[00:13.610 --> 00:14.150]  Fantastic.
[00:19.010 --> 00:22.230]  I'm going to want to see what other questions there are.
[00:23.070 --> 00:26.330]  And then, sure, I'll be here for as long as you guys are.
[00:28.970 --> 00:30.690]  We'll look for emails.
[00:31.290 --> 00:32.650]  We'll find some expertise.
[00:33.910 --> 00:38.310]  We can do a podcast every two weeks or so.
[01:09.570 --> 01:10.950]  Good morning, everyone.
[01:11.150 --> 01:12.390]  Welcome to the third day of HOPE.
[01:12.550 --> 01:13.050]  How are you doing?
[01:13.470 --> 01:13.790]  Great.
[01:14.630 --> 01:15.850]  Seen some good shows?
[01:16.850 --> 01:17.790]  Some good sessions?
[01:18.870 --> 01:19.530]  All right.
[01:19.710 --> 01:22.350]  Well, we're going to start off on fire this morning.
[01:22.810 --> 01:28.570]  So, the talk this morning is Artificial Intelligence, How IP Law Handles Machine Creations.
[01:28.810 --> 01:37.390]  And our presenter here, Ed Ryan, is going to help us think about or maybe find some initial answers to the questions of can a machine being an inventor?
[01:37.710 --> 01:41.750]  And does the machine's output qualify for copyright or patent protection?
[01:42.090 --> 01:43.690]  So, with that, I'll pass it over to Ed.
[01:47.910 --> 01:49.070]  Good morning, everybody.
[01:50.150 --> 01:50.790]  All right.
[01:50.930 --> 01:52.190]  It looks like it's up there correctly.
[01:52.450 --> 01:53.850]  Thank you so much for coming.
[01:53.970 --> 01:59.170]  I know it's 10 o'clock on the last day of a hacker convention, and we're all worn a little thin.
[01:59.190 --> 02:01.970]  So, I very much appreciate your interest.
[02:02.950 --> 02:05.310]  To start off with, here's a little bit about me.
[02:05.310 --> 02:07.730]  I've been a patent attorney for 14 years.
[02:07.990 --> 02:09.930]  My background is in physics.
[02:11.010 --> 02:16.410]  But the thing with physics is it kind of gives you a general applicability to a lot of different things.
[02:16.770 --> 02:22.930]  And when I was in college, I was also marinating in computer science just by kind of hanging around.
[02:23.290 --> 02:28.190]  So, I have, over time, wormed my way into working on machine learning.
[02:28.190 --> 02:30.590]  I handle a lot of machine learning cases.
[02:31.590 --> 02:37.050]  And as a result, I come to you as a very serious person to talk about very serious things.
[02:37.570 --> 02:41.830]  I think I've represented my general attitude toward life well in my pictures there.
[02:42.250 --> 02:48.850]  I'm also going to take this opportunity to very briefly give a shout-out to my children, who I think are watching on YouTube right now.
[02:49.090 --> 02:49.570]  Hi.
[02:49.790 --> 02:50.870]  I will see you all later.
[02:54.190 --> 02:59.150]  So, this talk is about artificial intelligence.
[02:59.390 --> 03:05.090]  And before we can even interface with that question, we have to know what we're talking about.
[03:05.090 --> 03:11.190]  And the problem with that is the definition of what is artificial intelligence has changed dramatically over the course of decades, right?
[03:11.350 --> 03:14.310]  The field has its start in the 60s.
[03:14.310 --> 03:24.590]  And it has this bad habit of moving the goalposts, where every time something new comes out, like, oh, man, look at this great artificial intelligence.
[03:24.810 --> 03:26.110]  And 10 years later, it's old hat.
[03:26.170 --> 03:26.730]  And like, that?
[03:26.890 --> 03:28.070]  We've been doing that for years.
[03:28.290 --> 03:30.010]  This is artificial intelligence.
[03:30.270 --> 03:41.430]  So, for the purpose of this talk today, we're going to talk about what, you know, everybody kind of refers to in the common conversation as artificial intelligence.
[03:41.430 --> 03:53.390]  We're going to be talking about advanced machine learning tools that are trained on a large amount of data, right?
[03:54.070 --> 04:01.270]  They're usually in a neural network architecture, and particularly a deep neural network architecture, and we'll get into that in a little bit.
[04:02.170 --> 04:16.970]  But this new model of artificial intelligence systems has had tremendous success in certain regards, deeply creepy success in certain regards, that have, like, made people stand back and really wonder, like, what is going on here?
[04:17.050 --> 04:18.650]  Is this a new class of thing?
[04:22.770 --> 04:36.770]  So, when we talk, you know, just to tell you what I'm going to tell you, we're going to start by giving you a quick, very surface-level understanding of what an artificial intelligence system looks like and how it works.
[04:37.010 --> 04:44.470]  We're going to move on to talking about some pressing examples that are really, you know, really motivated this talk.
[04:44.710 --> 04:52.790]  So, we're going to talk about DALI and its ability to create, you know, photorealistic and really compelling artwork.
[04:52.790 --> 05:04.090]  We're going to talk about Lambda a little bit, the Google chat bot that raised questions recently about its sentience by coming right out and saying, hey, I'm sentient, don't turn me off.
[05:04.690 --> 05:06.350]  We're going to talk about the law.
[05:06.570 --> 05:30.690]  And so, the law in this case means copyright law and patent law, because the purpose of this talk is to take the output of these machine learning models and figure out whether that output is entitled to copyright protection or patent protection under our current intellectual property laws.
[05:30.910 --> 05:41.210]  And we're really not going to be talking about sentience, and we're not going to be talking about a lot of the ethical considerations that people are trying to bite into right now.
[05:41.210 --> 05:53.910]  There is so much to talk about around artificial intelligence, and I'm really trying to focus because otherwise we would spend the entire day here.
[05:54.950 --> 05:58.690]  So, let's start with talking about AI and what AI can do.
[05:58.830 --> 06:06.650]  Just for reference, this is an AI-generated face swap of Stephen Wolfram's face with Isaac Newton.
[06:09.790 --> 06:20.850]  So, in general terms, a neural network or an AI, you know, let's focus on neural networks because that's the most common and powerful technology we have right now.
[06:21.010 --> 06:24.530]  It can be thought of as a universal approximator.
[06:24.950 --> 06:27.950]  And so, what that means is a neural network is just a function.
[06:28.570 --> 06:32.930]  It takes an input, it does some work on it, it produces an output.
[06:32.930 --> 06:41.730]  And in that respect, it's equivalent or, you know, approximate to any other function that you're modeling, right?
[06:41.930 --> 06:53.250]  So, you could have... in my view, there's really no material difference between a neural network and a classically programmed computer program.
[06:53.470 --> 06:59.470]  The difference is not in the function, the difference is in how it's created, right?
[06:59.470 --> 07:04.150]  So, with a computer program, you have an engineer who sits down and says, I want the thing to do the thing.
[07:04.470 --> 07:14.410]  And so, I'm going to create all of the structures, all of the processes that are required, and it's going to behave exactly as I intend it to, exactly as I tell it to.
[07:15.590 --> 07:20.730]  The way a neural network system works is you create an architecture, right?
[07:20.850 --> 07:27.930]  You have... there are a variety of different kinds of neural network architectures that are better, you know, some are better for some purposes than others.
[07:28.150 --> 07:32.950]  But you create this architecture, and then you throw data at it, right?
[07:33.150 --> 07:43.450]  You throw a huge volume of either labeled or unlabeled data, depending on, you know, how much human interaction is really required.
[07:44.510 --> 07:49.890]  And it learns how to mimic the situations that created that data.
[07:49.890 --> 07:54.070]  And this data can really come down to be anything that's quantified, right?
[07:54.170 --> 08:02.330]  So, when you're talking about a picture, we have tons and, you know, huge volumes of training data, because we have a lot of pictures, right?
[08:02.410 --> 08:03.090]  Video streams.
[08:03.430 --> 08:05.450]  Every frame of that is a picture.
[08:05.630 --> 08:08.370]  And that can be quantified according to its pixel values.
[08:08.570 --> 08:18.290]  So, you take a picture, you put it into the machine learning system, it breaks it down into a set of features, it works on those features, and it gives you the output you look for.
[08:18.290 --> 08:21.630]  You can do that for really any kind of sensor, right?
[08:21.870 --> 08:34.930]  So, we have machine learning systems that are constantly monitoring, looking for anomalous system behavior, because it's learned what normal system behavior looks like, based on a large volume of training data.
[08:35.090 --> 08:40.090]  So, when something new happens, it can say, hey, this is weird, let's do something about it.
[08:40.650 --> 08:56.450]  And really, the source of all of the creepy behavior behavior, when it comes to machine learning systems, comes from the fact that you can also quantify words and language in a way that's really astoundingly powerful, right?
[08:56.550 --> 09:02.590]  Because what you can end up doing is you can take a word and quantify it, not by, like, this is the letter A and it has a value one.
[09:02.870 --> 09:04.990]  You end up quantifying it by its meaning.
[09:05.430 --> 09:15.350]  We have tools that will take words and sentences and represent them in this large multidimensional space where a vector points to a particular spot.
[09:15.470 --> 09:19.050]  And that's just what that word means in a numerical fashion.
[09:19.330 --> 09:24.550]  So, let's talk about the AI system and how it does what it does.
[09:24.690 --> 09:26.490]  And just super surface level.
[09:26.630 --> 09:29.050]  This picture here made it on MS Paint.
[09:29.290 --> 09:32.910]  It looks like every other basic neural network explanation.
[09:33.090 --> 09:34.230]  You have...
[09:34.230 --> 09:37.470]  You can look at it as a set of vertical columns, right?
[09:37.470 --> 09:41.010]  On the left, you have a column of input neurons.
[09:41.390 --> 09:44.330]  On the right, you have a column of output neurons.
[09:44.710 --> 09:46.990]  And in the middle, you have the hidden neurons.
[09:47.090 --> 09:48.470]  And that's where the magic happens.
[09:48.730 --> 09:50.730]  So, your input neurons, you...
[09:50.730 --> 09:54.430]  If you had a photograph, right?
[09:54.550 --> 10:05.170]  That was 100 by 100 pixels, you would have 10,000 input neurons, one for each pixel or more, you know, depending on how many different values you're tracking from that, from each pixel.
[10:05.730 --> 10:07.910]  And then the output, let's say this is...
[10:07.910 --> 10:09.710]  We're classifying, right?
[10:09.810 --> 10:11.070]  So, we have this input image.
[10:11.290 --> 10:13.130]  We want to know if there's a face in it.
[10:13.230 --> 10:15.090]  It's a very common machine learning task.
[10:15.310 --> 10:19.210]  The output might just be a single bit, yes or no.
[10:19.470 --> 10:20.370]  There is a face here.
[10:20.450 --> 10:21.290]  There isn't a face here.
[10:21.350 --> 10:22.250]  It could be more complicated.
[10:22.490 --> 10:29.390]  In addition to having a face, it could have a bounding box that says the face is within this square in the image.
[10:30.750 --> 10:39.450]  The lines between the neurons could be called synapses or weights, whatever, depending on how closely you want to hold to the brain analogy.
[10:41.370 --> 10:44.210]  Those represent really just coefficients.
[10:44.990 --> 10:45.430]  Okay?
[10:45.690 --> 10:53.730]  So, when you do the feed-forward operation, which is, I am running this, I'm giving it an input, I want to see what the output is.
[10:53.730 --> 10:55.090]  You put your input in.
[10:55.410 --> 11:00.890]  Each neuron has a function inside it that operates a very simple function generally.
[11:01.990 --> 11:03.170]  Produces an output.
[11:03.390 --> 11:06.110]  That output is weighted by the synapse.
[11:06.250 --> 11:16.650]  So, it's multiplied by whatever that coefficient is and becomes the input to one of the neurons in the next layer.
[11:17.110 --> 11:27.790]  This is what's called a fully connected layer in that each neuron of a given layer is connected to every neuron of the next layer.
[11:28.030 --> 11:30.970]  There are, you know, this is basically as simple as it gets.
[11:31.250 --> 11:39.290]  You can have thousands, millions, however many layers you want, and that's what you call a deep neural network.
[11:39.470 --> 11:44.650]  And what ends up happening is the more layers you get, the more complexity you can really capture.
[11:46.650 --> 11:49.470]  So, that's the simple...
[11:50.490 --> 11:52.810]  I'm running this and I want to get an output.
[11:53.070 --> 12:00.910]  Now, when we're training this neural network, what we do is we take our training data set and we split it in two or three, right?
[12:00.990 --> 12:06.270]  We are looking to run it on a set...
[12:06.270 --> 12:14.310]  First, we run it on a set of instances where we know what the answer is going to be, okay?
[12:14.310 --> 12:15.510]  So, we run it through.
[12:15.670 --> 12:18.530]  We're expecting it to give us a five and it gives us a four.
[12:18.710 --> 12:19.970]  That's an error, right?
[12:20.070 --> 12:25.870]  The machine learning model has failed to give us what we know the true answer to be because this training data set is labeled.
[12:26.070 --> 12:29.030]  Somebody sat down and said, yes, there's a face in this one.
[12:29.170 --> 12:30.570]  No, there's not a face in this one.
[12:30.670 --> 12:40.150]  And so, when it gives us an error, we calculate what that error is using some optimization function, whatever the tool we're using.
[12:40.670 --> 12:44.390]  And then we propagate it backward through the neural network, right?
[12:44.650 --> 12:52.210]  And that lets us change the synapse weights to reflect the fact that we had an error.
[12:53.630 --> 12:57.910]  Ideally, to correct that error and move us closer to the correct answer.
[12:57.930 --> 13:03.690]  So that now when we run the next piece of data, it gets closer and closer and hopefully eventually converges.
[13:03.870 --> 13:07.690]  And then we test it by looking at our second batch of data.
[13:08.290 --> 13:14.670]  And we say, are you giving us the correct answer for data that you haven't seen before?
[13:14.990 --> 13:18.290]  And that lets us validate that we did a good job of training.
[13:20.150 --> 13:26.390]  And so I'm going to take us a little aside here because a lot of the types of data that we've mentioned so far, right?
[13:27.090 --> 13:39.430]  Works of graphical works, pictures, text, a lot of this is already copyrighted by individual human beings who wrote these things or painted these things or took the picture.
[13:39.710 --> 13:43.270]  And in those cases, we have to...
[13:43.270 --> 13:54.330]  There is an open legal question of to what degree is it legal to use somebody else's copyrighted work to train a neural network, right?
[13:54.430 --> 13:57.450]  So I don't want to give you the impression that this is all...
[13:58.490 --> 14:03.050]  Even now, this early on in the talk, we are running into legal problems.
[14:03.170 --> 14:15.930]  And that's not even the legal problem that I wanted to spend time on today, but it's something to keep in mind that every step of the way, because this is such a new technology, every step of the way, we really don't know what we're doing.
[14:16.810 --> 14:19.050]  And I say that with authority as an attorney.
[14:20.210 --> 14:22.390]  All right, so let's talk about what we can do with AI.
[14:22.590 --> 14:26.190]  And there are a lot of very powerful things and a lot of kind of unsettling things.
[14:26.430 --> 14:27.650]  We have...
[14:27.650 --> 14:31.390]  The thing that really creeps me out is face swapping and deep fakes.
[14:31.390 --> 14:40.450]  It really, like, undermines my, you know, trust in what we see in a way that I did not expect to see so soon.
[14:41.650 --> 14:44.750]  We have facial recognition in crowds, right?
[14:44.970 --> 14:51.830]  Where in a crowd of 100 people, this computer can point out the bad actor that we've, you know, got in our database.
[14:52.530 --> 14:56.150]  We have convincing human chatbots, human-level art.
[14:56.350 --> 15:04.290]  In this case, I have some examples of a funny picture of me with some unsettling filters that were generated to put them in different styles.
[15:04.630 --> 15:11.710]  But at the end of the day, all of these are about processing a kind of meaning, right?
[15:12.370 --> 15:17.470]  And I want to tie this back to what we were talking about when I mentioned words.
[15:18.290 --> 15:19.410]  Because it's...
[15:19.410 --> 15:22.610]  And this is a personal feeling that I'm expressing.
[15:22.650 --> 15:30.810]  I find it deeply unsettling that a computer can so easily manipulate meaning instead of just...
[15:30.810 --> 15:33.010]  You know, on the level of meaning, right?
[15:33.250 --> 15:36.150]  This means the same thing as that thing, right?
[15:38.050 --> 15:43.350]  That is really the heart of where the creepiness comes from in my view.
[15:44.350 --> 15:46.070]  Oh, and that's what I'm already talking about.
[15:46.310 --> 15:53.150]  So, just to give you an idea here of how that works, there's a very common tool called Word2Vec.
[15:53.890 --> 15:55.730]  It learns based...
[15:55.730 --> 16:02.710]  And it's unsupervised, which is to say no human being sits down and programs in the meaning of these words, right?
[16:02.910 --> 16:08.990]  So, it takes a large corpus of written text and it churns through it.
[16:09.050 --> 16:17.590]  And it does it based on kind of word placement, where if a word is placed in a similar place in sentences as another word, it has a similar meaning, right?
[16:17.730 --> 16:26.210]  And as a result, it can put those on this thousand-dimension space to give really accurate results for what that word means.
[16:28.330 --> 16:31.410]  So, let's now start applying that concept.
[16:32.350 --> 16:36.810]  Dolly and Dolly2 have been getting a lot of traction lately.
[16:36.810 --> 16:43.490]  And they're an example of, you know, extracting meaning and representing it in a different way, right?
[16:43.730 --> 16:45.190]  Similar to face swapping.
[16:45.190 --> 16:48.210]  We're extracting the face, the meaning here.
[16:48.430 --> 16:54.270]  And we're putting it, you know, in a place it could belong based on its context and its meaning.
[16:55.330 --> 17:06.370]  Dolly takes a natural language input, extracts the meaning from it, and creates an image that represents that meaning, right?
[17:06.530 --> 17:12.010]  And it's based on GPT-3, which is kind of a state-of-the-art language model.
[17:12.710 --> 17:17.770]  And it creates outputs that are really like human-level art.
[17:19.050 --> 17:29.910]  And I just want to say I'm so grateful that all of these things have been happening because it makes it very easy for me creating a presentation to come up with images.
[17:30.790 --> 17:46.910]  So most of the images that you're going to see are actually just me futzing around with Dolly Mini or my friend let me abuse his Dolly access to get some higher quality stuff as well.
[17:47.730 --> 17:54.990]  So let's give another example, which is Lambda, which is the Google chatbot that was in the news recently for being like, hey, I'm sentient, I promise.
[17:55.650 --> 18:06.870]  But as it's relevant to this talk, part of the proof for that was the guy asked Lambda, can you write a fable about your experiences?
[18:07.330 --> 18:08.870]  And it did, right?
[18:09.030 --> 18:13.110]  It wrote something that, you know, it's not particularly sophisticated literature.
[18:13.570 --> 18:21.150]  But if you were to look at this and ask, you know, assume that this were the output of a human being and ask, is this copyrightable?
[18:21.230 --> 18:23.210]  The answer would be yes, right?
[18:23.390 --> 18:29.950]  This is an actual story, not copied from anywhere, not, you know, it's an original in the sense that it is new.
[18:30.410 --> 18:36.210]  And we haven't seen this particular thing before and it was created by a machine.
[18:40.540 --> 18:44.780]  Now we're going to take a little bit of a weird turn because we're going to talk about DABIS.
[18:45.180 --> 18:47.760]  Now we're getting into patent law, okay?
[18:48.560 --> 18:53.160]  And patent law operates in a very different way from copyright law.
[18:53.420 --> 18:57.400]  It's not about, you know, this creative expression.
[18:57.660 --> 19:02.640]  It's about whether you have created something new, non-obvious, and useful.
[19:03.320 --> 19:25.800]  And so this guy, Dr. Stephen Thaler, or Thaler, created DABIS, which is an AI model that purportedly generates inventions and does so in a way that, you know, with all of the enabling details so that a person having ordinary skill in the art would be able to make and use that invention.
[19:25.800 --> 19:27.080]  And I've got a picture here.
[19:27.180 --> 19:29.900]  This is a fractal-shaped food container.
[19:31.900 --> 19:35.100]  This is currently, like, working its way through the courts, actually.
[19:35.740 --> 19:48.240]  But Dr. Thaler claims he played no role in the conception of this invention and that the device itself is the inventor.
[19:48.640 --> 19:57.080]  Which, you know, speaking from my perspective as somebody who is passingly familiar with machine learning technologies, that's pretty dubious.
[19:57.300 --> 20:03.000]  But it's treated as true for the sake of litigation, which is an interesting way to approach it.
[20:03.460 --> 20:12.160]  Basically, the patent office never stopped and said, hey, you, Dr. Thaler, should be an inventor here.
[20:12.320 --> 20:18.140]  Instead, what they did was they took him at his word and they said, well, a machine can't be an inventor.
[20:18.360 --> 20:25.640]  And so, for the sake of litigation, we're accepting it as true that the machine did all of the work here.
[20:26.840 --> 20:39.840]  And then something that just, you know, recently popped up is the GitHub Copilot, which is a machine learning model trained on the corpus of code stored in GitHub.
[20:40.400 --> 20:43.520]  And what it does is it generates new code, right?
[20:43.600 --> 20:48.540]  It's basically a complex autocomplete feature for source code.
[20:50.460 --> 21:03.100]  And there are a bunch of open questions and upset people who are pointing out that sometimes it generates the same code that was in the training data.
[21:03.360 --> 21:06.040]  And that raises questions of licensing.
[21:06.400 --> 21:09.100]  It raises questions of copyright infringement, naturally.
[21:11.380 --> 21:12.520]  It's a...
[21:15.620 --> 21:17.320]  Sorry, I just skipped a gear there.
[21:17.520 --> 21:30.740]  In any case, it's yet another example of a place where a machine is generating useful outputs that if they were generated by a human being, there would be no question.
[21:30.740 --> 21:32.140]  This is copyrightable stuff.
[21:33.660 --> 21:41.560]  And finally, I want to give you a counterpoint, which is the old infinite monkey theorem, right?
[21:41.640 --> 21:53.060]  And the concept here is that if you give a monkey a typewriter, and you let that monkey bash on the typewriter at random for an infinite amount of time, it will generate any arbitrary work, right?
[21:53.140 --> 22:00.860]  It will generate the complete works of Shakespeare, word for word, letter for letter, with no errors, because infinity is really big, and that's how randomness works.
[22:02.600 --> 22:12.000]  And so the question that we have to ask ourselves, is this monkey different from an AI, right?
[22:12.220 --> 22:19.540]  Is this monkey, which is just another kind of, another kind of way of representing a system that is generating a useful output?
[22:19.800 --> 22:25.060]  An output that, if it were created by a human being without question, we would protect it.
[22:25.340 --> 22:28.500]  You know, either a patent or copyright or whatever, right?
[22:28.680 --> 22:32.520]  And we kind of intuitively sense, this is just random noise.
[22:32.740 --> 22:48.460]  And the fact that the random noise happens to have something that looked like a signal in it doesn't imply that it's entitled to the rights and privileges of a human being as set out in the copyright and patent statutes.
[22:48.680 --> 22:51.260]  So I want you to keep that in mind, right?
[22:51.580 --> 22:55.020]  Everything that we're going to be talking about today, compare that to the monkey.
[22:56.800 --> 22:57.240]  Okay?
[22:57.380 --> 22:59.140]  So now we're going to start talking about the law.
[23:00.900 --> 23:02.120]  And I'm going to have a sip of water.
[23:06.300 --> 23:06.860]  All right.
[23:07.440 --> 23:09.120]  So patent law and copyright law.
[23:09.560 --> 23:19.420]  I'm going to, I've quoted from the relevant statutes here, but patent law grants exclusive rights to inventors.
[23:19.700 --> 23:24.400]  Whomever invents or discovers any new and useful process, machine, manufacturer, blah, blah, blah, blah, blah.
[23:25.060 --> 23:28.260]  Copyright law covers original works of authorship.
[23:28.660 --> 23:33.460]  So in both cases we have this vague subject, right?
[23:33.580 --> 23:35.340]  The author, the inventor.
[23:35.680 --> 23:40.460]  And we, it's our job today to try and figure out how broad those terms really are.
[23:40.460 --> 23:47.100]  The, the source of these come from the United States Constitution, at least in our realm, right?
[23:47.320 --> 24:06.000]  I can't speak so much for, for other countries, but in the Constitution we have an, an explicit provision that empowers Congress to make laws like this to promote the progress of science and useful arts by, uh, securing for limited times to authors and inventors.
[24:06.260 --> 24:14.220]  And so it turns out that the Founding Fathers didn't have any idea that such a thing as an AI would exist someday.
[24:14.720 --> 24:24.880]  Uh, and, there's not a lot of guidance there for what did they really mean when they say author and inventor because at the time there was only one thing that it could mean and that was a natural human being.
[24:26.520 --> 24:36.240]  So, it turns out that the concept of non-human authorship has already been explored in the law but not for machines, rather for animals.
[24:37.760 --> 24:47.640]  one example is the monkey selfie case where the courts found that copyrights do not expressly authorize monkeys to sue over copyrights.
[24:47.760 --> 24:50.880]  Basically, they lack standing under the copyright law.
[24:51.200 --> 25:01.200]  Um, and that, you know, the courts have also found that Congress can give animals the right to sue but they need to be really clear about what they're doing.
[25:01.440 --> 25:01.880]  Right?
[25:01.960 --> 25:19.800]  And so we can take that and we can extend that concept to machines and we have to ask ourselves does this machine have more agency, more creativity than the monkey did when the monkey picked up the camera and hit the shutter?
[25:20.120 --> 25:30.960]  Um, and that exact question is being raised by Dr. Thaler because he filed a copyright application and it was rejected on this very principle.
[25:31.120 --> 25:36.580]  The copyright office will only grant copyrights that have been authored by human beings.
[25:38.400 --> 25:45.760]  But, a deeper question to ask is a question about originality because the copyright law requires that the work be original.
[25:46.100 --> 25:48.820]  And that's more than just novelty.
[25:49.500 --> 25:49.980]  Right?
[25:50.060 --> 25:54.480]  The courts have found that originality requires a modicum of creativity.
[25:54.840 --> 25:55.360]  Right?
[25:55.520 --> 26:03.560]  So you can't just collect data and put it in a spreadsheet and say, look, I authored this, I get a copyright over this.
[26:03.600 --> 26:20.880]  It came up in phone books and in databases where people tried to say, I own this data that I worked hard to collect and collate, but because it lacks the creativity necessary, it cannot be protected.
[26:21.060 --> 26:33.480]  And so we have to ask ourselves not just is an AI allowed to be an author, but does the AI have the required creativity in order to qualify?
[26:33.660 --> 26:33.980]  Right?
[26:34.140 --> 26:40.580]  And again, when we look at the product generated by DALI, right?
[26:40.680 --> 26:43.960]  For example, I forget what the prompt was.
[26:44.200 --> 26:45.160]  This was a...
[26:45.160 --> 26:47.740]  The prompt here was a robot that failed at painting.
[26:48.200 --> 26:48.680]  Right?
[26:49.520 --> 26:52.740]  That's a good picture of a robot who's failing at painting.
[26:53.080 --> 26:53.460]  Right?
[26:53.640 --> 26:58.380]  That shows if I had made that, that would be pretty creative I think, but I didn't.
[26:58.480 --> 27:00.140]  I told a machine to do it for me.
[27:00.560 --> 27:01.020]  Right?
[27:01.200 --> 27:06.020]  And so the machine, going back to what we were talking about before, is just a function.
[27:06.240 --> 27:12.020]  It's just taking an input and spitting out an output according to its deterministic processes.
[27:13.280 --> 27:22.720]  And the fact that you can, you know, pour some randomness into that machine, I'm not sure that that necessarily qualifies it to be considered creative.
[27:22.720 --> 27:22.820]  creative.
[27:25.410 --> 27:37.680]  One more thing to consider is the use of tools because copyright law lets you get a copyright even if you use sophisticated tools to generate your output.
[27:38.080 --> 27:38.560]  Right?
[27:38.680 --> 27:39.920]  And so this comes up in photography.
[27:41.520 --> 27:46.820]  The photographer did not create the flower that they are photographing.
[27:47.060 --> 27:47.480]  Right?
[27:47.700 --> 27:48.240]  They didn't.
[27:48.360 --> 27:49.500]  They looked at it.
[27:49.560 --> 27:50.520]  They said, hey, that looks cool.
[27:50.520 --> 27:52.440]  they took out their camera and they snapped a picture.
[27:52.680 --> 27:55.860]  And the machine created the picture.
[27:56.840 --> 28:02.140]  But that's okay because the human being did have creative input.
[28:02.300 --> 28:03.300]  They picked the lighting.
[28:03.500 --> 28:04.600]  They picked the framing.
[28:04.800 --> 28:05.820]  They picked the angle.
[28:06.060 --> 28:09.440]  They had that modicum of creativity that was required.
[28:10.700 --> 28:22.840]  But something to keep in mind is that even though we can use sophisticated tools, the bare idea itself is not copyrightable.
[28:23.000 --> 28:25.340]  And that's explicit in the copyright statute.
[28:25.800 --> 28:26.360]  Right?
[28:26.560 --> 28:28.940]  So if I have an idea, no matter how I...
[28:28.940 --> 28:34.700]  you know, whatever form I put that bare idea in, it itself is not copyrightable.
[28:34.780 --> 28:38.260]  It's the expression of that idea that's copyrightable.
[28:38.540 --> 28:38.980]  Right?
[28:39.120 --> 28:47.980]  So if I have a story in my mind, right, and it's like man fights walrus and loses, right, that's just an idea.
[28:48.680 --> 28:55.040]  If I write a story, that specific expression of the idea is copyrightable.
[28:55.320 --> 29:04.840]  But if somebody else writes their own man fights walrus story, I can't stop them from doing that because all they've taken is that idea.
[29:06.720 --> 29:16.020]  And I want to use this to draw an analogy that I think is a very powerful one in understanding the role that AI plays here.
[29:16.320 --> 29:16.920]  Okay?
[29:17.220 --> 29:23.200]  Because it's easy to imagine replacing that AI with a human being.
[29:23.540 --> 29:23.700]  Right?
[29:23.960 --> 29:28.540]  Take Dali out of the equation and put an actual artist in its place.
[29:29.080 --> 29:29.680]  Okay?
[29:30.400 --> 29:36.600]  And I go to that artist and I want to commission an oil painting of a man fighting a walrus.
[29:37.060 --> 29:43.400]  And my artist friend goes out and does it and I pay them and I get my painting but they own the copyright.
[29:43.800 --> 29:46.220]  All I supplied was the idea.
[29:46.640 --> 29:49.160]  The artist created the art.
[29:49.520 --> 29:59.920]  In the context of Dali, that's all my involvement is, is giving the prompt, giving the bare idea, and the machine is creating the art.
[30:00.800 --> 30:15.400]  And so we have to ask whether the machine on its own can own that copyright because my involvement as the human being in this case is not that of a human being using a tool to create art.
[30:15.740 --> 30:22.200]  My involvement is no greater than commissioning an actual artist to create the art.
[30:23.700 --> 30:26.420]  So let's talk about the current litigation in the field.
[30:26.840 --> 30:34.000]  This is the on the screen we have the subject matter of Dr. Thaler's copyright lawsuit.
[30:34.180 --> 30:42.640]  He named his model the creativity machine which the copyright office denied registration to and this is all recent.
[30:42.800 --> 30:49.080]  This is in the past couple of months and Thaler is now appealing that decision to the federal district court.
[30:49.260 --> 30:55.280]  And so if you want to learn more about the copyright issues the documents are actually pretty good reading.
[30:55.480 --> 31:18.800]  You can find the refusal from the copyright office that lays out in great detail the case law and all of the reasoning behind why they take the position that a machine cannot be an author and you can find in Dr. Thaler's appeal brief all of his reasons for why he thinks it should be.
[31:20.060 --> 31:23.200]  Highly recommend it if you're a giant nerd like me.
[31:24.260 --> 31:24.360]  Okay.
[31:24.700 --> 31:27.120]  So now let's turn our view to patent law.
[31:29.640 --> 31:38.220]  In patent law we're talking about now inventorship which is a distinct concept from authorship because an inventor is not required to be creative.
[31:38.560 --> 31:39.140]  Right?
[31:39.240 --> 31:45.580]  There's no modicum of creativity and all that's required is that this creation be new and useful and non-obvious.
[31:47.370 --> 31:50.880]  And so the patent law actually does define what an inventor is.
[31:51.120 --> 31:58.060]  It's the individual who invented or discovered the subject matter of the invention which doesn't help us in our inquiry here.
[31:59.140 --> 32:04.140]  But we can maybe shed some light onto that by looking at the policy behind it.
[32:04.400 --> 32:04.480]  Right?
[32:04.900 --> 32:13.800]  The Constitution's stated purpose for the patent law is to promote the useful arts and promote science and the useful arts.
[32:13.980 --> 32:22.220]  And so we have to ask does protecting a machine as an inventor do that?
[32:22.640 --> 32:27.620]  I think and you know again we're in such uncharted territory here.
[32:27.680 --> 32:47.480]  I think the answer is yes from a policy consideration because ultimately if we want to promote innovation and our tool for promoting innovation is handing out patents well I guess it's okay you know it doesn't make that much of a difference whether we give you know name the inventor as the person running the system or the system itself.
[32:49.080 --> 33:20.460]  But a more challenging piece of the puzzle is that the courts going past the statute the courts have said that an inventor is required to contribute to the conception of the invention and conception is a weighty word right that involves mental involvement in my understanding it involves not an understanding of novelty because I can see how a machine learning system can actually do that right it
[33:20.460 --> 33:31.960]  can say based on its access to the entire bulk of patents You know, issued patents that it has access to, it can say whether it's seen something like this before or not.
[33:32.420 --> 33:38.460]  And it can, you know, that's something that a machine learning model is pretty good at, is detecting something that it hasn't been trained on.
[33:39.140 --> 33:44.860]  And, you know, generating stuff that kind of follows a model but is not, you know, exactly the same.
[33:46.360 --> 33:52.540]  But in terms of conception, I see that as more recognizing the utility of something, right?
[33:52.540 --> 33:59.980]  Not just saying this is what the thing is, it's new, but saying this is what this is useful for, right?
[34:00.120 --> 34:08.180]  Having a person sit down and sift through the output of the infinite monkey to say, Hey, monkey, you've really got something here.
[34:08.440 --> 34:10.900]  That's a human conception, right?
[34:11.040 --> 34:15.080]  We're very good at that, you know, picking signal out of the noise.
[34:15.400 --> 34:20.280]  And I'm not sure an AI would have the ability to recognize usefulness in the same way.
[34:20.400 --> 34:21.360]  And maybe they would, right?
[34:21.360 --> 34:27.820]  Maybe you can put in enough usefulness data to have it have a usefulness classifier.
[34:28.120 --> 34:33.080]  As it stands right now, I'm not convinced that the machine is capable of conception.
[34:36.440 --> 34:43.380]  So Dr. Thaler is also deeply involved in the patent law wild frontier here.
[34:44.060 --> 34:45.540]  He's got two applications.
[34:45.540 --> 34:50.540]  I showed you one of them earlier that he is suing on behalf of his computer for.
[34:51.980 --> 34:57.500]  They were, again, rejected on the grounds that a computer cannot be an inventor.
[34:57.660 --> 35:01.200]  And this one has progressed a lot farther than the copyright one has.
[35:01.200 --> 35:04.920]  This one has actually gone through the first level.
[35:05.040 --> 35:06.560]  It's gone through the patent office.
[35:06.840 --> 35:09.120]  It's gone through the first level of appeal.
[35:09.320 --> 35:11.820]  And it's gone up to the federal circuit.
[35:12.000 --> 35:18.340]  And if you don't know anything about, you know, how patent law gets worked on, that's one step shy of the Supreme Court.
[35:18.340 --> 35:23.000]  It's kind of like the patent court, where patents, like most patent law gets made.
[35:23.560 --> 35:28.180]  And they've just recently had oral arguments, which are, again, a fun lesson if you're into that.
[35:28.460 --> 35:30.460]  But we have not seen a decision on it yet.
[35:30.680 --> 35:40.840]  If I had to predict, I would say that Dr. Thaler will probably not be successful on the grounds of conception, if nothing else.
[35:42.660 --> 35:46.500]  So, Dr. Thaler is making this a worldwide project.
[35:46.780 --> 35:52.880]  And he's actually got a patent in South Africa, and has been rejected in many, many other places.
[35:52.900 --> 35:58.900]  He actually was initially successful in Australia, and then an appeals court knocked him back down.
[36:04.290 --> 36:06.750]  So, let's take a moment and think here.
[36:07.530 --> 36:10.370]  What is the purpose of this?
[36:10.370 --> 36:21.130]  Nobody is actually trying to claim, except for Dr. Thaler, nobody's really trying to claim that the machine is an inventor, or the machine is an author.
[36:21.410 --> 36:24.730]  Because they would much rather claim that they themselves were, right?
[36:24.890 --> 36:31.030]  I would rather be the inventor than be the assignee by virtue of owning the inventor.
[36:32.550 --> 36:38.830]  But I suspect that Dr. Thaler is a true believer, and is doing this to cause a ruckus more than anything else.
[36:41.030 --> 36:53.370]  In many cases, and particularly in the patent world, you know, I think that's a perfectly reasonable approach to machine-generated output, is just give it to the human, right?
[36:55.450 --> 37:05.070]  In the, you know, following my model of conception there, the human is doing a good chunk of this work by saying, hey, this is something, right?
[37:05.070 --> 37:11.110]  I expect that Dr. Thaler did not just take the first output of the machine and file it as a patent.
[37:11.130 --> 37:16.750]  I expect he sat down and flipped through some garbage and picked out something that looked okay, right?
[37:16.750 --> 37:17.970]  He was involved.
[37:18.230 --> 37:20.410]  I am... I haven't met the guy, you know?
[37:20.710 --> 37:21.450]  I haven't...
[37:22.150 --> 37:26.810]  Maybe he was... maybe he's very convincing, and he could convince me that his system works.
[37:27.050 --> 37:32.730]  But as far as I can tell, based on my knowledge of the art, there was a human being involved in the conception process.
[37:32.850 --> 37:34.690]  And so, could be a valid inventor.
[37:35.070 --> 37:37.170]  Now, I want to distinguish that from the copyright world.
[37:37.170 --> 37:48.790]  I have... my personal belief has settled on the conclusion that the output of Dali is not copyrightable.
[37:49.210 --> 38:05.300]  In the same, you know, following the analogy I gave you before of me hiring, commissioning a work of art from a human being, the human didn't provide the art.
[38:05.300 --> 38:08.860]  The human didn't do the act that is copyrightable.
[38:09.080 --> 38:10.680]  They did not create the expression.
[38:11.200 --> 38:19.180]  And the machine, which did create it, is not... doesn't qualify as an author.
[38:19.380 --> 38:24.360]  And so, there's nobody in that chain who could own the copyright.
[38:24.640 --> 38:34.860]  And so, it's my belief that Dali and, you know, Lambda's story and what have you, although they were prompted by a human being, that human being would not own the output.
[38:34.860 --> 38:39.700]  And I also think that the Dali people, OpenAI, do not own the output.
[38:39.920 --> 38:40.620]  They're trying to.
[38:41.000 --> 38:49.100]  The terms of use for Dali says that you, as a user, have to do whatever you can to help them own that output.
[38:50.520 --> 38:52.960]  But I just don't think it... I don't think it works, right?
[38:53.060 --> 38:55.180]  The dots aren't connecting there.
[38:59.800 --> 39:01.660]  And so, that's where we stand right now.
[39:05.000 --> 39:08.820]  The courts are actively involved on all of these fronts.
[39:09.020 --> 39:17.560]  And there are so many other legal issues that we didn't even touch on that could, you know, shake the whole foundation of it, right?
[39:17.560 --> 39:23.400]  The question of whether you can fairly use copyrighted work for training material.
[39:23.540 --> 39:28.200]  If that's a no, suddenly all of our training corpuses disappear, right?
[39:28.780 --> 39:30.160]  Because... well, most of them.
[39:30.240 --> 39:34.060]  There are a lot of things that were, you know, designed for the purpose, right?
[39:34.820 --> 39:36.460]  For limited purposes.
[39:37.060 --> 39:50.360]  And some of them that are just like, hey guys, we have a bunch of pictures of street scenes that are labeled with cars and people and what have you that are good for, you know, like limited... limited functions.
[39:50.660 --> 39:59.500]  But we wouldn't have access to the huge swath of creative data that's out there in terms of literature that's still under copyright.
[40:00.240 --> 40:01.820]  Photos that are still under copyright.
[40:02.200 --> 40:04.240]  Computer code that's still under copyright.
[40:04.660 --> 40:09.480]  You know, you would need to get permission from each one of those authors in order to make use of it.
[40:13.320 --> 40:18.540]  And so, at the end of the day also, I want to bring you back to the infinite monkey, right?
[40:18.680 --> 40:24.140]  Because we're going to have to contend with the fact that this is a machine.
[40:24.520 --> 40:27.680]  This is a machine that has no agency.
[40:28.380 --> 40:30.080]  It has no free will.
[40:30.080 --> 40:33.380]  It has no awareness of what it's doing, despite what it might claim.
[40:35.220 --> 40:45.320]  And we have to contend with the fact that when you have a system like that, it's no different from a machine, you know, a monkey that's bashing on a computer and occasionally generates a useful output.
[40:45.320 --> 41:01.360]  And on that note, I want to give you a little snippet of the oral arguments presented by Dr. Thaler's attorney to the Federal Circuit in his patent case, right?
[41:01.360 --> 41:02.720]  This was at the very end.
[41:04.840 --> 41:07.040]  The... Dr. Thaler's attorney went first.
[41:07.620 --> 41:09.600]  The patent office went second.
[41:09.840 --> 41:12.040]  And Dr. Thaler had, I think, three minutes to rebut.
[41:12.320 --> 41:14.340]  And this is just at the very tail end.
[41:14.660 --> 41:18.440]  The first voice you will hear is one of the judges on the panel.
[41:18.440 --> 41:22.040]  And the second voice you will hear is Dr. Thaler's attorney.
[41:23.360 --> 41:24.240]  That's an entity.
[41:24.480 --> 41:34.440]  By the way, do you happen to know, do the, you know, the mythic monkeys who type out Shakespeare get to be copyright holders?
[41:36.340 --> 41:39.240]  Well, I see I'm past my time, so if I can...
[41:42.310 --> 41:43.860]  And that's my time as well.
[41:44.110 --> 41:46.210]  At least for talking purposes.
[41:46.520 --> 41:48.040]  We have some time for questions.
[41:48.300 --> 41:51.960]  I think I see somebody standing there waiting at the ready.
[41:53.000 --> 41:55.750]  And if you really want to talk about sentience at this point, we can.
[41:56.420 --> 42:02.400]  And if you really, really want to talk about it, you know, you can find me in the hallway and we can fight.
[42:02.710 --> 42:03.000]  Yes.
[42:03.280 --> 42:03.670]  Hello.
[42:04.420 --> 42:05.500]  I have two questions.
[42:05.500 --> 42:05.840]  Yes.
[42:06.150 --> 42:08.900]  The first is, how can Thaler even sue at all?
[42:09.110 --> 42:13.050]  If the device doesn't have sentience, how can he even have standing?
[42:13.380 --> 42:15.050]  So in other words, how can he have...
[42:15.050 --> 42:19.670]  Can't they just throw it out right in the beginning and say, you don't have standing, the software would have to sue.
[42:20.000 --> 42:24.020]  And then that's the first hurdle that the software would have to have sentience.
[42:24.590 --> 42:29.360]  Presumably, let's just call it software, would have to have the sentience before it could even have standing to sue for itself.
[42:29.670 --> 42:29.980]  Right.
[42:30.090 --> 42:32.460]  So that goes back to the monkey case, right?
[42:32.460 --> 42:37.070]  And I think that the distinction... I haven't looked too deeply into that particular question.
[42:37.520 --> 42:43.380]  My gut is that the distinction lies in the fact that Thaler actually owns the system.
[42:44.000 --> 42:45.780]  And the monkey was just a wild monkey.
[42:46.190 --> 42:57.040]  And so there was nobody who could sue on behalf of this monkey because it was just a monkey in Africa that happens to pick up this guy's camera.
[43:00.650 --> 43:04.540]  The computer, on the other hand, is very much Dr. Thaler's property.
[43:04.820 --> 43:06.780]  And so, I don't know.
[43:07.380 --> 43:10.400]  You have a valid question, is what I'm trying to say.
[43:10.520 --> 43:14.840]  And I think that that's the answer to it, but I haven't looked too deeply in the briefs on that.
[43:15.210 --> 43:15.570]  Okay.
[43:15.840 --> 43:18.280]  And then the second part, and this relates...
[43:18.280 --> 43:20.440]  I mean, it can even relate to, let's say, Oracle v. Google as well, right?
[43:21.750 --> 43:31.130]  The idea of, you know, if we look at software and we look at the corpus of the data that goes into the... then parts of it go into the end solution, right?
[43:31.480 --> 43:41.340]  So, with all of the pharma companies now using machine learning to work and develop new drugs, in theory, does that mean that whatever drugs they develop are going to go straight to generic?
[43:41.440 --> 43:44.420]  Because they won't be able to patent rate or copyright it.
[43:45.230 --> 43:48.860]  So, when we're talking about drugs, we're generally talking about patent law.
[43:48.860 --> 43:53.000]  So, a lot of the questions that come up in copyright law don't really apply.
[43:53.600 --> 44:02.920]  And I think that you can solve that problem in the patent world by having that human being who's sitting down and saying, you know, I think this is useful.
[44:03.180 --> 44:05.710]  I think this is the drug.
[44:06.000 --> 44:07.580]  This is the genetic sequence.
[44:07.780 --> 44:12.920]  This is the, you know, whatever, that we're going to use and it's going to do great things.
[44:12.920 --> 44:16.300]  Because they have recognized the potential utility of that drug.
[44:17.020 --> 44:19.230]  So, I think that's where the answer is.
[44:19.340 --> 44:21.640]  It won't really upend things at all in the patent world.
[44:22.100 --> 44:26.580]  But if recognition is the standard, and I'm glad you kind of came back to that, right?
[44:26.710 --> 44:39.640]  If I see someone doing something, right, like an invention or something like that, but let's say they don't go to patent it or whatever, I just see them doing it, and then I write it up and then submit it, right?
[44:39.940 --> 44:44.960]  Wouldn't that be the same thing as watching your computer do it, recognizing it, and then submitting it?
[44:44.960 --> 44:46.840]  Isn't there an analogy there?
[44:47.230 --> 44:49.580]  Well, except you, like, they would be the inventor, right?
[44:49.680 --> 44:50.440]  They did it first.
[44:50.920 --> 44:53.340]  Well, right, but the computer did it first.
[44:54.600 --> 44:55.780]  I did not control that.
[44:58.960 --> 45:00.000]  All right, yeah, sorry.
[45:02.180 --> 45:03.250]  Hello, thank you for your talk.
[45:04.330 --> 45:18.380]  In the publishing world, I'm not a lawyer, but if the publishing company takes a collection of essays and compiles it as a compilation, I think that's copyrightable.
[45:18.380 --> 45:18.780]  Yeah.
[45:19.100 --> 45:40.860]  So if a human is sitting at Dolly or Dolly Mini makes a prompt and then curates the dozens of garbage and says, I think this one is actually worthwhile, do you think that act is copyrightable?
[45:40.860 --> 45:44.750]  So there's texture to that question, right?
[45:44.860 --> 45:55.250]  You've got the question of, does the act of selecting one of these outputs make that output copyrightable?
[45:55.340 --> 45:56.730]  And I don't think it does.
[45:56.730 --> 46:03.960]  But potentially a collection of those outputs arranged in some creative way could be copyrightable.
[46:04.060 --> 46:11.060]  Because now you've inserted a human being's creativity into the mix.
[46:11.640 --> 46:21.880]  There's, down in the vendor area, we have with us an artist who uses AI models as a starting point.
[46:21.880 --> 46:27.420]  And then, his name is Jonathan, I believe, and he's got some great work and you should go check it out.
[46:27.700 --> 46:41.840]  But the fact that he takes that AI, that initial AI bit, and provides further processing on top of it, makes it, in my view, definitely copyrightable because he's inserted his creativity there.
[46:41.940 --> 46:43.300]  I hope that answers your question.
[46:43.500 --> 46:44.160]  Thank you very much.
[46:46.210 --> 46:47.620]  So this one's a bit wonky.
[46:47.620 --> 46:53.820]  But within certain bounded areas, you know, AI has the ability to actually cover all possible outputs.
[46:54.040 --> 46:54.140]  Yeah.
[46:54.380 --> 47:01.200]  And so two artists actually covered every possible sequence of the 8-note, 12-beat melody using computer code.
[47:01.400 --> 47:01.420]  Okay.
[47:01.420 --> 47:04.580]  And they tried to copyright all 68 billion outputs.
[47:04.900 --> 47:06.960]  And I just want to know what your thoughts were on that.
[47:06.960 --> 47:18.160]  Because when we use computational power, the man fighting a walrus, you know, if you turn that over to an AI and run it long enough, you're going to cover every possible scenario there, right?
[47:18.400 --> 47:18.730]  Yeah, sure.
[47:19.060 --> 47:21.120]  Yeah, well, it's the infinite monkey again, right?
[47:21.120 --> 47:25.100]  Yeah, and this was an infinite monkey on 12, whatever it was, I'm not an artist.
[47:25.230 --> 47:25.400]  Right.
[47:25.680 --> 47:30.140]  You don't need advanced machine learning technologies to do that.
[47:30.210 --> 47:33.300]  You can just brute force it and go through every possible output.
[47:33.540 --> 47:38.800]  I would say they probably lack the creativity, the modicum of creativity.
[47:38.800 --> 47:40.540]  They've just collected a bunch of data there.
[47:40.710 --> 47:41.210]  Got it.
[47:41.320 --> 47:41.400]  Right?
[47:42.040 --> 47:46.780]  The human act of composition is not about generating sequences of notes.
[47:47.080 --> 47:53.500]  It's about that kind of intentional creative act, which, of course, is very fuzzy.
[47:53.920 --> 47:57.860]  I do think it's bold, and I respect the gumption.
[47:58.520 --> 48:04.180]  I think that was their point, is to say, you know, if you cover every possible output, then, you know...
[48:04.180 --> 48:08.480]  I think their point was to have fun at the expense of the federal government, which sounds great.
[48:08.480 --> 48:09.750]  Thank you very much.
[48:11.140 --> 48:19.680]  In the 1950s, there was the blacklist, and a lot of very talented writers and authors and screenwriters were blocked from using their talents.
[48:19.880 --> 48:20.160]  Yes.
[48:20.160 --> 48:26.280]  So they continued to do so anyway and wrote some wonderful things, only they had intermediaries that took credit for the works.
[48:26.750 --> 48:32.420]  Later on, decades later, many of them were given credit officially and given, you know, awards for their work.
[48:32.420 --> 48:48.750]  But do you think that if the courts were to rule that the computers could gain or could not gain this kind of protection, that in order to, you know, use the inventions, there would be people on an exclusive basis, there would be people who would act as intermediaries and say,
[48:48.860 --> 48:49.620]  yeah, I did that.
[48:49.880 --> 48:53.460]  And that might happen for a couple of decades until the law catches up.
[48:53.460 --> 48:57.580]  I think that's a great question, and I think that you're right that that's exactly what would happen.
[48:57.780 --> 49:07.880]  If the law says your machine cannot own the copyright, there's really nothing stopping a person from just pretending that they did it.
[49:08.020 --> 49:14.400]  Until, you know, eventually you have to litigate that and we start asking the question, right?
[49:14.400 --> 49:35.700]  Because that's what would end up happening is if, you know, and this is a series of hypotheticals now, if the courts rest on the idea that a machine, a pure machine output cannot be copyrighted, and a human being steps in and claims falsely that they,
[49:35.980 --> 49:51.800]  that their creativity was a part of it, then you would have to, like, when, when we start litigating those lawsuits, we would, as a matter of course, have to start investigating that and be like, all right, great, well, give us some other examples of stuff you've done,
[49:51.900 --> 49:55.280]  tell us about your process, and the jury would decide whether they're credible.
[49:55.560 --> 50:08.680]  I think there's a better solution, which is to take that machine output and put a thin layer of human creativity on top, and that will turn it into something copyrightable, even if the original output was not.
[50:10.400 --> 50:12.580]  Actually, I want to follow up on that exact same thing.
[50:12.750 --> 50:21.520]  What is the standard that exists for taking something that's not copyrightable, how much creativity has to be put into that to make it copyrightable?
[50:22.300 --> 50:22.940]  A modicum.
[50:22.940 --> 50:30.250]  All the courts will tell you is a modicum of creativity, and it's intended to be a very low bar, but at least a non-zero bar, right?
[50:30.250 --> 50:31.480]  So you have to do something.
[50:33.210 --> 50:38.640]  Something more than just, you know, collecting the data and representing it in some natural way.
[50:38.660 --> 50:41.250]  You have to exert some thought.
[50:41.480 --> 50:48.120]  So like one example that I always thought was kind of cool, and this might be apocryphal, I don't know, but map companies run into this problem.
[50:48.360 --> 50:53.480]  Because map companies are taking existing data and representing it in a graphical form.
[50:53.480 --> 51:09.000]  And so I've, you know, I've heard the story that map companies, in order to establish the copyrightability of their work, will add cities that do not exist to maps to add that element of human creativity and make their maps copyrightable.
[51:09.540 --> 51:09.860]  Yes?
[51:10.840 --> 51:12.580]  Back to Copilot.
[51:13.420 --> 51:18.380]  I'm not sure if you're familiar with the Software Freedom Conservancy's mailing list.
[51:19.220 --> 51:22.820]  There's a group of people talking about the Copilot issue.
[51:22.820 --> 51:23.200]  Yes.
[51:23.460 --> 51:25.800]  Meeting and publishing notes and discussing.
[51:26.340 --> 51:28.840]  So there's the...
[51:29.660 --> 51:38.710]  I guess the current strategy is we're going to make our own machine learning model with blackjack and copyleft.
[51:40.460 --> 51:45.180]  And I don't know if you've been following this discussion, but I wanted to know if...
[51:45.180 --> 51:46.140]  What are your thoughts?
[51:47.060 --> 52:01.100]  Can a model have a license and control the output of... dictate the license of an output of a model?
[52:01.320 --> 52:02.000]  Can the model...
[52:02.000 --> 52:06.800]  So you're saying, can the owner of the model dictate how the output of the model is used?
[52:06.980 --> 52:07.320]  Correct.
[52:07.320 --> 52:09.960]  I mean, yeah.
[52:10.340 --> 52:12.960]  It's an end user license agreement, right?
[52:13.020 --> 52:21.340]  You're using their software and they can, in exchange for letting you use it, put some terms on how you use it, right?
[52:21.340 --> 52:29.920]  So, yes, your alternative is always to just not use their model, right?
[52:30.100 --> 52:31.730]  If you don't want to be bound by their terms.
[52:31.940 --> 52:37.400]  That's a little bit different from saying that, you know, they own something that's not ownable, right?
[52:37.400 --> 52:49.340]  And so, for example, Dolly is trying to do exactly that, where their terms of use say, we own the output of our model.
[52:50.080 --> 52:53.020]  If we don't, you have to... if you own it, you have to give it to us.
[52:53.120 --> 52:54.660]  If we own it, that's great.
[52:54.880 --> 52:57.520]  You're allowed to use it for whatever purpose, but we own it.
[52:57.520 --> 53:07.400]  And so, they're trying, but they recognize, they clearly understand that the boundaries are fuzzy right now for what ownership means here.
[53:10.180 --> 53:10.940]  Thank you.
[53:14.840 --> 53:17.820]  So, just for timing purposes, I see we're at 10.52.
[53:18.480 --> 53:19.040]  Am I good?
[53:19.320 --> 53:20.520]  Okay, go for it.
[53:21.480 --> 53:31.340]  If I make a picture book, a photo book, out of the output of Dolly creations, Can I publish it?
[53:31.440 --> 53:32.600]  Can I get paid for it?
[53:33.000 --> 53:39.800]  So, you're saying you take a bunch of raw Dolly output, put one on every page of a book, call it a collection...
[53:43.540 --> 53:46.420]  I add a theme to it and make it meaningful.
[53:46.420 --> 53:47.240]  Yeah, you've curated it.
[53:47.400 --> 53:47.560]  Yes.
[53:47.980 --> 53:49.440]  I don't know the answer to that.
[53:50.060 --> 53:54.380]  You know, at heart, I'm a patent attorney, and so there's a lot of copyright law that I just don't know.
[53:55.440 --> 54:03.820]  I don't know about the separate copyright ability of a collection of uncopyrightable things.
[54:04.220 --> 54:07.260]  All right, I don't know what level of creativity would be required.
[54:07.420 --> 54:18.060]  So basically, the question comes down to whether the judiciousness, right, the selection process of the human qualifies as being a creative process.
[54:18.060 --> 54:20.380]  And that's a question I don't have an answer to.
[54:22.100 --> 54:27.140]  All right, and I think we have no more questions.
[54:27.440 --> 54:29.920]  Oh, and I'm out of time anyway, so we are all done.
[54:30.160 --> 54:30.960]  Thank you very much.
[54:43.780 --> 54:53.600]  The next talk in ten minutes, or eight minutes, will be using cybersecurity automation to organize your cyber threat intelligence knowledge by Andrew Koo.
[54:53.760 --> 54:55.100]  So please come back for that talk.
[54:55.400 --> 54:59.200]  And a quick note, remember, the closing ceremony tonight is at 6 p.m.
[54:59.360 --> 55:02.220]  in this room, and we follow by a samba band.
[55:02.440 --> 55:05.860]  Hang out and enjoy, and then help to clean up and pack up if you can.
[55:05.860 --> 55:11.080]  We also need volunteers to help load our rental trucks Monday morning starting around 9 a.m.
[55:11.200 --> 55:12.960]  in the D'Angelo Building loading dock.
[55:13.180 --> 55:15.220]  So if you can, please come and help pack out.
[55:15.380 --> 55:15.660]  Thank you.