What is Artificial Intelligence?

On May 4th at the Kuumbwa Jazz Center, Santa Cruz Works hosted AI for Knuckleheads, a panel discussion about topics in artificial intelligence. Despite a new variant on the loose, the turnout for this event was robust. In a comfortably full house, the audience was voluntarily masked and spaced. Snacks and beverages in the outdoor courtyard drew those inclined to mingle and socialize before the discussion began.

The panel participants included: Beth Ann Hockey / Senior Principal Data Scientist at LivePerson; Jim Whitehead / Professor of Gaming at UCSC; business leaders, Craig Vachon / CEO of AI Redefined ("AIR") and David Urbanic / SapientX.

Craig Vachon posed broad questions to the group to get the discussion going including:

• What indicates success in your professions?

• What is the end goal?

• When will AI be able to understand the complexity and nuance of human language in a given context?

• Are we responsible for the actions of the AI we create?

Artificial intelligence includes many specialized algorithims which process large sets of data. The panel discussed various types of AI as well as the considerations underlying their use and application. Some mentioned were: natural language programming (NLP), machine learning, deep learning, neural networks and computer vision.

Additionally, topics of responsible AI were addressed. Including the lack of explainability sometimes resulting in what’s called the “black box problem“ – when the inputs and outputs of a model are known but what happens in-between remains unknown. And the danger of data bias – when data is chosen in a way that is not representative of real-world data distribution.

For a more in-depth experience of AI for Knuckleheads, watch the video or read the transcription below.

Transcript

[00:00:36.830] - Craig Vachon

Thank you so much for thank you so much for coming. And joining us tonight. We're going to talk about artificial intelligence, and we're going to talk about what it means today, and I hope what it means in the future. And we're going to talk about our responsibilities as AI practitioners. We're going to talk a little bit about what the Ramifications are about those responsibilities. But before we do all of that, we're going to start with just sort of an introduction of ourselves. And if we may, Bethany, would you mind four to five minutes on who you are and tell us a little bit about your firm?

[00:01:29.870] - Bethan Hockey

Yes. My name is Bethan Hockey. I started off getting into AI way back when I started seeing interesting things about trying to talk with machines. And that got me going to the University of Pennsylvania, where I got a master's degree in computer science and a PhD in linguistics. And right out of school, I worked for NASA, and we put a conversational system onto the space station to help astronauts do procedures. I got to be the third most famous computational linguist in the world for two weeks after that. Yeah. I variously had my own consulting company where I worked with a variety of clients, including Ford Motors, a lot of smaller startups. I taught at UC Santa Cruz. I taught spoken dialogue systems, and I worked at intel for quite a while. And we put out sports glasses in collaboration with Oakley that were made for long distance cycling or running. And it was like you had a coach next to you. That was the model. So now I'm at Live Person, which is a company that you may not have heard of, but that you may use pretty often because we put out a platform that supports a lot of big brands to communicate with you through Webbased chats messaging.

[00:03:05.560] - Bethan Hockey

And now we're putting out some applications that are spoken. I think that's about it. About me. I'm a senior principal data scientist at Live Person.

[00:03:21.370] - Craig Vachon

Cool. Thank you, Jim. You're up.

[00:03:28.720] - Jim Whitehead

All right. Thanks, Greg. All right. Yeah. So my name is Jim Whitehead. I am a professor of computational media at UC Santa Cruz. I helped create the computer game design program there. And I have to say no one is more shocked than I am that we now have somewhere around 700 students at UC Santa Cruz studying games. And one of at this point is four different games programs. So pretty amazing journey from a Word document on a computer 15 years ago. In my research in my lab, I'm interested in an area called procedural content generation. So these are computer algorithms that create scenes or levels for computer games. But in recent years, I've been getting interested in how do you use these algorithms to create data for testing autonomous vehicles? And the insight there was, if you can have an algorithm create a synthetic city for use inside a computer game? Well, it should be possible to make synthetic streets and synthetic intersections for testing autonomous vehicles inside a game like environment. So we've done that with a project called Junction Art where we make moderately realistic looking intersections. We've also been recently working on creating models of synthetic people to test autonomous vehicles.

[00:04:51.200] - Jim Whitehead

And it's kind of a tricky problem and kind of a funny one in that the idea is for the pedestrians to act safe until sort of the last minute and then kind of Dart in front of the car to make it as difficult as possible. So it's like synthetic pedestrian with a death wish? Not really, because if they just jumped in front of the car, that wouldn't be a good test scenario because then nobody can recover from that. So the car wouldn't have to do that. So it's got to be like close to being impossible, but not really. And so we're aiming for that little sweet spot there. So I have two PhD students working on that in two different ways as well. And then also I've been very interested in natural language processing for software testing. How can you use what they call large language models to take advantage of the knowledge that's hidden away inside there and kind of tickle them in some way the output testing information that you can use? So I'll stop there. But anyway, I'm really glad you're all here today and glad you came out.

[00:05:57.810] - David Urbanic

So I'm not a former NASA employee or a professor and I've kind of come to this from a different I'm looking for a common thread. I got a degree in public policy from Stanford long back and worked for some politicians locally. And you can look it up. It was just a crazy house because politics isn't really where you want to make your money. I don't think I got into tech working for Boreland, did tech support because my daughter was young and it was an eight to five job, ended up taking about 80 or 90,000 customer service tech support calls across a bunch of different products and went into some different startups. Ar VR with a John Scully startup, built the first customer service team at Netflix way back when they were shipping DVDs, databases for UC system imaging. I did build some software that NASA uses and through it all I was always amazed by the tech I got to ride the software wave and the Rad software wave and the web commerce wave and the Internet wave, et cetera. But the takeaway from me was always that there's all this power and data and the more we have of each, the harder it gets to use because your menu trees get longer and suddenly everybody's got to be an expert to just use consumer software, right?

[00:07:30.510] - David Urbanic

It wouldn't it be nice if we could have the systems work our way instead of vice versa? And that's what led me to intelligent user interfaces and then to AI, and then from AI to conversational AI. And the idea that it would be nice if you could speak to a computer like a friend and say, I really want to dim the lights, or I really want to filter this image, or I really want and let it figure out which features to use instead of you having to become an expert on every different piece of technology you want to use. And at safetynex, that's what we do. We don't just produce a one size fits all solution that can speak a little Shakespeare and pick a song for you. We tailor systems to meet specific needs and specific use cases and very high accuracy. So if you want to do something really well and you want to do it without a lot of training, safety index can do that, and that's what we do.

[00:08:47.450] - Craig Vachon

So what's the end goal of your professions? I mean, when do you decide we're a success? That's it.

[00:09:04.950] - Bethan Hockey

For language. We've been working on this for a while, and when AI started out back in the dawn of AI time, we thought language will be easy, and chess, that'll be hard. But it turns out that you can do chess with just brute force algorithms. And the brute force thing hasn't really panned out quite as well for language, because it turns out language is really complicated, and all of us are really good at it, and we get really good at it after not that long, because there's something special about us that we're not quite figuring out enough to program it into a computer. But we're getting better. So I think success to me, success would be if we could make systems that actually converse usefully with people that gave value to people for helping them with things and were more natural. I agree. I don't think that the people should have to learn how the system works. Ideally, the system should just work the way you expect it to. And when you talk to something, the way you expect it to work is like the way it is when you talk to another person.

[00:10:22.420] - Craig Vachon

But Bethany, that kind of sort of begs the question of like half the time my wife and I are talking to each other and we don't understand each other.

[00:10:30.590] - Bethan Hockey

But in human conversations, over 60% of the conversation, even when the conversation is focused on a task, is devoted to keeping that conversation in sync. And this is not something that we currently do in our artificial dialogue systems and our systems on machines. There's a whole bunch of things. Entrainment is a word we use for when you sort of settle on the same terms for things with each other. There's grounding that's when you say something another person says. Now you think that you're talking about the same thing, and you might be right. Even right. When was the last time you talked to a computer system that did either of those things? Nobody. Right. So that is not something that you're seeing commercially. And that would certainly be a big improvement because then you could recover from mistakes. Right. So when the system goes wrong now, it says, I didn't understand you, please try again, or it shunts you off to an agent or something. Shunting you off to an agent isn't so bad, because then at least you can talk to a person who can actually do conversation really well. But no success would be being able to incorporate some of these things that people actually do so that conversations with machines could be more flexible.

[00:12:02.370] - Bethan Hockey

You could recover from mistakes. It would just be easier. And you wouldn't have to train yourself how to talk to this thing.

[00:12:11.040] - Craig Vachon

You would just talk will we ever be at a point that we are unaware that we're talking to a computer?

[00:12:19.730] - Bethan Hockey

It's already happened. It's already happened. There was a Google system. And if you want me to, I'll comment on what I think about doing this particular thing. But what happened was they built a system and they handcoded it really carefully and they sent it off to call beauty salons and make appointments. And you might have heard about this in the news, and they did not announce that they were a machine and they fooled a lot of people. And you can do that if you carefully, carefully engineer it by hand, you can do that.

[00:12:58.590] - Craig Vachon

And right now we can do that in limited scopes and really narrow.

[00:13:03.100] - Bethan Hockey

Yes.

[00:13:03.770] - Craig Vachon

Okay.

[00:13:04.380] - Bethan Hockey

Basically, the narrower you go, the more detail you can put in.

[00:13:08.770] - Craig Vachon

Yeah, Jim, success looks like right.

[00:13:13.870] - Jim Whitehead

Well, success for a professor always is like the next grant or the next student. But for this testing of autonomous vehicles, I think one of the things that really motivates me is so much of this work is being done in a very small number of cities, mostly in the US. And it's a big world, and there's a lot of really funky roads out in the world and a lot of ways of pedestrians interact with roads in ways that we would just find mind boggling in the US. I have two students from Bangladesh.

[00:13:51.750] - Craig Vachon

I lived in India for two years. I challenge anyone to understand that road system.

[00:14:02.250] - Jim Whitehead

Another student of mine is from India. And yes, they're just always like virtual cows wandering the roads and people like taking shortcuts across essentially an interstate, because otherwise it would be an extra five minutes on your walk every morning. There's things that happen in these other places that don't happen here. And I can guarantee that the current generation of vehicles is in training for that. And even here, there's some very weird things that are happening with Tesla vehicles. So one example is there is a Boulder in Yosemite that loves Tesla's, and so there's like a road as it goes into the park. I'm sure you may recall it. So the road kind of goes and then one Lane goes straight and the other Lane kind of takes a turn. And if you're in a Tesla and that left Lane, turn off the automatic and grab the wheel because it gets confused. And then there's a Boulder right there. And no less than three Teslas have kissed that Boulder. So it's just like it's a weird situation. It's this long tail. And so we're just hoping to find ways to automatically explore that long tail and also automatically kind of generate situations outside of those in the US, so that when this autonomy gets ready and when it gets outside the US, it isn't a big, rich city kind of thing.

[00:15:33.710] - David Urbanic

Well, Bethany mentioned entrainment and grounding. I would add a word, a term centrality, when I find that the system is focused on me instead of me being focused on the system. And we talked about that a little bit when I can say, computer, I need a fish tank for a whale, and it just starts thinking about how to do that. I think that we've arrived, although I shouldn't be talking to a mouse at that point, but maybe that's a reference that's too old for some people.

[00:16:14.490] - Jim Whitehead

I'm happily dating myself with it.

[00:16:18.510] - David Urbanic

Okay, the nerds unite. So I think it's not just about being at the center, but having the experience be less about data and more personal. And that has to involve context. And so when you can trust the system to protect your profile and maintain some information over time, suddenly everything you do is more valuable. And I experienced this a little bit with recommendation engines and your newsfeed, and it used to always just be annoying. It's like I'm getting a packaged version specific. For example, if you want to have a good day, the first thing you do when you start up your system is search for something that you find entertaining, because then every graphic in any website you go to is going to be about that. I don't know if it's surfing or swimsuits, whatever, it's just amateurish. So I've seen that improve a little bit, but really my interest is in language. And when something can speak to me and maintain the context and understand what I'm talking about, that's I think when we've arrived.

[00:17:36.130] - Craig Vachon

So I really love the discussion around context because I think AI's largest weakness is its inability to understand context. So that I'm open about this. I am the CEO of a company that's focused on bringing context to AI. Start with name is AI Redefined. With that said, today's AI is enormously narrow. We have Facebook engineers creating AI for engagement, and that is enormously successful. They're seeing 30 and 40% higher engagement rates, which turns into higher advertising revenue. So they're happy from a shareholder's perspective, but that AI has done so by creating polarization and amplification of hate, and arguably has made living as a human being in various parts of the world, including here, more difficult understand context and what are you doing about it?

[00:19:13.170] - Bethan Hockey

Okay. One thing is we have to be clear about what we mean by AI. We've gotten into a world now where people think that machine learning is exactly identical. In particular, deep learning. That was the best branding ever, calling it that. Machine learning is a technique, and there are lots of kinds of machine learning. So when will AI get it right? Ai has been going on for a long time with some machine learning in it, but not lots of other things. And people do AI. So people are designing what that AI does, and people are thinking about what it's useful to use AI for, and people are deciding which techniques to use for doing the AI. I've built systems that were hybrids of machine learning and rule based systems, things that did inference, and it did use context. We did use context. We saved in terms of a conversation like, you have a better conversation if I remember what you said 5610 turns ago. Now, there's a lot of systems out there that don't remember what you said on the last turn. That's not so good. But the systems that I was building, even the one we sent to the space station way back in 2005, kept all of the dialogue history, and we referred back to that dialogue history as the context of that conversation so that we could tell that if you said it, we knew what you meant by it, or if you said that last one, we knew what that was, too.

[00:21:07.710] - Bethan Hockey

So I think the question of context is not can AI do it? But do people want to do it? If you're getting some commercial value about doing a kind of surfacy classification kind of AI, and that's working for you, you're not going to worry about the context so much.

[00:21:27.610] - Craig Vachon

We're going to have a great discussion tonight.

[00:21:40.330] - Jim Whitehead

To some extent, I agree with Beth in that you look at today's autonomous vehicles. They actually are doing a pretty decent job of looking at pedestrians and their surroundings. They generally see them. They generally do a pretty decent job. So it's not like they're context free. And there is a great deal of situational awareness for these vehicles right now. So for some systems, maybe it's a big matter of scope, like how much context, how much more is necessary. But I think there's other areas where there's this whole area called robot process automation, which is about kind of making kind of scripting languages for automating things that you see on the web. And by and large, they don't have much context about what they're doing at all or much context about web pages, decision tree type. Yeah, they're not very smart and they don't have any real understanding of the pages. So they're kind of doing these very direct pointers into the content and then everyone's kind of shocked when it's like they're brittle and they break and they can't recover easily or an organization makes a change and then they can't react. I think there's an example of an entire business segment worth billions of dollars where there's very little context and a lot more to be done.

[00:23:09.530] - David Urbanic

Additional inputs, sentiment, and additional aspects to context, like an ontology. If we can not just keep track of what you purchased last or what you said last, but also how you're saying it and how that's changed over time, then we can respond to you quickly if you're speaking quickly or politely if you're speaking angrily. A lot of this comes from my background teaching customer service agents. When the person says, I've had it with you people, I just want to cancel my account. You don't say, what's your problem? They don't have a problem, you have a problem. Stop. But also the ontology idea, the idea that there are layers of meaning. And if you ask for an agent, you're asking for a person and a person is a mammal and that's an object and it's not a sports licenser broker. I keep coming up and meaning because I'm really in the language world and Sapiex is very much in the language world, but I used to be in imaging and the idea of the differences in radiology versus art restoration. And what does a spec mean? The more we understand about or the more the system is recognizing about the world it's dealing with, the better.

[00:24:42.840] - David Urbanic

And yeah, we do that and we can do that. And it's not questioned that we can do that, but do we really bring it all home? I think we need a company that's focused on that.

[00:24:56.210] - Craig Vachon

I love the discussion of ontologies because one of the investments that I've made is a company called Easyap and Easyap is focused on pharmaceutical ontologies, natural language generation company. They have their claim to famous. They helped pharmaceutical companies. The large ones shortened the time from data lock on phase three human testing to FDA applications from 28 weeks to a single week. And that didn't generate any more revenue for the pharmaceutical company. But it sure brought it a whole lot faster. And this was highly successful for moving COVID through the FDA approval process significantly faster.

[00:25:46.090] - David Urbanic

Covid or the vaccines.

[00:25:47.550] - Craig Vachon

Excuse me, the vaccines. Thank you. Yes. Did I also mention I'm a spy novelist? You just threw me a great idea. Exactly. Yeah. Ontologies are kind of like a lot of the big players in NLP and in AI in general are looking at people to add on ontologies on top of GPT-3, for example, for specific instances, let's have a pharmaceutical layer that understands medical writers, jargon that understands medical doctor's data input and then can create a narrative on that data. Your thoughts, Bethany, were if it's working, it's good enough. No, I'm sorry.

[00:27:03.870] - Bethan Hockey

My thought was that a lot of commercial organizations decide that if it's working, it's good enough.

[00:27:10.130] - Craig Vachon

Okay, so what's success? Because if I were to ask, what was your question? It was a great question. I want a fish tank for a whale. I want a fish tank for a whale. At what point is a chat bot going to be sophisticated enough to either realize that as a joke or realize that that is something that they can service? Well, today I'd argue they probably can't.

[00:27:49.410] - Jim Whitehead

So I'll take the counter here and say, so maybe a chatbot can't. But one of my favorite. I can't believe this actually is a thing. It isn't magic. Ai Systems is one called Dal. Edalle. And just very recently, version two came out with Dolly. You just give it something like a fish tank for a whale. That would be a perfectly excellent Dolly prompt. And it goes in and it will generate an image for you that I'm going to guess will probably look an awful lot like a whale and a fish tank.

[00:28:26.580] - Craig Vachon

There's a great Reddit section on Dolly and what it can do. I think Mark Adams sent us all sort of show me two wooden mannequins taking a selfie was the question asked of Dolly. And it was two artists wooden mannequins taking a selfie. It was brilliant.

[00:28:51.810] - Jim Whitehead

You say, give me a chair that looks like an avocado, and Lo and behold, there's a chair that looks like an avocado. It is really remarkable. And it just type in search from Dolly to the top web page will come up and there's all sorts of examples there. So I guess I would say there's an example where like, man, you can put in almost anything and it doesn't have context, but it's sort of like it's one of those fakes until it makes it. And it's pretty good a lot of the time. So I don't know now, I guess previous conversations, Beth, you were saying like, yes, but the remaining 5% is the difference between something that's fun and something you might want to put in front of customers. So.

[00:29:32.290] - Bethan Hockey

I think there's a difference here. There are systems that do. In particular, the deep learning algorithms are really good at kind of surface the sort of perceptual level features. So they're good at doing things that are like vision or like hearing. They're not as good at doing things that are like reasoning. One of the reasons I think that that's true is because a lot of the deep learning algorithms only handle sequences or grids. So a sequence might be a string of words that you say, a grid might be the pixels in a picture. Right. And they can do some things about that. But if you get into worrying about what the words mean and how they relate to each other, now you get into a lot harder situation because that requires some structure that a model that handles only sequences or grids can't really do. So then you have to do something different. Right. But some of these tasks where handling a grid or handling a sequence is all you really need, then those will work well. These models are really great for doing automatic speech recognition. This is sort of like hearing it takes an acoustic signal and it turns it into its best guess of what the words are.

[00:30:58.870] - Bethan Hockey

And those things have gotten a lot better using these kinds of models. And all of the vision tasks have gotten an awful lot better using these kind of models. But I think the sort of deep reasoning tasks have not gotten as much better or not better at all using these models.

[00:31:17.650] - Craig Vachon

I spend some of my day I'm a partner in a Paris private equity firm called Next Stage, and explaining to some of the nontechnical partners there that AI can be trained to spot cancer cells in a pancreas, but move it 3 CM over to a liver and you have to retrain the entire model. Is that a goal? Should we be highly specialized? Should we be looking for? I'm not talking about AGI artificial general intelligence yet. Not yet. But when do we take the next step so that the AI is smart enough to know this is a liver versus this is a pancreas, and I might need both, especially as patients have cancer that travels between the two. Thoughts?

[00:32:23.850] - Bethan Hockey

Well, I think that you end up with a hierarchical structure. Right. I think even with people if you're talking to a person about some topic and you suddenly change topics, that person doesn't actually do that well, understanding what you're talking about. If you just change topics suddenly without warning, if you've got a radiologist that thinks they're looking at a pancreas and you slip a liver in there like it might take them a little processing to figure out what's going on. Right. So I don't think there's anything wrong with collecting up specialists and having some sort of a hierarchical structure. So maybe you have a specialist that does pancreas and you have a specialist that does livers, and you have another specialist whose job is to decide whether we're looking at pancreas or livers.

[00:33:16.450] - Jim Whitehead

Although one of the lessons that seems to be coming out in the last few years with these large language models is bigger just really actually is better. And so in this particular case, if you were able to get like an incredibly large data set, maybe 100 million images of cancer and non cancerous livers and pancreas and lungs and whatnot, and then put this into an increasingly large model with millions, tens of millions, hundreds of millions, billions of variables, it does seem like if you have skillful people doing that work, bigger does result in better results and you get kind of more generalizability across it all. And so there's definitely a lot of concern right now that who can afford bigger is better. That Dolly paper that I love, we did back of the envelope estimate of, like, how much money are we looking at when we read this paper? And it's like, it's probably $30 million for one paper.

[00:34:23.470] - Craig Vachon

Ai who did Dolly was funded with $4 billion.

[00:34:27.280] - Jim Whitehead

Yeah, startup $4 billion. And I think there were like 15 some odd people on the authors list, each of whom are probably making 400K plus plus probably $10 million of computer time, plus. Then there's all the other people around them, like making the company work. So for just one paper. So, yeah, that's a game universities cannot play right now. So bigger is better, and the winners are people with the pockets. That kind of speaks to a certain kind of accumulation of power and wealth and capability in the hands of people who can pull together that degree of capital.

[00:35:07.510] - David Urbanic

Two quick thoughts. It seems to me that you can have a society of mine, a community of hierarchy of specialists. You might also entertain the possibility of a meta system that can identify what's happening in different circumstances and generate some value in that way. But it's not really my area. The other thought is I'd like to ask Dolly to draw something like colorless green ideas, sleeping furiously, and just watch it go nuts.

[00:35:53.510] - Bethan Hockey

So I want to go back to bigger is better. That's true. Like, bigger usually is better. But the conferences that I go to have started having sessions where people talk about ethical AI. And one of the things that we talk about is whether you should keep making the models bigger and bigger and bigger at the cost of those size of CPUs and that amount of electricity and all of the cost that goes into doing that.

[00:36:21.280] - Craig Vachon

Which is a perfect segue. Thank you. Ai is a human construct to help humans. The humans that are creating AI, it seems to me, have a hard time explaining AI. They're not sure, especially with some of these really large models, how AI is making its correlations. And by the way, AI makes correlations, not causations. Right? We can all agree on that.

[00:37:08.300] - David Urbanic

Okay.

[00:37:09.490] - Craig Vachon

So is explainability an important aspect. I'll tell you, some of my customers at AI Redefined, they're large engineering companies doing preventative maintenance for very large water works and saying if this water main goes bad, it's going to cause this much pain, this much resource. Their biggest challenge isn't should we use AI? They love to use AI. Their biggest challenge is the insurance company going to reimburse them for making this decision without being able to explain how that decision was made.

[00:37:57.850] - David Urbanic

One quick related thought is that it's kind of a rigged game, because as soon as something is explainable, the verse excludes it from AI. It's like, well, it's just engineering that happens.

[00:38:16.510] - Bethan Hockey

The other thing that happens is that when you get something to work in AI, it gets its own name. That's why you have machine learning and you have computer vision and you have reinforcement learning. Yeah. As soon as it kind of works, then it gets its own name. Right.

[00:38:34.190] - Craig Vachon

Jim? Is AI explainable? Will it ever be?

[00:38:41.210] - Jim Whitehead

What do you mean by an explanation, Craig?

[00:38:44.930] - Craig Vachon

We practice it.

[00:38:51.090] - Jim Whitehead

So lack of explainability in and of itself is not necessarily a problem. You go out into the fields next to UC Santa Cruz. I cannot explain to you why that plant is exactly there and that Bush is exactly there and why that bug is exactly there. It's a big system. There's a lot of emergent capabilities in that system. We can sort of explain it as a whole. Is like that's grasslands and coastal Sage scrub ecosystem. So there's an explanation but doesn't get down to the weeds. And so with AI, like all of these AI architectures, there are papers that explain what the network architectures look like, and there are even explanations for why they think they're working the way that they do. And there's a lot of work on visualizations to kind of give us some insight as to what's going on. So I think it's not quite true to say that there is no explanation at all. And so maybe it's more like, well, what explanations do we need? How much explanation do people need for decision making when working with that AI in some particular context? So those are, I think, some of the kind of interesting things with autonomous vehicles.

[00:40:12.310] - Jim Whitehead

There's like this huge question right now of how do you give some sort of visibility to pedestrians of what the car is thinking about what to do next in a real car, you look at the driver if they're looking this way or that way, that tells you something, but you don't have that ability in the car. So there's a very simple level of explained ability that would be very helpful.

[00:40:38.170] - Bethan Hockey

Sure. As Jim said, there's definitely people working on getting more explainability. The deep learning models, especially, are pretty opaque, and you can't tell exactly what's going on. I mean, people do have ways of probing them. So that's how you discover that your model that's supposed to detect polar bears doesn't detect them on green grass in the summertime, and anything else that's on white background is a polar bear. You get things like that happening. You also have the same thing that has always happened with statistical systems where garbage in, garbage out applies. So if you say, why did this model go wrong? I mean, a kind of not very helpful maybe, but correct answer is the data did it, but when the data is so big, like, can you pinpoint what did it? Maybe not.

[00:41:45.050] - Craig Vachon

All right, last question before we ask the audience for their questions, as technologists as AI practitioners, are we responsible for the actions of the AI we create so Interestingly, Tesla autonomous vehicle has taken one route. The answer is fuck no.

[00:42:19.210] - Jim Whitehead

Just that beta on it.

[00:42:21.490] - Craig Vachon

And MercedesBenz has said, well, yes, as we chatbots have been, there's lots of stories of natural language processes that have been thwarted and taught racially tinged language or misogynistic language. If you have Mal intent, you can really mess with these things to be further Mal intent. Right. We as technologists building AI, are we responsible for the results?

[00:43:04.410] - David Urbanic

I would say yes to the extent that tool creators generally are. So you build a hammer, somebody hits somebody over the head, it's not the intended use. But if your system is excluding people from lending opportunities that humans wouldn't exclude them from, and you've contracted to provide a system that would do what's expected? I think so, yes. I think it's about setting expectations and delivering.

[00:43:41.650] - Bethan Hockey

Sometimes those lending systems exclude people from lending opportunities that humans would have also excluded from lending opportunities. Right, sure. So the reason that the system is doing it is because when you're doing systems trained on data, it's always about the past. The data is the past. It's not aspirational so if in the past a particular group was discriminated against and you dump in that data, your system is going to do that same discrimination. And so it's not just did we design it to do that, we have to be aware of what the data is that's going in. And so we need probably way better tools to be able to probe what the data has in it, because otherwise you're just perpetuating parts of the path that you maybe don't really want to keep doing.

[00:44:40.470] - Jim Whitehead

Yeah, I'll certainly second that. If you ask this Dolly Two system, give me pictures of a lawyer sitting at a desk. They are all male and they're all white. You ask it to give you pictures of flight attendants. They're all female and like half Asian as well, because I guess Singapore Airlines, I think so. It's all coming from the data and it's all bias in the data. Now, the researchers kind of pointed this out. The researchers provided these examples, but there are some debates ongoing, which is like, is it sufficient to just say, hey, this is a thing, and now we're working on Dolly Three, or is it like, is the responsibility, like, hey, this is a thing, and we should really solve this thing before we move on to the next version? There's not really any consensus on that right now. And I think you can make arguments both ways. And one argument would say maybe if you kind of do the next version of the system, you trained it in a slightly different way, maybe with just some differences in training, like this goes away along with improvements and everything else. The other one says, no, people are going to start making use of this right away.

[00:46:01.000] - Jim Whitehead

This will get used in all sorts of unintended ways. You really want to solve these right away and be able to have techniques for addressing this bias. I can see it both ways. It's such an early stage technology. I think AI is just one example of a wide range of early stage technologies. How long do you let a technology go and develop and kind of explore? Before you really say, like, no, you need to deal with all these problems that are being created by it? How long of a leash do you let it go before you say you have to solve that? And I don't think there's societally a well understood answer to that. 100 years too long, one year too short. Somewhere in between.

[00:46:47.490] - Bethan Hockey

The company that I work for, Live Person, we actually have signed on to fairness and NLP efforts, and we actually worry about this. We put out a platform that our customers who are big brands use, and they don't want to have horrible bias or mishaps, and we don't want them to have that. So we worry about bias. We're presenting a paper at a conference on looking at whether when you mask personal information, whether we can perform equally well on various kinds of names. And we found out we don't we don't perform as well on certain kinds of ethnic names. And so we're working to make the model correct that. And this is one of the reasons that I work at Live Person is because I feel that way about it, too. And Live Person is a company that puts fairness in AI and responsibility in AI pretty up front and center.

[00:47:47.970] - Craig Vachon

Good. All right. Questions from the audience. Yes.

[00:48:01.630] - Attendee 1

You gave us an example of how they go to Net. Engineers are finding that hateful content drives engagement. I just wanted to ask you to talk a little bit more about the connection between bad outcome and context. For example, you might have meant that context can change the kinds of predictions that a model will make. You might have also had in mind that context will change how we would evaluate predictions, and you might have something else in mind.

[00:48:43.380] - Craig Vachon

More the latter. To be blunt, I think the challenge that I have with Meta's approach to their engagement algorithms is this discussion we had at the very end, which was our engagement algorithms did exactly what we programmed them to do, which was increased engagement. The fact that the unintended consequence was that they were promoting hate speech, amplification of misogyny and the like was the unintended consequence that the AI wasn't smart enough to realize it was doing. And it likens me back to the 1982 McKinsey Consulting was approached by some Midwestern farmers, and they were complaining about weeds in their irrigation pond, and these weeds were causing the pumps to foul. And Mackenzie took a very narrow approach and said, we can solve this with the importation of the Asian carp. They imported the Asian carp into the US into these irrigation ponds. It solved the fucking problem perfectly. There was not a damn weed left in those irrigation ponds. And, oh, by the way, it took 24 plants and animals to extinction. And today we spend north of $13 billion a year trying to eradicate the Midwest, because every time it floods, it floods outside of the irrigation ponds.

[00:50:25.210] - Craig Vachon

That's the unintended consequence that I worry about with AI, because it's so bloody and arrow and so effective and so damn effective that you have this AI, this unintended consequence technology.

[00:50:44.930] - Attendee 2

Being advertised. There's an application called Replica, which is being advertised currently, especially on applications like Tik Tok where it's advertised as being a thing replacement for replacement. What is your moral or even darkly curious opinion on AI being used? Slowly?

[00:51:39.110] - Craig Vachon

I'm not familiar with this company, but I am familiar with in Japan, because of the aging population, they've created a series of AIS that help people with loneliness as they age. So it's a great question.

[00:52:00.810] - David Urbanic

Eliza and Pet rocks. So anthropomorphizing anything can become good company. And it's one thing to apply that ethically to provide a companion. I think it's another thing to take advantage of that, to distort social development. So what are we going to do about it? That's a little more difficult but bad.

[00:52:35.730] - Bethan Hockey

I think one thing we can do about it is better education. So I think something like this is great because I think everybody should understand at some level how all this stuff works and what it can and can't do and what you might want it to do and what you might not want it to do. And I think that people are going to build things and people are going to put out disinformation. And your best protection against those things is knowing how to think about it and understanding a little bit about how it works.

[00:53:08.710] - David Urbanic

I totally agree. You know, what works against that is when we have a bunch of hype where people say things reputable people like, what was it? Was it GPT that might already be a little bit conscious? That's a good reaction when we coopt words and create smoke instead of educating it's just worsening the situation.

[00:53:33.150] - Bethan Hockey

Okay. But I've had kids say to me, that sounds like nonsense when they hear stuff like that. So we need more kids who hear that stuff and say, that sounds like nonsense. I'm going to look into that a little more.

[00:53:58.870] - Attendee 2

Situation. It actually shows you when having conversation. It has its own system set up to keep its own database of answers to questions.

[00:54:35.450] - Craig Vachon

I know this really shitty novel. It's called The Knuckle, head of Silicon Valley, where they use this weapon of mass persuasion, and it uses a very similar technique, actually.

[00:54:49.510] - Bethan Hockey

I hate to break this to you, but personalization, which is a really hot topic right now, works pretty much exactly like you just described. You keep a history of what you know about people, and you use that in the conversation. Well, that is an issue. I think we aren't doing that except in situations where people positively opt in.

[00:55:20.770] 

Female.

[00:55:21.630] - Attendee 3

We use machine learning for failure prediction. And the higher the stakes in the game, not only monetary, but personal damages, insurance or death in making decisions, people going to jail or not doesn't have to be only technical. I see a huge burden and the need for a process to establish a quality metric or a way how to be able to send it the first time I'm right. Well, when you read 90%, how do you ensure that the decision recommendation expressed is true comes back to the liability question. I think there's a huge open field. I can't even literally easily say, oh, it's just a matter of.

[00:56:28.350] - Craig Vachon

Whether it's a new friend or whether it's an inverter in a solar field. Yeah.

[00:56:35.550] - Bethan Hockey

Evaluation metrics are actually a really hard problem. Right. Because you have to explainability issues. Well, it comes to explainability and it comes to what counts as success, and it comes to what features you should consider in evaluating it. So that's actually an open research question is how to do evaluation, certainly in natural language processing and language AI. It's an open question, and there are a lot of research going on in that. And there are papers and conferences on that. Every conference.

[00:57:10.830] - Jim Whitehead

Yeah. I was just going to Echo that kind of how good is good enough for even a traditional machine learning based system? Yeah, it's really context specific. Sometimes getting a prediction 90%. Right. When humans can only do it 60%. Well, like, that seems magical. Right. And then there's times when 90% is like, well, that's horrible. Humans get it like one in 10,000 error kind of thing. A lot of it depends on you just have to kind of know the use situation. And a lot of it is just by talking to people who are users and asking them what do they think of the results? And there's kind of sometimes a trade off between, like, is it better to be very right but much less of the time, or is it important to be, like, just kind of flag things more often? But you're right less of the time. And so there's often a trade off in systems with that as well.

[00:58:13.950] - David Urbanic

Exactly.

[00:58:15.330] - Bethan Hockey

Yeah. You have to know whether you want accuracy or whether you want coverage or whether you have to have both.

[00:58:21.150] - David Urbanic

But also part of its discovery and part of its definition. I mean, if you're going into a use case and you're providing a solution, at the very least you should be specifying what you consider acceptable and how you're going to measure it, because then you can be held accountable at least to that standard, and then you may have to evolve that over time.

[00:58:52.110] 

Up.

[00:58:59.590] - Jim Whitehead

Yeah.

[00:59:00.470] - Craig Vachon

One of the investments I have is a company called Humanity Health, and they're using AI to promote longevity this question of like, they get asked it all the time because it's an expensive monthly subscription. They're like, well, how do we know whether or not we've actually exactly humanity health? But by the way, all of these companies that I'm discussing are hiring. If you're interested in jobs in these spaces, please talk to me.

[01:00:00.410] - Attendee 5

It's machine learning, it's rulebased transparency. But every client I have thinks this magical AI is going to go do some deductive reasoning on a very complex problem and come up with an answer that for a couple of million Bucks, you will get there.

[01:00:22.440] - Bethan Hockey

When will you come clean? I don't promise that.

[01:00:30.690] - David Urbanic

I don't think the Emperor ever came clean. I think somebody in the audience points at them and says, naked.

[01:00:39.550] - Jim Whitehead

I guess I'm torn in that. The hype bubble around AI is pretty extreme right now. And there's a lot of people kind of spreading AI Pixie dust over things that are like five if statements and call it a day. That said, man, the advances in our ability to do object detection, the ability to do classification of things and images, these large language models, as opposed to the AI winter, which I feel like was kind of justified. There has really actually been some tangible progress on some very hard problems recently. And so I think at the end of the day, there has been a pretty big shift in capability over the last ten years, even five years. But extracting value out of that, putting that into a particular context, figuring out whether you want something that's super sophisticated and recent or whether you just want something older, but it's more reliable and explainable and you can engineer it better. There's still a lot of craft and like, extracting the value and doing something real with all of this stuff.

[01:01:58.370] - Attendee 5

How do you actually come up with enough.

[01:02:07.590] - Craig Vachon

So the way we do it with satellite inspection, the reason we use synthetic data is you don't put up hundreds or thousands of satellites. Right. And so you can create synthetic data that will encompass 70 or 75% of all of the possible scenarios that satellites might encompass. But then you actually have to add humans who add context to that environment that can think creatively beyond machine learning. And then you have to have a system that allows for the plumbing for humans to interact with that model.

[01:03:17.930] - Attendee 5

Of people that are smart enough to understand this conversation today.

[01:03:22.300] - Attendee 4

So can I have a question about that? It sounds like you almost really want a code in nuance. Right. So when I drive my car over 17, I come to a curve. I typically will pull into the inside of the curve. Right. Or if I'm driving next to a big truck, I tend to approach to the other side. My car does not do that when it's in autonomous mode.

[01:03:45.470] 

Right.

[01:03:45.840] - Attendee 4

I get to a curve. And if my car is in autonomous mode, all of a sudden I feel like I'm going to run into the guy that's coming ahead of me. So how are you able to craft in nuance to be able to have autonomy, make it more humanlike?

[01:04:03.050] - Craig Vachon

I think you have to add humans, right? Exactly. If you look at Stanford University's AI 100 it's a five year study of the AI industry. Their final paragraph of their summary statement this 186 page document says that we as a species have only progressed when we combined efforts when we contribute as a species, as a race together and that AI needs to do just that as well. That AI's journey is not to replace humans but augment humans to make our lives, human lives more beneficial, more impactful. If you want to read something about how to add their specifications are you got to add humans back into this because it's the only way to get to beyond synthetic data.

[01:05:16.860] - David Urbanic

Beyond this nuance it's seemingly the only approach and to that point I've seen a lot of the service and support solutions lean in on agent assist and a lot of the natural reasoning solutions lean in on collaboration.

[01:05:40.090] - Jim Whitehead

I'll be a little contrarian here and say I love adding the human element in I've studied all sorts of symbolic AI systems. The lesson that just keeps biting me over and over again is more data does just make things better. So for your specific situation of how to tell the car to be more on the inside I think that's just like feeding a bunch more trajectories specifically from highway 17. So that sees enough cars that are going on the inside of the curve and it'll learn it and you make the model bigger and it'll dedicate 20 variables to like the Laurel curve and highway 17 and it'll just be in there, right. So bigger models and more data does solve a lot of these problems but it hasn't solved it with the rock and Yosemite yet. This gets back to that very big piece of data. How much is enough? So I don't know. I think there's an information theory answer here that there's probably something information theoretically interesting about that compared to their data set by sorry, we really do have to wind this up but these guys are available to talk to you off stage here but I know Coomb was saying you know, 08:00 I thought but please mix it up in the audience and I want to just thank everybody for coming out tonight and remember stay safe.

[01:07:27.010] - Jim Whitehead

Thank you.

[01:07:36.930] - David Urbanic

Well done. Yeah.