AI is the New Electricity - Dr. Andrew Ng

thank you thank you um you know it's really looking forward to speaking with you today um as you know I work in AI and AI is changing the world and I think the way that AI will change the world will be through a lot of the work that you are doing building the small startups today that will be the giant companies of the future you know the title of this talk was Aires the new electricity and I think that about a hundred years ago as we started to build electricity into the United States that change industry after industry right everything from manufacturing agriculture healthcare transportation communications was transformed with electricity I mean if not for that chassis but this would be a completely dark room right service really transformed the world and I think AI technology especially machine learning but other things too is now reached the stage of maturity where we see a surprisingly clear path for its also transform every industry so what I want to do today was um as I was preparing this presentation I spent some time thinking over what might be most useful to you and I thought I'd do something different and do a whiteboard presentations I had over how I think some of you might be able to use AI in your startups I think they're going to project to the big screen right someone has drawn does and this is a camera person okay awesome thank you so um a lot of basic ideas of AI have been around for many decades so question of of analysis people ask me hey tanker why is AI finally taking off now so let me draw a picture but that is the picture I draw when people ask me that question and if someone asks you the same question you could draw the same picture for them if on the x axis we plot the amount of data you have for a problem ah then what's happened in society over the last twenty years is with the rise of the Internet the rise of mobile devices we've just accumulated more and more and more data from all the problems for example once upon a time in the supply chain you order something the fact that you ordered you know a notebook that would be recorded on a piece of paper and maybe that piece of paper is somewhere in the filing cabinet in Philippines today the fact that he plays in order for something or you order the TV is much more likely to be a digital record and a computer so the digitization society created data and now if on the vertical axis I plot performance then what happened was for the earlier generations of AI for the traditional machine learning algorithms right traditional ml for traditional machine learning it was as if for the early generations of algorithms which is regression support vector machines may be decision trees it was I said they didn't know what to do of all the data you can now feed it but in industry after industry we've been seeing that when you train a neural network so we train the small neural network and n stands for neural network then the performance sort of gets better if you train a medium-sized neural network and if you train a very large neural network then the performance sort of keeps going up and this means a couple of things one is well if you want to get the best possible performance or if you want performance all the way up there you kind of need two things one is everything helps if you have a lot of data and maybe big data is one of the buzz words people sometimes use and second is you need to computational power to Train maybe the big enough neural networks in order to take advantage of that data although thanks to the rise of GPU computing I think that more and more people including startups now have access to you know the whatever 1/2 doesn't or maybe a couple of dozen GPUs that you need in order to take advantage of the data sets you might be able to acquire now there's a kind of hype about AI and um I want to share of you just what's really working and what's really useful despite all the hype and and the fact that what transform industry after industry almost all the economic value creates it through modern AI and I think also the best opportunities for all the for many of your startups is through one type of AI which is learning a derby or input to output mappings and this is called a supervised learning but um as an example for a long time we've had algorithms that could input an email and tell us if it's spam or not right 0 1 this is supervisor I think a de piel input output um I once did some work on loan applications where you input a loan application that's a and map it and output B that says will they repay this loan 0 1 and this turns out to be a good way to decide what those to approve there's a good decent business the most lucrative application of this on the planet today I I believe is online advertising where the large online ad platforms all have an algorithm that inputs an ad and some information about you and outputs where you click on this ad because for all of you for the large online advertising platforms every clickers money and so there's a very large financial incentive to build the best possible system then show you the ad that you have a 5% chance of clicking long rather than a 4% Johnson clicking on maybe not the most inspiring application IV I may be but but certainly incredibly lucrative um speech recognition input an audio clip I'll put a text transcript and you know some of you will have I guess Amazon echoes or like Google homes or like Apple Siri like devices in your homes the main reason you're willing to set these devices today is when you say a lexer you know whatever Alexa what time is it the chance of the Amazon echo waking up is very high if it responded to you only 50% time you won't use it the consumer experience just unacceptable but we're willing to use speech recognition devices like Amazon echo or Google Voice Search or buy to voice search because speech recognition accuracy is really high so we actually still live in an era where for a lot of applications having great AI technology matters because if you're sweet recognition systems accuracy is 99% the users might love it if a sweet - accuracy was 95 users might not be willing to accept your product and so we still live in an era where having great AI teams really get the best performance out of the technology matters to the user acceptance and online advertising as well right every clickers revenue so the ability to show a slightly better ad is incredibly lucrative and maybe this is why I live in the era where the quality of the AI technology really matters um one of other examples input and electronic health record and output this is owned one of my students on a navantia Stanford did this right and put your health echoing estimate the chances of someone you know mortality in the next year this helps steer people to palliative care self-driving cars one of the key components is input a picture and maybe the radar lidar readings and I'll put the position of other cars right so um that's not driving companies I guess I'm on the board of drive dog AI and and I'm really lots of self-driving car companies this is a key component where again the quality of the technology matters mentioned I'm working a landing AI which is helping manufacturers transform using AI so one of the things we do is input the image of a part you know picture of a consumer electronics device that you may have in your bin in your home or and then that you may carry with you and output you know is there a defect right and again this turns out the machine learning albums needed to do this is really tricky but we can do this you can carry out inspections awesome automatically and this is very valuable so this is just learning A to B mappings right the supervised learning but when you find the right business use case and we find it by business context this turns out to be incredibly valuable now um I like to cut through height for machine learning and AI and because there's a lot of hype um turns out that you read in the media of a lot of different types of AI technology and so you hear about I'm gonna little bit technical I guess supervised learning that's learning ATP mappings transfer learning which is um learn from one toss learn learn to recognize casts and use what you've learned from recognized accounts to read x-ray images because pictures of cats around and there are x-ray images you hear about unsupervised learning learning without label data and you hear about reinforcement learning which is great for playing Atari games and playing your various sorts of games it turns out that the PR you see is not a long in proportionate to the short term opportunities and the economic value created the value created as you go down this list rapidly decreases and so I actually think that many of the best options but you will be in supervised learning although the AI deep learning work changes so fast is entirely possible that there'll be new inventions and this order will flip around in two years or five years but at least today I think the media reporting is not proportionate to the actual value created although we like future technology so in the future maybe in two years this warning will change when other ordering insight that that may be useful is on most of the hype is on a I applied to unstructured data and that means things like images and audio clips and text because we're very human we like it when the learning algorithm seems to understand language right um because we all know unlike pictures but it turns out that if you look at the economic value created there's actually um let's see structured data so structured data means you know your database of transactional records who visited one website whether they purchase or how many minutes was this phone call those are examples of structured data and once again it's fun to read news articles about you know computer vision recognizing cats that's unstructured data image data but at least today there's actually um more economic value created through machine learning applied to structured data like your boring database records although I think that things on unstructured data like speech recognition would create new categories of products like the smart home speakers um you know I want to share with you one perspective on AI and if you remember just one thing from this talk remember what I want to share of you NIC's um with the lessee smell a lot of times you know looking at the rise of the internet right rise of the internet as the previous wave of as one earlier previous wave of Technology of disruption and one lessons I learned was on if you take your favorite shopping mall right here with my wife like Stanford Shopping Center right but and and and let's say you build a website for the shopping mall and website shopping mall can sell stuff on the website this does not turn the shopping mall into Internet company right this is really obvious today but they were actually retailers maybe five years ago that said things like no amazon has a website week on the website amazon sell stuff on the website we sell stuff on the website it's the same thing but of course is not the same thing so what's the difference we're solving well first website versus Internet company the difference is have you organized your company to do the things the internet lets you do really well for example Internet companies will use the pervasive AV testing because you can and if you don't your competitors well and they'll learn faster we're used to short shipping times right so there's no shopping mall my ship a product every may rearrange things every six months internet companies we should have crowd out every every week or maybe even every day because we can and we learn much faster that way um we as internet companies have learned to push decision making down to the engineers and the product managers so you cannot have an Internet company where the CEO says everything and everyone just does what the CEO says because it's impossible to Co to know enough well at least when the company grows so this is why we push loads as you're making down to the engineers in the product managers because only they understand the tech and users well enough to make great decisions so this was the internet error the rise of the AI era I think a traditional tech company plus you know a few neural networks or a deep learning or machine learning or whatever this does not make the company company and I think one of the best opportunities for startups and why I think they'll be great AI starts as built today is just as with the rise of the Internet some incumbents did a great job transforming themselves at in Microsoft and Apple did a great job transforming themselves to be Internet companies there were also startups such as you know Google Facebook and Amazon and bite you and so on that did a great job rising up with the internet but the AI era my own biases and that teams at Google and Baidu but I think we're gonna fight you have done a great job transforming themselves to be great AI organizations not yet done still more work to be done right frankly all no one's done we're also trying to figure out what's the bulleted list that should come down here but I think that just like the last wave some incumbents Google by doing a few others I think Facebook Microsoft you know hammers are not great I think there'll be new startups as well that that figure these things out now we don't tell you know what are these things down here but I'll just mention a few strategic data acquisition you know one of the same about that later but in a lot of AI businesses is actually data even more than technology that creates a defensible barrier for your business and so for example you know in in China people speak slightly different languages in different regions so I've done things like launch a product on one location the acquired data use that to attack the neighboring location with the language slightly different use the data from first regions to go to the third region but we don't monetize any of this we think all this data we monetize in there totally a different location so both you know like competitors and I we've all done things that make no sense from a revenue point of view and this because we're playing this multi-year chess game in terms of data acquisition so I'm seeing the great a I organizations get really sophisticated in terms of thinking through these are multi-year data acquisition strategies this is less of a problem for startups but more for traditional companies or the large companies are doing a lot of work to unify the data warehouses because there's only of you know from instead of having like 50 siloed data warehouses political and single data warehouse that your engineers and your software can even look at it to create value I find that we're very good at spotting automation opportunities so for example one rule of thumb I've given to a lot of my teams is almost anything almost anything that the typical human can do in less than a second mental thought we could probably now or soon automate right so so I can't product managers rather are my teens Jonathan what are things that I could do in less than a second because those are often good candidates for automation in the near future using machine learning and then there's one very complicated category of a new job descriptions you know I think it was really I think Microsoft had pioneered the role of the program manager but in in in second Valley we have rows I proper manager a program manager and we figured out this workflow of how the right you know mobile ads or web apps right the product manager draws a wireframe and the engineer you know me tries to implement and it will figure out how teams work together in the AI era these workflows are breaking down so for example right well you guys know what a wireframe is right so you know product manager draws a picture maybe we were designing the Facebook app or something maybe a product manager it was a picture like this and then the engineer implements that right that's the wireframe but say you're building a self-driving car if a product manager comes to you and draws this picture and shows this picture to the engineer and says hey built this the engines say this is totally useless why are you drawing a picture for me so we're think about that in the AI era our product where we're training our product managers the PRD the new PRD the new product line this document is a data set together with an evaluation metric on the requests like them how to managers might take a dataset an engineer and say hey do you think you can get 99% accuracy detecting all vehicles less than 50 meters from our vehicle and get this whole product manager specifies and and so this I think is the is the new the job descriptions in the AI era and the way we break up toss is very different so I think that um you know 20 years ago we did not know I did not know I don't know any of us knew how important a be testing was would be for the for the a fellow Internet era I think that was still all of us even the great AI companies are still figuring out what are the things on this list on the right and I think some incumbents will do a great job freaking these out in transforming but hopefully I think that there'll also be some startups that will you know be the great companies of tomorrow because of this wave of technological disruption now um I just I mentioned one last piece of advice which is how do you get into AI right I should get all sam'l on how do you break into AI because the so many opportunities to use machine learning by machine learning telling the scares and technologists really difficult I was really surprised in the manufacturing realm it really forced us to use AI technology that was invented in the last two years to do some things we do so the an technologies is really complicated but um if you think about how you learn to code right you did not learn to code by jumping into I don't know the open source you know most mozilla code base with some some some huge project you learn to code by hello world and then learning to print numbers from one to ten and then eventually implementing quicksort or whatever so for people that want to get involved in AI I think that um you know online courses are really digital things like hoses we teach on Coursera is actually a very good on-ramp to learn about the foundations and then beyond that up unfortunate to mentor a lot of Stanford PhD students you pretty perfect predictable repeatable process but it's often a good idea to try to replicate existing results and if you do this research papers replicant publish results people tend to under Pichet how important is is when you show up you know with a new project or if you have an interview somewhere these things really matter and then finally you know by the way when when all of us interview open source we I mean when you have stuff on github right all of your written blog posts that really counts but there's a fairly predictable process for teams to get great at AI and so your bio me is trying to hire the Stanford AI PhD students or the Berkeley AI PhD students or whatever if you can but we now have great resources to help you learn these ideas to try to incorporate them so he has an electricity I find it hard to think of an industry that AI will not transform and for us to implement the societal change the only way to do so is for really all of you many of you to build the great companies of the future that that will help make the whole world better using AI thank you very much [Applause] [Music] [Applause]

Channel: Startup Grind

Views: 13,061

Rating: 4.8939395 out of 5

Keywords: startup grind, startup entrepreneurs, entrepreneur, entrepreneurship, billionaire, documentary, motivation, inspiration, how to, how to be an entrepreneur, how to start a business, startup company, startup tips, startup ideas, startup business tips, silicon valley startups, success story, startup stories, startup advice, technology, silicon valley, san francisco, venture capital, investing, fundraising, unicorn, future, innovation, startup news, technology news

Id: fgbBtnCvcDI

Length: 20min 32sec (1232 seconds)

Published: Tue Mar 06 2018