What Do We Need to Know About AI? | Jana Eggers, NaraLogics | BoS USA 2016

Jana Eggers, CEO, NaraLogics

We expect AI and Machine Learning to have a significant impact on the jobs we do – Forrester estimate 16% of current US jobs will be done by AI by the end of the decade (while creating 13.6 million jobs, this still reduces the workforce by 7%). AI represents a significant opportunity for software businesses to embrace innovation within the industry that will destroy some businesses and provide a competitive advantage to others. Jana considers where AI is now, the problems that it can solve and what companies need to know to start claiming the benefits. What do you need to get started with machine learning and AI in your own organization?

Slides, Video & Transcript below

Slides from Jana’s talk here

Video

https://businessofsoftware.wistia.com/medias/1dflxxav1a?embedType=async&videoFoam=true&videoWidth=640

 

Transcript

Jana Eggers: So, I going to talk about bringing AI onto your team. There’s a lot that’s discussed on this; we love to personify AI. It’s something you search for, AI, and this is the picture you get, it’s about the thinking, about being like a person.

I will try to go through my slides quickly so you can ask all the questions you want. I’m happy to talk about anything, I’ve actually been doing AI for over 25 years. I am older than I look, or so I’ve been told that, but I started in the late 80s working in super-computing and optimisation, computational chemistry. I did research on computational chemistry, I did NSF and research as well as research out at Las Alamos on computational chemistry. I went to the dark side of business. I was raised by a banker so most of my career after my research stint was in software.

I love software, my passion is taking high technology, the cutting edge, into business. So I was working on SaaS products before we had the word SaaS. We were not hosted with us and ASP and they had all these crazy names for us, but that’s what I love to do.

I was in Lycos in 1996, so my parents cried when I went to Lycos. I was working in trucking optimisation software and literally my mom was like, what are you gonna do? Trucks are real! They had lost hope with me a long ago when I became a mathematician, but that’s a whole other story…

Let’s talk about AI. I’d like to change your frame of reference from thinking about robots and what’s going to replace us, and think more about artificial light – and artificial light has done a lot of great things for us, it’s allowed us to explore new areas like the caves up there, it’s lit the path and allowed us to do things, horrible things like working too late at night. It’s messed up migratory patterns, there’s lots of bad things that have happened with artificial light.

I’m not saying forget all the bad things you’ve heard about AI, and only focus on the good. Artificial light has been very, very positive but it’s also had some downsides. I will talk about how you can get involved and help us be more aware of those downsides and fix them before we find out the bad parts about them. The biggest thing is AI didn’t replace the sun. I will go on record will jobs change? Absolutely! Is there gonna be something massive where we’re all just sitting on the beach while our robot overlords come and run everything? No, I don’t actually believe that will happen. I’m someone who is in that frame of belief. Do I think there can be bad things that happen? Yes, I worry more about robots starting the next world war by mistake, by a programming error.

That said, so let’s talk about this little light of mine and what we can do with it. I will give you a few lessons. Algorithms are introverts. We think a lot about robots, those are the extroverts, that’s that form that we’re used to seeing and thinking about for AI. Algorithms are actually introverts, they don’t talk a lot, they sort of sit in the corner and they are quietly thinking. This is a story from Stanford that I’d like to point this out, they were looking at how they could get people to clean up more after eating their breakfast in the cafeteria so what they did is they put a trash can on top of a Roomba and they noticed when people were getting up to leave and the trashcan would wheel its way over and it would stand right front of them – hint hint, throw your shit away, right!

So what was interesting despite the fact you have this moving object coming towards you, to hint that you did that, people still ignored it and walked away. What do you think got people to start using it? They made it wiggle [laughing] That seemed more real to people, rather than it just being a robot, they actually had it do a little twist. So don’t forget your wiggle in your algorithms. What is it that will get people to think about using this, what is it that’s going to get them to help you realise if you’ve done the right thing with your algorithm: are you solving their problem? So, design with the full expectations of interactions. Make a job description.

We thought about Clayton Christiansen, and he’s just phenomenal, and one of the thing’s that I love that is underplayed but is talked about at BoS is what’s the job your software product is gonna do? What is that job? So, think about what people pay us for, what are your customers paying you for? Maybe you do something for free, they are still paying you with their time. Think about what people are paying you for. Now do the same for your algorithm. Just like you did with the team member, write the job description for your algorithm, what is it solving, what is the value that’s bringing? And don’t say science! The science is the tool it’s not the end. Write that for your AI team member, make sure you have a job description as well.

Lesson 2, the algorithm and the data and the problem solving. To me it’s the chicken, the egg and the bacon. We can have bad chicken, bad eggs, it’s hard to have bad bacon, but seriously think about it. It’s really hard – people say if I have this data what algorithm should I use? Or if I have this algorithm what data should I use? Both of those play off each other. A lot of people think it starts with the data, but oftentimes you can go find data if you have a great algorithm. Again, that algorithm, if you know, hey, this solves a certain problem, go look around your organisation for data that fits that algorithm – that’s why to me the data and algorithm are chicken and egg – but you can’t forget the bacon because that’s the value. It goes back to point 1, what is the job? Don’t forget the bacon when you’re thinking what data do I have access to, what algorithms can I find and apply. Those 2 things trade off. I hear way too many people saying I have this data, how can I solve the problem with it? I don’t know, does that data solve the problem you have? Maybe you need to look harder to find data. Or, hey, I know this machine algorithm with – you know once you have a hammer everything looks like a nail. That’s happening a lot with machine learning right now. People know this and they think it applies to everything and it doesn’t.

I will show you a few examples of that. This is automatically playing. Notice the iterations up there. So, this is tensor flow. Google was open, it’s the playground on tenser flow, and it’s awesome by the way. I picked one of their training datasets and look how in 55 integrations a 6 neuron, 5 layer deep algorithm came to this classification. 55 iterations, that’s great!

Now, I pick a different data set, same exact algorithm and it starts playing and it can never classify it. I let it go to 200, but you could let it go ad infinitum and it’s not ever going to. The point there is that algorithm and the data go hand in hand, it’s just like the chicken and egg, you can’t take them apart, they rely on each other. Now, same data, new algorithm, and I think it gets around 85 or so. That’s the point here, is chicken and egg, algorithm and data. Think about both of them and don’t be constrained by either of them. Tenser flow is terrific and gives you a great way to learn about AI. Go and play with it and see what happens when you configure things differently and you’ll start to understand, you’ll see how important the data and algorithms are together but again I will go back to it. Bring home the bacon! Bacon makes everything better; the value is what you’re driving for so don’t let the algorithms or the data distract you from that.

I have to have Ada Lovelace in here, I’m a nerd! What I loved about this is, a script is what you give to an actor, but a program is what you give to the audience. You’re going for the program, that’s what you’re giving the audience. The script is, in this case, data, the actor is the algorithm ,but what you’re going for is the audience. That was the inventor or Perl if you didn’t recognise him. So, lesson number 3 rocket science is the easy part.

I know I’m harshing on that algorithm part, but that’s the magic and that’s what people get so excited about it, and I mean dwhat Lacoon and Hinton and Bengio an dall those guys have done in the last about 5-7 years is truly amazing. It’s really revolutionised stuff that I was working on. Out at Los Alamos I was working on similar artificial neural nets and that was back in the early 90s. What they have been able to do to make those algorithms much more effective and more usable by people that weren’t using super-computers, what I used to do there, I would be requisitioning for somewhere between 1-3 months I would requisition about $1 million in compute time and that was cheap cause I was buying it from the lab, from ourselves. The fact that we can do that now for $500 is amazing and that is due partly to the hardware which is true, but also due to the efficiency these guys built in, in optimising these algorithms, so it’s really exciting to see but don’t get distracted by that part. Rocket science is actually pretty basic. It’s really managing a Space program that’s hard. This is from Voyager, if you don’t know there’s a disc on the Voyager cause everybody knows aliens can read discs. But there’s a disc there, and imagine what went into that. How do you encapsulate our humanity onto a CD ROM? How do you do that? Imagine all the thinking that just went into this one disc. 111 samples of what it means to be human on Earth. It’s narrated with Carl Sagan on there, it’s pretty exciting and this was done back in the 70s and it’s pretty exciting to think about that and what we’ve sent out, but that’s one tiny piece of what Voyager did. Think about all the computer software that’s controlling it, all the communications that went around that, so you got to think big with these AI systems and not get distracted with the one point about the algorithms and that’s what I hear about so much, it’s people debating algorithms. I think it’s an important component but what is more important is to take your whole Space program and think about what you’re trying to impact and that will make the decision of the algorithm a lot easier.

So, lesson number 4 Beware of Kasparov’s mistake. Do you remember when Kasparov lost to Deep Blue? Do you remember why he lost? He lost due to a bug in the code. What’s interesting is he lost it, it threw him off his game. What happened is, the computer made a move that he didn’t just understand and he said, I have to be missing something, I have to be missing something. I’m looking at the board and it doesn’t make sense and this computer is smarter than me, I know, because look how well it’s been playing. It chose that move randomly. It did. He wasn’t expecting a mistake, he wasn’t allowing himself to believe that he had actually found a bug in the code, that he had found the wormhole to go through, right? He said this computer is pretty damn good, I will take it for what I see and say that’s great, I got to figure out what it’s thinking.

So expect the bugs, especially in your data. And by the way, your data sucks, I promise. We deal with this all the time. My company, we do recommendation software, it’s a new type of AI. We use traditional techniques but we’re also using a new technique that’s really very focused on signal/noise and it’s from neuroscience, its from how neurons in our brains connect , I could talk to you about that later. But the big point is that we work with folks who will tell us our data has been cleansed, it’s amazing we have spent so much time, we have spent the last 10 years working on this project and as soon as we start using it and it comes up with recommendations and we have this ability… A lot of AI isn’t transparent, it’s a big problem that we have in AI right now, particularly Deep Learning technologies. It’s not transparent. It doesn’t tell you why it’s giving you this answer, just, ‘The computer said so’. And so we actually – that’s one of the things we do, it’s very explicit and it tells you why we’re giving the answer, and we look at the why and they say this answer doesn’t make sense. We look at the why and luckily, plenty of times they look at it and say actually I hadn’t taken it – that’s really good, I like that!

But most of the time when they disagree with it, it goes back to their data, it’s not our algorithm, and when I say most of the time I mean 99% of the time and the reason is people don’t realise they spent a lot time cleaning the data but they didn’t actually know what the data was used for, and don’t realise this department uses it differently than this department and they’ve overloaded fields, they’ve put lots of… and it was great for them, it worked perfectly, it was very clean data from that department, who by the way was QA-ing it themselves, and did your data go in her correctly? They’re like yeah, looks great, we love this, it’s terrific! They didn’t realise that now I have an impeded mismatch between those two things.

So, there’s lots of problems about that and it goes back to we have to know what we’re driving for. Do I have the right data for the problem I’m working on? A lot of times, like here, it was a bug in the software and it comes out wrong, I need to have people questioning that in AI and we’re not doing it, we’re really letting it go too much. Well the computer figured it out, the computer’s smarter than me, it’s not. Cause it’s based on your data and a human coding.

Does anyone know what this picture is? This is specifically for Mark. No one knows? The Roomba tracking poop around, the Roomba poop apocalypse, right? This is what a gentleman from Arkansas woke up to one morning and it was everywhere, it was on books in his bookshelf. Imagine if you know the Roomba, the spinny brush spreading poop everywhere. He was horrified, as you can imagine, but he wrote a very funny post about it. This happened this summer, this has actually been happening for years. The Roomba guys who were local, iRobot guys, are fantastic engineers, really top of the line, amazing engineers. But they can’t detect shit [laughing]. Now, I don’t know about you guys, but you get a little, tiny bit of dog poop on your shoe, you smell it for 3 days. One of the most humanly detectable things in the world, we can’t teach a robot to detect. They’ve tried, it’s not like they haven’t tried– it’s not like they say we tell people not to tell their Roomba loose with their dog. They do do that, but they haven’t figured out how to and you can come up with lots of ideas but nothing is reliable.

I’d like people to understand where we are with AI, and this is a really good example. We’re not as far as you think so we can do things that are really good at computations, we can do that really well, but a lot of things we can’t detect and humans are good at detecting. Always think of that, think do I have, does AI or the data, do we have the capabilities of detecting this? You may want to scope it differently and say we are only going to detect this amount but make sure you’re questioning this, otherwise you will get caught up and have something that ended up failing.

Think of Microsoft Tay, the horrible incident. You may not know that’s actually a chat bot that they had running for several years in China and it’s loved. People tell it good night and they say it’s one of their best friends, they interact with it 3 times a day. It’s amazing! You’d think they know all of this, well think of the different culture between China and US. 4chan doesn’t exist in China. 4chan is what brought and made Tay a genocidal maniac and that could have been solved if you had thought about it in advance. The way they trained Tay to speak like a millennial, they could have also trained her to avoid. You can have negative set training too. If anybody starts talking about Hitler, don’t do this. There’s plenty of sets out there, there’s a name for that, yeah, Godwin? I mean I’m not saying Microsoft is stupid but it’s easy when we have success we forget to think about does that success apply in this situation? Zach before was talking about that. Are you really sure you don’t have a false dawn? They didn’t think about how is this then gonna replicate, and replicate in a different culture. So, this is hard stuff, don’t underestimate the complexity. Make sure you bring in expertise, and it needs to be expertise in a lot of things and that doesn’t mean expertise outside of your organisation. One of the things I tell people is be careful about handing this outside because you’re the person who knows your game the best. You’ve got to be a coach, you’ve got to be the mentor; it doesn’t mean you have to understand how to play every single position, so make sure you are judging that with those outcomes in mind.

So how do you let yourself shine? Algorithms are introverts, remember that, build a job description. Results are bacon, make sure you are focusing on results. Start small, but make sure you have a big vision. I’m not saying don’t scope your products down when I say it’s not rocket science, make sure you’re focusing on the rocket science in context and make sure you have that big vision cause we get focused on the whole thing we forget about the program we’re doing, make sure you learn from Kasparov and expect mistakes and shit happens, right? Make sure you have the hazmat suits ready.

I will leave you with Rob – he wrote about a year ago in Fortune an article saying, it’s called the Algorithmic CEO – his point was we’re entering a new era. Before you had to have CEOs that were sales people, they were the lead sales people promoting to Wall Street to customers, and get people engaged. We went through an era of operational efficiency and that was what was important for CEOs , you have to know it and how to lead a company and we built a whole bunch of leaders that knew how to run companies. He said we’re entering a new age, the algorithmic CEO. The CEO needs to know how to run the company driven by algorithms, it’s the single biggest change in today’s business world. He focused on how do we use algorithms to personalise, not just to an individual but to our customers and how do we make sure we’re tailoring to them. It’s a great read and recommend it but that’s what we’re talking about here, algorithms driving business change and he said those who learn how to leverage it are going to be the winners.

What’s the problem with AI? I hoped I’ve helped you today and this is my ask of you. I don’t think any of you were alive then but in 1935 – you may have heard there was a New Yorker cartoon, a drunk man under a street light and policeman walks up and says, ‘What are you doing?’ and the drunk man says, ‘I’m looking for my keys’. The police man says, ‘Did you drop them here?’ He says, ‘No, I dropped them 2 blocks that way but the light is better here’.

So AI has been the realm of – I’m pointing the finger at me here, I’m a mathematician – mathematicians, physicist and computer scientist. It’s been a very high technology leading edge since the 50s – the leading edge since the 50s, it’s taken us a while to get there. The reason why it has come of age now is really data and computer power.

I was lucky the stuff I worked on I could get the chemical readings I needed to from very sophisticated equipment out there and had access to the super computers. So I had the data and computer power that I needed but most people didn’t and it wasn’t until the internet came around and we had all that data, sharing and making it accessible, we need that for researchers and we’ve been able to make it usable for business and it’s exciting. That said, we’re still the drunk under the street light, we’re focused, as scientists we’re so happy we have data and compute power we use that hammer on everything we want to pound down.

We really need more people coming in, we need neuroscientists – this is one of the reason I joined the company I did, it saw something different than what we were doing before and something new that had value and I got excited by that. We also need product owners so most of the time machine learning is a piece of the engineering department that is a layer underneath the rest of engineering. They are like a support system to the rest of engineering. We have got to break that mold and bring product owners up for the machine learning particulaly itself. We need information architecture and think about UX – we need to start bringing other fields in, we have got to think about the business problems and funding. I don’t mean the drunk funding that’s going on right now. You say you’re deep learning and people puke money at you. It’s disgusting. It worries me there’s another AI winter coming up. We need ethicists and lawyers to figure out what we’re doing here, do we understand it and we’re able to show…

One of the reasons we get selected by our customers is really for compliance because we can give that explicit reason why we are making a recommendation, and for compliance reasons that it matters and it should. We need entrepreneurs like you and also the suits. We need everyone. We need you to get involved in AI, you need to pay attention to it, please don’t leave in the realm of the people like me. But we also need people not like you, it really is about diversity with AI. This is a full team effort so get involved with AI and interact with it. So it’s time to shine, I leave that to you guys, can I have 20 minutes for questions? [clapping]

Q&A

Audience Question: I’m an inside sales leader and I’ve been reading about what this technology can do. Where do you see this technology being able to help sales organisations?

Jana Eggers: Great question! What’s most exciting to me is you know how much research salespeople do. They spend a lot of time researching and directly targeting their customers and that’s a lot of the times of why they’re so successful because they can build the relationship with the customer and imagine if they have that digital assistant that’s collecting relevant information. We’re working with the government on this as well, think of the intelligence analysts. How do I make sure the right data for my problem right now is coming to me? For salespeople it’s about recommending what customer to call, why to call them and the information to share with them about it. I won’t have a robot calling them, it’s them analysing and going, you know, I know them and this should be tweaked this way. But imagine information coming into them, very personalised to them, and it could be about – our sales leader at our company is big into sports. It’s natural for him to talk about sports so imagine when the data comes to him it’s personalised in sports, what’s happening in that city in sports is something that he can bring out. He can relate to someone there but it’s personalised to him, and another person who’s much more technical can have more technical information come in. It’s not just the customer and what the customer is doing that is important, it’s also how is this person enabled?

Audience Question: What are the barriers between investors and AI focused products and what barriers will break down first in that relationship?

Jana Eggers: So, I don’t think there’s a lot of barriers right now. VCs are very excited about this space and doing a lot of investing. I’m worried about is they invest somewhat in the same thing, it’s a hammer, and I worry we won’t be able to drive that much value, we’re pounding the same nails so how do we get more diversity in what they are investing in– and that’s what I think will break down, we will get saturated in one are and not enough in more diverse areas.

Audience Question: I feel like the term AI is being abused so much in the SaaS world and it seems every investor only I feel like a lot of SaaS companies are thinking about this stuff when their actual core product is lacking. What advice would you give to SAAS companies that are thinking about it? Where’s the balance? Should you even touch it if your core product is lacking or should it be done in parallel?

Jana Eggers: Unfortunately, too many people rushing to get it into their product just so they can say they have AI and because no one on their team has experience with it, they hire someone that did a summer project somewhere in Machine Learning and they grab some open source tools off the shelf and this person doesn’t know the data. They may know the algorithm but don’t understand how that algorithm applies in different cases and they don’t necessarily know the business. So it’s really, yes I think people need to get involved in this as I think there will be great advances where software will be better. If you did some form of AI, there’s some kind of intelligence we could apply we couldn’t before. That said you have to think about it in the context of everything. So where do we trade it off? You trade it off at that better point. You said there’s thing I know I can make it better but what about the unknown unknowns? Make sure you are investigating those, understand those and think like where can machine help people make a better decision or help my operational people make a better decision. This is what we do very often. Our customers are the global 100 so we’re talking about the high end data and they come in typically saying we need to get involved in AI, how do we get involved in AI? And we spend time with them, finding a problem that fits AI well. So it’s, that’s what you need to do, is go in and do some research and where do I find the right data or need to do that? I have some examples, I have this good one but I can’t tell you cause it’s launching in the 1st. I wish I could – whisper it right there. Cover your mouth and lean down and no one would know.

It’s such a great example of a big company thinking about where they can apply it and really changing how they interact with the customer. They have a channel sales model and this will be something that I truly believe will bring them much closer. I say they’re not an arms length from their customer, they’re really a football field – they came up with something that I believe will bring them much closer to their customer. Think of it on the sales or operational side or how do I interact more with the customers? I would be shocked if there weren’t some opportunities that it can be applied in a contained way that you can start getting used to it and then build that in out. I get asked this all the time and it’s related. How do I hire these machine learning people? I don’t believe you have to hire machine learning people, I don’t think there are enough machine learning people out there that have a breath of experience that can integrate into your team well. I would hire smart engineers but make sure you challenge them to understand the different algorithms. You’re asking how is this applying to business, why is this the right algorithm, why is this the right data? I don’t think there is a new breed of engineer that needs to be built, I think it will come naturally. But I think people worry too much hiring machine learning experts. It’s needed for people like Google, Facebook or Amazon but most people’s problems don’t need that right now. We’ll get there, but don’t let that hold you back.

Mark Littlewood: What should a top software company thinking about where are the AI developments that people are thinking – how it will affect these people? What should they be doing to stay ahead of the game?

Jana Eggers: It’s a good question because the challenge is the hype that’s out in the industry. What I’d say is back up and think about where your most important decisions, where is the place where you get a lot of data that you argue about, and where is the place where you don’t see that you have lots of combinations of things. One of the examples I gave at one point is we spend, take an hour long meeting, we spend 55 minutes debating the data and where it came from and what it means and 5 minutes on the decision. Where is that AI can help you figure out that 4 different decisions that you need to make that have that information more readily at hand at hand and it’s hard because of the transparency and AI is not quite there yet, and think more about that, where do you debate the data more and it will figure out more combinations that you thought off. You didn’t run 15 reports – think about go, it ran 20k reports, and then it’s bringing you the top 4 you should think about, and then let’s discuss and debate that. Again you have to know your data provenance and all that. Get started and also don’t be afraid to get started with smaller data sets. AI is improving with the amount of data that it needs, and this is where we have a hammer, everything is a nail. Even deep learning is getting better with smaller sets of data which is great! We can start smaller, we can understand it better and advance to larger algorithms.

Mark Littlewood: It’s getting there, it’s hard. Everyone is interested in this stuff and try to think how is it relevant.

Jana Eggers: I wanna be fair about that, it’s very exciting what’s going on, it’s just not quite the automated magical level that’s being described.

Audience Question: Great talk! Thank you! Do you believe in singularity happening in the next 20 years?

Jana Eggers: No, not at all. I think we’re so far from that. Right now and I just did this calculation based on the experts in July but right now most of the experts, if you look around it would happen in 35 years. I think they are off by at least another 35. I will be dead and won’t be saved by the singularity but could I be wrong? Absolutely! I know where we are with this and I know the pace it’s moving at and I know where people are focused, I don’t think we’re anywhere near. Yay for us! Sucks for our grandkids! The compute power, like I said I wish I can remember, I had a really late flight in last night, but the difference in numbers in the energy it takes between computers and us to analyse a situation and make a decision is unbelievable and that alone is going to really hamper the efforts towards, hey we’ve got a super intelligence.

Mark Littlewood: Cool brain fact as well, 30% of the energy that you eat is about keeping your brain going. I had no idea.

Audience Question: When the Roomba finds the dog shit, that’s where the singularity starts. So we’re all busy, excited and hyped about AI and we’re getting infused with that ad-nauseum in the media. Do you have recommendations to cut through that and find out what’s at the core for businesses, it could be blogs, books, conference, but finding stuff that’s really actionable and less theory?

Jana Eggers: Yeah, I wish I did and I do get asked this all the time. It’s easier for me cause I’m in this and I can throw the articles out and I should tweet and say this one’s good and this is not, but I don’t have a good recommendation. I’m excited O’Reilly is doing their first AI focused conference at the end of September and I what I really liked is, there’s lot of conferences and they tend to be on the research side. What I love about what O’Reilly did, for disclosure I’m speaking there so I’m biased maybe, was they specifically said the talks have to be about applications to business, so it’s not about the research but about how do you apply it to business. My talk is on project managing an AI project.

Audience Question: Thank you for your talk! Do you think the current approach, what we could say, narrow AI is the right way to get to consciousness or singularity?

Jana Eggers: I think it will get us understanding, again I think a lot of this is us understanding where the technology is, so I do think it helps us understand pieces. I talk about the Space program, there were a lot of pieces to understand to get that Space program. How to eat an elephant one bite at a time. So I think AI allows us to do that, but we can’t lose focus on the end and I do think there’s a lot of shovelling into our mouths and not realising, oh, I need a different taste – that sounds really bad to eat an elephant – but we’re just eating the same thing and going this is good cause everyone says it’s good. We need to have much more diversity there so I’d say narrow AI isn’t bad as long as it’s not the same AI going against different problems, we need to make different types of AI – this is a cool brain fact I didn’t know before – the neurons in our brain, they thought they were specialised for a long time until a guy here at MIT who is one of our advisors actually took the neuron from the retina of a frog and it put it in its ear and it actually could perform the function of hearing, which is amazing. So it’s the same cell and it’s not specialised. I do think there’s a lot of general – even narrow AI can be applied across multiple things, but there are still some things we need to learn to do differently. And our cognitive functions are different from our perceptual functions so think of it that way so it’s not just the neuron itself but it’s the way that’s wired into the brain and the function that it has.

Audience Question: Thank you for a great talk! Looking at it from the point of view of business, giving your talk the formula came into my head, data x computing cycles x the math and the science of AI equals the ability to generate value from all that . Data is getting a whole lot cheaper, computing cycles are getting cheaper, to your point, AI experts are pretty dear. Is there an opportunity given that the first 2 are dropping down by orders or magnitude to be able to apply simple AI to get more of a near term arbitrage on it, looking from a business point of view, perhaps not best practice for science.

Jana Eggers: That’s fair, I do think there are some problems that can be solved – the example I was gonna give, these folks are using some really more commodity image recognition, and I don’t think it was naturally thought of by them or in their space, that oh, that would actually be useful for us. So I think there’s opportunities like that. Think about chat box, natural language understanding, that is another example that has gotten to the commodity stage, that’s saying nothing about the companies that are specialising in that and we need to improve it but the basics are there and are available freely. The one place I would argue with you on the data. Data is cheap but you get what you pay for and I’d really be careful about, do you know that data, do you know where it’s coming from and do you know everyone who generated it and do you understand it? I would put a little more emphasis on the data than you did there, I don’t think good data is cheap. I think data is cheap but not good data. There is a trade-off there but I think you’re heading in the right direction and there are a lot more simple problems we can solve if we just look at our problems differently. I have to solve this one and use machine learning to do it.