Machine Learning Engineer reacts to State of AI Report 2019

Mornin'! So I just stumbled upon this, which is the State of AI Report 2019, and I figured what a great way to spend a Sunday morning— yeah, it's still morning— going through it and checking it out. So it's pretty recent. I don't know how I haven't found it before today because I kind of like to stay up to date on these things, but I'm gonna go through it and see what, see what kind of little golden treasures are within it. And I guess I can react to it from like a state of a machine learning engineer, right? So that's, that's my role, machine learning engineer, and I, but my focus is on machine learning and health. So we'll see if there's, if there's anything to do with that in here. All right, what exactly is it? So we believe that AI will be a force multiplier on technological progress in our increasingly digital, data-driven world. This is because everything around us today, ranging from culture to consumer products, is a product of intelligence. I agree with that. So I'm not going to read through all of it word by word. I'll let you do that if you want to just check it out straight away. There'll be a link in the description. But let's, let's pop open this SlideShare. You know what's crazy? LinkedIn have like a, like, I'm not sure. I don't know how like SlideShare has become kind of a thing on LinkedIn. I know it's been around for years, but it's like PowerPoint on the cloud. So here we go. State of AI Report 2019 on June 28th. It's about a month ago. It's the start of August at the moment. We've got some authors. Thank you very much to Nathan and Ian. We got some definitions. Artificial intelligence, a multidisciplinary field of science whose goal is to create intelligent machines. This report is going to be broken down into 5 main topics. So we've got research, Talent, Industry, China. So China, I guess China, that's a good thing. China can have its own topic because it's so large, right? 1 point something billion people. And then politics. Beautiful. And this is the second edition of it apparently. So we've got State of AI 2018, the inaugural State of AI 2018. Maybe this is going to be a yearly thing. Thank you. Thank you to all these people who've made it. Let's go through definitions. So I'm not going to read through all of these again, but we might as well do AI as a broad term, right? Because it's, it's, it's hard to talk about things when you don't have— it's much easier to have a conversation when you have an agreed upon definition of something. So let's just read it with the AI definition. A broad discipline with the goal of creating intelligent machines as opposed to the natural intelligence that is demonstrated by humans and animals. Yeah, we're pretty intelligent beings. It has become a somewhat catch-all term that nonetheless captures the long-term ambition of the field to build machines that emulate and then exceed the full range of human cognition. I, I'd agree with that too, right? And so essentially the goal of AI is to build a machine smarter than us so it can help us. Well, at least that's, that's what I think of it. And then you've got machine learning, reinforcement learning, deep learning. I'll let you check those ones out yourself. We got algorithm, an unambiguous specification of how to solve a particular problem. It's just a way of breaking down a larger problem into step-by-step, and usually you can do it with code, or actually just a written set of instructions like a recipe could be considered an algorithm. Different types of learning, unsupervised learning. This is good that they, they define it at the start because machine learning and artificial intelligence can be hard to get into without knowing what what exactly these— like, natural language processing can be a, can be a pretty scary term if you don't know what it actually is. Scorecard. What did they predict last year? 2018 predictions: breakthrough by a Chinese AI lab. I guess green means correct. Chinese labs win ActivityNet. I'm not actually sure what, what that even is. Train ImageNet in 4 minutes. That's pretty crazy. Now, I'm not sure what the exact ImageNet score was when it originally started. I'm, I'm guessing it's a lot more than that. But I mean, fastai, I think, did it in 18 minutes with, with cheap hardware or relatively cheap and accessible hardware. But to beat it by like 4, like a multitude of 4, that's pretty insane. DeepMind RL StarCraft 2 breakthrough. I wrote an article about this about like 2 years ago, I'm pretty sure. When I was so excited about DeepMind getting into reinforcement learning with StarCraft. And that was correct. AlphaStar beats one of the world's strongest StarCraft II players 5-0. A major research lab goes dark. Yeah, that was OpenAI's GP2 model, right? So they released, they released this big language model, or at least said how powerful it was, showed what it could do, but didn't release the model. So they kind of, they kind of went against the OpenAI name, right? So So it's like closed AI. There was a bit of a meme going around at the time. The era of deep learning continues. That's an orange light. You're right. How do you, how do you actually measure whether it's keeping going? I'd say it is. Drug discovered by ML produces positive clinical trial results. This is kind of thing I'm interested in, in this like machine learning and healthcare. These are all red. I won't worry too much about those. We only care about what's right. Nah, I'm joking. But yeah, this is pretty hard, right? So if you want to discover a drug, you got to go through a whole bunch of clinical trials. I'll let you look into those. Section 1, research and technical breakthroughs. Reinforcement learning conquers new territory. Montezuma's Revenge. OpenAI achieves superhuman performance at Montezuma's Revenge. Now I think I haven't had much experience in reinforcement learning. The only reinforcement learning problem I've ever worked on, we failed at it. It was trying to teach Sonic to play for an OpenAI competition, trying to see, teach Sonic, an agent to play Sonic, and it failed. So I'm kind of biased with reinforcement learning. I'm not very good at it, right? So when I hear these reinforcement learning things, is that, wow, that's, that's pretty cool. Montezuma's Revenge, I'm not actually sure what that game is. RL conquers new territory, StarCraft II. We saw that at the start, that was DeepMind beating 5-nil a world-class player, 5-nil. That's actually pretty insane, right? But I think there's a few constraints, right, that the DeepMind model was able to bypass. So, I'm not sure 100%, but I'll link some paper or something below that you can read a bit more about it. I think I read the blog post from DeepMind that their algorithm could see the whole game space of StarCraft II, whereas a player, if you've played it before, you can't really see the whole map because it's clouded out. So did it actually beat it? I'm not entirely sure. Yeah, there you go. Due to various constraints on the action space. More RL. RL conquers new territory. Krake 3 arena capture the flag. Human level of performance is achieved by having multiple agents independently learn and act to cooperate and compete with one another. That's pretty insane, right? So you're teaching two computer programs to play with each other and then cooperate to be better than a single algorithm. That's like, that's where it's getting really valuable rather than a single agent. And remember, agent is just like a computer program playing a game. That's how I understand it. And without explicitly being told how to play that game, it's just learning from the environment. And so being able to make one agent is pretty phenomenal. But being able to make multiple agents like multiple computer programs to just play together. I mean, if you remember back in the days, like when you play a video game and you'd have like computer programs playing against it, they were pretty bad. So what I'm imagining is that in the future, they're going to be pretty damn well good, if not way better than you. You got more RL. OpenAI Five improves even further. I believe this is with Dota, Dota 2. Yeah, Dota 2 playing bot. 94, 99.4% win rate over 7,000 online games with 15,000 live players. This is insane. So 2017, it beats in a 1v1 match. 2018, a 5v5 best of 3, it lost 2 games. So it lost. April 19, 2019, OpenA5 wins 2 back-to-back games versus the world champion Dota 2 team in a livestreamed event. That's actually insane, right? So look at that progress. 2017, winning in a simplified match. A single player bot wins in a simplified match. Then 2018, a team of bots, a team, right? So not just 1v1 anymore, a team lost 2 out of 3, not too bad. And then April 19th, so 6 months, right? OpenAI Five wins 2 back-to-back games versus the world champion Dota, 2 team. Lots of progress in RL. We got more, a bit more information about the OpenAI. So essentially what I've noticed in terms of these reinforcement learning algorithms getting better is the compute power. So compared to the August 2018 version, the April version, so the 2019 one used 8 times more compute and OpenAI OpenAI released a blog post recently on artificial intelligence and compute. Long story short, it requires a lot of compute to build any kind of— and what does it mean by compute? So as in computing power. So say, for example, the laptop I'm using has a computing power of 1. You might as well say to build this, this AI version, you might need a computing power of 10,000, right? Maybe even more. I don't know the exact math, but let's just say your common laptop is 1. To build an AI agent like this, you might need something that's 10,000 or 100,000. There's a good blog post by OpenAI that you can check out on, on AI and compute. What's next in RL? So there's a lot of RL, a lot of RL. Training a single robot using play to perform many complex tasks without having to relearn from scratch. You're right. So the goal. Yeah, of course, the goal is to train a single robot using it to perform many complex tasks without having to relearn from scratch. Beautiful. RL. So more RL stuff. A lot of it is at the moment, RL breakthroughs you saw have been like in games and that's kind of like a controlled environment. But if you look at like this sort of stuff, like real, real world stuff, like trying to manipulate something with a hand, That actually turns out to be really hard. So there's a lot of progress, a lot of progress to be made there. It's actually pretty exciting, right? Imagine if you couldn't use your hand, but then you could teach a robotic arm using your own brain how to operate. That would be fun. We've got more RL curiosity-driven exploration. How can agents learn to solve tasks when their reward is either sparse or nonexistent? Yeah, that's a, that's a tough one, right? It's the exploration versus exploitation problem, which is a massive one in reinforcement learning, which means should you dedicate effort to exploring new, new ways of doing things or should you exploit the things that already work? So this is like a paradigm in reinforcement learning, right? It's like, which one should you do? Do more? Hard problem to solve. And I'd argue it basically probably isn't, isn't solvable because there's always going to be a trade-off of some sort. What's next in RL? More RL stuff. Reinforcement learning. Moving research into production environments. This is— yeah. So I, out of my experience of being a machine learning engineer, this is something I haven't actually seen, right, is using reinforcement learning at scale in terms of like in a production environment. But I think— what's this saying? Facebook releases Horizon, the first open-source end-to-end platform that uses applied RL to optimize systems in large-scale production environments such as Messenger. Well, face— looks like Facebook is using RL in production. Let's, uh, let's check this out. Facebook Horizon. Wow, okay, this is This is more than a year old. Horizon, the first open source reinforcement learning platform for large-scale products and services. An end-to-end platform built on PyTorch One that is designed to jumpstart RL's transition from research. This is pretty cool, right? Like Facebook, I love the fact that a lot of the stuff, a lot of the world-class stuff that we, you see in this field gets open sourced. Now it's one thing to open source something, it's another thing to be able to use this. So whether, a small startup has the compute capabilities of doing, of actually taking advantage of this, is another story from just having access to it. But the fact that it's there means that the opportunity is also there. And go back. Lots of RL stuff, that's what I'm noticing. So that's pretty cool. That's moving RL into production is, is going to be very exciting rather than just being in like a gamed framework, a game closed environment framework. ML for life science. So this is what I'm really excited about, right, is using artificial intelligence in the health field, particularly because my formal degree— so I'm self-taught machine learning online, right? And through a bunch of different books and whatnot. And I'm still constantly learning, but my formal education is in nutrition and food science. So I'd like to be able to use machine learning or artificial intelligence to figure out what should I eat? That is my goal. But this is, this is a step forward in terms of using machine learning in biological fields. Two deep CNNs, convolutional neural networks, work in concert, so work together, to significantly outperform prior state of the art far earlier than expected. So, AlphaFold predicts de novo. I'm not entirely sure. I should know what de novo means, but I think it's, it's a stage of protein structure. So, when a protein begins to form, it goes through different stages. There's like tertiary and something else. I don't know them all. And 3D structure of folded proteins. Could that be valuable? I think so, right? Because if you can predict the size of protein, then you could probably predict where it could bind to, what it could be used for. Let's go back. Where are we at? So, two neural networks working in tandem. This is actually getting really exciting, right? Interrupt this broadcast, take a sip of coffee. Because if two neural networks can work together, it's like many heads are better than one, right? So when you make decisions in a committee— my brother was actually watching a video that individuals trying to predict something is usually wildly wrong. But then when you average the decisions of, say, 100 people, in this case, maybe two neural networks, you get better overall decisions. Does it always work? Probably not. But This is really cool to start seeing neural networks working together, like kind of like GANs. We'll probably see GANs in here in a second. A big year in NLP. Totally agree. 20,000, 2,000, 20,000. 2019 was kind of the year of NLP when you have things like BERT. Oh well, BERT came out in 2018. The transformer-like architecture, Allen Institute's ELMo, OpenAI's transformer. FastAI, Ruder and Howard's UMLFit. That's actually really a really good model. We actually use that in production in my machine learning role for a natural language processing task, transfer learning in NLP, essentially. So yeah, I think this is what it's saying. Transfer learning from pre-trained language models ushered the ImageNet moment for NLP. Now, the ImageNet moment, what was that? Well, I think it was like at 2013 or something like that where ImageNet, where it got surpassed like in terms of human, human level performance versus machine learning performance. Machines outperformed humans for ImageNet. And I think that's starting to happen for NLP. And that's due to transfer learning, right? Which is you take a whole bunch of— in NLP, you take a whole bunch of text like Wikipedia, all of Wikipedia, not just Wikipedia articles, all of it. So billions and billions of words. You have a computer algorithm go through all the words and you figure out the relationships between different words. So, like, for example, I took my dog to the vet. You figure out which word is more likely to come next in a sentence. That's what a language model does. So, in that sentence, I took my dog to a vet, the word vet is more likely to appear than the word car, for example. So, you just do that at a very large scale. You use that information, fine-tune it on your own niche. So maybe you want to automate law, law cases. And so you have all this language, you have law cases, you build a classifier on top and you make your predictions. That was transfer learning in NLP summed up in a nutshell. But Fast.ai have a great blog post on it. But yeah, NLP, lots of progress. Yeah. Preach. So here we go. Pre-trained language models. This is what OpenAI's GPT-2 model does. Remember, that was the one they didn't really release fully. So that's kind of like OpenAI's, closed AI's GPT-2 model. So you would put in a sentence and then you can try this online. Actually, you put in a sentence. In this case, Miley Cyrus was caught shoplifting from Abercrombie Fitch on Hollywood Boulevard. And then the model, because it's pre-trained on a whole bunch of text, like unstructured text online, it's just read basically the entire internet. It kind of can figure out what words should come next in this sentence. So here's what it tried to complete it with. The 19-year-old singer was caught on camera being escorted out of the store by security guards. The singer was wearing a black hoodie with the label Blurred Lines on the front and Fashion Police on the back. Scroll down for video. So that's pretty crazy, right? Oh yeah. And then you can try it out. You can't see this. How do I hide? How do I hide that? Yeah. Try it out here. Hugging Face. transformer.huggingface.co. Hugging Face are amazing. Shout out to Hugging Face. If you want anything to do with NLP, you should be checking out Hugging Face's GitHub. A big year in natural language processing. Researchers start sniffing glue. I love these, these titles. These are great. Whoever made this. The two people that made this. I've already forgotten, but I'll thank them at the end again. And start sniffing GLUE. So GLUE is like a benchmark, like it's— I don't know what it exactly stands for, but essentially it's a collection of natural language processing tasks, right? So things like maybe— and I might be wrong— like, here we go, spanning logic, common sense understanding, and lexical semantics. Long story short, there's this benchmark for researchers of how they can evaluate their NLP models because it's good to have like a benchmark and human baseline is 87, right? So if you or I did the test, it'd be good if we got 87. But 20th of June, 2019, it got surpassed. And which, which model did that? I don't think it says. A new benchmark, Superglue, has been introduced. Good name. More NLP. Facebook show multilingual data to make translation. So this is great, right? Facebook, Facebook have a lot of data, right? And they need to translate a lot of stuff. Federated learning and growing interest. Yeah. So this is really valuable as well because federated learning is basically like spreading the ability for a machine learning model to work rather than just on a centralized computer, like on a big server, like hosted by Google or Amazon or something like that. Federated learning is like distributing it across, say you had 100,000, no, a million, billion, a billion smartphones or something like that. Your machine learning model could learn little bits and pieces on each phone and then combine it. So, that's what federated learning is. Still, still very hard to do, as you could imagine, but Google is making process, progress with it. And I guess, yeah, they would be very interested in that because they have a whole bunch of Android phones out there with lots of keyboards. Developers can express a new data type. We'll start going through it maybe a little bit faster. Increasing emphasis on data privacy and protecting deployed systems, ML systems, from attacks. Yeah. So adversaries can manipulate data collection. Yep. Corrupt model and tamper with its outputs. Yeah. So this is a really tough one, right? So you think about the amount of data that is flowing through the internet right now. And pairing that up with cybersecurity, like, is— so can cybersecurity match the speed at which data is growing and being used? I hope so. It kind of has to, right? Otherwise it'll be a big, big, big problem. And so what adversarial is, is kind of like, say your machine learning model is predicted like you're a self-driving car. Can you fool the self-driving car vision with some sort of image so that the car does something unexpected? Right. And I think there was an example of using something like on the Tesla steering wheel, you could put an orange on the steering wheel and the car would think that you are actually holding it. That's just one example, but a better one would be something that would fool the actual camera, not just the steering wheel. Deep learning in medicine. Yes. This is what I'm very excited about, right? Is actually I've been reading this book. It's called Deep Medicine. Can you see that? Yep. Great book. How artificial intelligence can make healthcare human again. So long story short, lots of, lots of great like proof of concepts and breakthroughs, but it's really hard to put this stuff into an actual clinical environment. So diagnosing eye disease, this is really, really cool. So because eyes like you can, kind of take a photo of them. This uses a 3D U-Net. Oh, that's just the architecture. Okay. This map paints the eye structure according to expert— I'm not going to pretend I can pronounce that word— expert performance on referral decisions. It can be easily adapted to various imaging machines by only retraining the segmentation network. So this is really exciting stuff. Detecting and classifying cardiac— another hard word to pronounce— arrhythmia using ECGs. Cardiologist-level performance is demonstrated using end-to-end deep learning trained on 54K patients. That is insane. This is only going to get better, right? Like, as— and it's not one thing for the machine to just do way better than a cardiologist, right? I think the real value is going to be in the synergy. So the machine helping the cardiologist to do their job and vice versa, right? The cardiologist going, hey, from my 20 years of experience in this, I think the machine may be thinking about something different, or the machine vice versa, right? The machine going, hey, I think you're missing this. Long story short, there's a lot more data out there on ECGs than a single person can evaluate. Machines can help. The bigger the dataset, the better the model. Well, that's usually true. 600, above 600K chest X-rays have been published to boost model performance, but dataset issues remain. Yeah, see, this is always a problem, right? You'll have a lot of, a lot of open source data, but the problem again is, is like the compatibility issues. So maybe one hospital takes X-rays in a certain way and another one does it in a certain way. And then that's just only X-rays. And then you've got other things like tracking patient records, like their blood vitals, their other metrics and stuff like that. So there's a lot of data out there, but it's really hard. A lot of your time and effort will be used to sort of put that together. But once it's in a standardized format, machine learning actually works incredibly well because if there's patterns there, it's going to find it. Deep learning in medicine. Neural networks decode your thoughts from brainwave. This is Neuralink 101. Phenomenal. Really interesting to me as well because my father has Alzheimer's disease and dementia. So if neural networks and brain-computer interfaces could be used to improve his quality of life, that would be phenomenal. Neural networks can restore limb control for the disabled. How exciting is this? Far out. Machines learn to synthesize Chemical molecules using neural networks with Monte Carlo tree search to solve retrosynthesis by training on 12.4 million reactions. Wow. So when I did chemistry in, in university, I failed my first year of chemistry. So maybe I can use machine learning to help me get better at chemistry. AutoML, evolutionary algorithms for neural network architectures and hyperparameters. I think this is going to be a massive area of explosion in the, in the next couple of years is AutoML. Because you think about it, a lot of deep learning techniques require like hands-on focusing and adjustment. So like tuning hyperparameters, tuning layers, etc., etc. But it makes sense to use compute, more compute power to tune those things for you because at the end of the day, you can, your brain can only think of a certain number of different combinations. So I think what this is saying is that Google's AutoML, prior AutoML work optimized hyperparameters on neural network architecture individually using reinforcement learning. And again, remember, reinforcement learning is very computationally expensive. RL systems require a user to define an appropriate search space beforehand. So you need to define what the algorithm is going to search for rather than just letting it search. And yeah, see, 100,000 GPU hours. So that is just insane. 100,000 hours on GPU at $2 an hour, which is about that standard price or maybe a bit more expensive than usual. That's $200,000. An alternative learning evolutionary AI framework, LEAF, is used— is to use evolutionary algorithms to conduct both hyperparameter and network optimization, ultimately leading yielding smaller and more effective networks. That is actually really exciting, and it's outperformed Google's AutoML. Okay, I think this is, this is my prediction for the next 2 years. AutoML and things like LEAF. Let's look up LEAF. Um, what's it called? Learning Evolutionary AI Framework. So my prediction for the next couple of years, let's say 2 to 4, 2 to 3 years is think services like AutoML and LEAF and take over. Well, building, let's just say building neural architectures from scratch, like handcrafting gets less and less important as these things take over. So your ability to build a neural network from scratch, okay, it's important maybe to understand how it's done, but your you will likely hand over that capacity to something like this. And look how recent this is, April 9th, 2019. So like 3, 4, 4 months ago. I'm gonna read this paper. Not now, but later. So I'll bookmark that for later. I'll leave a link so you can read about it too, 'cause LEAF sounds exciting. We're only 34 out of 136 slides through. Maybe we'll skip a few. AutoML, Designing Research Constrained Networks with real device performance feedback. The pace of CNN-based automated architecture search is accelerating. Facebook ups Antiverse Google. And I'm always excited about seeing these things, right? Big tech companies taking on each other. And here I am, the small little guy sitting in my room. These guys have way more compute power and money and knowledge than I do. Oh, this scary thing. So Google demonstrated a multi-objective RL-based approach, MNASNet, Neural Architecture Search. I believe that's what it's called. To find high-accuracy convolutional neural network models with low real-world inference latency as measured on the Google Pixel platform. So 74%, 76 milliseconds accuracy. So that's really important, right? Is, okay, you can have these massive neural networks, but how do they perform on a smartphone? Because ultimately that's where people are going to start to use them. Not everyone has a warehouse full of GPUs to access on demand. Yet. Um, but Facebook— so they got Google at 76 milliseconds, 1.5 times faster than MobileNet. Facebook comes out and achieves 23.1 milliseconds latency and 420 times smaller search costs. Wow. Okay, so this is where I see a really interesting battle. Right. So Google is fundamentally, in my eyes, a hardware company, right? So you look at their revenues and a lot of it comes from ads. But that ad revenue and all the, all the stuff that Google do with AI wouldn't exist without the hardware that they operate on. Google Cloud wouldn't exist. They need— so their long-term game is building better hardware to improve the software that they offer. Whereas Facebook is not a hardware company. Right. Facebook is a, is a software company, is a technology company. So because they don't have access to the hardware, right. So Google make the hardware like TPUs, GPUs, cloud storage, etc. So they can use their own stuff to train incredibly long, incredibly big models because after all, it's their own hardware. But Facebook likely has to rent this stuff off Google and Amazon. Fair, fair being it might be a reduced cost because they're buying it so much at scale. I don't actually know where Facebook get their hardware. I'm pretty sure I heard that they could not buy enough GPUs from NVIDIA because they couldn't make them. So what Facebook have to do is figure out, and because they operate at such a big scale, they need to figure out smarter ways of doing things. So this is really exciting, right? You'll, first of all, you'll see Google come out with something that they've built this new state-of-the-art model and it's trained a ridiculously long time on Google's own proprietary hardware, but then Facebook will come out a couple of months, a couple weeks later with like the same thing that Google just did, but done much cheaper, cheaper. Not always the case. This is just a little trend I've noticed. Done much cheaper because Facebook don't have the proprietary hardware. Interesting little battle going on there. But once again, these two companies have billions of dollars and the best engineers in the world. And so it just means that people like you and me have to work harder. State-of-the-art in GANs. GANs continue to improve. Yeah, GANs looked terrible in 2014, but very exciting. Now they're just looking insane. Like, that is— that to me is a real picture. You've seen those photos of people's faces. There we go. Look at that. Now they're starting to generate whole people. This is really funny. So, me and a friend, Michael, shout out to my boy, we used to work together. He— we had this idea, right, just taking all of Joe Rogan's YouTube videos and podcasts and then creating, generating just a completely new Joe Rogan podcast on whatever we liked. Or we would go on the Joe Rogan Show, but we'd just be talking to an AI version of Joe Rogan. And it sounds like people have done just that. And Joe actually retweeted on this. So I just listened to an AI-generated audio recording of me talking about chimp hockey teams, and it's terrifyingly accurate. It's entirely possible. But yeah, I think this is going to be big over the next few years, is like generating fake videos and fake, fake text and all that sort of stuff as GANs get better and better. Learning 3D objects from a single image. That's actually insane. 16,600 AI papers over the 25 years shows immense growth in publication output with machine learning. Again, machine learning and re— RL. So RL is everywhere, right? It's getting popular and popular. 10 times more papers. And so this is another thing to remember, right? Look how many papers are coming out. Thousands and thousands, probably hundreds daily, right? So you can never keep up with, with the, with the stuff. With everything. So if you're going to learn machine learning, my advice is to, to be patient, learn what you can, focus on the fundamentals, code and math and probability statistics, but don't be too stressed about trying to keep up with every little single thing. Some beautiful people will make a report like this and kind of give you like an overview of it at the end of the year. Section 2, talent. Wow. We're only up to Section 2 and we've been going for way too long. Let's speed this up. Google is the dominant force. We all kind of knew that. Google, MIT, Stanford. Yeah, these all make sense. Big dog universities and big dog tech company. 80%, 88% of 4,000 researchers. Most of researchers are men at NeurIPS. The balance maybe is kind of improving. Not entirely sure. Compensation of senior engineers at large tech companies approaching $1 million. Yep, that's, that's a lot of cash. But remember, you got to adjust this for cost of living, right? Most large tech corporations are based in big cities where it costs a lot to live. So maybe you get paid a million, but you have to pay a lot in rent and just living expenses. At the other end of the spectrum, there's a huge growth in $1.47 an hour data labeling jobs. Wow. Yeah, that's capitalism in a nutshell, right? You'll have senior software engineers earning $1 million plus, and then you'll pay $1.47 an hour to data labeling jobs. So that's kind of the way you build an AI company is you have a lot of really smart engineers in one country, and then you outsource your data labeling service to another country, pay ridiculously cheap prices like $1.47 an hour. For people to label your data. You really well-paid engineers use that labeled data to create an AI algorithm, and you then sell the AI algorithm as an API or something like that. I believe that's what Rev.com is based on. Smart, taking advantage, I'm not sure. I don't know the whole economics of, is that actually a lot of money in China? I'm not entirely sure. Pioneers of neural networks, so they should. These are our idols. Thank you for all the work that you do. Got Yann LeCun, uh, Geoffrey Hinton, and Yoshua Bengio, boys, along with Turing. So Turing, he's not here, but rest in peace, Alan Turing. Europe publishes the most AI papers, but China's average is growing. Yeah, China is kind of exploding. I guess that's why they have a, their own section on this report. New $1 billion investment in computing at AI— computing and AI at MIT, shifting gears for new generations. Yeah, I think this is really important, right? So yeah, this guy who donated $350 million cited his concern was that the US was lagging behind China. So that's, that's the thing, right? At the end of the day, progress in anything, you need money, right? You can have all the excitement in the world about about AI and its potential. But excitement doesn't pay for things. You can have, have all the drive in the world, but you still need gas, still need cash for these things. University course enrollment in AI is growing. That's amazing, particularly in China. I probably won't go to university for AI at the moment. I need to— I'm more interested in applied versus research. I'm happy to let smarter people research things for me. We go here. Gender diversity of AI professors and students is still far off being on equal footing. Yep. Again, this is— I totally agree. Women in tech is a great thing, but of course the balance is still not 50/50. Male-dominated AI applicant pool in the US. Talent survey. More talent survey. Trend of hiring AI researchers into companies slowing down. Not entirely sure. Industry. Beautiful. Global venture capital investments in AI themes grow at a clip to reach above $27 billion per year. That's actually kind of insane. Big tech companies continue to gobble up AI-first startups. Yeah. So this is, this is like kind of like the bigger get bigger, right? Look at these companies: Amazon, Facebook, Microsoft, Google, Apple. If you build a great AI startup, right, and you're doing some great work, chances are if it's really good, you'll probably get offered to be bought by these big companies so that they don't have to do the hard work that you've done. They can just give you the cash to keep building upon it and then they own all the IP. Companies we featured in last year's State of AI. Look at all this funding. This is actually insane. A billion dollars, basically. Billion, $500 million, $200 million, $500 million, $2.3 billion IPO. I haven't even heard of Neuro. What's Neuro? Neuro. Neuro.ai, delivering the future of local commerce autonomously. Okay, that's pretty cool. So you could order something on Amazon. I wouldn't be surprised if Amazon bought this in the next couple of years. Buy something on Amazon and this bad boy delivers, delivers it to you within half an hour. Let's check their Instagram. They're a very Instagram-friendly company. Like, look at this thing. That thing is just— yeah, who wouldn't want that delivering their groceries to them? I do. So yeah, they got— that's probably why they got close to a billion dollars. RPA is increasing. Cleaning robots are taking over. 2009, this Boston Dynamics robot could barely walk. Now it's doing parkour. Robotics are getting insane. ABB, not sure what ABB— oh, that's a company investing $150 million to build the world's most advanced automated flexible robotics factory in Shanghai. That's crazy. Bright Machines. Look at these assembly lines, right? They're just— I can't even imagine. I've never seen one in person, but you've seen those videos that are like slowed down. They're just Insane. US factories are installing record numbers of robots. This makes sense. Amazon's probably leading this. There we go. Amazon is massively scaling its physical infrastructure. Yep, makes sense. Amazon are just blowing up everywhere. They simulate robots on the cloud. And this carton wrap, bad boy, you throw, throw packages in, cartons come out. 700 boxes per hour, 5 times human throughput. That's insane. Self-driving cars are now a game for multi-billion dollar balance sheets. Yeah, I watched this great talk by George Hotz the other day that self-driving cars are like largely a scam. It's kind of, kind of, you'll see these self-driving car companies, some of them will get like, like $450 million in funding and then all they're building is like a shuttle bus that drives around like a little community. Great talk, but again, self-driving cars, a lot of money. Healthcare. Woo-hoo! This is what I'm most interested in, most excited by. The US FDA cleared 3 AI-based diagnostic products in the last 12 months. IDX for diabetic retinopathy detection from eye scans. Imagen software to detect wrist fractures. MaxQ software to prioritize The clinical assessment of adult non-contrast head computed tomography, CT, cases. Pharma companies, of course, this is ML in drug discovery. Nutrition, boom! Using genetic metabolic, metagenetic and meal context information from 1,100 study participants to predict Individual's metabolic response to food. Oh my God, this is exciting to me. I want to answer the question, what should I eat with AI? Who's doing this? Zoe. Twins respond differently to a high-carb muffin to the same set meals. Yeah, so this is right now, current nutrition advice focuses too much on the food. Eat this, not that. When it should be focusing on the individual, right? Because twins respond differently to a high-carb muffin, right? So, this is Glucose, glucose continuous. So one twin had a high glucose response to a muffin and one twin had lower. I need to check out Zoe. Oh my God. This looks like exactly the thing I want to build. Fueled by science. They have a research study. Careers. Available positions. Production data science engineer, Python. Excuse me, I'll be back in a sec. I got to apply for this. Okay, we're back. That is amazing. That, that has literally given me butterflies to see, see this. That's what I want to work on. Patenting and AI patents increasing. Computer vision is the most popular area. Yeah, but we didn't see much computer vision in this, in this, uh, this report. It was a lot of, uh, RL and NLP and robotics, but not much computer vision. I mean, maybe that's been exhausted. I still think there's room to improve. Still very early days for monetizing hosted AI services. This is actually very interesting. So Amazon's overall compute storage and network software, so like using Amazon just to do your compute, is a lot of money. 25— is that $25 billion? Yeah, I think so. But their ML infrastructure and like their APIs, not that much, nowhere near as much as the overall service. So yeah, this is, this is interesting, right? So even if you do build an AI hosted service, will you get— this is Amazon, right? So the biggest, one of the biggest of big dogs in cloud computing, and even they're not— 11 million sales in 2018. Interesting. Lots of AI hardware. Yeah, this is, this is a big thing, right? Because at the end of the day, it comes down to do good AI work, to do good computing work. It comes down to the hardware. Like, it almost doesn't matter like how good you are at software. If you don't have the hardware to do what you want to do, you can't do it. So hence why cloud, cloud services are just growing insane because it kind of gives you and I and other companies access to the best hardware in the world. Of course, it's going to be Google going to use some of their own hardware and keep it proprietary and that sort of stuff. But to do good AI work, you need good hardware. So that's, that's going to be a growing industry for, for years to come. AI hardware. Yep, dedicated AI chips are going to, to explode. More AI hardware. I don't know enough about hardware and compute stuff. All I know is that the graphs always goes like this. And yeah, here we go. Um, smartphones. Yeah, so that's, that's a big thing, right, is how can you, how can you make the, the hardware on smartphones to utilize AI as well as like a dedicated GPU or something like that? Because after all, we all have smartphones, but not everyone has like a full-blown GPU. AI hardware, more hardware. Google and Nvidia throw their hats. To apply AI computation to the 40 trillion gigabytes of data generated from connected devices by 2025. Yep, so they're going to be making custom chips, of course, for AI. 5G is going to make AI insane. Yep, totally agree. Imagine, because 5G allows the data throughput, smartphone can compute with such a large amount of data. It's exciting times. Politics. I don't really know much about politics. We'll probably skip through this. Who should decide how AI is developed and deployed? I would agree with general public and don't know or no answer, because if you've seen those videos on like, uh, Mark Zuckerberg or like, um, Sachin Panda getting interviewed by, uh, legislators and judges and whatnot, it's kind of, uh, yeah, I'll let the video speak for themselves in terms of if they should be making decisions on, on AI. So this is a big investment by Merkel, this lady here. Wonderful, I think, president or prime minister of Germany who wants to invest a lot of money into building AI in Germany. Again, I think this is like a global thing, so I'm not as much swayed by like having AI in America or AI in country X. I think AI is a global problem. Finland, it's 1% AI scheme, training 1% of its population in the basic of ML. I totally agree. I think there should be like data. Data should be a subject in school, right? So just to understand how data can be used, what machine learning is, what artificial intelligence is, because at the end of the day, it's going to touch every single industry. So much like everyone should understand the basics of English and communication and language and mathematics. The same thing should be with data. Weaponizing NLP, tough problem. OpenAI's GPT-2 model, if you've read some of that, like the ability to generate text and have that out as fake news. You need to build another AI to figure out what's fake and what's not. Deepfakes, again, this is generating those fake videos like Joe Rogan. China, you can pay for stuff in China with your face. So they're all on board with the facial recognition. Chinese internet companies expand into farming. At the end of the day, people got to eat, right? China's spending a lot of money. R&D spending by Chinese cos is growing 34% year on year, but US companies still account for 61% of global tech spend. Yeah, so the, the money that they're spending doesn't really match the money that they're, they're getting. This is a very interesting graph. China's publications on AI papers is increasing, increasing to catch up to the US. And if you, if you plot the fit lines, you see here this green one is China. It overtakes the US in like 2020, which is basically over the next year or two. So the research coming out of China is exploding. Predictions for the next 12 months: there is a new wave of new startups applying recent breakthroughs from NLP research. Collectively, they raise over $100 million in the next 12 months. We did a lot of NLP work at where I used to work, so I, I totally agree with this. This is a— 2019 is the year of, of NLP. Self-driving technology remains largely R&D. I totally agree as well. Self-driving has been exciting for like the last 5 years. Like, you watch any AI panel talk about self-driving cars, there's always that issue of, uh, uh, the local— that's the local minima of every AI panel is self-driving cars and the ethics around self-driving cars. Watch George Hotz's talk on self-driving cars if you want more about that. Privacy-preserving ML techniques. Are adopted by a non-GA Google, Apple, Facebook, Amazon, Microsoft Fortune 2000 company. Yes, privacy is very important. Institutions of higher education establish purpose-built undergraduate degrees to fill talent void. Yes, I think this has been a while coming. It's kind of why I made my own AI master's degree, right? So in my university, when I went to study AI, didn't even have an undergraduate for AI. I had to make my own. Um, so I totally agree, this is going to start happening. Has to happen sooner rather than later. If it doesn't, just study online. Google has made a breakthrough in quantum computing hardware. This is going to change the game, right? So this is Elon's law. Hardware, as I said before, hardware decides what you can build, right? You need the hardware. So this is a good prediction, I think. Google, they have the money, they have the talent, they have the hardware resources. Um, so their hardware, their compute, their quantum computing hardware breakthrough will trigger the formation of at least 5 new startups trying to do quantum machine learning. Totally agree. We had a team, some incredible— we called them the Quantum Boys where I used to work, working on quantum using reinforcement learning to develop, to design quantum hardware. Again, they talked in language I didn't really understand. But yeah, this is Elon's law. Find out if things are possible at the atom level, right? So quantum, right? Atoms. And then use— put the atoms into the place you want them to go. Elon Musk's law. Check it out. As AI systems become more powerful, governance of AI becomes a bigger topic, and at least one major AI company makes a substantial change to their governance model. I think you're already starting to see this with Facebook. They're kind of changing the way they do things. Good or bad, not entirely sure. But yeah, I would largely agree with, with these predictions. Conclusion. That's a great report. Thank you for writing it. Big shout out to Nathan and Ian. Thank you so much for writing it. I had an absolute joy going through it. I'm probably going to go through it again. Um, and I've got a lot more to research. Like look at my, I've got like 4 or 5 things here to just read through. Neuro, and there could be another, another 20 or so that I could have pulled out from the, from the report. We've got 3 more slides. Let's finish them off. Conflicts of interest, authors. Again, massive shout out to Nathan and Ian. Great report. I look forward to reading the next one. As always, keep learning, keep creating. Just a quick note before you head off. If you did watch this video, and you enjoyed it and you found some things interesting and you maybe read the report, but you didn't quite understand a lot of things and you thought, "Mow, I could never do that." I'm here to tell you, don't think that. If you want to get into learning artificial intelligence and machine learning, it's not going away anytime soon, right? So, my advice, if you want to get into it, you can. Be patient. It's very broad, right? So, explore, have fun with it. And most of all, remember, learning anything new is hard, right? So it's going to take time. So if you want to get into AI or machine learning, get into it. Don't let things like this or the terms that you hear hold you back. You can start learning. The resources are there online. Get after it.