AI Fire Hose - #1: Igniting the Future with Ash Bennington & Mikhail Voloshin

0:00If you've been considering futures trading, now might be the time to take a closer look. The futures market has seen increased activity recently, and Plus 500 Futures offers a straightforward entry point. The platform provides access to major instruments, including the S &P 500, NASDAQ, Bitcoin, natural gas, and other key markets across equity indices, energy, metals, forex, and crypto. Their interface is designed for accessibility. You can monitor and execute trades from your phone with a$100 minimum deposit. Once your account is open, potential trades can be executed in two clicks. For those who prefer to practice first, Plus500 offers an unlimited demo account with full charting and analytical tools.

0:45No risk involves while you familiarize yourself with the platform. The company has been operating in the trading space for over 20 years. Download the Plus500 app. Trading in futures involves the risk of loss. It is not suitable for everyone. Not all applicants will qualify. Hey, everyone. If you like this podcast, go behind the paywall to get privileged access to the smartest minds in finance. Visit realvision.com slash RVpod and use the promo code podcast10. That's podcast10 to get 10 % off our essential membership for the first year. Join the Real Vision community and learn how to become a better investor.

1:24And now to the top analysis of today's markets. What's going on, guys? Welcome to the first episode of Real Vision's weekly AI firehose, where we drink from the pounding stream of everything happening this week in the AI space. I'm Ash Bennington, joined by Mikhail Voloshan. Mikhail, always a pleasure to be with you, dude. It's always a blast. You know, the field of artificial intelligence just moves so darn fast. You and I talk about this stuff all the time. So let's share some of what we chat about with the general world and see what their reactions are, you know? Yeah, obviously, this is the first time we're doing this show natively for YouTube and other platforms.

2:04So we should talk a little bit, I think, up front about who we are, what we do, and how we came to this. I'm Ash Bennington. I'm a host at Real Vision. Mikhail, talk about your background, because it's just so impressive in terms of this field and the relevancy to all the conversations we're about to have. No, shucks. Well, I've been in the field of artificial intelligence for basically as long as I can remember, ever since I watched Short Circuit and Terminator as a kid, two very different perspectives on the future of AI. In college, I studied neuroscience. My graduate work was on aquatic invertebrates and their very, very simple nervous systems.

2:48I left that program to go work at Microsoft Speech Recognition Labs. uh in the time since then i've worked at military contractors on communication devices uh for encrypted protocols i've uh i've worked at google uh for their on their double click team and i apparently have a very noisy little co-star right here that wants my attention that sounds like a cat it is a cat she'll be hopping up here on and off during the course of the presentation um she uh she definitely wants to get in on this action um the nowadays I run a consulting company called Mighty Data Inc. that deals with machine learning and artificial intelligence related problems in mostly in the financial and bioscience spaces.

3:36And I'm also a published writer for, you know, I can't not make a plug for my novel Dopamine over here. It's about the Russian mafia, computer hacking and Dungeons and Dragons. I hope you get it and like it and read it and think it's awesome. All right, Mikhail. Yeah, let's talk a little bit about what we're doing on the show. You know, this came out of a series of conversations that we had on Real Vision, just an incredible response from the Real Vision audience. You and I have been having these conversations back and forth. And we just wanted to bring it to our viewers and our listeners, because I think there's just so much that's happening in the space right now.

4:14It's really hard to get your head around if you're a layperson, if you don't have a background in this stuff, to understand the hype from what's real, to understand where we are in terms of the actual implementation cycle of some of these technologies, and really just to get a sense of all the things that are happening quite literally around the world in the AI space right now. That's how I think about what we're doing here. What were some of your goals for having this conversation on a regular weekly basis here on Real Vision? Well, look, when we're dealing with an exploding field like artificial intelligence, it's really easy to get lost in the headlines.

4:47And so you'll read a headline that suggests that the world is about to change tomorrow. And it turns out that if you read the article, the nitty gritty actually isn't all that exciting. Or if you or the vice, the opposite often happens to where like some little passing detail by some quote in some press release turns out to have major, major impact down the road. And so as a layperson, it's really hard to be able to tell one from the other. So I'm hoping to be able to curate some of the content and to be able to sort of sift the wheat from the chaff and see what actually might pan out versus what's just a flash in the pan.

5:24Hey, talking of which, I know we've got a couple of stories that we've both been excited to talk about, about human computer interface, brain interfaces. I mean, this is some really interesting, sort of very high tech, very cutting-age stuff. This is some really sci-fi, cyberpunk material right here, and the stories that I've chosen here are real and are indicative of progress that might seem like a total breakthrough if you're not otherwise familiar with the research that has been going into it. The first of them is a company called Unbabel. They've produced a device that allows for you to type at a pretty fast rate without actually moving any muscles or with moving muscles in a very, very microscopic, like imperceptible level.

6:18This isn't quite like reading your mind, but they find that it might be really, really helpful for sufferers of Lou Gehrig's disease or ALS. And it does speak of ways to interact with the computer that are just now breaking out. So let's try to explain what this is. It's the idea of EMG. This is electromyography. This is that nerve, not rather nerve impulses, but muscle impulses. Talk a little bit about how this works, what some of the constraints, limitations, and benefits of the technology are so let's uh let's start by talking about what it's not uh there's eegs which are electroencephalographs which are these big helmets that you sometimes see with like lots of you know electrodes all over them and those read people's uh brain waves but the data that you get out of them is very very fuzzy and they're not directly subject to conscious control uh by the operator so uh they're not really particularly amenable to operating a machine uh the other thing that they're not is uh what's called chronic implantation we'll talk about that in a minute but uh that's when you actually slice open the skull and insert an electro a microchip with a grid of electrodes on them um what this is is something very different uh it's non-invasive um it is a sleeve what that what non-invasive means for those without medical backgrounds it It means you don't have to physically, surgically implant anything.

7:47You can literally just walk up to the device and use it. Exactly. You don't cut anything open. It doesn't go inside your body. It just goes on your skin. Now, what it's doing is it's an armband that's listening on the neuromuscular junction between the muscles of your arm and hand and your peripheral nervous system. So when you move your arm, your brain sends signals down your spinal cord, which go to your arm nerves. In the words of invader Zim, humans don't have an arm control nerve.

8:23And then that electrically stimulates your muscles electrochemically, and that's why your muscles move. In sufferers of ALS, the nerves are just fine, but the muscles themselves don't respond to the stimulus, so nothing actually moves. However, you can use these electrodes to listen for these motions. Now, what this company Unbabel has built is a device that listens for how the patient would have moved their arms had their muscles actually been working. They then feed this to an artificial intelligence that is integrated with a large language model. And LLM is the same kind of AI that's being used in GPT.

9:11GPT is a form of LLM, basically. So what it's doing is something very similar to autocomplete or... You know the software on your phone where you swipe your finger around, and based on your swiping motion, it figures out what you were intending to type? even if you only approximately swiped near the keys that you had intended to hit. So it's basically like a predictive algorithm that looks at the keystroke input and then puts together the probabilistic determination of what the next word should be. Exactly. And because it's hooked to an LLM, the probabilistic determination is much, much better than any previous development.

9:50It's a lot better at knowing what word the person is going to type next. Now, the patient does need to be trained to use this. The AI and the patient sort of need to work together to learn how to interpret the patient's intentions. So there's going to be a learning curve on this, but I guess the upside is that obviously this isn't something that's going to be reading your thoughts from across the room. You have to very much cooperate, indeed train the language model to use your particular movements in order to correctly interpret what those words should be. Exactly. But, you know, after all is said and done, you know, not only is this going to be useful for sufferers of ALS, which is just an absolutely horrific disease, and like this could be, this could allow them to talk in a much more fluid, natural manner.

10:46Just to put things in perspective, Stephen Hawking, probably the single most famous uh sufferer of als even more famous than uh lou gehrig whom the disease was for a while named after uh he used an eom sorry an eog device which i think stands for electro optical graph um and he could type it about uh two words per minute that means that means it measures the the movement of his eyes and interprets the movement of his eyes in order to form a series of characters and ultimately words that allowed him to speak in a synthesized voice that we've all heard on this exactly um this technology came from uh the fact that sufferers of ALS the uh the eyes are usually the the last voluntary muscle to go um and they are um like they were originally this research came from just looking at a grid of letters and they could see which letter the person was staring at so Hawking could type at about two words per minute um Halo the device being released by this company unbabble allows people after being trained to type at about 20 words per minute so 10 times stephen hawking's uh rate and they the company unbabble is looking to get a consumer grade rate of about 60 words per minute uh and like once they hit once they hit 60 they'll release commercially and then this will become viable for the consumer Republic.

12:15And that's a huge, that's a huge improvement. It's a huge number, a huge improvement and a dramatic difference. One of these things where a difference in quantity really changes the difference in quality. This is not the only news of the week where your dual background in AI and neuroscience is helpful. I understand there's another story about BCI brain computer interface. This one's really, really cool. This comes out of UC San Francisco. Um, and it features a woman who got paralyzed by a stroke, I think something like 30 years ago. And a researcher at a research team at UCSF implanted a grid of electrodes on a chip into the motor cortex of her brain, which was designed to measure the motions that she would send to her larynx and tongue and lips in order to enunciate words.

13:07We're going to take a quick break and be right back with more of the Today's top analysis on the Real Vision daily briefing. Have you ever wanted to trade Bitcoin but haven't dared try? With Plus500 Futures, you can trade crypto without the hassle of opening a wallet. With just a few clicks, you can register and start practicing with their free and unlimited demo. See a trading opportunity? You'll be able to trade it in just two clicks. Feel ready? You can move to real money with as little as$100 once your account is approved. And the great thing is that in addition to crypto, Plus500 gives you access to a wide range of instruments.

13:41S &P 500, NASDAQ, gas, and much more. Explore equity indices, energy, metals, Forex, and beyond. With a simple and intuitive platform, you could trade anytime, anywhere. Experience the fast, accessible futures trading you've been waiting for with Plus500. With over 20 years of experience, Plus500 is your gateway to the markets. Visit us.plus500.com to learn more. Trading in futures involves the risk of loss and is not suitable for everyone. Not all applicants will qualify. Plus 500. It's trading with a plus.

14:17Here, let me just read this. This is directly from Neuroscience News. This is the clip that we're talking about. Chang, that's the lead neurosurgeon on this team, implanted a paper-thin rectangle of 253 electrodes onto the surface of the woman's brain over areas his team had discovered are critical for speech. The electrodes intercepted the brain signals that if not for the stroke would have gone to muscles in her tongue, jaw, and larynx, as well as her face. A cable plugged into a port fixed to her head connected the electrodes to a bank of computers. This really does sound more like the sci-fi aspect of literally having a brain implant that you can then use to actually create synthesized speech.

14:59This is, you know, again, it's just almost a level of sophistication higher uh, EEG versus EMG. I mean, this is the idea of a port with a cable running into your head is, you know, something that's, that we've seen from the matrix and from, uh, you know, ghost in the shell. And so this is what we refer to as chronic implantation. Uh, that chip is in there. It's, it's not coming out, you know? So the, um, what this is doing is again, the, uh, working on the electrical signals coming out of the brain, the difference between this technology and the armband one is primarily that it's measuring not the neuromuscular junction, but actually the originating signals in the brain.

15:44Now, in this particular case, the AI was monitoring her brain activity and correlating it to phonemes that would have corresponded to what she had spoken had the muscular activity actually been able to activate. And over several weeks, she was able to work with a team of AI specialists that were able to train this algorithm to recognize, yes, she would have said an E at this point, she would have said a P at this point and so on. There are 39 phonemes that can be used to encode any word in the English language. And after this several week training period, she was actually able to she was actually able to communicate in more or less real time.

16:29Not only that, but also make facial expressions on an avatar in a virtual reality environment that could move its face in a way that her natural face no longer could. What an incredible story and such an amazing opportunity for people who have suffered from these chronic diseases, from illness, from injury. I mean, just an incredible moment to be alive as we talk about all of this new technology and its ability to benefit humankind. Okay. And now for something completely different, let's shift gears here a little bit and talk about money, specifically OpenAI losing money hand over fist. This is a story that both of us were looking at this week.

17:07What's your take on this, Mikhail? Oh, man. You know, OpenAI is a private company. They don't have openly disclosed financials. But word on the street is that they lost over half a billion dollars in 2022. too. How it's possible to lose money when you're as huge and as popular as OpenAI might be something that might not be entirely intuitive to folks. Well, Amazon lost a lot of money. You just keep growing the company and you try and grab market share while not caring at all about your margins. Let me just read this. This is a quote from Business Today. This is the actual numbers here. OpenAI, the AI research company behind the popular language models like ChatGP2 and DAL-E2, has reportedly doubled its losses to$540 million in 2022 due to soaring development expenses for its chatbot.

17:59The company is now looking to raise as much as$100 billion in the coming years to fund its goal of developing artificial general intelligence, that's AGI, and advanced AI capable enough to improve its own capabilities. By the way, Mikhail, before AI starts improving its own capabilities, this is a great time for you as an AI engineer with these rising costs, I guess, for those input costs are coming from human labor, I would imagine, to sell your services. My consulting business has been doing well. Let's just leave it at that. You know, I will say that, like, so a lot of people might not understand the financials behind what it takes to operate these models.

18:42So what happens is when I write software that for a client or for my own needs that uses GPT, GPT is a cloud service. and every single word that I send to GPT and every word that it generates back to me costs me money. It's something like one and a half cents for every 10 ,000 words. It's not a lot, but if you're processing lots and lots of documents, like, for example, let's say if you're trying to give it the entire collected works of Shakespeare in order to have it try to create a new play that Shakespeare would have written. You know, that adds up. You know, it can be quite a lot. So that's how OpenAI makes their money, is that people literally pay them for that.

19:37But OpenAI has some substantial costs. First of all, OpenAI has hired some really expensive engineers, and the folks that develop these models are themselves very well reimbursed. How much have these rates increased since chat GPT has become so much a part of popular culture? Are people earning double? Yeah, I mean, is it like... Double or triple is not an unreasonable estimate. There's a caveat to that, which is that a lot of folks who were in data analytics and statistics and other fields that are tangentially related to machine learning have been rebranding or retooling their skill sets in order to specialize on machine learning and natural language processing because of the increased demand.

20:31Yeah, and you would too. Yeah. So there's been an increase in supply. And in addition, there's been an increase in demand. So the field is on fire right now. So even with supply rising, you're still seeing rates that are double or triple. Exactly, exactly. So you've got, so OpenAI's hardware costs are substantial. Obviously, the equipment that runs these large language models is ridiculously expensive and getting more so because of the demand for it. By the way, talking of which, I want to talk about NVIDIA's blowout earnings. Yesterday, this is a huge story in the space, and it's a huge story more generally in finance.

21:17I want to just read this to you from Bloomberg. NVIDIA said sales will be about$16 billion in the three months ending October. If the gain holds, it will mark a record. Analysts had estimated just$12.5 billion, still a huge number, but a massive increase. According to data compiled by Bloomberg, NVIDIA's results last quarter blew past projections as well, and it approved an additional$25 billion in stock buybacks. The stock's been on fire. It was up this morning, no surprise. You know, a lot of people might not know this, but the math that runs a lot of these large language models, So these LLMs are, they adhere to an AI data structure called a neural network, which is basically a set of, let me describe it this way.

22:07It's a set of equations that are loosely based on a very rough understanding of how neurons in the human brain interact with one another. So I don't want to overstate what a neural network is. It's very, it's only barely related to actual biological neurons. in concept only. So why do they call it that? What's the reasoning behind it? Why is it called the neural network? It's actually one of the oldest AI models that's ever been built. And it was originally put together by a psychologist who was studying some of the quote-unquote wetware or hardware of the human brain and thought, hey, I wonder if you got a machine to do this.

22:53It was a man named Frank Rosenblatt, who I believe built his first prototype and i want to say 53 something like that 1953 and uh this is a machine it was an analog computer that would uh that had a set of dials representing the strength of connection weights um you know xx uh synaptic weights quote unquote uh and the machine actually had motors that would reach back onto its own dials and turn its oh you know it would adjust and recalibrate its own settings so some really like diesel punk era sci-fi back at the time now he called it these things started to be called neural networks for the simple reason uh that he was inspired by uh neural by the principles of neural connectivity and the way that actual neurons uh encode data and record memories um but since since that time we've learned a lot more about the actual biology of neurons.

23:51And more importantly, the equations that drive neural networks have sort of done their own thing and have gotten refined and optimized in their own directions. So they've kind of diverged, but we still call them neural networks due to that sort of underpinning of co-evolution. We're going to take another quick break and be right back with more of the day's top analysis on the Real Vision Daily Briefing.

24:18such cool stuff mikhail talking about the interface between human beings and machines there's some interesting questions that are coming up right now about copyright law about your ability to protect intellectual property that's been generated from artificial intelligence this is really fascinating to me because it touches on a whole lot of areas of intersection that i'm interested in the arts technology uh drama film i mean it's just a really interesting open space. Talk a little bit about what's happening there. So there's been a major development in court cases where it's been ruled that AI-generated art is not copyrightable, which means that if you produce a show or even a character completely by machine, then you cannot copyright that material.

25:06And what that means is that you can't make money on it. The real-world impact of this is that it's a major deterrent for Hollywood. This means that they can't copyright AI-generated movies, which means that they're not going to make them, which is probably good news for a lot of writers out there. Well, here's the really interesting thing, and I'm fascinated by the Hollywood intersection here. Let me just read this from the Hollywood Reporter. In March, the Copyright Office, this is the U.S. Copyright Office, affirmed that most works generated by AI aren't copyrightable, but clarified that AI-assisted materials qualify for protection in certain instances.

25:46This is really interesting to me. An application for a work created with the help of AI can support a copyright claim if a human selected or arranged it in a, quote, sufficiently creative way that the resulting work constitutes an original work of authorship, it said. So this is what's really interesting to me. If you go and you go into ChatGPT or some other LLM, large language model, and you tell it to create a play, create a screenplay, write something, write a poem, and then you go in manually and you adjust it. Maybe you make some changes. You change the name of a character. You change a couple of circumstances.

26:25And then here's the weird thing that I've been thinking about. You feed it back into the AI to improve it again. You've kind of co-evolved that script, that poem, that novel, whatever it is that you're creating, along with the machine. And therefore, according to this interpretation from The Hollywood Reporter, at least quoting the Copyright Office, it sounds as though that may be something under U.S. law right now that is, in fact, copyrightable. I mean, this is a really weird zone we're entering. There's a lot of room for, there's a lot of gray space, let's put it that way. And, you know, they kind of had no choice other than to say that you have to allow AI-assisted work for one simple reason.

27:09Where exactly do you draw the line between where the AI helps and where it doesn't? Like, if you write a screenplay and it has a spell checker, and, like, that spell checker is technically a form of AI. Like, do you now lose your copyright because you, like, fixed separate and changed the E to an A? and now you've lost copyright on that work, that can't work. That can't happen. Grammarly would be another example that I run all my emails through before I send them out, make embarrassing grammar mistakes. That's AI. So it seems to me that it's going to have to be almost like a modicum or a scintilla of originality.

27:44And effectively what that means is that the big broad headline that works created entirely by AI are not copyrightable, in practice, winds up being almost completely nullified. It seems as though it's almost the opposite of what the headline would suggest. By the way, we should say that this is an article that comes from the Hollywood Reporter. That's like maybe reading about green energy in a newspaper called, you know, Coal Miner Today. You're probably going to see a bit of a bias toward creative professionals. By the way, I'm not unsympathetic to the case there, but it is just a certain perspective that the Hollywood Reporter would be likely to have.

28:20So it's worth keeping in mind that the law always lags behind technological capability for a large number of reasons. We see this in telephones today where it's illegal to have a bot that's automatically sending out text messages. But you can circumvent that by having a bot generate the message, give it to a human being to press the deliver button, and then the human being, like their physical finger, hits send. That's still fine. so uh basically as long as there's some meat bag that's still hitting the send button you're okay the um well we meat bags are very grateful well as now how that integrates with what if the meat bag is paralyzed uh and is using a chronically implanted electrode in order to hit send that becomes a whole nother matter yeah it's really interesting uh one other thing that i wanted to talk on today is this news coming out of facebook about a new AI assistant.

29:22I think it's called AI Llama that's going to help with code development. This is something that, again, is right in your wheelhouse because it intersects with your background as a coder and an AI. Talk a little bit about what you think about the potential for these tools. What's their opportunity set and what's the limitation look like today? Well, AI has been writing code for a surprisingly long time now. One of the earliest applications of GPT was in fact the fact that it cranks out code just as surely as it cranks out English. So GPT's background was to, it was originally, all of these LLMs were originally conceived to be translation software.

30:06They were originally built to translate from French to English, that kind of thing. And it was found almost by accident that they're really, really good at quote-unquote translating from a question to an answer. So what researchers found was that the task of converting a set of words to another set of words actually encapsulates this functionality. So what happened with code was when they trained GPT, they trained it on an extremely large corpus of text that they pulled from the internet. They trained it on something like 45 terabytes coming from social media like Facebook and Twitter, coming from Wikipedia, coming from all of the collected works of Project Gutenberg, I think, and so on.