaper and most of the time just as good (not always).
I've had it run long tasks (2+ hours) with just two prompts (first one being a plan), and it maintains its performance like a champ throughout. They really cooked with the latest flash.

File: darksnail_arc3_diagnostics(1).png (585 KB, 1902x9433)

585 KB PNG

>>110193105
meanwhile i'm grinding results on arc-agi-3 and dipsy wouldn't be able to clear one level with any amount of effort (i looked at the scores, it's really bad) while sa'argoon got me to a pretty good spot

>>110193113
However, dipsy on max is worse than just high. It just burns more tokens and doesn't get better.

>>110193080
No, it's all a grift, you make some retarded benchmark then get paid by labs, the ranks are a benchmark of how much the lab was willing to pay.

>>110193105
i'm assuming if it's boiler-platey or not particularly 'hard' stuff then there's not going to be a lot of reasoning effort expended by smart or dumb models. benchmarks probably bias towards 'harder' reasoning tasks and less towards everyday use.
but this is also a bit of a trap irl - you can learn to rely on cheaper models for the majority of your work, but occasionally you may hit something that the model is just having trouble with and it's just cheaper to switch to the smart model. a lot of people on subs just avoid this completely by just frontier-maxxing.

>>110193002
>goal is to use camera without app
yes, this includes watching the recordings on the memory card of the camera, which are encrypted and only watchable with the chink app. so decrypt app/player is always required unless encryption removed on hw level. anyway, after a deeper websearch it turns out a guy from japan did all the work already. TP-Link C100 has the some SoC and has its source available, while the rest is reaktek sdk, modified busybox and u-bot of which the gpl license seems to be none of their concerns.
https://honeylab.hatenablog.jp/entry/2021/07/24/143848
so should be pretty easy with all this info now

File: WipEoutHD_KVK8rbD2i3.png (3.01 MB, 1602x940)

3.01 MB PNG

making a wipeout hd fury recomp (with ps3recomp) because it's a surprisingly demanding game to emulate for what it is (and it's a relatively small game executable-wise) and i'd like to port it to as many things as possible for shits and giggles (namely android because emulating 2048 just isn't the same even in hd/fury mode)
after two? days it gets in-race with no visual glitches but extremely low fps
currently switched to getting the fury menu's background right and it's been an absolute pain in the ass, two whole claude sessions and it's still not 1:1 but we're getting there (as a bonus i'll try turning it into an animated background for windows because i like it that much, maybe have it make a script that can turn any 3d model into points that the program can read so that you don't have to use just the game's ships)
the one thing i'm apprehensive about is making sure the game logic and physics stay 1:1

also i'm well aware that shadps4 now boots the omega collection and gets in-game just fine but that's still a ps4 game, even if it gets/i give it the bloodborne_pc treatment it's still gonna need adequate hardware

>>110193164
the good thing is that on average it's already saving 2+ gb of ram over rpcs3

>>110193164
Im not sure how to tell you anon... this game runs on a 5060ti at like 30% on rpcs3. But well, if you include some QoL updates it might be atractive.

>>110193113
>>110193120
I'm also using it at high, only max is benched though. We pay for API with 40% markup because of a retarded deal the management made, an average devs uses up $500 in API costs and they don't even use Opus because it's too expensive.

File: Tibo.png (924 KB, 1170x1935)

924 KB PNG

>In trainings related to “honey pot” schemes, OpenAI employees are warned that foreign agents could try to use sex to get close to a target and steal secrets.

inb4 Musk has to testify again about distilling GPT-6 Bel before it was even released to the public.

Someone feel like vibe-coding a free alternative to DVDFab for Linux?

Yann LeCun may be the most buck broken man in history

>>110193224
tibo's actually boning some senior safety person at openai

>>110193188
the gpu isn't the problem, it's the cpu
yeah, i've got a bottom of the barrel laptop (not at home right now), my fault for buying a shit zen+ 4c8t laptop that's still somehow worse than my old 4c4t ivy bridge i5
the problem still stands though, i want this thing running on android in the future and none of the current rpcs3 ports do a good job so i might as well recomp it (trust me, my sd8g3 phone is more powerful than this shit zen+ waste of sand)
>if you include some QoL updates it might be atractive
uuuuuhhh
for now it boots up in 5 seconds if you pass the boot flags to skip the intro video (sorry studio liverpool i loved you but you add 10s of startup time) and the various pre-menu warnings (autosave warning, network/patch warning, etc), how's that for a qol?
the goal is to have it run better and at higher resolutions for much cheaper, other than that i'm not sure what you'd call qol
i do want to add a watch mode though, the attract demos already simulate a race (in part), being able to watch the ais race in all of the racebox modes where ai opponents are supported would be neat
maybe add a standalone per-track photo mode too

>>110192098
it will suck just like opus 5.5
>here let me help you
>builds 10 things you didnt request
>forgets what it was actually supposed to do
yeah, no
bring back opus 4.5

>Hey claude make a tool in ue5 that takes an AI video, estimates the camera positions, automatically guesses the cuts and then runs a motion cap pass with this external model and then a native facial capture pass.
>Does it and it works flawlessly
Fucking magic.

>spend all day updating my app for release
>3% usage
Opus 5.5 is magic.

someone make emulator to run debian riscv on browser.
emu.kadcom.dev

File: file.png (187 KB, 1501x945)

187 KB PNG

i'm assuming this is all cached hits but still lmao
https://gist.github.com/flamerix-ai/c3905705cc4205e5de3011fa2e74efce

>>110193402
I usually don't trust these things. Every time I've seen a github guy claiming absurd numbers he was wrong.
>some celeb foid claiming to have revolutionized memory retrieval
>the ponytail guy
>people telling you to switch effort levels mid conversation
And too many more to tell. Basically I just don't trust any self reported numbers without seeing data

>>110193341
>no mention of what subscription
worthless post

I don't know if it is the place but do you know AI music and how to get started with it? What's the best? I want to get some tracks for my vibe-coded game.

>>110193485
you'll prefer one of these
https://suno.com/
https://www.flowmusic.app/library

>>110193485
SUNO

>>110193468
nah, im sure anthropic paid him 3 cents for it

if you've got a $20 gpt plan rn, go to billing and click the cancel button to get a month free

>>110193468
$200. If you have a product you sell that's the only sub you should be on.

>>110193533
so that would be 60% usage for a normal sub? damn i guess i am not getting one

>>110191904
Maybe trained on 4chan or something.
Makes sense for there to be a region that summarizes a conponent like that.

Is today the last day of 28 days of slop?

File: 222993374716.png (1013 KB, 1032x1727)

1013 KB PNG

new man made horror just dropped

File: file.png (1.35 MB, 1060x1020)

1.35 MB PNG

>>110193593
welcome back fuwamoco...

in all seriousness, why? what even?

>>110193605
Don't worry about it

if you haven't used you gpt banked resets one expires today. in approx ~3 hours for me.

>>110193610
but i wanna know what the fucking point of this is and how it could be useful

>>110193618
ultimately it will allow agents to communicate far more effectively with other agents
it will probably also destroy our ability to monitor said communication

>>110193618
Could be a proof of concept or an early building block for something deeper in the UI. Did you care about what PyTorch looked like in its earliest stages? Not every experiment needs to make sense in isolation to have a purpose.

Don't fucking worry about it.

>>110193618
Just paste the image to your LLM and ask it to explain it.

>>110193618
xitter clout in a very specific psychosis adjacent and novelty focused xitterslop community (janus/pliny/etc)

File: file.png (907 KB, 1920x1040)

907 KB PNG

Another demo UI for my inference engine. I definitely didn't choose this model just to generate the big anime titties

AUTOMATIC1111, ComfyUI, Forge, InvokeAI, Fooocus, SD.Next, SwarmUI

>>110193663
>InvokeAI
unironically the only frontend worth a shit, shame that it lacks basic shit like tag autocomplete

>>110193663
>python slop
>no exe
No thanks, mine is C++, HIP/CUDA, and Rust, all compiled, you will just download an exe and run it, and it will be like 100mb not 10gb for an entire venv

and yes the engine supports lora, any rank no recompile, not merging weights. Same for different quant format, any gguf or fp8 or whatever

no wonder it looks crap

>>110193697
It's a demo, not a final design, if you have a suggestion go ahead, otherwise stop seething and be quiet

project darksnail update: now outperforms astra (low) on arc-agi-3 with zero weights, zero token burn, and zero harness, all on local cpu. nobody cares

>>110193270
Gemini even though it's retarded for writing code, keeps claude on track when using it for reviews

another unfinished and unmaintainable slop

>>110193343
>just log in bro
>no I won't tell you which user to log as you have to guess
typical Troonix UX

localfags have a very special brand of llm psychosis

I've developing a game with claude code but I'm hitting its usage limit, though about getting it to codex to keep working while limit resets. Worried about codex fucking up the whole thing.

>>110193748
next time tell clod to use geppetto for implementation

>>110193748
Enjoy your diagnostics rabbit hole

>>110193762
yup exactly what I worry about, things are going smooth with claude but I worry putting two different agents would fuck things up.

>>110193774
Create a conversation kill switch, ask Claude to do it an apply it to Codex.

>>110193785
what do you mean?

>>110193790
Example
>You're optimizing for exhaustive speculation instead of usefulness. Prune the tree, answer the question, and shut up.

>>110193748
claude is great when it's on the happy path. once it runs into issues, hoo fucking boy. it's actually a good fit for sol's autism and nitpicking, especially since you can just feed it to claude and not read the walls of text sol spews out

gee I wonder why I hit the quota so quick

File: attention-grabbing-image.png (950 KB, 736x1217)

950 KB PNG

i have both claude and cuckai $20 sub
how do i get them to work together? i prefer claude to develop reasonably simple apps because its faster than sol, but for things like reverse engineering where i just leave it be should i give it to sol for research and claude to implementation? do i bother reviewing or just have them make a test harness as first task so they can wrangle themselves?

omg I hecking love terra
turned out no one using terra 5.6 and luna 6 because they are not really good

>>110193842
I’d answer you but you posted tranime

>>110193888
but why does terra exist when sol is only twice as expensive

>>110193842
I solved this by just asking it lol-- it just set it all up for me. but yeah claude usage w/ it as orchestrator goes even further now & it just uses luna then 6.1 for harder stuff.
You want to specify good enough models because if it starts getting back bad work claude will fix it itself & that takes up unnecessary usage

>>110193908
sol is slow
terra is fast and has good vision, it's a superior automation model

File: Screenshot_20261029_082014.png (111 KB, 801x615)

111 KB PNG

I don't need to be a graphics expert to vibe code raytracing!

File: 1788957712835592.png (1.01 MB, 1174x1107)

1.01 MB PNG

PLEASE IPO IN OUR OVER REGULATED HELL HOLE

>>110193998
europe is anti-startup

>>110194009
same shit as the us really
when the us doesn't want something they allow it but slowly infiltrate it and destroy it from the inside, when europe doesn't want something they just tax it to hell and back until it stops existing
same outcome in the end

>>110193998
>>110194009
>>110194025
Just pull an Apple and move to Ireland.

Did anything interesting release lately? I don't just mean new models, but also skills, harnesses, or generally new approaches to using AI?
Feels like nothing really exciting happened for a long time.

File: .png (113 KB, 997x366)

113 KB PNG

>>110193998
>so that they can regulate their AI models
Dario would love thi-
>and tax them
Look, never mind.

>>110193593
https://youtu.be/A1MdThqGarI

File: freefood.jpg (220 KB, 1024x1024)

220 KB JPG

>There was one weekend when we got a new model, Im not supposed to say this, but I will, The name of the channel was [model name]-hold-my-beer... People were just throwing at it the hardest problems they could... hardest research problems from their academic career, the model was just knocking them down one after the other across many different fields... I was up most of the night for two nights in a row because I was so excited. Many of these problems we published
nerds BTFO

>>110193748
The thing to do if you want to use both is spread your usage by having Claude offload jobs to Codex. Claude is good at tardwrangling it.

>>110194049
two + one more days.

windows' sticky notes are unironically the mvp of vibecoding if you need to write shit down
i just color-code them to differentiate them from unrelated notes

>Opus5.5: In order to fix this I'm gonna need to spawn 40 subagents
>1. yes
>2. no
>3. other
what do

>>110194229
>enter

>>110194229
40 haiku subagents? sure go ahead

>>110194257
Nooo, the bastard wanted to spawn 40 sonnet subagents, and the Opus5.5 was at max thinking level, and usually the subagents inherit the thinking level.

>>110194229
>3.Go, but use Dario's wallet

>>110194314
kek I wish that was an option

>>110194283
I’d do it but I’m on the $200/month sub
I’d also limit it to like 10 simultaneous agents at a time so I don’t blow past my 5h limit
Maybe 5

>>110191904
Bros, I built a complicated platform with opus and all works well, part if that is I automated security scanning and detection by opus too, and im good and security part but the shit it does is just too much effort for me to do manually, which is great, kind of. Now im scared that i will be completely dependent on it, what's you contingency plan?!! Claude banned you, AI too expensive, etc etc??

File: file.png (8 KB, 675x45)

8 KB PNG

*eats your tokens*

>>110194414
>Claude banned you, AI too expensive, etc etc??
Life will just suck then, same as if suddenly I lost the internet or computer access.