This is only because Stable Diffusion was able to use techniques openly published by the likes of DeepMind, Google Research, and OpenAI. It even directly used some open source code from OpenAI. It wouldn't exist otherwise. DALL-E 2 was published in January 2021. I'm doing generative music AI (melody only) and there is quite a bit of real research required to make progress (I've decided not to publish for now).
If it helps, Sequoia also bet on WhatsApp, Google, and a whole host of other iconic tech companies (as well as duds). I guess, with FTX, they made the mistake of drumming up the founder too soon (btw, doing so isn't specific to Sequoia; almost all other VC firms do this. It is priming up the market, as it were).
As VCs gain more traction and became more accessible (on social media, esp), it only serves to their detriment when they make a mistake, even if it is a honest one. I don't think VCs care that much, tbh, given the power imbalance that having that much money brings.
I don't think any Sequoia fund has returned less than 8x in the last 20 years. That's better than the total US market or the S&P 500 from the low of the Great Recession to the inflation-fueled peak of the pandemic (7.1x and 6.3x respectively), which is a cherry-picked, unprecedented 13-year stretch compared to the worst-performing fund's 10 years. FTX-sized failures are expected and part of their operating model.
Well it's not just the pro-AI articles, its most new articles on the web today. It has been for some time now, google top results for IT questions (for example) are always GPT SEO spam sites and you don't figure it out until after reading a bit and seeing it's pretty generic advice and answers..
I could see a valid use for generative assets in game dev. Some open world type game with different landscapes / backgrounds that are all autogenerated.
The trillion dollar answer is Personalized Advertising: your images trained on DreamBooth or similar and used to generate advertising of you and your associates (friends and family) using various products in a happy manner, or even using competitor products in an unhappy manner. Celebrities beside you stating how wise you are using said product, beer commercials with the latest sports heroes sharing drinks with you, and major release feature films with you appearing as an extra for at little as $1 more.
Yes, it is going to happen quickly. As long as the immature pornographers can be kept at bay. There is far more revenue in advertising than jerking off, porno morons.
FYI: I created a feature film quality actor replacement system aimed at Personalized Advertising over a decade ago with an earlier version of this generative AI technology. But investors (both VC and angels) could only see porn as a viable application, while I offered presentation after presentation of how Personalized Advertising is a holy grail of advertising revenues. Perhaps it will happen this time. Immaturity can only last so long, right? Maybe?
Indeed, personalized advertising to match your psychological profile as much as possible, with just the right incentives/brain-hijacking. Then, why not just do this for all "ads" for all industries (politics is one of the most aggressive currently), and, then, why not all information one is exposed to? Let's get extreme, and look at the diplomacy-playing CICERO, that can employ strategic reasoning for how to best shape ads for you. A/B testing will becoming A/(infinity) testing until you've made a "purchase".
Such balkanization is inevitable in our current Capitalistic society. How it manifests could be varying degrees between Orwellian and Disney Innocence. Now is the time one can influence such developments. If one does not act now, you'll miss out on shaping this medium, as later you're the medium.
You're awfully ungrateful to the "porno morons" that are often first movers on innovative new tech where more traditional business models (being generous with that definition here) play it safe with cowardice
In this instance, they ruined the perception of a technology, where at this time we could have compelling media exploring the shift in narrative when the viewer is integrated into the media itself. That's media of a form never seen before, and it has society positive applications. Education at many levels is just one positive use, and for autistic children it can be a critical entry means. When actor replacement is applied to x-rated media you get a lawsuit engine, and despised by larger society.
Not trying to underplay how impressive the tech here is, but could you expand on what you see as the ($) value of something like this? From my perspective, these look like not-so-interesting stock-photography-level photos, nearly all of which have clear tells that they're AI-generated.
I think this is not necessarily the case. At least one of the generative models (DALL-E?) has a license that you own the images it generates from your prompt and can copyright and license those images.
sure, stock photography may just disappear as a business model.
while SD is freely downloadable, it isn't energetically free to run inference,
so for example (and as an off the cuff speculation), opportunities could arise to provide execution of these models from places where the cost of energy is lower
we can never confidently predict what the business models will be when such an impressive new technology arises.
Well even the free models aren't free at scale. And if you're paying for the inference anyways you might as well pay for one in a model which was specifically truncated to your needs.
To provide some perspective here, SD can be retrained on comparably small samples to generate images for very specific needs (example see [1]). So there is certainly a bunch of $ to be harvested by generating very good training sets for industry needs, say for instance mockups in architecture, creating logos and brand styles or just nice looking designs for consumer products.
How do you monetize fashion photography of fake people using fake products? It would be like putting someone else’s products in your ads. I would guess that advertisers want pictures of people wearing their own products?
as starting point for people to work on when creating graphical content. This will save a lot of time for everyone and allow non professionals to create their own assets.
I suspect "let the public type in prompts and generate funky images" isn't going to be a significant market.
But with improvements to the models I can see it fitting into a "integrated into professional creative tooling" business model. Generate a rough "first draft" based on a prompt and edit. It simultaneously keeps a human in the loop to filter some of the model's zanier outputs, makes sure the output coheres with some design vision, and also people using these features professionally are much more likely to actually pay for what will be fairly expensive-to-operate features.
I'm excited about generative models, buuuuut I will echo the other thoughts here re: VC hype. The last cycle of VC investments turned out to be full of hot air and no substance. The kind of hype around generative models feels a lot like crypto: lots of breathless excitement combined with few details. We are simply being promised that this technology will usher in some kind of golden era but nobody can seem to point to killer use cases. There's so much hand-waving around actual applications that it threatens to generate a hurricane from the sheer amount of air being displaced.
As someone who works in the AI/ML field I've explored this stuff and it's definitely promising for some use cases and can potentially be transformative for some niches, but I do not see this as world-transformative at this point and considering where said breathless hypeness is coming from I'd encourage caution in over-exuberance.
[edit] My sincere hope for this new cycle of tech hype is that we actually create some businesses of worth. I do not expect this to come from VC-world, who seem to have spent the last decade building wholly incompetent businesses and foisting them on public investors (or in the case of crypto and ICOs, just fleecing retail investors directly).
My hope is that developers can actually use these models to do valuable things that people will pay for, as opposed to shovel yet more money into yet more furnaces to prop up absurd valuations around comparably tiny revenue.
There are some serious copyright implications if the initial design of anything is going to be done by an "AI". Because copyright cannot be granted in AI work, any subsequent work based upon the AI's original creation will have to be scoped out from its relation to the original work, not leaving much for copyright to protect. This is a business concern for any business that relies on copyright to ensure their materials are not copied (pretty much all businesses).
these results are so impressive that rather than constraining the technology copyrighting may be exactly what ends up being disrupted
for non-photorealistic work, or if it becomes virtually indistinguishable from photography, how would you know whether an image was expensively sourced from a human artists/photographer or cheaply generated by an AI? how would we enforce copyrights in relation to AI if anyone can lie and claim "oh yeah, i paid a wetware neural net to do that, no problem.."?
None of this is a problem in copyright litigation today, in which anyone can copy anyone else's work and lie about it. They will have to use evidence to prove whatever claims they made. The difficulty isn't in distinguishing between the works.
we aren't talking about "copyright litigation today", we're hypothesising about how deeply disrupted the current arrangement could be by ai content generation
in particular, the argument that "ai generated works is not copyrightable" in support of it being untenable to use it for business purposes is challenged by the observation that "we won't be able to tell where it came from anyway"
>we won't be able to tell whether where it came from anyway
You won't be able to if you look solely at the output. There are ways to check audit and verify the process by which said output was produced. They are also gamable but the cost of gaming them is not insignificant.
> You won't be able to if you look solely at the output.
i think this is the scenario in question, if someone is lying about the source of an image
> There are ways to check audit and verify the process by which said output was produced. They are also gamable but the cost of gaming them is not insignificant.
this is interesting to know about. do you have some zero knowledge proof arrangement in mind here?
As I already explained to you, which you did not seem to appreciate, this is not already not a problem in copyright litigation because if your defense is "it's generated by AI" then it is your burden to establish that convincingly. You can't just say "well, you can't prove that I didn't copy it", it will not be convincing to a judge or jury. The plaintiff will make allegations that the defense will have to rebut. Saying "it's hard to tell where it came from" doesn't rebut copying.
we're definitely talking past each other at some point here (i didn't propose a defense of "it's generated by AI" for example)
- i've worked for a company which acquired its art content from a foreign outsourcing business.
- the foreign outsourcing business of course hired human artists, albeit at much lower wages than is possible where the company i worked for was located.
- the company i worked for launched a product containing this art content and never encountered a copyright challenge.
if AI eventually produces content indistinguishable from what humans can produce, what prevents these foreign outsourcing businesses from quietly pretending they still hire humans? the true origin of the content can be trivially "laundered".
I'm aware you didn't "propose" it as a defense, but that's how it would work in a litigation.
>if AI eventually produces content indistinguishable from what humans can produce, what prevents these foreign outsourcing businesses from quietly pretending they still hire humans?
The fact that they don't and if it is discovered, it will mean all of the work generated by them is not copyrightable and therefore everyone is free to copy it. They'd also have to lie to the copyright office and the businesses they were selling their services. The only thing doing the work here is the lie.
Right now I can go to another place and find another photographer who will recreate something for less. Technology might make that more trivial, but it doesn't change the fact that as a business opportunity it'd only work by having to massively misrepresent oneself.
Ostensibly, if I was a business and I was contracting these kinds of services, one thing guaranteed to be in the contract would be a clause that they authored the works, obtained the copyrights, etc, lawfully. And generally in a contract they would represent what the business represents (we are an illustration shop that sells images that we authored... etc).
So to more specifically answer your question - "what would stop them?" the same exact repercussions as would right now (which is what I intended to make clear in my initial response to you).
> The fact that they don't and if it is discovered, it will mean all of the work generated by them is not copyrightable and therefore everyone is free to copy it. They'd also have to lie to the copyright office and the businesses they were selling their services. The only thing doing the work here is the lie.
you and i have extremely different impressions of how effective regulatory repercussions are at actually constraining behaviour.
circumventing laws and regulations really does do a lot of the work in the economy. vast amounts of plagiarised code run in software we touch every day. artistic work gets crudely lifted. i don't see something as powerful as ai content generation being constrained by dictums about what can be copyrighted.
What regulators? These are all private actors using contract law and copyright law. They are only as effective as the potential gain in winning a lawsuit and I never suggested otherwise. As I pointed out, already, yes people do copy now. I think your belief is mistaken that the triviality of creating the content is the fulcrum in this scenario. The fulcrum is profitability. Having to operate extra-judiciously and at the bottom of the barrel in order to make your grift isn't an exciting proposition now, and it's not going to be in the future.
Animated series generated from words and text, no labour needed. Just make the characters, the base models, a few demos. South Park has a staff to do all the animation - no more - now they'll just tweak a few things.
You will see some amazing, raw creative stories come out of this that are going to be entertaining in ways we haven't fathomed. That may be a bit of a meme.
I suggest it may fundamentally alter low-grade creative content. Basically, second tier fiction I think will be a total commodity. So cheezy romance novels, fantasy - I'll bet in 10 years AI is as good as most mediocre authors.
They are already generating news articles. Bloomberg has an AI creating articles. Basically to 'verbalize and organize facts'.
One thing coming I think nobody is talking about is lighting and it will change everything. Lighting on set is a massive pain, everything revolves around it. Being able to use AI for lighting will change production costs dramatically.
All of the immediate applications I envision outside of entertainment are nefarious, sinister. I don’t want to share these ideas, but I know if this stuff has occurred to me, it will be implemented eventually.
"Generative AI can make these workers at least 10% more efficient and/or creative: they become not only faster and more efficient, but more capable than before. Therefore, Generative AI has the potential to generate trillions of dollars of economic value."
What utter bull - at a time when inflation is on the rise (7-14%), you aren't going to unlock ANY market potential, in fact you are going to shrink the market, significantly.
Prepare for impending market collapse and massive market failure...
This is a short term shock to a long term trend. Absent a nuclear cataclysm, tech and productivity are always going to grow.
Investors go long and capture the outsized returns. The narrative and potential of AI/ML is incredible. It stands to be as disruptive as smartphones and the internet were.
As for your thesis on the current economic picture, much of the market still thinks there's an easy out just around the corner. You can make a bet on that and find out.
Sequoia was very late in getting into crypto and then piled money into FTX without any diligence, and then wrote glowing articles on SBF promoting him as a savior of humanity:
The expectations they have for 2030 are frankly delusional.
“In 3-8 years we'll have a machine with the general intelligence of an average human being. The machine will begin to educate itself with fantastic speed. In a few months it will be at genius level and a few months after that its powers will be incalculable.”
—Marvin Minsky, 1970
>The expectations they have for 2030 are frankly delusional.
Well, they do start with the delusional "copilot writes 40% of your code" microsoft marketing bullshit which everyone who has ever used it knows is not true, so it really is "garbage in, garbage out".
83 comments
[ 1.7 ms ] story [ 181 ms ] threadSince the postscript mentions that the article was co-written by GPT-3, perhaps this is one of the generated lines? :)
(compared to the prior SOTA for generative ML in images i.e. GANs, diffusion models use far more compute-heavy and are much slower!)
Sure, Sequoia wants to hear all about your great startup proposals.
And it will (rightly) reject 99%, and finally fund a "known team that can execute" kidding not kidding
I'm doing vocals! Can't wait for the seams to come together.
Although, I'll remain skeptical of anything brought to us by the same Sequoia who also wrote a breathless SBF puff-piece that aged like fine milk: https://web.archive.org/web/20220922164619/https://www.sequo...
As VCs gain more traction and became more accessible (on social media, esp), it only serves to their detriment when they make a mistake, even if it is a honest one. I don't think VCs care that much, tbh, given the power imbalance that having that much money brings.
> We also generated illustrations for this post with Midjourney
Something about the writing style is just off.
GPT-3 can generate pretty convincing anti-AI copypasta.
- GPT-3
Please Sequoia, for we are only knaves, let us play
As massive VCs are starting to take note, it's starting to feel a bit like the crypto scene. Interesting tech without much real-world application.
If done right, image generation and augmentation could be a powerful and valuable feature.
a few in the community are diving deep on game assets: https://twitter.com/emmanuel_2m
Yes, it is going to happen quickly. As long as the immature pornographers can be kept at bay. There is far more revenue in advertising than jerking off, porno morons.
FYI: I created a feature film quality actor replacement system aimed at Personalized Advertising over a decade ago with an earlier version of this generative AI technology. But investors (both VC and angels) could only see porn as a viable application, while I offered presentation after presentation of how Personalized Advertising is a holy grail of advertising revenues. Perhaps it will happen this time. Immaturity can only last so long, right? Maybe?
A further balkanization of reality.
https://twitter.com/rainisto/status/1597206841813274626?t=7B...
The future is going to be a mish-mash of the past?
If everyone can just run their own prompt against free models, wouldn't it be more likely that no one pays for it anymore?
while SD is freely downloadable, it isn't energetically free to run inference,
so for example (and as an off the cuff speculation), opportunities could arise to provide execution of these models from places where the cost of energy is lower
we can never confidently predict what the business models will be when such an impressive new technology arises.
To provide some perspective here, SD can be retrained on comparably small samples to generate images for very specific needs (example see [1]). So there is certainly a bunch of $ to be harvested by generating very good training sets for industry needs, say for instance mockups in architecture, creating logos and brand styles or just nice looking designs for consumer products.
[1] https://lambdalabs.com/blog/how-to-fine-tune-stable-diffusio...
But with improvements to the models I can see it fitting into a "integrated into professional creative tooling" business model. Generate a rough "first draft" based on a prompt and edit. It simultaneously keeps a human in the loop to filter some of the model's zanier outputs, makes sure the output coheres with some design vision, and also people using these features professionally are much more likely to actually pay for what will be fairly expensive-to-operate features.
I'm excited about generative models, buuuuut I will echo the other thoughts here re: VC hype. The last cycle of VC investments turned out to be full of hot air and no substance. The kind of hype around generative models feels a lot like crypto: lots of breathless excitement combined with few details. We are simply being promised that this technology will usher in some kind of golden era but nobody can seem to point to killer use cases. There's so much hand-waving around actual applications that it threatens to generate a hurricane from the sheer amount of air being displaced.
As someone who works in the AI/ML field I've explored this stuff and it's definitely promising for some use cases and can potentially be transformative for some niches, but I do not see this as world-transformative at this point and considering where said breathless hypeness is coming from I'd encourage caution in over-exuberance.
[edit] My sincere hope for this new cycle of tech hype is that we actually create some businesses of worth. I do not expect this to come from VC-world, who seem to have spent the last decade building wholly incompetent businesses and foisting them on public investors (or in the case of crypto and ICOs, just fleecing retail investors directly).
My hope is that developers can actually use these models to do valuable things that people will pay for, as opposed to shovel yet more money into yet more furnaces to prop up absurd valuations around comparably tiny revenue.
for non-photorealistic work, or if it becomes virtually indistinguishable from photography, how would you know whether an image was expensively sourced from a human artists/photographer or cheaply generated by an AI? how would we enforce copyrights in relation to AI if anyone can lie and claim "oh yeah, i paid a wetware neural net to do that, no problem.."?
in particular, the argument that "ai generated works is not copyrightable" in support of it being untenable to use it for business purposes is challenged by the observation that "we won't be able to tell where it came from anyway"
You won't be able to if you look solely at the output. There are ways to check audit and verify the process by which said output was produced. They are also gamable but the cost of gaming them is not insignificant.
i think this is the scenario in question, if someone is lying about the source of an image
> There are ways to check audit and verify the process by which said output was produced. They are also gamable but the cost of gaming them is not insignificant.
this is interesting to know about. do you have some zero knowledge proof arrangement in mind here?
- i've worked for a company which acquired its art content from a foreign outsourcing business.
- the foreign outsourcing business of course hired human artists, albeit at much lower wages than is possible where the company i worked for was located.
- the company i worked for launched a product containing this art content and never encountered a copyright challenge.
if AI eventually produces content indistinguishable from what humans can produce, what prevents these foreign outsourcing businesses from quietly pretending they still hire humans? the true origin of the content can be trivially "laundered".
>if AI eventually produces content indistinguishable from what humans can produce, what prevents these foreign outsourcing businesses from quietly pretending they still hire humans?
The fact that they don't and if it is discovered, it will mean all of the work generated by them is not copyrightable and therefore everyone is free to copy it. They'd also have to lie to the copyright office and the businesses they were selling their services. The only thing doing the work here is the lie.
Right now I can go to another place and find another photographer who will recreate something for less. Technology might make that more trivial, but it doesn't change the fact that as a business opportunity it'd only work by having to massively misrepresent oneself.
Ostensibly, if I was a business and I was contracting these kinds of services, one thing guaranteed to be in the contract would be a clause that they authored the works, obtained the copyrights, etc, lawfully. And generally in a contract they would represent what the business represents (we are an illustration shop that sells images that we authored... etc).
So to more specifically answer your question - "what would stop them?" the same exact repercussions as would right now (which is what I intended to make clear in my initial response to you).
you and i have extremely different impressions of how effective regulatory repercussions are at actually constraining behaviour.
circumventing laws and regulations really does do a lot of the work in the economy. vast amounts of plagiarised code run in software we touch every day. artistic work gets crudely lifted. i don't see something as powerful as ai content generation being constrained by dictums about what can be copyrighted.
And they can be narrowed in focus.
You will see:
Animated series generated from words and text, no labour needed. Just make the characters, the base models, a few demos. South Park has a staff to do all the animation - no more - now they'll just tweak a few things.
You will see some amazing, raw creative stories come out of this that are going to be entertaining in ways we haven't fathomed. That may be a bit of a meme.
I suggest it may fundamentally alter low-grade creative content. Basically, second tier fiction I think will be a total commodity. So cheezy romance novels, fantasy - I'll bet in 10 years AI is as good as most mediocre authors.
They are already generating news articles. Bloomberg has an AI creating articles. Basically to 'verbalize and organize facts'.
One thing coming I think nobody is talking about is lighting and it will change everything. Lighting on set is a massive pain, everything revolves around it. Being able to use AI for lighting will change production costs dramatically.
What utter bull - at a time when inflation is on the rise (7-14%), you aren't going to unlock ANY market potential, in fact you are going to shrink the market, significantly.
Prepare for impending market collapse and massive market failure...
https://www.statista.com/statistics/225698/monthly-inflation...
https://ycharts.com/indicators/us_inflation_rate
If you are so good at seeing the future, you should be the richest person on earth. Surely there's a way to short the market.
Investors go long and capture the outsized returns. The narrative and potential of AI/ML is incredible. It stands to be as disruptive as smartphones and the internet were.
As for your thesis on the current economic picture, much of the market still thinks there's an easy out just around the corner. You can make a bet on that and find out.
https://web.archive.org/web/20221027181005/https://www.sequo...
Sequoia coming into AI now, with the same kind of grandiose proclamations doesn't sit well.
"Copilot is generatin 40% of code in projects where it is installed" - is that any true, to start with something? What are these kind of projects?
“In 3-8 years we'll have a machine with the general intelligence of an average human being. The machine will begin to educate itself with fantastic speed. In a few months it will be at genius level and a few months after that its powers will be incalculable.” —Marvin Minsky, 1970
Well, they do start with the delusional "copilot writes 40% of your code" microsoft marketing bullshit which everyone who has ever used it knows is not true, so it really is "garbage in, garbage out".