Janell Shane, it's really great to talk to you.
I I reviewed and looked at a lot of books over the past couple of years and yours, your book you look like a thing and I love you Which went on sale last year and I bought my local bookstore was great.
And it's not the only thing that, you know, does it.
There's also I think a lot of people are familiar with your AI weirdness blog, which we've written a ton of stories about at CNET, which is a really wild repository of like.
Bizarre seeming like bizarre comedy posts, but they're actually from AI projects and generated in all sorts of interesting ways.
The book is more of like a an introductory primer to AI which I thought was fantastic with a lot of your sense of humor and illustrations, but also really pointing out a lot of the limits and possibilities.
But thanks for Talking with me, I guess, I was curious how you got involved in writing about weird AI?
Yeah, well, I got involved, kind of just through experimenting with funny generated stuff.
So I think the first thing I generated was recipes cuz I'd seen somebody else have one of these neural nets so t was imitating recipes and [INAUDIBLE]
To hilarious, exactly my kind of humor as it turned out there's something about having a recipe and it filtered back through this thing that doesn't understand them, but it's imitating parts of them.
And then, okay, well, why don't we add peeled rosemary and treaded bourbon I mean.
So that is how I got into it is just kinda these.
I'm going to generate more silly recipes and then kind of digging into a why is it making these particular mistakes and then people would ask me to comment on You know, publications that have just come out or research results.
Well, I guess I better read this journal article then so I can know what people are talking about.
And yeah, kind of.
The questions kept coming up, people had questions and it seemed like, not just people visiting my blog.
But lots of people across all different swaths of society.
We're having these pretty fundamental questions about what do we even mean when we say AI?
I was curious about your thoughts on that and the role of people in that like I got the sense reading this and looking at the way you you curate and develop a lot of these that, these lists start out as bizarre and comprehensible.
They kind of grow like these plants, and then you curate Pieces of them and present them.
So it seems like a combination of a, an AI process and there is also a human curation of that that works in symbiosis.
So I was curious about like that process for you [INAUDIBLE]
Yeah, they definitely curation is an important part of presenting neural net generated stuff is interesting.
So and I've seen plenty of blog posts and things out there where people will say, well here's a list of like 200 things that this neural net generated.
And it's harder to read.
You really need Some kind of structure to that list or you need them presented in some order or with some kind of, explanation of what it is we're seeing or what did I the curator find interesting about this that I want to then share like, what's the point of me sharing this with people?
And tied in with that curation of the results, of course is also just like What do I want to actually try to train a neural net to generate and what would make an interesting subject?
What and what might the results be?
And even now, especially that we have some Choice as to, you know, the level of sophistication of the algorithm as well.
Do we always want to throw the most capable neural net at it or are we looking for something that's a little bit glitchy here?.
A little bit more hate to think of something from 2016 as retro [LAUGH] But in a world of neural nets that's, that's where we are and I've had people say, We remember back when these things would spell fart all the time because they didn't know that that wasn't the [LAUGH] Yeah, they didn't know that that was a word that you should avoid and i don't really get that anymore in the neuromaths that I've been doing in the last year or so.
And so there's already [LAUGH] You can think of these different Levels of competency and you know how long you're going to train these things for and are you going to mix together different data sets will that give you a kind of interesting result?
And you can almost kind of liken it to The effect of photography on painting, and, looking at Impressionism, and Cubism, and different ways of going for a non realistic result is especially what technology is capable of producing a pretty realistic mirror.
Yeah, it feels like for the everyday person who's not who may be thinking about AI either in terms of something that could become, you know, self conscious or something that would be that would take over people's roles and things in that science fiction way.
The other part that I feel like people see a lot are filters.
You know, in the sense that Bring up that photography element, things that people may or may not realize are AI playing with all these transformational filters on camera apps or other apps on their phone.
I definitely see that there.
Do you do you think Is that the beginning of like a whole Style of art that's still is gonna be blossoming further.
Like, I'm sure there are, people look at AI fiction and there's some element of music.
I'm curious what other ways you might see this explode in terms of filtery types of exploration tools at home.
I think definitely having this weirdness as a filter, and as a thing that you can do and an expression mode, I think we're, we're going to be seeing more and more of that.
And, you know, and it's been really fun to see the way people do this with visual art.
So there's There are artists who now will scan in their paintings and do transformations on them or will you know, there is a, you know, buy a bag of lettuce and take picture of each of these leaves of lettuce and then clean them Neural net to start generating glitchy letters.
There's been some really interesting work with music as well so not trying to produce for example metal that sounds like any earthly band.
But what happens if it's slightly glitchy and now it is present doing, 22nd symbol solos or we have this thing saying unintelligible words or screaming for app, you know length that no human lungs could do.
So there's there's definitely a lot of depth in this kind of unrealistic stuff that we can get back from neural nets.
It's not, I guess it's growing now as a visual art style, and I may be in music too.
I know I follow a lot of people that look at using GaNS and those types of make it I would love you to talk about that to explain people what that is but the way that art kind of We saw it with Google's deep dream where if first.
Where you'd see these images on Twitter where dog faces will be appearing and landscapes and snails and strange things.
And you start to see these like, it becomes an art style and aesthetic all of its own.
Do you see that Do you see that idea emerging in like music right now in pop music or like, even in film making or, where people might start incorporating that.
Yeah, and I think it's as as the tools get More sophisticated, more easy to use.
I think that's when we'll start to see these kinds of neural net transformations of weirdness sort of encroaching into these other
In these other areas, so filmmaking, for example, is a bit of a tough one, just because we're dealing with high resolution, we're dealing with lots of frames, and you know you, you'll see people playing around with video shorts using deep fake type things.
And so that's really interesting, but to Make a longer film that uses this as a sort of a dominant effective really plays with it.
I mean there's really exciting things to be done like I would love to see some kind of film be The gorgeous color that uses neural net generated objects, this sort of aesthetic that you get when you look at the images generated by something like began, where you know there's all these different Categories you can and if you look at it attempts at for example clock.
The clock itself is usually is kind of clock shaped but it's a little melty looking the letters are completely legible in the hands of the clock or like branching off into all these different you know, cut almost like feathery hands and it's Neural nets attempt at reflecting a real clock and the lighting is often much better than the geometry and like.
So you get these really beautiful images, sometimes hyper realistically textured and they make no sense when you step back and look at them and that aesthetic is.
Is really interesting like I would love to see a film play with that or some other of these weirder neural net effects.
Let's have a let's have a film full of cat people.
Yeah That's amazing.
So yeah, some of the some of those photos that you post, you know, or generated images.
It's It is like almost like an optical illusion in an impressionistic way where first you see what seems to be created to be like a cat.
You realize that the pieces don't line up and that there's no like actual cat there.
It's like cat impression and it's like a hint of ear or fur.
And then you can't resolve it when you look closer and that's like a it's not something I would expect to perceive.
It's an interesting thing that tends to show up in a lot of those images.
Like a suggestiveness that that you realize is not there when you get closer but is there when you get the like the the general sense of it.
In your book there there are a number of things that like really came out and and spoke to me And I think one that we talk about a lot in AI people again going to thought of what AI and its limits are.
It's such a great book discussing the limits.
The breaking points like you're really playing with breaking everything and seeing where things succeed and that it's so data dependent algorithm dependent and then As a result, it's really dependent on on people and, and the biases that that can generate from that and how to that it's a tool to be used and to be tweaked and curated and I know we talked about biases a lot now AI and have been for a number of years but But the sort of tool [UNKNOWN] of it that it's like it's what you make of it and how you use it.
I think that was one of the things that I for sure wanted my book to accomplish is that I was seeing a lot of indications that people were thinking, Yeah, you throw AI on this project and we'll come up with the right answer.
It's smarter than people or it will You know, the AI said this is the answer.
It's based on math, that must be correct.
And so I really wanted to say, Hey, you know, where are the limits?
How smart are these things?
how likely are they to Get the wrong impression solve the wrong problem, copy the humans without realizing what part of human behavior they should or shouldn't copy.
And, especially this sort of idea that you can tell an algorithm to be moral or tell an algorithm to value human life or to not hurt people.
The, the three laws of robotics or any sort of laws of robotics are not something that you can train today's algorithms on.
But you get people interacting with these different products or With policy as if, yeah, we, you know, all we have, why don't we just tell it to be moral?
[LAUGH] And so I really wanted to say, Okay, now that's like saying, you know, we're going to tell the fire to be moral or something I you know, it's not a thing that you can do.
It's a it's a force of nature that we have to work with
Google IOL last year.
I know they had a few people.
Talking about there was a talk about artists working with AI together.
One working with with his work he she's been working for years with robots, drying and co developing.
You know, this this work that's a little bit of both of them.
And I was thinking about that as like a really exciting future way to think about AI.
Like self driving car manufacturers are almost talking about that now to when they think about it, it's not self driving so much as you in a relationship with the sensing mechanisms.
And I was curious how, what sort of future you imagine with that.
And in the case of self self driving cars, I think.
We, as builders and users of these algorithms are maybe starting to get a more realistic picture of how independent they will be, and maybe coming up with better tasks for them to do and say, okay, rather than put this thing in charge of driving a car and handling any weirdness that the road can throw at us, maybe We have it in charge of just like, watching our blind spot or watching our mirrors and maybe helpfully telling us that they think we humans have missed something.
So, this sort of narrowing of scope is is really playing better to what Machine learning today can do.
I mean, we're definitely dealing with narrow artificial intelligences that do best on really narrow problems.
And so, I think, you know, we've had a few few bits of technology a few startups, spectacularly fail because they tried to build something that was too sophisticated that was really ahead of what today's Technology could do
What, right now obviously like I, [UNKNOWN] We're all stuck at home and we can't go very many places and living in this new landscape like do you feel that the your thoughts on AI are changing?
Are there certain things coming to the forefront now that Are posing new possibilities or new threats or or challenges than before.
You know doing more of our interaction over voice and video there are new Threat models just for spoofing out there.
So we've seen papers saying, Hey, did that it would be possible to insert a deep fake into somebody's zoom chat like I don't know why we would want to do that other than To prank someone, [LAUGH] Into thinking that, I think it was in this case Ilan Musk has joined your call by mistake or something.
So there are these different, cases where, there can be different Threads there's different possibilities too, so if you see people playing with zoom backgrounds for example, and the zoom background, it's relying on a single lens.
It doesn't have depth information.
It is using AI to figure out like where you stop and your background begin and so
Having AI do more customization of the interface between people and technology.
Having these virtual places that we can join, have a bit more realistic
Next to them, there's a lot of AI heavy lifting in augmented reality things like that.
So, you know having you know now if you search for shark on Google there's a little thing that says do you want to in your room in 3D and.
It'll render that.
And so there's possibility as well now that we're [LAUGH] Doing a lot more interaction that's mediated through things that I mean not so many people don't know that we have so much AI in our phones already, right now that are running all these fun little games and.
Effects of things are making our cellphone photos look way better than they have any right to.
I mean, so much of that is AI.
So yeah, it's really interesting to have that kind of interleaved a bit more into our lives.
I guess I have a question about do you have any thoughts about AI's ability to help with with everything right now?
Do you see possibilities in where AI could could provide an unexpected benefit?
I mean, we're already looking at that in health tech, I think in terms of, I don't know if that's necessarily the application.
Have AI in quite the same way but between like looking at people whose proximity to each other or even in terms of studying patterns and behavior, more so than even necessarily a cure, but Do you see possibilities and optimistic things about where AI could be leading us next?>> I mean, there are probably some really useful ways where some of these tools can help.
I think one thing we have to be really cautious about is they know there is this definite desire to help among people who have skills to build Applications to use AI for stuff and I, you know, there, we really need to make sure that what we're building what we're working on is something that's actually going to be helpful when it's actually needed by the people who are in the field.
You know, testing people working with patients like is this actually something that we need?
And or is it even possibly harmful if we're trying to, you know, you see examples of Yes, this will screen people for COVID-19 by looking at pictures for them of them.
For example, we've trained this algorithm like little it really is there any basis for this or By handing out these decisions on whether somebody has COVID-19 buy from their voice or their face and, could we be giving them really bad, dangerous information that's going to cause people to do, unfortunate things and are too rely on this.
Data that's not worthwhile.
So I think there's a lot of danger there as well to, you know, build and release some really harmful algorithms just from meeting an overabundance of a desire to help and not much Caution to talking to talking to experts.
So you know there's I'm seeing a lot more of who though don't build that no think about this or how is this going to be helpful or even this is going to this is going to hurt people don't release this.
I'm seeing actually a lot more of that than I'm seeing.
stuff that is being used responsibly.
And it could be because the responsible uses of things take a lot longer to develop because there's all of this work that goes into building something that's actually good and Actually called for.
Yeah, we're more and more bad, but I feel like the world is based on so many predictive models now that it's the, the power of data seems terrifying, but.
But that's where we're and we're all trapped at home.
And you know, thanks so much for the time and I hope that people look at your, your blog, but also, this book is great if you're interested in learning about AI, and you want to get a start and there's so much that I didn't know and then I'm still learning And I think gets asked a lot of questions that I'm still thinking about.
Thanks, I look forward to following more of your, yeah, Weirdness blog, and what you're writing about next.
Great, thanks, appreciate it.
Nice talking with you.
Nice talking with you, too.
Thanks, see you soon.