Hacker News new | past | comments | ask | show | jobs | submit login
Nvidia’s Implicit Warping is a potentially powerful deepfake technique (metaphysic.ai)
216 points by Hard_Space on Oct 17, 2022 | hide | past | favorite | 145 comments



Not sure I want to live in a world where billions of people still believe what they see when these deepfake technologies mature.

Not sure I will care either, because by that time I will have been locked in my living room for weeks playing my customized VR role playing game where Han Solo and I wreak havoc in the Far Cry 5 universe with my Dovakin powers while being chased by a all-female bounty hunter crew commanded by a 1982 Phoebe Cates.


This is why I think it’s irresponsible to not give the public access to these early models. I think a slow, increasing, exposure to the absurd/false will work out much better than the isolation desired by the orgs who are trying to “protect” us, because that isolation will end up being temporary, resulting in a step rather than a ramp.


There needs to be far more criticism of the increasingly closed development of these models in what is essentially becoming an AI arms race. Like Sam Altman and his team at OpenAI (more like ClosedAI) who were more aware of this than anybody yet sold themselves out for the highest bidder. What a joke of a “non-profit.”


I’m just jealous that I’m not able to sell out.


Right, the earlier the general public is confronted with these technologies, the better!


It's so bizarre to me how much commentary I see to this effect, do people forget the times we live in?

You can fool more people with a fake article and $1000 in Twitter click farm spend than you could ever dream to with a deepfake.

Like we've already seen entire elections undermined with basic internet trolling, the problem is already here, but somehow people are overlooking that and fixated on the most fanciful version of it?


Plus, for some reason humans as a group tend to prefer the fake perfect over the real good. Autotune perfects bad playing and singing, we use audience brain responses to tune the editing of movies, CGI over practical effects, clickbait over quality titles, zinging 1 liners more important than the novel they describe.

One the one hand, you can't fault the masses for buying the goods they're sold, and on the other hand you can't fault the sellers for maximizing the apparent quality of their products, but somewhere inbetween all of that broad mindedness and beauty there has to be something at fault for creating a world where actual human beings are not good enough to participate as equals and the playing field is pay to win on a scale that boggles the imagination.


I first encountered that idea as "supernormality"

https://en.wikipedia.org/wiki/Supernormal_stimulus


My thought process: if spreading disinformation without evidence is so successful right now, imagine how much worse things could be in the future when people are spreading disinformation with “evidence”.


I have a comment below about it, evidence is actually contrary to the goal.

If you bring evidence you're introducing a place for a counter attack to apply leverage.

If instead you make completely baseless claims on obvious false pretenses, you've actually made things more difficult to counter because only trivial counterproofs exist, which have to be dismissed to believe the false claim in the first place.

-

Take COVID vaccine deaths for example. Imagine I baselessly say that the medical field is lying and 50% of COVID deaths are actually vaccine complications.

For someone to believe that, they must completely distrust any official numbers on COVID deaths... so once they've fallen for the lie, how do you convince them otherwise? The only counterproofs are the trivial to find sources of data that they already had to dismiss to believe me in the first place. Suddenly I've implanted a self-enforcing lie that entrenches its believers against anyone who isn't in their echo chamber.

The root of all this is straightfoward enough: there is nothing stronger than requiring someone to disbelieve their own beliefs to counter your disinformation. If you add a deepfake, you've added something outside of their belief system to attack, so you're weakening the attempt. People simply do not like to be wrong about things they think they've figured out.


Makes me think of https://en.wikipedia.org/wiki/Big_lie: the use of a lie so colossal that no one would believe that someone "could have the impudence to distort the truth so infamously."

And yes, I do not think adding a deep fake of some big shot saying or doing X would work better than just repeating the lie that they say or do X.


If you used a Biden deepfake you'd more likely want it to be of him tripping on something than admitting allegiance to the lizards.

> Imagine I [...] For someone to believe that, they must completely distrust any [data]

Do you think this is this like a 419 scam where saying something a bit outrageous sorts out the gullible and bypasses the wary or do you think that your claim can somehow hijack a credulous person long enough so that they make that mental recategorization of the data sources and are stuck?


The former gives you a springboard for the latter

The people falling for the obvious nonsense are self filtering just like people falling for obvious 419 scams.

But as you grow the base that believes in your disinformation, you gain real people who are fully convinced of these things, and the effect of that is a force multiplier.

People talk, and if people's self-held beliefs are the strongest reinforcement, the second strongest is those we surround ourselves with. If someone falls for this stuff and starts talking to the spouse, now someone close to them is pushing this agenda. People can start nudging their friends to be more skeptical.

It's not going to be a 1:1 conversion: a lot of people close to them will push back, but remember, this is all based on absolutely no proof, so it can twist itself to fit any box. People can moderate the story to avoid pushback: "Oh you know I'm not an anti-vaxxer... but I did heard that vaccine has a lot of complications", and maybe they connect that to a real article about a myocarditis case, and now maybe they're not pushing my original lie of "50% of deaths", but I've planted an suggestion in a rather moderate person using a chain of gullible people.

And something especially effective about this is the fact that, while the most brazen aspects of disinformation hit less intelligent people hardest (https://news.ku.edu/2020/04/28/study-shows-vulnerable-popula...)

Once you start to make inroads with increasingly better educated groups via the network effect, they tend to not want to believe they're wrong. Highly intelligent people can be more susceptible to some aspects of disinformation in this way: https://www.theguardian.com/books/2019/apr/01/why-smart-peop...

That lends itself to increasingly authoritative figures becoming deeply entrenched in those campaigns, leading to things like... https://wapp.capitol.tn.gov/apps/BillInfo/Default.aspx?BillN...

-

Overall I've said this before, everyone is dreaming of AI dystopias rooted in things like deepfakes putting us in a post-truth era, or AI gaining sentience and deciding it doesn't need humans...

The reality is so much more boring, yet already in progress. We're starting to embed blackbox ML models trained on biased or flawed data into the root of society.

ML already dictates what a large number of people are exposed to via social media. ML is starting to work its way into crime fighting. We gate access to services behind ML models that are allowed to just deny us access. How long before ML is allowed to start messing with credit ratings?

And yet getting models to "explain" their reasoning is a field of study that's completely behind all of these things. You can remove race from a dataset and ML will still gladly start codifying race into its decisions via proxies like zipcodes, after all it has no concept of morality or equality: it's just a giant shredder for data.

Right now a glorified bag of linear regressions is posing much more of an effective danger than T1000s ever will. But since that's not as captivating instead we see a ton of gnashing of teeth about the ethics of general intelligence, or how we need to regulate the ability to make fake videos, rather than boring things like "let's restrict ML from as many institutional frameworks as possible"


> But since that's not as captivating instead we see a ton of gnashing of teeth about the ethics of general intelligence, or how we need to regulate the ability to make fake videos, rather than boring things like "let's restrict ML from as many institutional frameworks as possible"

It’s not only not captivating, it’s downright inconvenient. If I’m at a TED talk I don’t want to hear about how ML models (some of which my company has deployed) are causing real world harms __right now__ through automation and black box discrimination. If you read Nick Bostrom’s Superintelligence it spends laughably little time pondering the fact that AI will likely lead to a world of serfs and Trillionaires.

No, people want to hear about how we might get Terminator/Skynet in 30 years if we’re not careful. Note that these problems are already complicated by ill-defined concepts like sentience, consciousness and intelligence, the definitions of which suck all of the oxygen out of the room before practical real-world harms can be discussed.


It's never about what someone said. Proof of that is completely unnecessary. Everything will be internalized as long as it's from your side, no questions asked. Deep fakes won't make that much worse.


We already have people throwing articles and YouTube videos where the title says something like "holy shit X is happening!" and everyone runs with it as if it were evidence and then if you actually read the article or watched the video the evidence is basically "actually nothing happened" or hersay alleging that X/10000 almost happened.


> You can fool more people with a fake article and $1000 in Twitter click farm spend than you could ever dream to with a deepfake.

Strong disagree. Viral content is way more effective.

And also, imagine what you can do with $1000 in ad spend and a deepfake.


You're not disagreeing. I didn't think I needed to connect the dots that you're paying the click farm to go viral.

> And also, imagine what you can do with $1000 in ad spend and a deepfake.

Way less. A deepfake detracts from your mission. If you deepfake Joe Biden saying "I'm going to destroy this country as instructed by my masters" and you actually gain traction, suddenly you've thrust your subversion into the spotlight and start reaching people adversarial to your goal.

News sources are going to start digging to find where the clip came from, the White House is going to respond, the video is going to start getting analyzed to death.

-

If instead you register a dime a dozen domain like "america4evernews.com" and write a crackpot article about how Joe Biden is actually working for our enemies and you have it on good authority that he's being controlled by puppet masters who want him to destroy America, you'll find an army of people who want to believe.

They don't need sources, they'll avoid sharing it with the "sheeple who believe MSM" and stick to their echo chambers. It's a strictly better outcome.

People don't seem to understand that modern misinformation is not about fooling everyone, it's about fooling the right people. You're goal isn't to reach every set of eyeballs, it's to reach the eyeballs that are easily fooled so that they come in conflict with those who are not, thus entrenching both sides against each other.

-

In some ways it's like a virus: if it's too strong it draws attention to itself early and can't spread easily. If instead it causes subtle symptoms that compound, it can spread very widely before it's even noticed, and use that infected base to expand even further.


Remember that video that showed Nancy Pelosi supposedly 'drunk', but it was just slowed down and pitch shifted? It was so easy to tell it was faked, but it still spread like wildfire.


however the story of withholding technology, education, etc, for profit (of some kind) is much older than this particular latest development.

if you ask me, this 'tradition' is the essence of imperialism (or just one amongs other techniques necessary to have an empire)


Some of these videos are particularly fascinating in how the head position is not exactly replicated but does look more like the body language I'd expect from those people. Like, try to note the sequence of facial expressions and head angle that Barack Obama goes through, then watch Angela Merkel and Winston Churchill. They're not 1-to-1.


We should stop calling it deep fake. And start calling it facial masking. And realize that this will now be kiddie tech in the future.


Deep fake is a great term. Facial masking hides all the other uses of deep fakes. You can fake anything with these techniques based on deep learning.

The term is just spot on.


That's exactly why there should be a distinction made. The term is too broad for a specific subset of deep fake technology that has matured far beyond its fun monster phase.

Facemask and soundmask are good terms for a democratized warping of visual and audio data.


Those terms seem a little broad to me. If the news reported that Barack Obama's likeness was used in a "facemask", nobody would know what happened. If they said it was used for a "deepfake", most people still probably wouldn't know what happened, but at least they could look it up and get a general idea.

Considering how facemasking and soundmasking have multiple contextual meanings, I think deepfake is the perfect word. If you need more precision, then you can use the phrase "visual deepfake" or "audio deepfake".


I agree with this mostly in the sense that we need a specific word without any other context that defines what these are. I'm not sure that "deepfake" is that word but it definitely fits the bill based on what we're looking for. Deepfake has the advantage of being a word that's already in common usage that a good number of people recognize as "Oh, that's a fake done by a computer" so adding "audio" or "video" in front of it also fits the requirement of being easy to distinguish without being cluttered by other meanings or contexts. While we may be able to come up with a better word (in the same sense that a "tweet" is now unambiguous in context), I think it would be a misstep not to take the snowball that is "deepfake" and continue to use it in this context especially as a means to educate people on fake, computer-generated audio and video. People can barely keep up with tech buzzwords as it is so the fact that "deepfake" is already something in the public consciousness means a lot.


Deep Fake tech needs to puncture the public's consciousness ASAP, diluting terms will be counterproductive


Wouldn't that be a voicemask?


Within VFX the terms are actor, object and background replacement. Rarely mentioned in any of the deep fake literature is the requirement for a non-facial skin skin tone correction for most deep fakes to begin to look natural. Everybody ignores the fact that you nor I have the same skin tone, and that needs to be touched up for any face replacement to start to look correct.


A good VFX artist can also do a hell of a better job than any of these deepfake automated platforms. And has been able to do that for a long, long time now. So while deepfakes becoming more available to everyone worries me the fact that we have not yet seen tons of well done face swap videos done by malicious actors paying a decent VFX artist makes me wonder how big this will really become.

That being said for smaller countries with less internet access (and less education) this has already become a big problem (lots of cases in elections in Africa) so I think once again it comes down to education: we must inoculate people to bullshit like this by giving them the tools to spot it and future things like it.


Yep, I was an actor replacement specialist at Rhythm & Hues for several years.


Hopefully you didnt have to be involved in that very fun massive lawsuit


The "Deepfake Puppetry" term used in the article seemed to me like an improved term, an excellent and accurate description. It is using deepfake technology, and the "puppetry" term both describes accurately what it is doing — animating a target image like a puppet — and includes a vivid description to the nontechnical audience.

"Facial Masking" is not bad either, quite accurate, but I'm not sure it provides as broad a description than "Deepfake Puppetry".

Either way, it is important to get a term that is both accurate and resonating with the audience to ensure that the general public understands the serious potential damage for this technology (as with any powerful tool, can be used for good or evil).


Puppetry is a better term!

It not only captures the face but also the body, and maybe the voice. Facial masking is more programmer friendly.


It's already kiddie tech. Just google deepfake maker and you get dozens of hits for free services to make passable fakes.


It only works if the background is continuous and fairly static - you can't easily synthesize a complex background if you have no reference material. You can of course extrapolate the visible background using other AI that has knowledge of real surfaces; however if the background has familiar content (like a sign) that is not visible in the reference frame(s), it's unlikely to fool anyone no matter how well the face is animated.


> it's unlikely to fool anyone no matter how well the face is animated

Call me a cynic, but I believe humans, myself included, are easy to fool. If the face looks good enough, plenty of people will fail to spot the Invisible Gorilla.


It must be trivial to use one of these approaches to generate a removable background and render the background with something else, like a static image, video, or 3d software


> And realize that this will now be kiddie tech in the future.

Oh, absolutely. And we're not talking decades here. This tech will likely be ready to rock within a few election cycles


I think with enough money / professionalism it's ready to rock already; a lot of the deepfakes I've seen so far have been done by "amateurs" (that is, individuals, not big companies). The last big companies' work has been the star wars prequels that used CGI, which was still in the uncanny valley imo.


God what an awful way to hijack scrolling on such an interesting article.


Totally agree. I got dizzy within the first couple of seconds scrolling down the page.

Why do people think it's a good idea to circumvent the scrolling behavior that the developers of the browser probably have spent hundred of hours perfecting over the years?


Funny, it's weird on both scrolling wheel and gesture on mouse pad... The scrolling wheel felt like the scrolling would get stuck at certain points, while the gesture was... Like a weird speed? It didn't feel the same on my laptop for both, and that's even weirder than I expected.


Yes. It was so frustratingly broken I had to stop reading and do something about it. Running this in the (Chrome) browser console fixes it for me:

  window.removeEventListener('wheel', getEventListeners(window)['wheel'][0].listener);


There should be a way to block all these wheel-hijacking via uBlock filters.


    metaphysic.ai##+js(aeld, wheel)
Done.


Even if there was today, Manifest V3 will take any ability to do that away


Not in Firefox!


That fixed mouse wheel scrolling for me in Safari's JS console as well, thanks!


Yeah. Read a sentence, tried to scroll, immediately left. Nope.


They fixed it! It was an unexpected change to one of the plugins they were using (wrote them to let them know).


I just gave up reading it. Good job devs!


I didn't see any! Did they remove it?


No, it's still there in Chrome. It doesn't seem to be there in Firefox though.


It's there for me in Firefox. With a mouse if I spin the wheel then it starts an animated scroll. However, I cannot send any more scroll inputs until that animated scroll has finished. So with two scroll wheel spins in quick succession, the second one is ignored. This only seems to be the case with two spins in the same direction; reversing the direction while the animation is happening works as expected.


The scrolling is still "floaty" in Firefox - sort of like operating a boat (not fun for a website)


No issues on my Firefox 105 on Linux + uBlock Origin


Same. I think uBlock is the answer.


Hm, I wonder why it's not for me. Possibly my Firefox is old, or possibly it isn't using the GPU (current driver issue).


I was curious about this comment since I was not seeing this on Safari on Mac.

I switched to chrome, and that is just horrible!

I am curious why I am not seeing this on Safari (and apparently others on Firefox). I don't see any errors on Safari implying that some javascript is failing to load.


Devs were only using one browser to test their code. Nonetheless, if you are going to implement any mouse behavior hijacking, you should reconsider your life choices. Take a long walk down the park, have a cup of coffee, come back and... Don't do it.


oh I totally get that, but I would expect that they would be doing their testing on Chrome. I mean I get not liking chrome but ignoring it for testing is... a choice.

I would have expected the worse behavior on Safari... not Chrome.

Or am I just being nieve about frontend development? I really only do it for any personal projects so don't really have any insight into it professionally


I'm not sure what the issue is (using firefox and I don't have Chrome installed to test with) but I'm pretty sure the scroll hijack that people are complaining about, it is an intentional effect to make the site "fancy," which they only tested in Chrome, or something like that.

One side perk of not using Chrome is that there's a correlation between only testing on Chrome and producing code that we're better off not interacting with.


If this is the intended effect, its bad one.

I just find it interesting because I look at the apple product pages. While a bit janky they work as intended. But they are also applying that effect very deliberately and not to an article for some reason. So maybe that is the distinction.


I get it on Apple as well. It must be a janky implementation or something.


Yeah, quit your job over it.



Or https://addons.mozilla.org/en-US/firefox/addon/luminous/ for more fine-grained control (block individual JS events, not all JS from a given page).

Or uBlockOrigin in "medium mode" or higher incorporates JS blocking on a per-site basis.


Came here to say the same. It is so insanely aggravating.


I have never seen anyone praise or laud the benefits of scroll behavior overrides. It seems to universally be a negative if not hated, and always distracts from the content. Always.

And it required extra work to yield this negative behavior!

So how in the world does this end up happening? How could a team be so profoundly detached from reality? This site's behavior is so incredibly ill considered that I marvel that people worked on that, people approved it, people said publish, etc, without one person stepping in and asking what in the world they were doing.


On iOS I can usually open user-hostile articles like this in reader view, but even that is broken here.


I'm there and it has been fixed apparently.


Whatever it is my blocking plugins remove it.


Is it just me, or does every single AI innovation lately seem to be.. pointless?

Like, if I had the time and resources to work in AI, stuff like substituting faces and even stable diffusion would be about the last things I would ever work on.

What would I start with? Something more like the MS Office software of the 1980s, only automated. I would have real-world applications that, wait for it, perform work so I don't have to. AKA automation.

TBH this stuff exhausts me to such a degree that I almost can't even follow it anymore. It's like living in a bizarro reality where nothing works anymore. A waking nightmare. A hellscape. Am I the only one who feels this way?


> What would I start with? Something more like the MS Office software of the 1980s, only automated. I would have real-world applications that, wait for it, perform work so I don't have to. AKA automation.

But that's what Stable Diffusion et al is: automation. It's just not automation for what you spend your time on, but it is automation for what a countless amount of people spend their time on, doing stock photos/drawings/clip art for endless amount of articles and other generic videos.

I also think that the current "creative automation" comes from a perspective where many people have said for a long time that computers will of course be able to do boring, repeatable jobs like counting numbers and what not, but it will never be able to do the job of an "artist". But now, it seems like at least a subset of "artists" will be out of job unless they find a way to work with the new tools rather than against them.


I think it depends on if you are a designer or an artist. In my mind a designer works on a specification for a client and produces it, they are interchangeable, they dont put their signatures on the work because its not theirs, its the clients.

Artists however wont have to worry about ai. Ai music already exists but people still support mainly real artists on streaming platforms and go to concerts, because provenance matters for artists and it doesn't for designers. Could an ai make a warhol? Probably not because what made warhols art popular was that he essentially worked outside of the training set and provided something previously unseen. Machine learning is bound to the training set. You can make generic corporate bathroom art for hotels or filling empty picture frames with it, but there will still be real artists and galleries and concerts and museums, because often times people value the provenance of the artist much more than even the work itself.


> Ai music already exists but people still support mainly real artists on streaming platforms and go to concerts, because provenance matters for artists

I don't think that's what's going on. The top pop singers are generally already singing things written by other people against accompaniment written by other people. I think by far the biggest reason few people are listening to AI music is it's just not as good as human music yet?


Thats true, but people are still listening for the artist and not the talent team. Drake is famous for having his writers live in tents in the studio. I bet songwriters desperate enough for work to live in a tent probably are cheaper than making a contract with whoever is trying to sell this stuff to major record labels.


So if the talent team becomes partly AI-based, no change on public perception. They still listen to "the artist".


That would depend on if its cheaper to have an AI team or a real team. For example mcdonalds had all the tech to replace burger flippers by the 1980s if not earlier. Just retool a robot that builds Fords to hold a spatula. The reason they still hire people is because its cheaper than what the maintenance contract on such a robot costs.


That's a good point

Just like how "Calculator" used to be a job title for humans who manually performed calculations.


"Ai music already exists but people still support mainly real artists on streaming platforms and go to concerts, because provenance matters for artists"

I don't listen to my favorite music because of who made it but because I like it.

If music that I like started bring made by AI then it's listen to it without hesitation.

In fact a lot of music bring made today is already a collaboration between humans and machines and had been so for a long time.


But I get what zack is pointing at, it's automating the wrong stuff. It's like still having to work in a coal mine but thank god you don't have to take care of kids at night because someone invented autosnatcher.

A lot of world is grinding in pain due to extremely bad software and the money and brain power keeps pouring everywhere but there. Well not entirely.. there was a lot of money thrown on these bad applications, but it evaporated due to software services companies subpar engineering.


Thank you, that's what I was trying to say. Tech innovation today is almost always some variation of let them eat cake.

I had the fortune/misfortune of moving furniture for 3 years right out of college 20 years ago to support my internet business at the time. I saw how the vast majority of people toil away their lives to make rent and child support each month. That experience shattered my will to such a degree that I came out of it a different person.

My concern is that the divide between working poor and techie riche is now so vast that they can't even see one another. If the wealthy and powerful could see, they would invest less in profitable schemes and more in shared prosperity. But they can't. So wealth inequality continues to grow unabated, with AI being just another tool to profit from another's labor or eat their lunch outright.


Its a big question I think about regularly. I have quite a lot to say about it. I think the whole structure is subtle and it's not as simple as rich/poor divide. I'll be back.


These innovations aren't really pointless though? They may not be relevant to your interests, but they have practical applications and Stable Diffusion especially is already seeing a lot of interest from artists and people who need art but don't need something 'custom' enough to pay for a human. In both cases they are saving lots of 'basic' work that might have either been done by a human before or not done at all.

Plus, these are the things we hear about because they look flashy. There is plenty of work behind the scenes on applying these innovations to more 'practical' matters like automation.


Exactly the opposite here. The recent progress in text/image/sound generation is the first thing that's actually made me interested in ML/AI. If I could restructure market priorities so all of the data scientists working on ad tech and recommendation engines and virtual assistants were working on this stuff instead I'd do it in an instant.

I can also say from experience that extreme negative feelings like "other people are doing things I don't find interesting and it makes me exhausted and miserable" were, for me, a sign of clinical depression.


You're right, and I agree with your conclusion about depression.

The catch is that I feel most depression is environmental today. It's singularly exhausting to struggle while watching people who have the means to enact real innovation and change squander their potential on yet another gimmick.

The only thing that's really helped me was to realize that it's all a gimmick. Life itself is a divine comedy. We each determine our own definition of meaning since science can't provide one.

Using the book/movie Contact as an example, basically my philosophy has shifted away from Ellie Arroway (Jodie Foster) and more towards Palmer Joss (Matthew McConaughey). In it, he wrote a book called "Losing Faith: The Search for Meaning in the Age of Reason":

https://www.youtube.com/watch?v=HFcHpamkHII

This spiritual battle we're engaged in between science and faith has been with us since the beginning. It's crushing down on us harder and harder now as science works to stamp out the last vestiges of our individuality and humanity. That sentiment might not make a lot of sense to many on this site, but it's the daily lived experience of billions of people forced to spend nearly the entirety of their lives toiling under subjugation (working towards another person's life purpose) to survive.

You're also right about interest and motivation. I find AI to be perhaps the last frontier, since there's a chance it could explain consciousness and maybe even give us access to something like an all-knowing oracle or even a means to contact aliens. It's pretty much the most interesting thing there is, and why I got into computer programming in the first place. I'm just sad quite literally that I squandered so much time on running the rat race and never got a chance to contribute and "get real work done".


"640K ought to be enough for anybody." - etc etc etc.

It's becoming increasingly clear that many of these techniques that for the past 10 years that have been derided "pointless" or novelty now have real applications.

For one example, the automotive industry - ignore the hype on autonomy, computer vision is already delivering real benefits for active safety systems. I use github copilot every day - its not perfect but good enough to add value to my workflow. Apple's automated tagging of my photo library via computer vision allows me to discover hundreds of images of my life and family I'd forgotten all about. Stable diffusion can clearly replace an artist in some cases, ignoring the moral/ethical issues.

I'm extremely excited for future of all this, frankly. The first step into such a new paradigm is always hard - people made the exact same "home computers are pointless" arguments in the late 70s/early 80s. I don't think anyone agrees with that anymore...


> I use github copilot every day - its not perfect but good enough to add value to my workflow.

Wow I thought this copilot thing was just like a joke. People use it to write software? I'm really curious, now. Can you share examples of the code you're writing with it?


You can use it for just about anything that is text - it will complete configuration files, not just code, and even does a reasonable job a lot of the time writing method docs automatically. Thinking of it as just for code is already a narrow application. Sharing code examples is also of little relevance - just try it yourself and see if it fits your needs. The dollar/value ratio maps for my needs to avoid trips to docs or search engines for syntax confirmation - it may or may not for you.

If it gets it wrong sometimes I don't really care, all the time it takes is pressing tab to complete or not to complete if suggestion is garbage. It's just an extension to the tab key's functionality, which is why it's so easy to use every single day when it integrates to most text tools.

I also think if this is what we can have today, the future of code completion tools is very exciting.


> I use github copilot every day - its not perfect but good enough to add value to my workflow.

Could you please share an example? I saw a description of copilot but couldn't imagine what it might be useful for.


Neural processing is ubiquitous in phones (all the flagships do neural enhancement on the image, and iirc apple does gesture recognition neurally), and all Ice Lake and newer laptop chips (plus lol Rocket Lake), as well as all Zen4 chips have neural instructions too. NVIDIA's had it for 2 gens now. It's pretty well across the stack at this point. And it's getting used for different things in different places. Things like face recognition or gesture recognition are natural fits in a low-power environment, reduces power consumption of those features hugely. PCs and servers can do content generation or some other larger "inference" tasks.

Neural game upscaling is a huge win too. The neural-TAAU upscalers (DLSS 2.x and XeSS) perform a lot better than FSR still, even compared to FSR 2.0/2.1. And there will be other things they can figure out how to ML-accelerate as adoption continues, I'm sure. It enables some solutions to hard problems that don't have good deterministic algorithms, and you can run a lot bigger models than you can without acceleration.

also fabrice bellard (of course) wrote a neural compressor... https://bellard.org/nncp/

it's also likely going to be useful in level design and asset creation as well, although of course as we're seeing now with Stable Diffusion there's some interesting legal questions with content creation.

optical flow engine (not neural) is another big win for computer vision stuff, too, offloads a few tough tasks to hardware, like object tracking and motion estimation. hooking into that with zoneminder or something would be awesome.

I'm also stoked for shader execution reordering too. This is similar to what Intel calls "ray binning", it basically is a best-effort "re-alignment" of the thread (state,etc) to the most similar execution group for memory read alignment/coalescing. I think one or another raytracing implementation (Intel or NVIDIA) they were coalescing based on material.

Intel talks about theirs here: https://www.youtube.com/watch?v=SA1yvWs3lHU

Ada whitepaper: https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvid...

I am interested to hear what AMD is doing with RDNA3, supposedly there is a new RDNA3 ISA instruction for matrix acceleration but rumors are it's not a full high-performance matrix unit like CDNA and >=Turing? I don't know why you'd add an instruction without some hardware acceleration though. So maybe less than the other implementations... which is a little disappointing.


We see these negative articles because they get clicks and attention but there is actually a ton of legit uses for this. In the VFX industry for example doing a face swap (sometimes called a replacement) is a fairly standard practice. Car crash scene where a union stunt performer is being paid to stand in for an actor? Instead of finding one that looks as similar as possible to the actual performer you can have any competent compositor just replace the head of the main performer to the stunt performer. Giving said compositor more tools to do this faster and easier is great - maybe now they can just replace a few individual frames and let the AI do the rest and then tweak the end result as needed.


? Nvidia is a graphics company so they are automating work so that they don't have to, that being making graphics.


I think it's a move in the right direction - but I've never been a fan of convolutional networks.

There's significant potential value in networks that can regenerate their input after processing.

It can be used to detect confusion - if a network can compress + decompress a piece of input, any significant differences can be detected and you can tell that there's something that the network does not understand.

This sort of validation might be useful, if you don't want your self driving car to confidently attempt to pass under a trailer that it hasn't noticed - which tends to kill the driver.


You just have a very slim vision of what work people do. The tools listed could be added to photo, special effects and film editing tools and provide value.


Vision and language are two of the pillars of human intelligence.

Perfecting these two things opens whole universes of potential.

Ask a robot to “do the dishes”. It has to know what that means, find the dishes, find the sink, the on/off mechanism for the water. These are all language and vision tasks.

The balance, navigation, picking/placing etc seems like a minor subroutine fed by vision metadata.


I like that the term chosen by AI researchers for many of these applications is "transformer". That way, I can look forward to a future where transformers do my dishes.


"Autobots, wash up!"


"What is my purpose?" "You wash the dishes." "Oh. My god."


They aren't doing it because it's useful. They are doing it because it's easy and they take what they can get. Further, it's not farfetched to imagine that models that can understand and predict how objects and faces move is a stepping stone to more useful stuff.


You're not the only one that feels that way. I wish I could add something constructive beyond this.



Well we DO have GitHub copilot which I have yet to use


You're right, much of what I'm struggling with is my own inability to keep up with the rapid pace of change. I'm turning into an old man yelling at clouds, like Grampa Simpson.

Then I project when the answer is just under my nose.

It's good to be reminded from time to time that what we're asking for may already be here, we just need to realign our perception to see it. I truly believe that's what meditation/prayer and manifestation (magical thinking) are all about.


> Is it just me, or does every single AI innovation lately seem to be.. pointless?

... but, I think you have really missed the point! Maybe you think government is here to help rather than to govern minds too! And that what is shown in the news is a good faith attempt to relay reality!

If you are managing the world, companies, etc - perception is everything! If you are able to control what people perceive, and they receive everything via a screen, well, who cares about truth? The imagery, the ideas - that needs to convince... and that is pretty much all that you need to manage the masses.


Every new tech is a potential powerful dangerous tech, so it was fire and so it has been any new discovery.

I saddens me to see here these sort of posts.


> I saddens me to see here these sort of posts.

We should actually be much more proactive about technology: it is not an unalloyed good, and an unwillingness to consider its downsides leads to naive designs. Even leaving bad actors aside (which I don't recommend doing), what works well in a group of 10,000 users may not scale to one billion users. Where one scale has no effect on social cohesion, another scale may have a tremendously deleterious effect.

We've been doing these experiments in search and social for years now. Taking lessons from that and applying it to the next great wave of AI innovations seems like a Good Idea to me.


It is a fool's errand to try to evaluate the net societal impact of a particular technology, even with perfect predictive power.

Looking back, what technology would you have retroactively stopped? Tetraethyl lead? Perhaps. Nuclear? Cable TV? ANNs? The Internet? Drones?

Also, the financial incentives aren't aligned. Certainly it makes sense to hold back technology to avoid embarrassment if you're a trillion dollar company; you have more to lose than gain. However, if you're a scrappy startup, it makes way more sense to roll the dice.


> It is a fool's errand to try to evaluate the net societal impact of a particular technology, even with perfect predictive power.

This really isn't true. We regularly do cost-benefit analyses for business; we don't skip them because they're not perfect predictors. We do market analysis and all kinds of customer deep dives to perfect UI/UX and customer response. We've created targeted dopamine-delivery services that are continually refined to maximize impact.

All of this implies a certain kind of ability to evaluate. Looking at tradeoffs and potential uses of technology is well within our abilities, and we should do it. We should be more skeptical of human nature, look harder at the extremes and edge cases, and work towards mitigating the risks. Will it be perfect? No. Will it be helpful? Yes.


We have bans on chemical and biological weapons, nuclear weapons are carefully monitored and there are considerable efforts to prevent proliferation. In hindsight, I think people today would have prevented the nuclear arms race, as it puts civilization at risk, and there is no end date for that.


Bang on. To add fuel to the fire (heh), the fire mentioned by parent wasn’t perceived as deadly and then suddenly was safe. It was and is deadly unless the precautions our society has spent thousands of years developing are carefully followed. Exactly what we must do with technology. Except we don’t have thousands of years.


It heartens me to see here these sort of posts!

Because it means that finally, after decades of us being blinded by the wow-factor of technology, hackers and engineers - we who are responsible for creating the next wave of technology - are starting to ask grown-up questions about whether some things are such good ideas.

That healthy scepticism doesn't have to mean pessimism, or Luddite rejection of "progress". That's enormous progress of a different kind - a shift from purely technical progress to a better balance of spiritual and social progress.


We had that before -- e.g., your nerds had grown up on s.f. allegories about society, as well as general education -- but it was obliterated by the frenzied influx during the dotcom gold rush.

Since then, "the victors write the history books", including formative influences on children's thinking.

It's encouraging if we manage to collectively scrape our way back from that, and have genuine concerns -- despite the noise of fashionable posturing, "influencer" conflicts of interest, etc.


Good points about the more rounded origins of hacker culture. I had read most of Arthur C Clarke by 9 years old and already had complex misgivings and ambivalence about "technology".

> obliterated by the frenzied influx during the dotcom gold rush.

Something changed in mid 1990s, something that was more than just the commercialisation and Eternal September. Perhaps it was millenial angst but a certain darkness and nihilism crept in as if William Gibson stopped trying to imagine future dystopias because the future had "caught up" and we were living in it. I think that's when things actually stopped progressing. After that, everything that had been written as a warning became a blueprint.

> Since then, "the victors write the history books", including formative influences on children's thinking.

80s and 90s media moguls only owned the channels, and there was cursory regulation. When you own all the platforms, and devices that people use, and shape the content they see from school-age onwards it is no longer "media" but total mind control.

> It's encouraging if we manage to collectively scrape our way back from that

It may not be "back", but it may be somewhere better than here. Each generation finds it's own voice. I have some optimism in the kids since 2000 - it's like they're born knowing "The cake is a lie". They're just not quite sure yet what to do about it.


> Every new tech is a potential powerful dangerous tech, so it was fire and so it has been any new discovery.

Yet the toothpick is a little bit less dangerous than the atomic bomb.

We shouldn't blindly accept every new "tech" as "progress", the fact that we can build it doesn't mean it'll be beneficial for us


Dental issues have killed more than atomic bombs have...


Everyone has teeth. Not everyone has had direct exposure to nuclear explosions.


I generally agree with your point, but the difference is that the dangers of most new technology can be reasonably contained. Deep fakes will undermine people's trust in anything they see or hear, which is a severe negative. It's unclear how to minimize this kind of harm.


You see that as a negative, to me I think people need to be less trust full of things they see on the internet...

I grew up with the prevailing idea of "never trust what you see online", yet today many have deep and misplaced trust in online media, influencers, and personalities.

I think a little more distrust is needed


The fact that people think that they should have any inherent trust of anything you see or hear (especially online) in the first place is a severe negative. Misinformation has always existed and always will exist, the solution is education and not knee-capping development. This same attitude would've made the internet a far less open place than it has been if it had been applied to it at its start.


"The fact that people think that they should have any inherent trust of anything you see or hear (especially online) in the first place is a severe negative".

This is inherent of our species. The idea that you should distrust anything you see or hear is an alien concept and simply not pragmatic.

"Misinformation has always existed and always will exist"

False equivalence. The online situation is brand new. Any citizen able to spread massive amounts of fake news to lots of people on the cheap is a brand new capability.

"the solution is education"

No, it isn't. Every study shows that highly educated people are also gullible enough to be manipulated with fake news. Which makes total sense, as absolutely nobody has the time to fact-check and do a deep background check on the 5 zillion pieces of information they see on a given day.


By "the solution is education" I don't mean education in general but education on how to identify poor sources of information and enough reasoning ability to not fall for obvious misinformation. As it stands we barely equip students to recognize good sources for stuff they have to cite, let alone recognizing good sources for general information.


> Deep fakes will undermine people's trust in anything they see or hear

Not quite. Deep fakes will undermine people's trust in anything digital they see or hear. It bodes ill for technology not for people per-se.

In other words, those who should be most terrified by the implications of deep-fake technologies and AI are those most heavily invested in digital technology. I sense that in some ironic way, suddenly the "boot is on the other foot".


We tell children to beware fire, we regulate it in many public areas, we put fire extinguishers all throughout buildings, and even dedicate entire buildings for fire-fighting teams.

It's very valuable to discuss the potential dangers of new technologies and how we might mitigate them.


Seems like we did that because fires happened in cities and did real damage. What damage has AI done?

> AI might put false ideas in peoples minds

Like speech or writing or media?

> no no not the ideas, its the medium, the delivery method. you see the fakes will trick people into thinking the lies are real.

oh that's it? just deception + scale.


(To note, I work in machine learning, but not in this specific area and not with Nvidia.)

Who are you quoting in your arguments here? I did not make those arguments and I cannot find someone else in this thread who made those arguments. Perhaps you are creating a strawman to argue against?

Re: Fire: I wasn’t around for it, but it’s safe to assume people discovered fire was dangerous around the same time they discovered fire, wayyyy before cities existed and before humans decided to harness it.

Re: Social harms from AI: you should ask yourself about the harms that can come from automated decision making. We’re automating decisions for policing, for hiring, for delivering posts on social media, for delivering political advertisements on social media, etc. We’re using AI research for profiling, for improving bomb drones, for getting children to spend time and money on games and social media, etc. I’m sure you agree at least one of these are harmful.

Re: Social harm from ‘deepfakes’: It’s currently costly to create a convincing fake image, audio, or video. It’s easy to extrapolate that it will be easy to make convincing fakes in the near future. It’s easy to see that can cause harm, especially since people are already tricked by obvious photoshops and deep-fakes.

In the US at least, there are political attack ads rampant today that use altered media.

I find it difficult to find a generous interpretation of someone who thinks we should proactively shut-down any discussion about potential and current harms from AI.


Fire cooked our food, what are the benefits of deep fakes?


Low bitrate avatars, Zoom-ing in your PJs, virtual cosplay, video production (low budget for SFX), video anonymity.


All of this either doesn't matter or barely matters.


Low budget movie acting.


After reading the article, you're overstating the "dangerous" part here. It's far from a fear mongering blog. They mention it's potential for misuse - no need to be saddened by that


If it makes you feel better, this post is published on the blog of Metaphysic, a company that makes hyperrealistic avatars (aka deepfakes).


But it doesn't read "dangerous" anywhere?


Except an individual cannot affect the entire world from the confort of home with fire.


Are you familiar with forest fires?


You start forest fires comfortably from your living room? What a wizard!


Nor a deep faked video once everyone understands we cant trust anything we see anymore


"can't trust anything anymore" is arguably a worse place to end up.


I disagree. Blind trust allows for propaganda.


"can't trust anything" makes any kind of reasonable society impossible.

If you meant "trust some things if they fit some kind of criteria" you mean something different than "can't trust anything."


Cant trust anything you see in the press or on the internet, in case that was not obvious.


The UI on this website is horrid.


Thanks for the post, I do suggest changing the font weight and color to make it more readable.

Here's how it might look from the client side (computed style at Inspect->Style):

.elementor-875 .elementor-element.elementor-element-230321ec { text-align: left; color: black; font-family: roboto; font-size: 21px; font-weight: 400; line-height: 1.6em; }




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: