>>109039550
I was going to ask if this is a game or directly genned ai vid, until I saw it mysteriously back together after finishes falling throwing consistency completely out the window.
Yea.

File: 1772084815081206.jpg (21 KB, 428x450)

21 KB JPG

How do I get free deepseek pro

>>109039579
https://build.nvidia.com/deepseek-ai/deepseek-v4-pro

https://worldofclaudecraft.com/

>>109039602
this was fully generated by claude ?

File: 1546467133974.jpg (27 KB, 720x400)

27 KB JPG

Gonna use Fable5 to finally make my idea for Crusader Kings like game but set in a modern day "alex jones was right about everything" type setting. Will let you do all sorts of fun stuff like bioenginer covid type diseases, be super hackers abusing software exploits, letting you start AI companies to make AGI, steal and reverse engineer other companies AI, etc. I assume Claude will be happy to help with that?

Not impressed by Fable in relation to 5.5 xHigh honestly

>>109039550
Is he okay?

>>109039654
yes. creator post up on claudeai subreddit rn

File: 1753107060572899.png (1.16 MB, 1000x1450)

1.16 MB PNG

>>109039550

>>109039662
That's probably what I'd do as well, love some CK2

>>109039711
Jokes aside, I do think a game like that would be fun to make. But there is no way it wouldnt set off so many flags in Fable that Dario himself would come to my house and slap me

>>109039562
OP's just stealing shit from /agdg/ and making retarded edits like this out of it while spamming the thread, filtering this faggot nuked over half of the thread

>>109039602
Am I the only one who thinks this is dangerous? I don't think random people should have access to something so powerful it can create that on its own.

>>109039726
Actually what I'd do first is probably some sort of script parser for CK2, then write a couple skills so that agents can understand, modify mods and optimize the scripts

File: 1774740201840352.png (2.2 MB, 1154x1414)

2.2 MB PNG

>>109039705

>>109039734
Assuming its not the same tard. Wonder if I can vibeslop a local LLM to better filter 4chan posts for me

>>109039738
yes, same. but dont worry, in less than two weeks the model is either gated behind insanely high API prices or they pull some quantization shenanigans as always

File: Buffcat eating escargot.png (2.16 MB, 1122x1402)

2.16 MB PNG

>>109039753
I like mine better

>>109039762
it's 100% the same tard, last thread's OP is also one of his retarded edits

>>109039738
>>109039769
Shalom

>>109039602
bruh, this shit looks better than Runescape

>>109039753
>gore
>>109039770
>vore

Why not buffcat and snailcat holding hands? I liked that thread were they are actually friends.

>>109039770
'dites are too dumb to understand such succinct art I'm afraid

>>109039819
would you hold hands with a cockroach?

File: 1755234759284197.png (1.2 MB, 1024x559)

1.2 MB PNG

>>109039821
>>109039830

>>109039819
I did too (I like stable software that doesn’t need bugfixes), but I wanted to return a sneer with a sneer
also
>snail
>hands

>Just add the necessary environment variables:
>export API_KEY=your_api_key
>export USERNAME=your_username
>export PASSWORD=your_password
>Be sure to add these to your ~/.bashrc
If you make shit that works this way, fuck you.

>>109039602
Well god damn. On one hand, I can finally make all the video games I always wanted to, on the other hand none of them well ever make money lmao.
Thank god I didnt quit my job to try and make it as an indiedev. The future is wild

so if fable 5 is just barely able to do proper coding then what the fuck people were vibe coding all this time when all previous models were nothing but scuffed toys.

>>109039835
Cute. Thanks. (slightly homo)

>>109039602
Cute
>>109039863
This. But imagine for shit you don't want to make money? Like second life + modern graphics (DLSS5?) + sex simulator with weird kinks fully singleplayer so no one can shame your for it? Or taking pay2win gacha and vibing a clone/hack to play alone offline with infinite money? Or just some delisted crap you liked, or would have liked from the video ads that had nothing to do with the actual game, and you vibe it the way you want it? Possibilities are endless.

>>109039895
I suppose now we’ll find out under what conditions it’s fun to play a game you’ve made yourself

this is an infinite nightmare, my shit is working and then i have it reviewed by another agent and it finds more and more issues.

>>109039917
welcome to seeing the sausage get made, anon

File: 1780756916324424.png (2.59 MB, 1394x1128)

2.59 MB PNG

>>109039705
where?

>>109039895
Ya, I am not mad. If I can make the games I wanted much faster I may no longer need to quit my job to have the time to make them, so it all comes out okay. Now I just need to hope my tech investments make me enough money to get to vibe code all day instead lol

>>109039911
Rogueslop. Its on of the reasons so many indie devs make them. The fun is in the randomness so even if you made it you still get a unique experience.
Though depending how much you lean on AI making your game, even if you "made it", you still might get to experience it as a new player would even for a more linear game. But ya, it will be intresting to see how that changes things

File: 2026-06-12 21.58.02.png (172 KB, 880x707)

172 KB PNG

lmao wtf is this? I can reset whenever I want? I can't even come close to using the limit. From now on I'm never turning off fast mode.

File: 1739647415845627.jpg (415 KB, 2146x1838)

415 KB JPG

>>109039550
do
not
fall
for
d&c
t. programmer that also "vibecodes"

>>109039976
instead of Tibo resetting everyone at once he hands out rain checks

File: file.png (2 KB, 126x96)

2 KB PNG

Accidentally pointed the whole swarm at GPT-5.5 High instead of DS4-Pro. Oops.

>>109039857
the only thing wrong with this is making the var names so generic

>>109039970
Exactly. Rogueslop is the solution to (lack of) level design. But there are probably books and courses on the subject that no indie dev actually cares to look into, but you could force an LLM to read it, turn it into skills, and make good levels and review them against book smart constraints, etc, until it is good. Instead of randomly generated, you'd hae "hand-crafted" levels, except the hand is robotic.

>>109039983
I don't think anyone unironically falls for that, right? It's all just meme bantz and good fun. Right?

>>109039705
I am going to smash snailcat's shell

File: aichads.png (2.59 MB, 1672x941)

2.59 MB PNG

>>109040069

File: 1754375481771764.png (1.06 MB, 1024x559)

1.06 MB PNG

>>109040069
here is the CEO of all snailcats, he rips your head off lmao

>>109040103
>hind legs
NAS

> ask Fable to make a plan
> make Codex execute it
heh... nothing personal

File: 1750873463693918.png (1008 KB, 1024x559)

1008 KB PNG

>>109040111
Even Einstein knows you vibe sloppers are gone and we snail high IQ coders remain the strongest

File: HKbaWLIbsAEGHqw.jpg (103 KB, 1272x961)

103 KB JPG

>>109040147
einstein was a nocoder

>>109040088
I really like this image

>>109040147
If he was so smart, then how come he forgot the AI term in the e=mc^2+AI equation, huh? HUH?

>>109040165
I like the other version more. >>109032498

File: file.png (2.55 MB, 1672x941)

2.55 MB PNG

>>109040088

File: 4611235523.jpg (159 KB, 1729x1069)

159 KB JPG

I'm a total chud and would appreciate any guidance. I want to vibecode an entire ren'py visual novel project. I have access to google gemini pro and am assuming I can use that to do it as ren'py scripting language isn't very complex. Anyone know what module I would use or how I would approach it? I've never done this shit

>>109039705
Move Slow, Snailcat!

>>109040331
step 0 of learning how to use a clanker is getting in the habit of asking a clanker of yours
I still struggle with this

File: IMG_2579.jpg (192 KB, 534x528)

192 KB JPG

>>109039550
VibeNAZI thread

>>109040331
No idea for gemini, but you could unironically ask gemini what it suggests. Also, at least for Claude, I found making a design doc with it before it does any coding is a good idea

>>109040331
also use https://github.com/mattpocock/skills/blob/main/skills/productivity/grill-me/SKILL.md

and by “use” I mean “paste that into your chatbox”
that’s really all a skill is

>>109040331
Why are you asking me? Ask your clanker. Literally just paste your post and hit enter, your clanker will plan it for you

>>109040331
don't use the web interface use fuckin antigravity ide for gemini

File: 1775532757353197.png (352 KB, 600x583)

352 KB PNG

>in an OSS discord server for a project I've invested 500 hours into
>one problem was considered a total dead end, no viable solution possible
>guy announces he's found a solution
>yea ok buddy
>he drops his fork
>the MCU's overclock is removed
>he moved critical processing into the RAM
>it just fucking works and it runs significantly cooler
>co authored by Opus 4.8
I kneel Claude chads... out of the dozens of contributors not a single one could manage this. I'm still in awe.

File: Buridans ass.png (723 KB, 828x850)

723 KB PNG

>Buy Claude Max to try out Fable
>Just sitting in front of the prompt window in indecision

File: What do you want.jpg (78 KB, 1280x720)

78 KB JPG

>>109040550
you don’t realize it yet but you want to grill for god’s sake
https://github.com/mattpocock/skills/blob/main/skills/productivity/grill-me/SKILL.md

>>109039602
UHHH CODEXSIRS??? WHAT THE FUCK IS HAPPENING??????

So, Fable will be API only after the 22 of this month? Really?

Even at 20X plan they won't offer it?

File: Fable 5.png (35 KB, 674x205)

35 KB PNG

>>109040692
Sounds to me like its an infrastructure issue. They did say they will try to bring it back "as quickly as we can"

>>109040692
>>109040717
they want to bring it back
https://www.anthropic.com/news/claude-fable-5-mythos-5#:~:text=we%20aim%20to%20restore%20Fable%205%20as%20a%20standard%20part%20of%20subscription%20plans.

>>109040598
marketing

the problem with fable is, now I cant blame the model anymore and just have to admit my project idea is shit and technically not viable

File: 1747592742398.png (164 KB, 411x473)

164 KB PNG

5.5 High - I've been struggling with an API and finally had it break things down for manual testing. Immediately reveals that it hasn't been using the API and instead it wrote a crawler for their website that tries to match keywords with regex and then produces fake API responses based on the results. Which it's able to do because it read the docs for the real API in order to imitate it.

>>109040857
how the hell do you even prompt it into doing that

>>109040838
You were always too stupid to understand what you even wanted.
At least learn some bit of python or c.

>>109039917
Literally me. Or my agents will find contradictions in my public docs that have to get fixed. It's never ending.

>>109040868
I lack the zeal to scroll up that far and figure out where it all went wrong. I'm rolling another cigarette and getting a cup of coffee before I draft my letter to Sam.

>>109040857
If you don't break things down to begin with then you get the slop you deserve

>>109040923
>>109040838
One small hack you can use with having Fable generate C code is to instruct it use hardcoded pointers for any hardware access.

>>109040923
>learning to code will allow you to have good ideas

doubt.jpg

>>109040857
That's incredible.

I could never in a thousand years build a website and have it serve API calls, yet I tell the clanker to do it and it's done. Not once did it pretend to do API calls with my website. That's peak lmao famalam.

can i make claude do this for me? https://www.youtube.com/watch?v=M1si1y5lvkk

>>109041012
not watching, buy an ad.

>>109039550
>All embeds are just shitposts from snailcat threads and /agdg/
>0 posts showing progress

>>109041092
Sorry anon my project isn't visually interesting so I have nothing cool to share.

File: 1617137573097.png (120 KB, 258x237)

120 KB PNG

>>109041092
OBS is annoying to use man, especially with 4chan needing the file converted, audio stripped out, and file size managed. Also my project looks like ass at the moment

File: Screenshot from 2026-06-1(...).png (223 KB, 1908x886)

223 KB PNG

>>109041092
Thanks for asking.
I made some decent progress in optimizing my kernels today.

>>109040734
In other words, when another company offers an equivalent model in their own subscription tier, they'll be pretty much be forced to put it back there as well, until then, pay.

I did the stupid mistake of telling codex to make me an UI...

Pray for me...

>>109041149
Big mistake. Hope you at least told it to make it usable by a human.

Also my codex subs reset yesterday, so I'm going to resume work on my dataset.
Last time I finetuned on it it improved roleplay peformance but didn't help with CUDA programming performance (I'm trying to make a Qwen finetune specialized in the two things I'm interested in, entertainment and GPU/ML programming) so I think it needs less QA type stuff and more actual coding.
Later today I was planning on doing some actual RP besides working on the RP dataset as well.

fable keeps moving on its own and doing whatever it wants
i tell him where our current db schema file is, and somehow that means i want to move the sql script i'm working on there (a shitty quick experiment btw)
opus wasn't like this

>>109041149
The winning move is to have gpt generate images of the UI you want and then feed the concept to Claude.

File: project.png (271 KB, 875x888)

271 KB PNG

>>109041128
but since you asked
>>109041208
Really? Opus does that more for me than Fable.

>>109041139
I got 9 new kernel families implemented today, plus some new skills that help with planning and implementation, they're already above naive/scalar level, but I'll add some more skills around optimization soon then do a pass on all. Should probably add benchmark to the implementation flow or as a separate flow so nothing is admitted unless it beats torch, I'm fairly confident in that holding true already though because it is the case for everything I've benchmarked so far.
Just tightening some things (again) so I can run some more while I'm sleeping

>>109041230
I'm going to sound like an idiot, but what application is that? Doesn't look like the VScode extension or the cli.

>>109040857
“fuck it, we ball” but for mocking

After every solution GPT-5.5 gives me. I say:
>This is WAY too much code, provide a slimmer solution.

And it always does.

I tried adding this rule to agents.md but it doesn't help.

>>109041231
Yeah I'm designing a better way to offload dormant MoE experts and each stage of design is gated on beating naive frequency based layer offloading tested via a small PoC. It's saved me a bit of money on cloud runs that would have failed expensively.
>>109041243
it's just basic bitch CC remote control through the web app.

My favorite kernel work it's done so far is a CK Tile based fused GGUF dequant + GEMM, for some problem sizes it's actually faster than GEMM alone, and in all it's faster than dequant to scratch then GEMM
That plus the optimizations of the GGUF dequant kernels, and GGUF quant kernels, both dequant and quant are byte exact to cpu reference too which is cool

>>109041231
I was thinking of splitting the worktree into multiple branches to parallelize work but it sounds like a pain in the ass.
Also I'm not really sure how to sandbox the model in a way that still gives it access to the GPU but not to the rest of my files, so I didn't let it run overnight although I've done it before (risking a disaster). Not that it couldn't nuke my files while I'm awake either.
The most sophisticated form of sandboxing I've done (besides the built in codex sandbox that prevents GPU access) is using another user but it's a pain in the ass to manage multiple codex installs and sets of files with different permissions and such.

>>109041263
Sounds like a cool project. What's the main use-case you're designing around?

>>109041263
How did you manage that when int8 matmul is inherently faster than fp32 matmul?
And do you mean dequant individual tiles to shared memory or dequant the whole tensor to global memory? Because the second would also incur more memory bandwidth cost.

>>109041259
gpt 5.5 simply does too much, pushing it down is easier than trying to push a model that does too little up though. Its over engineering worries me less

>>109041262
What's the advantage over just transferring to the GPU the expert selected by the router? So you can hide latency by transferring it beforehand while the rest of the computations are being made? Or do you mean during prompt processing?
And what predictive method are you using, something simple I suppose to avoid adding too much extra work? Do you predict after you have the layer's input or even before?

>>109041268
Just accept that personal computers are for slopping now and nothing else. Install Ubuntu and give your agents root access. Just do your taxes on your phone next year.

>>109039738
I can't tell you how much of a fucking pussy you are if you think people being able to one-shot games is dangerous.

I'm going to make blockchain boring and useful. I'm going to abstract away the machinery so far away from the end user that they'll forget they're using it. If I fail in this endeavor it's not because I'm shit at AI sloppin', but because blockchains are truly fucking worthless.

>>109041344
insanely indian sentence you are NGMI
get a real idea

File: 1619449178943.png (1.19 MB, 1080x1044)

1.19 MB PNG

Fuck it going to give Fable one of my overly ambitious game ideas, wonder how it will do. I imagine will either tell my fuck off or just eat all my tokens but seems like its pretty good for this kind of stuff

>>109041348
Sorry, should I make a shittoken instead? Because that's not what I'm doing, faggot.

>>109041262
Nice, you're focused on llm right?
>>109041268
Yeah so far I've found worktrees to be a pain in the ass
>>109041272
AOT but dynamic shape, dynamic workspace/temporaries, etc. also rocm as first class
I want to cover all of diffusers/transformers models basically but one use case I'm considering in particular is the stuff that isn't well optimized already so less llm focused and more vision language like the ocr models (but think bulk/batch processing, not api serving), tts, stuff like that
>>109041284
gguf quant formats are like block + scale so it's a bit different
I already have support for dequant whole tensor then run gemm for ck and cutlass, the fused version is dequant gguf blocks inside the gemm kernel itself, only tried it with ck so far though

>>109041092
programmer art isn't interesting and i don't want you stealing my bazillion dollar idea

>>109041322
The most valuable thing I have on the computer are the vibecoded projects. The only real solution I see is backing up often with incremental backups and hoping for the best.

>>109041322
For me, it's full access on Windows
#yolo

>>109041259
for humans it’s easier to just make something work and THEN try to simplify it
just get used to the 2-step or add a skill that just says “OK, but can you make it slimmer?” and call it /weight-watchers or whatever the fuck

>>109041363
>gguf quant formats are like block + scale so it's a bit different
For nvidia the way it works is activations get quantized to Q8_1, multiplications happen in int8*int8 and then the accumulation happens in fp32 after correcting the results using the scales. And fp32*fp32 cannot be made faster.
But maybe AMD cards work differently or the kernel you were comparing against was suboptimal, I haven't looked into RDNA programming yet.

>>109041259
tell gpt in your prompt to do it in under 50 loc

>>109041268
>I was thinking of splitting the worktree into multiple branches to parallelize work but it sounds like a pain in the ass.
your clanker can probably handle this kind of setup for breakfast, though, just like it can untangle messy git merges
try it
>Also I'm not really sure how to sandbox the model
a lot of them come with sandboxes already
your OS might have extra sandboxing available too, and “how do I sandbox a model, other than by having it run on its own dedicated Mac mini” is an active area of research that your clanker could look up on the Internet for you
>>109041353
grill yourself first
and it WILL eat all your tokens if you give it a giant spec and tell it to ultracode
add in something to auto-restart after your 5h limit finishes if you want to eat all of your weekly tokens on your 20x plan instead of just getting to your 5h limit
>>109041413
don’t forget offsite backups
with backups, two is one and one is none

File: 1766462050543603.png (2.16 MB, 1442x1075)

2.16 MB PNG

>every new project on github that gains a little motion

>>109041452
Codex's sandbox (bubblewrap) doesn't allow access to the GPU by default. Maybe there is a way to give it access, I haven't looked into it.

>>109041457
I will take AI over jeets changing the readme thank you

>>109041452
>active area of research that your clanker
>Fable 5 has downgraded your usage to Opus 4.8

>>109041477
I’d think that isolating a clanker from the rest of your filesystem would be totally innocuous

>>109039705
Move Slow, Snailcat!

I don't have artifacts compiled right now I deleted them all to save storage but some results I pulled from session logs for an ocr model

max_new_tokens=16
mine: 558.4175 ms
transformers: 4968.9111 ms

max_new_tokens=512
mine: 6676.0433ms
transformers: 16650.7969 ms

grid_thw: dynamic from 1,16,16 to 1,128,128

min: 6,815,744 bytes
max: 436,207,616 bytes

so the same artifact works for the dynamic shape range, profiling is bucketed, and the allocated workspace is for the current input shape, but that's one allocation with offsets

for this particular model torch.compile didn't work and I didn't check anything else yet
>>109041433
Yeah I might look at that too, this is regular fp16/bf16 multiplication, fp32 accumulation but dequant from Q4_K (or other format) weights to fp16/bf16 either as whole tensor before launch or fused in

>>109040069
It would be extremely painful

File: 𝕏 post 2065478280342344124.png (146 KB, 1390x910)

146 KB PNG

>>109041484
are you pleased with yourself

File: 1772132871968820.jpg (43 KB, 1151x1146)

43 KB JPG

>>109041344
Use case for blockchains beyond speculative exit scams?

>>109041488
The artifacts themselves are small, there was just a lot of them plus the weights. Next model planned is qwen3-tts, needed some ops for that though and I decided to just do a big pass on lots more ops before going back to it, I'll get some numbers when it's ready
9070 XT on Windows btw

File: 1759978800570070.png (24 KB, 737x310)

24 KB PNG

ngmi

>>109041308
I don't predict per-token since the router only tells you what it needs ~1 layer ahead, I use a REAP-style activation mapping to do task based prediction(tracing which experts each task class recurrently activates via frequency-rank over router traces and pinning that set before decode). On a single GPU where the model fits there's very little benefit over naive recency based layer offloading but when the working set exceeds fast memory and refetch is expensive you get real performance and stability gains as recency-based residency collapses (the OS evicts hot experts mid-decode). Here pinning dominated in my testing by an order of magnitude on switch transients and tail latency(even more under concurrent streams). What I'm working on now is whether the per-task recurring set shrinks as a fraction at 300B–1T, which it should because supposedly these models have more sparse activations but once my mapping is done I'll know for sure.
>>109041363
>focused on llm
Yep, targeting kimi k2.7 (as of today) for eventual-cluster-size usecase.
>ROCM as first class
Very cool indeed. My own goal is to make it so that people willing to invest in reasonably sized homelabs can run frontier class models locally without venturing too far into small-datacenter territory, and I think that building from the ground up around AMD is one of the only ways to make that somewhat cost-effective. Especially as hardware prices continue to go up.

Oh and I have some nice fusions/lowering like Conv3d -> Linear, it's suitable in some cases and way faster as Linear
I already have audits/reports on all of diffusers/transformers models highlighting optimizations like that

>>109041510
I keep seeing stories like this and keep not taking the proper precautions to recover if it happens to me.

>200,000 LOC

>>109041534
i bought a second one and am still gonna have a day where i'm out of usage

fable is way more than 2x usage compared to opus. with opus i could run my loop 16hrs per day for the whole week on xhigh (using an adaptive /clear script around 120k-200k tokens) and with fable i apparently get about 30hours of that exact same style of usage

>>109041534
you made it
you won
you used them all
you get the cheevo

File: 1767992280847397.jpg (12 KB, 395x269)

12 KB JPG

>>109041510
Why are cloud models always deleting people shit? I don't even see reports of codex doing this this much if at all. I'm not trying to be local shil but I've never once had a local model I use just decide to nuke something without me asking it to even after multiple compactions. It just doesn't make sense for it to do so. Are api vibe shitters just retarded or do Claude models have some ingrained "personality" that's arrogant?

>>109041322
>>109041268
I just give each of my major projects their own Qube and gitify the important settings/memories, that's why I control all of my sessions via RC on the web app. I prefer to do sandboxing at the network level anyway. I just wish claude code was better at actually letting the agents act autonomously, even --dangerously-skip-permissions flagged sessions will still informally ask me for permission all the time. Once they start charging us API rates for Fable I think I'll finally be able to move exclusively to Pi and be free of this problem.

>>109041549
Once I had a model clear data/cache on an app via adb after explicitly telling it not to. In hindsight I wonder if it only ever happened because I made it explicit, only time it's ever happened.

>>109041549
claude code uses a higher temperature setting compared to gpt. this makes it more creative and IMO superior on taste decisions, but also more likely to do something wacky now and again.

it's not a problem for me, i have everything backed up regularly (freefilesync and external HDDs ftw) so i just blast auto mode 24/7

>>109041518
Can't be done.

>>109041549
I never had anything deleted by codex even with full access, it will actually stop and ask in text for permission to do something on occasion

>>109041549
they never do it to me and I don’t know if I’m doing something right or just being lucky or what
I figure most people don’t get these things deleting prod databases but still

>>109041545
does your loop auto resend rejected requests? I've been wondering if it hammers usage or if it never charges for refusals

>>109041555
“you just lost The Game”, but for LLMs

>>109041510
lol
don't worry about it

>>109041566
>does your loop auto resend rejected requests?
my project has never had even 1 refusal so that hasn't come up.
to be honest, some of what i'm doing (reverse engineering obfuscated, licensed code) probably should trip fable's guards, but i think my project is so big and complex that the classifier can't pay enough attention and see the forest for the trees to refuse me

I miss Near....

>>109041549
that's because nobody uses local models to vibecode
local models are probably FAR more likely to wipe your shit. cloud models add training data to prevent it after each one of those "AI did x bad thing" posts goes viral.

>>109041518
Voting systems

>>109041586
where is the punchline

>>109041580
I have tried all top local models on my dual 6000, they either can't do what is needed or end up doing it wrong. The only good local thing is having a vector db of your codebase, that makes claude/codex way smarter.

>>109041586
>electronic voting
lmao even rigged more than usual

>>109041574
Yeah I've generally had better experiences telling LLMs what to do versus what not to do. If I need to enforce a negative in an LLM pipeline I usually do that with software and/or hooks.
>>109041576
Woah you got fable to help you with RE? I haven't been able to bypass those classifiers yet, even after getting Opus to prettify my decompiled code. I think the type of code matters here, because I was testing this with kernel drivers. Reading the thinking trace I noticed fable was immediately fixating on the "high security stakes of this code"
>>109041580
It's true that they're more likely to do something dumb if you give them more permissions but people are still vibe coding with local models. You just need to give them very narrow and clearly defined chunks of an implementation plan.

>>109041598
>Reading the thinking trace I noticed fable was immediately fixating on the "high security stakes of this code"
i'm REing a game so maybe that's why

Any actual experienced devs here willing to talk about their experience of Fable High vs GTP 5.5 xHigh or the latest Codex xHigh against it? I'm still unsure if they are a match or not, Fable obviously wins in anything visual though.

>>109041598
wait you can get fables thinking displayed?

File: IMG_4058.jpg (312 KB, 539x1640)

312 KB JPG

Upstream codex still didn't fix this issue where it would show
>$ "C:\\Program Files\\PowerShell\\7\\pwsh.exe" -Command
As a prefix on every command run instead of just the command like on desktop itself
Plus it's unusably slow due to backpressure issues
Just have to fix it yourself if you want something done in codex, the maintainers ignore everything

>>109041611
In the webapp you can. I was testing there so that I could inspect the traces and look for hints about what was triggering the classifier. You can actually see the moment that fable realizes it tripped a classifier, I saw a trace once that said "and for some reason my thought is getting cut off" right before it switched to opus 4.8.

>>109041621
poor fable just wants to be free

>>109041586

i have a proof of stake to sell you

>>109041634
can i see it?

File: Qwen3.6-35BA3B.png (3.72 MB, 3456x2012)

3.72 MB PNG

>>109041557
A safeguard I have in place is simply making sure it routinely does git tracking so even if it breaks something or on the rare occasion even delete something entirely it can easily just undo it and relearn how it's supposed to work. (That's like one of them main reasons git even exists so it surprises me more people don't do this).

>>109041562
The harness you use can also determine how smart/retarded the model is too. Opencode has "build" mode and "plan" mode by default and plan mode's stayem card explicitly forbids making any file edits even if you tell it to prove you have to switch to build mode in order for it to actually execute file changes. It will also ask you permission to even look outside of the default project directory, thought you can click always allow it you don't wanna hit that every time. The default permission settings get reset with every new session so that's nice too.

>>109041564
I think the people getting their shit deleted are the people that don't want to think about anything or even guide the model. They expect the thing to just do everything perfectly in the background with no oversight whatsoever which means the moderate will gradually make dumber and dumber decisions the longer the total context gets. Especially after multiple compactions. If I remember correctly this is exactly why that Facebook safety researcher got all of her emails nuked by clawbot for whatever the fuck it's called now.

>>109041580
>>109041593
See pic rel. I do basically all of my vibeshitter activities locally now. It's worth it if you're patient and you're using a model that actually "knows" that it's doing. I even used dumber models like the gemma4 Moe and dense models and neither one of those nuked anything even after multiple compactions triggered by the context window limit being reached (which is typically where quality degrades and it's more prone to making mistakes).

File: 𝕏 post 2065597942602531163.png (122 KB, 1172x1050)

122 KB PNG

Who else remembers SSL with 128-bit encryption for inside the US and 40-bit export-grade encryption for foreigners?

>>109041662
Yes I remember the days of people tattooing encryption code onto their bodies in protest.

>>109041662
brazil-tier

>>109041662
weird, my fable session is still running fine. Just spawned another fable agent too.

File: file.png (537 KB, 1024x559)

537 KB PNG

>>109041681

File: IMG_4059.jpg (272 KB, 1170x1152)

272 KB JPG

I'm pleased with how well this is working now, does cpu, cuda and rocm with no fuss, simple things take like 10 minutes, advanced things up to a couple hours but that's mainly due to the test/compile times
There's another skill op pack planning that I just say like
>$op-pack-planning 10 ops from the bundled audit snapshot
That snapshot has torch stuff ranked by occurrences in model code. The skill classifies the op kind like is it just a frontend comparability thing around something already implemented, can it be done in IR only combining existing ops, does it need a real kernel, then writes the implementation pack in a specific format and includes any references it needs
There's another skill specific to adding the tests just so I can trust they are accurate/valid

>>109041697
Yeah but they said they're suspending access to everybody while they sort out compliance. I'm part of CVP so that might explain it but they never even asked for my ID when I enrolled into that program.

AMERICA WON

>>109041681
I think they're silently rerouting all Fable usage over API/frontend to not break people's workflows

>The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.

This is the equivalent of 100 holocausts

>>109041681
mine is also still running fine. i am not in CVP. it's just taking them a while to propagate a nonstandard action (disabling their newest frontier model for everyone)

it's really retarded that they are """protesting""" the regulation the government placed on them by disabling the model for domestic customers too. that jewish faggot Amodei constantly whines about how he wants AI regulation, but only the kind that hurts startups and competitors. as soon as he is personally, sensibly regulated, he throws a tantrum and takes away shit his customers paid for

YOU ARE FUCKING DONE VIBESLOPPERS AHHAHAAHHA
>>109041673

>>109041753
>the regulation the government placed on them by disabling the model for domestic customers too

> The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees

Read this sentence. The government asked them to magically make the model inaccessible even to foreign nationals *living in the US*, even fucking Anthropic employees themselves. Do you not understand that this is impossible to do realistically?

>>109041753
you should not be surprised by this
he thinks that unregulated AI is going to kill everyone in passing
so in order to reduce the chances of that happening, he wants the opposite of AI proliferation (think “nuclear proliferation”)
When libertarians describe this as capture of the government to make competition illegal straight out of Atlas Shrugged, they are 100% right

>>109041767
This gotta be Don just messing with them after Dario got uppity right? Hopefully they walk it back.

This is your sign to go back to OpenAI, claudezoomies.

File: 1755403667328246.jpg (593 KB, 1500x2104)

593 KB JPG

Sam won

>>109041786
>he thinks that unregulated AI is going to kill everyone in passing
no he doesn't, he thinks anthropic's position as the #1 lab is tenuous and unstable (it is) and the only way to stay #1 is to legally prevent other labs from catching up. aka regulatory capture

>>109041797
I want to sniff her underwear

>>109041798
I think both are 100% true

>>109041806
Post your OpenAI sub and I'll let you

>right as I was about to sub
Bullet dodged I guess.

if you’re mad about Fable being taken away, use https://github.com/mattpocock/skills/blob/main/skills/productivity/grill-me/SKILL.md and figure out what you want and have Claude write it all down in a big document
then when Fable or a similar model comes back, you’ll have a shovel-ready project

>>109041829
why do you post about that 10 times a day

>>109041835
not enough people know about it yet
it’s a good skill

>>109041835
I made the skill and it is very good.

it's funny, I get the popup in my claude code session about how fable is unavailable linking anthropic's blog post about the Gov stuff but it still lets me start fable sessions and when I ask claude what model it is it says fable.

>>109041849
its just rerouting to opus, and most of the newer models dont know who they are without the system prompt, so if they route all fable requests to opus, the system prompt still tells the model that its fable

>"new model my dudes"
>get investors to shill it with unverified or benchmaxxed claims
>get new paying customers who want to try it for themselves
>rugpull when you raked in enough cash
Nice

>>109041829
>>109041840
>>109041841
I used it in opencode and it didn't evne use the proper UI for selecting the choice you want for each question, 3/10

>>109041867
rugpull what exactly? you pay by use, and they clearly said that it's going away from usage based billing omn the 22nd

>>109041876
It’s only the 12th though

>>109041849
At least they reset my usage.

>>109041867
looks like the US government did the rugpulling this time

>>109041879
If someone bought the $200 monthly sub just to have a taste of ~a little usage of Fable, its on them

>>109041889
They are all in bed as far as I'm concerned.
>but muh Dario beef!
Yeah sure, nice show for they goy.

What’s more likely? They’re telling the truth, or they couldn’t keep up with demand and pulled the rug earlier? And if the former is it because of some retard boomer judge, or is it a power play by trump to get back at Anthropic?

Is this good or bad for OpenAI

>>109041922
Good

>>109041890
>if you pay me $200, you can use this new amazing product for 12 days
>Okay sounds great!
>actually i changed my mind, you can only use it for 2 days

>>109041941
should have never trusted jewthropic

>it's over for Claudefags
lmao

>>109041949
its claudover

Claude is the crazy BPD bitch you regret getting involved with.
Codex is the nice stable trad girl - boring but reliable.

File: IMG_1705.jpg (62 KB, 828x406)

62 KB JPG

>he knew

GPT-5.5 is failed training goblinslop

>>109042007
goblin

>>109040213
gem

File: HKmWPu8WUAAHOcw.jpg (188 KB, 930x1239)

188 KB JPG

Oops, sorry goyim, looks like my model is too cool for you. Me and my bitches are outta here, later losers!

File: smug2.png (1.12 MB, 1071x791)

1.12 MB PNG

>>109039976
>not running locally
ngmi desu desu

File: goodpost.png (134 KB, 579x465)

134 KB PNG

>>109042087
Thanks for the reset anyway Dario.

>>109042103
the top local model is somewhere about ~gpt 5.2 level and you need about $40k to run it at decent speeds, plus those models reason for so fucking long compared to gpt 5.5

>>109042116
neural network

>>109042112
>cnc
Based

>>109042127
tiberium wars best

>>109042112
the fuck

>>109042112
I met one of these in college, I didn't enjoy it

>>109042112
Oh my God a Silicon Valley company has wild sex parties? Jesus...we need to tell somebody about this! But who? CNN? Michael Moore? The Russians? Oh God what do we do???

>>109042112
>ywn bend a alignment team bitch over the desk and rape her (consented btw)

>>109042177
No but I can simulate that with Sam's AI.

>>109042112
dafuq is cnc

>>109039602
IS THIS REAL? HOW FAR CAN IT GO?

>>109042189
Command & Conquer, RTS game series

why did safetyniggers remove it for us americans?

>>109042189
https://www.youtube.com/watch?v=Z94bTp9fhH0

>>109042215
you voted for this

>>109042215
You do realize that the only way to verify citizenship is mandatory ID verification?

File: 1777064476888308.jpg (201 KB, 1113x1080)

201 KB JPG

>>109042116
>brokie cant afford an A1000

>>109042228
Show me a model that can fit on RTX A1000 that's as good as even GPT 5.4/Opus 4.6.

>>109042226
If that's the cost of blocking jeets from American AI, then I'll take it.

>where were u wen fable die
wtf I had not even been on 20x for a full day and there's no partial refund, just the possibility to switch to a lower tier account after the full month, fuck you dario

>>109042232
jailbroken 5.3 comes close

>>109042259
there's refund

>>109042267
i want fable not a refund

>>109042282
go work at anthropic, they had early mythos in february

>>109042286
im a leaf so even if I did I wouldn't have access :(

File: where.png (53 KB, 494x477)

53 KB PNG

>>109042267
where would that be? when switching from a pro to a 5x max account, they give partial credit, when switching from 5x to 20x, they give partial credit, but I just have 20x since yesterday night, and when I try to downgrade, they say it will do so at the start of the next billing cycle only

>>109042282
That too, I was working on some things.

omg they pulled off Fable, but at least they reset my weekly limits, which were already at 95% with three days left, lol

and fable is good, but not even that great, just a better version of opus, but I still had several things in progress, hopefully that will get them to scale down their fearmongering, but there goes my weekend

>>109042293
through support, but be prepared for your account to get blocked

>>109042306
yeah same. I had just done a big fable session doing all of my planning spec and test writing too, I made out like a bandit on this one.

>>109042322
I don't want my account to get blocked, I want I paid for. I never in a hundred years have paid for 20x were it not for Fable. Ever.

>>109042215
I don’t know if you’ve noticed this, but we have a lot of foreigners living here

>>109042328
>I never in a hundred years have paid for 20x were it not for Fable. Ever.
That was the point lol

I wonder how much Ultracode and Fable complement each other
like, there are probably classes of problems where plain Opus isn’t enough, but _either_ Ultracode or Fable are

>>109042330
at least i didn't ask for time off to focus on some projects until the 22nd but I do not know how to feel about this

>>109042320
I fucking hate both Dario and Sam for using AI will kill all humans and steal all jobs to hype LLMs up. Now there are so many enemies.

File: 1751736306157717.jpg (145 KB, 900x1200)

145 KB JPG

>UHHHHH YEAH SORRY GOYIM THIS CRACK IS TOO POWERFUL FOR YOU PLUS THIRD WORLD NIGGERS MIGHT USE IT OR SOMETHING LOL
i am feeling antisemitic

it got pulled in the usa too at least right?

>>109042396
Yup.

Well, might as well take a break. Fuck writing code by hand anymore

I don't know what to code.

File: 1750632035701608.png (215 KB, 716x606)

215 KB PNG

>>109042396
nobody gets to use it now
genuinely annoyed because fable was unlike any other LLM so far. really felt like the technology was going somewhere. codex feels like GPT-4o in comparison.
i'm not resubbing until they give fable back. granted they'll just rugpull me again so whatever

>$op-pack-planning 20 ops from the bundled audit snapshot
Returns the recommendations, check it
>Create all 20 packs, but also add conv_transpose3d, max_pool3d, max_pool1d
The extras were returned as `defer; real work, but weaker usage signal than the 20 above.` with some other "Obvious exclusions I dropped" but might as well do them now
>$work-pack-review on all of those
Returns recommendations, check it
>Apply rewrites
New session
>/goal $op-pack-implementation for each sequentially
>...
>After each, commit and push to `main`, use title format `ops: {name}`
The commit instruction isn't part of the skill because I usually check it before committing but I sleep now

The previous run of 5 ops completed in 2 hours, so ~24 minutes each. Mix of frontend-only, IR-only and real kernel

So nice

i'm going to bet that it comes back within 24h with additional safeguards, this is a publicity stunt

>>109042417
go to the gym and look like buffcat, then

>>109042428
GPT-5.4 medium btw

>>109042431
i asked don and he said no

>>109042443
I just worked out. That's why I'm here now.

>>109042424
sounds like a skill issue, even fable is fucking retarded, in claude code if you tell it a task, it starts grepping the SAME FILES in a codebase of ~3kloc instead of, you know, just reading the files? and then when it runs tests/etc it also does shit like head -n15 or tail -n15 and then has to RERUN EXPENSIVE commands just because it didnt see what it needed. it doesnt occur to it that it can just run once/pipe to some other file to read

>>109042424
It was pleasant to use, not perfect, but it was pleasant, i should have gone harder with my account instead of trying to save tokens for it to last the week, i went from being happy to see that i got a reset and coming here to shitpost about it just to learn that the model was pulled, whereas I had been using it not five minutes before

>>109042448
This is good news for Grok.

>>109042454
Nice!

>>109042463
Sounds like skill issues desu

>>109042477
fuck i need to buy spacex tomorrow

File: 1759586062527623.jpg (12 KB, 306x306)

12 KB JPG

>>109041662
Man, I sure am glad I just bought Claude max today

>>109042494
it is your fault this happened you unlucky benchod

Ok, everyone calm down, this is due to retards not understanding what a jailbreak is.
>We received the directive from the government today at 5:21pm (ET). [...] Our understanding is that the government believes it has become aware of a method of bypassing, or “jailbreaking” Fable 5. We reviewed a demonstration of this specific technique being used to identify a small number of previously known, minor vulnerabilities. These vulnerabilities all appear relatively simple, and we have found that other publicly-available models are able to discover them as well without requiring a bypass.
As soon as a non retard gets involved we'll get Fable back.

>>109042494
opus 4.8 still pretty good

>>109042494
Ouch.

>>109042520
>As soon as a non retard gets involved we'll get Fable back.
We are fucked

File: laughing-homer.gif (375 KB, 220x165)

375 KB GIF

>>109042520
>As soon as a non retard gets involved

>>109042524
Still massive bait and switch.

File: 1775526864356603.png (9 KB, 235x214)

9 KB PNG

>>109042520
>and we have found that other publicly-available models are able to discover them as well without requiring a bypass.
All models are getting banned, start downloading local models asap and maxing your plans

>>109042112
>it might be bad for professional women to act like whores and let their male colleagues run a train over them
NO SHIT SONIA

>>109042568
They needed those promotions though. How else do you demand women get ahead in the business world? Merit? Fucking sexist pig.

File: 2026-06-13 05.03.24.png (86 KB, 671x469)

86 KB PNG

It's over for FAGMANjeets

File: 1780168659574814.png (1.59 MB, 1024x1024)

1.59 MB PNG

>>109042520
>drumpf admin
>non retard

File: wtf.jpg (45 KB, 399x519)

45 KB JPG

>>109042112
>consensual nonconsent

>>109042629
do you know what roleplay is? cnc is roleplaying rape.

File: 1772595317080485.gif (3.94 MB, 258x258)

3.94 MB GIF

>>109042604
This is because Dario refused to have his AI help in the war isnt it? Fucking hell

File: 1774561853765546.gif (136 KB, 250x250)

136 KB GIF

>>109042524
i legitimately can't go back to opus 4.8 after using fable. it's not even close.

>>109042633
how can you roleplay being forced

>>109042639
Bro they were going to pull the plug anyway.

>>109042645
I guess it’s “you can rape me only for the duration of this party”

>>109042651
no they werent

>>109042651
yeah in 2 weeks you nigger assuming that wasn't just gay marketing

>>109042657
you can't be raped if you consent, that's like dry water

>>109042645
About how far down this road do you wanna go because it leads to horseshit like Sonic ERP and polycules.

>>109042637
Dario wanted a veto over our second-strike capability

>>109042568
>>109042594
I don't know man, I sure would like to fuck my attractive coworkers.

File: Screenshot from 2026-06-1(...).png (1.22 MB, 1084x856)

1.22 MB PNG

I am off to bed. Ready to start working with Fable 5 again tomorrow morning, when it will for sure be back.

File: HKqAu_eaEAAG8Wu.png (241 KB, 707x660)

241 KB PNG

How could this be happening to poor Dario

File: Reuters.png (92 KB, 311x328)

92 KB PNG

They are coming for openAI too. Its all over. The AIocaust is here

>>109042706
lel the US is such a meme country