/tech/ - Technology

>>34521
The fuck do you mean AI-generated

>>34520
is adobe even doing anything at this point? i think it remains to be seen if this will ever even reach full feature parity with all the adobe suite, will the artcraft guy really pony up the massive costs and front all the managerial labor to handle all the dumb little edge cases and bugs a decade old software has had to work around, just to promote his moviegen studio? or will he just silently leave this as a half finished slop abandonware as soon as people stop talking about it.

>>34532
i mean remember the AI coded browser that just imported a bunch of presto libraries and stole a bunch of mozilla code just to end up as an abandoned experiment? or the claude C compiler? i strongly suspect the photoshop clone is only remotely feasible because there's a bunch of image manipulator software around to reference.

>>34532
When I first heard about this supposed clear-room reimplementation of the Adobe suite I was super confused, because I could only find articles about ArtCraft. But clearly ArtCraft is nothing like Photoshop or After Effects. It's just a shitty UI over some image and video generation AIs. I don't get the hype.

>>34534
artcraft is the videogen UI, photocraft is the photoshop replacement, effectcraft is the after effects replacement, vectorcraft the illustrator one, you get the idea

https://github.com/storytold/photocraft

for once the fireship guy actually has a good takeaway on the vibe coded adobe suite: you can reverse engineer anything now. i have maintained for like two years that AI is a poisoned apple and the only possible outcome for it is unprecedented capital destruction. i dunno why are people fretting over tech feudalism, my read is that silicon valley is going to fucking implode, every single avenue for profit is being threatened right now, with a technology they themselves promoted.

>>34537
I kept thinking to myself surely they've got to be able to do something about IP.
Then I realized no developing country would ever respect it - the gigs up.
If not now, then given exponential growth it will be soon enough.
Anyone got a way out of the riddle?
Spend the last throws of the economy on reorientation before it's too late?

>>34502
Thing is if they don't make the models they lose anyway.

>>34538
>Anyone got a way out of the riddle?
bomb china to recobble a monopoly over high tech industry

>>34537
People were observing this the moment people began making silly, shitty mashup mods.
What you're ignoring is that this technology is completely centralized and surveilled, which gives the powers that be an out to contain it; one that will probably nerf the capabilities of the LLMs, make IP law even worse, and make it less accesible.
That said, >>34538 makes a good point that the scenario I laid out is an American problem and other countries won't give a fuck about US IP law, and will continue to use vibecoding to reverse engineer everything they can.

>>34540
>What you're ignoring is that this technology is completely centralized and surveilled
the thing is that there is an international competitor providing access to a comparable technology. not only that but they also freely share the blueprint for it as well. to the extent that you can "contain" LLMs (each new model gets jailbroken in mere hours) the real issue is that the burgerreich was unable to cobble a monopoly around an emerging technology. this limits what they can do in two ways: capping LLMs turns into a really delicate business decision because it immediately harms your revenue and drives your consumer base elsewhere. and also you need to nerf capabilities worldwide, across every nation and LLM provider, even though it's rapidly commoditized. soon not only the chinese and french will have capable models, even enthusiasts will be able to run local models capable of doing reverse engineering in their local hardware.

>>34543 (me)
all that to say, what really gave me the sense that we were really witnessing the death knell of silicon valley wasn't the mashup mods, which internally rely on extensive modding tools that have existed for a while, so they really arent that impressive, it was deepseek. it really should have hammered stocks even harder back then, but i guess it will take even longer for the penny to drop.

>>34537
> PhotoCraft's PSD support is a standalone crate written from Adobe's public specification
What was actually reverse engineered? I thought the proprietary file format was, but it is actually public? So what was reverse engineered?

>>34547
In many cases vibe-coders just point to a feature and hope the LLM replicates it. This is not an actual decomp.

>>34538
>Anyone got a way out of the riddle?
violence

>>34547
>What was actually reverse engineered?
officially, nothing.

>>34551
And actually?

>>34552
they're probably doing this >>34548 and the LLM is probably aping features from FOSS projects in its training dataset, just due to sole volume and the way it works from some videos (e.g. photocraft doesn't do real time rendering when messing around with the curve adjustment, but photoshop does, which you'd think it would do if it was lifting up features wholesale from the adobe source code)

if nothing else, this dumb artcraft shit will help normies cozy up to AI a lot faster than anyone expected

>>34554
most people will form their idea of ai based on the most high-profile use cases so for them it's a machine that creates unemployment, induces psychosis, makes ugly advertisements and so on. people who are pro-ai for non-stupid reasons should be boosting good uses of it in a way that the naysayers can actually sympathize with

>>34555
> it's a machine that creates unemployment
ultimately people will normalize AI because AI is unable to overcome its nature as fixed capital

>>34547
if there is a spec, then implementation is trivial. And there is a spec of PSD files, because you know, if you want other programs to use your file format (make it "standard"), then spec is necessary.

Said spec: https://www.adobe.com/devnet-apps/photoshop/fileformatashtml/

>>34552
in theory it should do "clean room" implementation, where one LLM writes a spec based on data publicitly available (no no decomp), while other LLM implements a program based on that spec. Since the spec writing is most likely incremental and the guy behind it claims to have an actual experience, the code review is partially manual. This also why the suite doesn't have all of the features of ps as >>34553 points out.
It also steals a shitload from FOSS projects but not directly.

>>34557
Even GIMP can open PSD files, what exactly is novel here?

>>34559
the only novel thing is that it's aping all the UI across all the adobe creative suite, which is presumably what is locking people to that specific provider - switching to photocraft should have no learning curve in theory. there was a gimp branch that used photoshop's UI, but afaik there was nothing for illustrator, premiere and after effects as well.

>>34559
Nothing, that is precisely the point. Making a program PSD-compatible is not a big deal.

CoreWeave stock just went down the shitter cause it turned out that their situation is somehow much worse.

>>34565
ed zitron moment

>>34566
Tbf it has extremely high short interest. Everyone's aware that it's shit.

>>34567
yet people on xitter kept clowning on ed zitron because CRWV didnt immediately flop and die ten seconds after IPO last year, so you'd be surprised.

btw it's also a reverse cramer moment, i always wonder if jim cramer's employed just out of sheer pity

not nearly as significant obv but character.ai filed for bankruptcy this week, it used to be (might still be) one of the bigger "chat with your ai gf" type sites

>>34568
Important thing to keep in mind once Anthropic goes public in November (if they don't pussy out again) I guess. They might not have the "charismatic capitalism" effect of Musk's that keeps both SpaceX and Tesla alive, but I'm still predicting a brief jump in stock price at the start where you'd then have a few people prematurely declaring that AI isn't a bubble.

>>34570
it's a lot bigger than analysts are letting on i think, because the brunt of chatgpt users have a similar use case to c.ai

Wrote an article about AI… like everyone else.
I'm critical so I'm not going to advertise it here.
It was a moderate position:
Should not be thought of as an occult magic-box.
Rather a completion engine from training.
Something like a "copier" without a legal original.
Mass destruction looming, no big deal.
Thought it was maybe getting "out of hand".
At least when we use it without an idea of its return.
Maybe we haven't any idea what we're doing.
No real call to action.

>>34574
bitch is that a haiku

>>34575
lol, My five readers are lucky the whole article isn't written like that.
Just delete all the space between the ideas and call it a poem.
Not sure why thought you all uniquely ought to enjoy this punishment.

chatgpt is so based

>>32796
AI? more like rebranding–already-existing-chatbots-from-like-15-years-ago-as-quote-on-quote-artificial-intelligent

>>34583
already a stupid take three years ago, unforgivably delusional in the present

the next big thing is yapping about RSI, even though in practice it just means modest self-improvements to LLMs which in the large scheme of things amount to nothing. people will really lose the plot about RSI, just like they did for the math shit, and the hugging face hack. i guess the "LLM rights" crap didn't stick so they're going for a rebranded version of the singularity.

>>34585
Even if RSI as they describe it is possible all it means is an AI becomes maximally efficient on its hardware, optimized intelligence per watt or whatever metric they use. Hardware is a fundamental bottleneck, and even if they start to invest in new architectures and substrates that stretch the intelligence just a bit further, plateauing is inevitable.

this crappy xeet of a VC visiting anthropic is doing the rounds across AI and fintwt circles, and it's pretty notable because VCs are so retarded and unable to comprehend what they're seeing that they can't help but give the game away, so i'll spare you the details, because it's written by an obviously stupid person, but the salient bits i think are relevant regardless because they show what frontier labs are thinking, and are about to demand in terms of regulation:

> frontier AI labs should do is release open source models, maybe hosted and verified at firms like AWS.

> have the deployment of open source models be hosted at firms like AWS who verify these models and assume liability for their use.
so what are they planning to do is regulate LLMs at the infrastructural level. this will absolutely collapse the hyperscaler market, like just totally destroy it, presumably they already imagine it will be the case, which is why only AWS is mentioned. my prediction is that it will also be done through an international treaty. this will pit frontier labs directly against nvidia and jensen huang, at least temporarily so, and i dont know how long it will take to get the ball moving, it needs some political finesse the incumbent currently lacks, so i assume it will be done post-midterms, when the US has a dem majority and the US is perceived as a "sensible trading partner". i don't see how this would extend to local models, but that's where their mind is at currently.

>This would basically kill the use of Chinese open source models used in American companies, which removes an important threat vector (very hard to root out malicious code in these).
so the new scaremongering line that will be doing the rounds is that open weight models have "malicious code" in the weights, not sure how that's going to fly, or how are they planning to demonstrate that's the case without also throwing shade at their own models

> give AI labs an antitrust shield (as Schiff and Banks have proposed banks.senate.gov/news/in-the…) so they can coordinate on safety without raising antitrust concerns
pretty self-explanatory, i think people had picked that up from the "slowdown" but frontier labs are planning to operate as a cartel pretty openly.

>We are probably facing a wave of cyber and other attacks soon.
massive false flag incoming, in case you were not convinced by the hugging face attacks.

the rest is trash regurgitating the garbage frontier labs peddle
https://nitter.netbub.com/arpitrage/status/2110473485030129919

>>34592
>Some evidence for this is code per Anthropic employee has gone way up
TIL when you're a frontier lab and start suffering from the same problems as GitHub due to vibecoders, it's actually hecking RSI rather than a liability.

>be me
>trying to rip the google docs of an abandoned 4chan project for archival purposes
>look up if there's a url format for this
>first result is a reddit thread
>reply says it works with something chatgpt told them
<tells others to ask chatgpt as well
>mfw
So either the site is completely botted now or normoids are so brainfucked by llms they can't even give a coherent account of what they're doing anymore. Second result was some indian guys SEO blog, which had the answer buried under three paragraphs of boilerplate. I hate the modern internet.

>>34599
Unc still using Google? We all ask chat (the LLM, not the omnipresent second-person) nowadays fr fr.

>>34599
What was the project?

>>34600
Come on you're baiting. We're past peak "chatgpt how should i feel about this", right?
>>34602
Quoting the last changelog entry from 2022:
>Version 8.5.0.2: Sapros commits suicide after an entire decade of seething with no game to show for it. Fenoxo and Savin continue to make $30k a month
I like looking at dead or dying projects, because you can make some pretty definitive value statements about them.

>>34599
Normies have been living on autopilot forever.

>>34585
i'm admittedly clueless around how rsi is being theorized in terms of modern llms; seems to me you can't really trust the ai to not just fuck itself up. ai can write perfectly serviceable and functional code but the reliability of its answers it not even close to 100%.
i am however talking about autonomous rsi of course. if it's just ai writing code into anthropic's repo under human instructions and supervision would be that significant, that's already how coding works now.

>>34591
yeah there's this too, what is the actual expectation for what the rsi would amount to? it's not gonna change its own training or weights.