About time, he better never do this again.

File: HMGXjRhXQAA2yBq.jpg (415 KB, 2210x2528)

415 KB JPG

>>109172049
https://x.com/AnthropicAI/status/2072106151890809341

File: 1768695938114216.png (153 KB, 320x394)

153 KB PNG

save me sam altman

>>109172066
>he better never do this again.
He got everything he wanted, ie. complete regulatory capture. Every LLM company now has to ask the government commission on AI (with Anthropic as an advisor) for permission to release models

He never even wanted to release Fable, it was his employees who wanted that. In his vision, fable-level models are only for price-gougeable government & enterprise customers, not consumers - and the goyim should be forbidden from releasing fable-level models that can compete with his

>>109172086
cmon dario enable access

>>109172099
ok i will, sorry

>>109172105
ty appreciate that

>>109172089
This is still a massive win for Sam.

>>109171959
That's really cool, ive made some finetunes of smaller models but I want to try my hand at a larger one so I'll have to look into that.

New China model https://huggingface.co/meituan-longcat/LongCat-2.0
Will be interesting to compare to GLM 5.2, GPT5.5, Opus 4.8, and hopefully Fable once Dario takes his thumb out of his ass and re-enables the damn model.

File: file.png (51 KB, 893x495)

51 KB PNG

>>109172136

anyone here mess with devin is it any good. just tried to get it to refactor some repos but i'm confused on some of the docs it says it can do "desktop automation" what does this mean? is it able to automate testing/development on gui applications?

>>109172136
is it better than fable?

>>109172086
FABLE BACK
FABLE BACK

>>109172158
What? Fuck no, absolutely not, get away from that scam. It cannot do anything of interest that cannot be done with other freely available and superior tools. Their pricing as it folds in access to providers is SHIT and the tool they sell you is *nothing*.

>>109172136
>>109172150
nigga no one care fable be dominating the headlines this week

they really got to launch fable twice

>>109172171
what is the best option for parallel multi-agentic stuff with desktop-level automation going beyond the terminal or ide environment then? i only just discovered this was a thing and was curious. the last time i even tried this was with claude desktop but it wasn't that good and that was a couple years ago.

>>109172150
>worse than GLM-5.2 and over two weeks behind
better luck next time, Xi

File: 1755500321835128.png (29 KB, 975x427)

29 KB PNG

>>109172176
except this time it's a neutered version of Fable 5.0 that perform far worse on evals (even though it's confirmed they internally have Mythos 5.1) and you don't get any usage via subscription. enjoy paying $600/day if you want to use it, and don't even think of doing anything even tangentially related to reverse engineering or security

>>109172172
I'll give a shit about fable if they ever release the weights or at least stop safetycucking so hard. Why are you so eager to give your money, data, and dignity to a man who sees himself as your babysitter?

>>109172190
because I need the power I NEED IT BADLY

>>109172184
codex desktop has by far the best computer usage. it works best on Mac though unfortunately. on Windows it can't do background PC usage

>>109172189
still important that they release it again so chinaman can continue to distill it.

>>109172184
Claude Desktop is fine, Codex App is fine, Pi is my preference, OpenCode is good, though Pi and OpenCode are TUI, not GUI. Pi can be upgraded into whatever the fuck you want.

>>109172196
>because I need the power I NEED IT BADLY
Do you need it badly enough to pay $600/day lmao

I feel like a lot of teenagers here are going to, for the first time, realize how expensive it actually is to use an Anthropic model at API prices

File: snailwhoreslaughing.jpg (720 KB, 1448x1086)

720 KB JPG

>he still thinks he's gonna get access to Fable

>Your subscription ends in 3 days
Nah, I don't think I care.

>>109172150
second best logo behind the DeepSeek orca, I’ll give them that

>>109172184
I switched to oh my pi fairly recently it's bretty nice

File: clauron.jpg (398 KB, 2644x1600)

398 KB JPG

>>109172196
There is only one lord of the models and he does not share power.

man, why are indians so salty about API prices? just stop being poor and set up your local rig.

>>109172229
I've been considering oh my pi, mind sharing your thoughts on it so far?

Anyone try out the new sonnet yet? I am traveling and haven’t had time today to give it a few tests

>>109172265
really like the advisor feature, having a second model validating the work takes a lot of the effort out of keeping the clanker in line. interface is fine, doesn't keep the pi simplicity but that comes with the feature bloat.

>>109172189
that's a good thing though, letting any joe doe have access to such level of power was a terrible idea

File: 1763619015918331.png (279 KB, 500x400)

279 KB PNG

>>109172317
>that's a good thing though, letting any joe doe have access to such level of power was a terrible idea

>>109172033
Didn't I tell you to remove yourself from the gene pool??

>>109172297
So far it's worse than GLM5.2 but costs significantly more, which makes it pointless. It is a considerable step above old 4.6 Sonnet though. I wouldn't be surprised if it's punching in 5.4 / Opus 4.6 territory, which would put it on par with GLM-5.2 just at a higher price point.

>>109172310
That's one of the most appealing features to me, something I could add to Pi admittedly but why bother if I can just try omp. Thanks, anon.

>>109172196
Wasn't Fable refusing to do anything cyber security related to non in-group plebs like (You)? Any supposed amazing capabilities are nerfed by it's baked in safety sloppage

>>109172343
I’m not interested in cybersecurity, though
Fable didn’t get spooked by anything I wanted to do

>>109172343
wtf are you guys using it for that you run into shit like this?

>>109172317
>Implying they don't see you as peasant nigger-cattle either


You realize they especially look down upon you right?

>>109172348
>Fable didn’t get spooked by anything I wanted to do
it has an entire new set of safeguards (see "Fable-5 mitigations" strings in claude code) so enjoy it refusing you tell you how many Rs are in strawberry since that might help you build a bioweapon

>>109172343
>non in-group plebs
Fable isnt usable by anybody for security work, unless by in-group you mean orgs on the Mythos list but technically thats a different model.

>>109172323
>>109172353
I have clearance, that should be the least you need to have access

>>109172363
What kind? Viewing military related information? The kind of clearance you need to fuck with nuclear lunch codes? The kind for working at specific organizations?

>Hur dur I havea clearance

Is "my dad works at Microsoft" tier vague posting.

>>109172363
1) if you have security clearances and are posting about it on 4chan, you won't have them for long
2) shittons of people have security clearances and it does nothing to affect whether you get Mythos 5.1 access. a tiny number of Glasswing "partners" (paypigs) get Mythos 5.1

>>109172363
Nobody cares, cope harder though.

>>109172370
Reminder that GPT 5.6 Sol will be Fabe tier, but faster and much cheaper

File: IMG_3068.png (2.69 MB, 1402x1122)

2.69 MB PNG

>>109172049

>>109172376
>1) if you have security clearances and are posting about it on 4chan, you won't have them for long
Nta. Clearance holder cuz I'm part of the military. For the love of God please stop talking like some naive newfag. To be granted a clearance worth a damn you need to be put under investigation and depending on the type they might also ask you a bunch of questions like where you've been the past 5 years, whether or not you have any relations or communication with foreign nationals, and will also ask you for multiple references, bug set references for information about you, and then also make phone calls to any friends and family members they can get a hold of. The feds do not give two shits about what sites you browse or participate on as long as you're not leaking confidential or above info. They usually don't give them out to retards unless they are government nepo babies.


>t. clearance holder and chan poster for several years. Reinvestigation was semi recently.

>2)

Pretty much everything you said there was correct. Only a very select group of silicon valley rich fags are allowed to even say that they have access to it

>>109172370
>dox yourself
No

>>109172376
I'm not saying I have any special access because of my clearance, I'm saying it SHOULD be limited to people who do.

>>109172378
(You)

>>109172390
interesting, four-leg snailcat seems better than 2-legs even though a proper snailcat has 0

>>109172413
See >>109172409

>>109172413
Oh come on, don’t you want to be famous like people get when they’re arguing in War Thunder forums?

>>109172413
You've never used it so you have no business claiming it's half as good as cargo cultists claim to be, let alone claiming other people should have no right to use it but you apparently do.

>>109172413
>I'm saying it SHOULD be limited to people who do.
why would that be the case LOL
you're literally meaningless to these companies unless you can give them billions of dollars

>>109172419
So we are in agreement?

>>109172422
lol

>>109172423
I'm not even talking about Mythos, I'm talking about Fable 5, which I very much did use.

>>109172426
I'm talking matters of national security; money comes after that.

>>109172422
I'm >>109172409
Having a clearance doesn't automatically mean you have access to the juicy shit. That Air Force idiot leaking that information to the discord had less to do with him having a clearance and more to do with the Air Force having beyond dog shit security (they have no business claiming they're smarter than anyone else. They're just pampered LOL)

File: fusion.png (309 KB, 2450x1454)

309 KB PNG

has anyone tried this OpenRouter fusion thing yet?

>>109172450
yeah

>>109172433
No? And throwback has not demonstrated concretely that their shit needs to be gate kept from normal people. It's just Dario fear-mongering and I'm still convinced fatal 5 was only banned because he pissed off Trump that badly and the orange man is known for holding grudges. I suspect it's basically his and hedseth's way is saying "buck up at us and we'll make an example of you". They were made an example of. Have we all conveniently forgotten the shit is a bubble?

>>109172328
But it's an American (#1 freedom) model.
It can't be worse than a Chinese (communist) model.

>>109172465
it will burst in 2 more weeks sister

>>109172450
Didn't someone already proposed this a couple months back? I think they called it "llm council" or something and they got made fun of for it on xhitter

>>109172468
And I'm sure it'll be profitable in two more weeks and the codex team will eventually constantly doing resets in order to not piss off their crack addict customers because it'll be so profitable

>>109172476
You are in the wrong neighborhood, luddite

>>109172470
Pewdiepie did this council thing months ago

>>109172480
Oh trust me I'm a rampant vibeshitter and use models to fix existing code that maintainers are too lazy to fix themselves (or they simply forgot it exists). I'm also realistic and I'm not tied to one provider like an idiot

File: file.png (2.54 MB, 1047x1356)

2.54 MB PNG

Me when a filthy nigger luddite runs his filthy dicksuckers in my presence

>>109172488
>>109172499
truly the machine gods have favored us this thread

>>109172499
See >>109172324

I'm not getting any hopes up. Yes they'll restore access but there's going to be a catch. Probably gated behind API prices.

>>109172470
It's been done several times over, there are providers whose whole shtick is doing this. You can even DIY it with a Pi extension.

File: 1757527193759952.png (2.54 MB, 1448x1086)

2.54 MB PNG

I have been architecting how to make luddites help society use AI without hitting energy issues, I think I came to a pretty good concept, planning to send this to Sam and Dario tomorrow.

>>109172518
is there an extension to get pi to see and manipulate desktop environments or GUIs?

>>109172518
So it's just "sub agents" with extra steps?

>>109172523
Several. That's usually the answer to any "is there a pi extension" question.

>>109172534
It certainly can be. Fusion doesn't do anything that couldn't already be done with well-orchestrated agents/subagents, it's just taking steps (and some control) out of your hands. Plenty of people already use parallel agents for all sorts of things, including this three-headed approach you see with Fusion.

>>109172551
How easy is it to maintain Pi tiptop for claude and codex? I have my own harness but I'm starting to wonder if I should switch.

>>109172534
you have a judge who takes the input of n+1 researchers and and evaluates their answers and sends the best one or a synthesis to a summarizer who gives the final output.

>>109172586
but how is that better than ultracode?

>>109172557
It's really down to preference, how you work and how you do it. It's lightweight, but it receives frequent updates, and I find it extremely easy to work with. GPT5.5 is my main for work, it was very easy to connect my subscription to Pi and it's never given me an issue. I like having a bare-bones install and keeping my agents files, skills, and prompt templates local to given project. I like how easy it is to implement custom extensions for it, and often do at a per-project level that keeps my base install clean. I do most of my work from a terminal and I find pi to fit my workflow very comfortably. It's quick and lean and that's what I want from it.

>>109172586
Okay so how does it decide which answers the "best"?

>>109172609
that's what +1 is for, it sends it one more time for that

>>109172604
Big caveat, it's all node, and I understand why but I still roll my eyes at it.

>>109172609
same way any judge judges a bunch of attempts at solving a problem, really

>>109172609
it can't

>>109172049
this is a good sticky

>>109172666
if you don’t continually polish your Markdown files you’re not gonna have a good time
we aim to practice here what we practice in our projects

Oh, so Fable ban was lifted. Also Sonnet 5.0 was released, is it worth using, or should I just use Opus until Fable is back?

>>109172701
If you just blindly use Opus for everything Sonnet is just going to be a cost-cutting measure that you don’t need

>>109172599
different models have different strengths and weaknesses, you get (sorta) the best of both worlds on a lot of tasks when you combine models. The judge is mostly just a way to synthesize the panelists' responses for your own convenience. In my testing I've found that this mixture of agents implementation is really good at breadth tasks, where naive union of responses is enough to get performance gains, but bad at narrow and deep tasks where the judge must actually judge which panelist response is the best without a deterministic correctness oracle.
So stuff like research is a great use-case for mixture of agents but complex code refactoring is not(unless your refactor goal is just a deterministic target like runtime or code graph complexity metrics).

>>109172409
what if I'm an avid war thunder player

>>109172736
I just had an idea, what if openai and anthropic merge into a single company and make fadex 6.0?

>>109172766
>fedex

>>109172766
they'll reach FAGI in no time

>>109172222
I don't think Anthropic ever recovers from that to be honest.

is there a hierarchal form of context caching and uncompressing? like if you fill up your cache what happens? does it overwrite? what if you want to avoid that and need all information present?

>>109172780
server rejects your request, most harnesses will automatically /compact then retry.