Samim Winiger: Welcome to Ethical Machines. We are your hosts…

Roelof Pieters: Roelof.

Winiger: And Samim.

Pieters: Ethical Machines is a series of con­ver­sa­tions about humans, machines, and ethics. It aims at start­ing a deep­er, better-informed debate about impli­ca­tions of intel­li­gent sys­tems for soci­ety and indi­vid­u­als.

Winiger: For the sec­ond episode, we invit­ed Jack Clark, the world’s first neur­al net­work jour­nal­ist, report­ing for Bloomberg, to talk with us. Let’s dive into the inter­view.

So how long have you been in the US for now?

Jack Clark: I’ve been here for about three years now. I moved here to join The Register, where I wrote about AI and data­bas­es. And then I got hired by Bloomberg, and so now I’m help­ing to cov­er AI for us as well as doing more tra­di­tion­al enter­prise com­pa­nies.

The way I think of it is, neur­al net­works are going to be fun­da­men­tal to a very large amount of the AI that we’ll see and expe­ri­ence for the next few years. So I fig­ure if I report on any com­pa­ny or indi­vid­ual mess­ing around with this tech­nol­o­gy, then I can get a view of a good sec­tion of the AI world as it expands. And from a sto­ry point of view it’s very fruit­ful, because both lots of research hap­pen­ing and it’s cre­at­ing some fas­ci­nat­ing exper­i­ments and prod­ucts that we can write about.

Pieters: So what do you think is the best approach to explain­ing these com­pli­cat­ed top­ics?

Clark: I think you have to read every­thing. I spend about one or two hours a day read­ing the preprints as they come on arX­iv, read the papers and study that. And then try to turn it into an anal­o­gy. Because no one knows what a neur­al net­work is out­side of acad­e­mia, but every­one knows that when you’re an extreme­ly young child you’ll like, pick up a flower and stare at it for sev­er­al hours. And it’s this ear­ly form of learn­ing that chil­dren do that is anal­o­gous to what we’re try­ing to do with some of these sys­tems.

Winiger: You’ve heard major fig­ures in deep learn­ing and beyond blam­ing jour­nal­ists late­ly for over­hyp­ing the issue, for mis­char­ac­ter­iz­ing. You’ve heard a lot of ugly words being thrown around. How would you con­front these very regard­ed fig­ures crit­i­ciz­ing jour­nal­ism in such a broad way?

Clark: Ninety-five per­cent of the peo­ple mak­ing these crit­i­cisms have spent years study­ing arti­fi­cial intel­li­gence. They have a tech­ni­cal back­ground. They prob­a­bly have a PhD. They have years of expe­ri­ence of look­ing at very com­plex tech­nol­o­gy and com­ing away with objec­tive and sort of applied appli­ca­tions of it. 

Most jour­nal­ists don’t have PhDs in machine learn­ing. From the journalist’s per­spec­tive you know, I’ve had to spend sev­er­al years read­ing a lot of lit­er­a­ture and doing a lot of math­e­mat­ics and try­ing to teach myself this stuff and I’ve made that invest­ment. It’s dif­fi­cult to find the time. So peo­ple like Demis Hassabis or Yann LeCun or Jeffrey Hinton, one of their main respon­si­bil­i­ties and of their pub­lic rela­tions depart­ments should be to spend a lot of time with jour­nal­ists and make sure that they they just edu­cate the jour­nal­ists about how this stuff works. It yields a media where you have an under­stand­ing of it. Because they have to be gen­er­ous with their time just as we do in our writ­ing of their sub­ject.

Pieters: So what are things which kind of hap­pen, be it like unex­pect­ed or just inter­est­ing to you. What are the kind of most excit­ing things?

Clark: The two very excit­ing things to me, one is mem­o­ry sys­tems? You know, we have seen in recent years every­one has start­ed to look at long-short term mem­o­ry again. The appre­ci­a­tion for the role the hip­pocam­pus plays in con­scious­ness, that same obser­va­tion is being applied to AI to give us sys­tems that can do long-term rea­son­ing and multi-part pat­tern recog­ni­tion.

The sec­ond one is about rein­force­ment learn­ing being com­bined with deep learn­ing in robot­ics. I mean, you saw just this week Fanuc invest­ed a stake in Preferred Networks. Preferred Networks have been doing rein­force­ment rein­force­ment and deep learn­ing applied to robot­ics plat­forms. They read their neur­al Turing Machine paper, they read the q-learning paper. They’ve also read work out of Berkeley’s lab from Pieter Abbeel and Sergey Levine on end-to-end visuo­mo­tor pol­i­cy train­ing.

So that already has cre­at­ed a sit­u­a­tion where Fanuc has invest­ed mon­ey, ABB has put mon­ey into Vicarious. And we have some star­tups I’ve just heard about who are all doing this. So this is excit­ing, you know. Especially after we saw the DARPA Grand Challenge, those robots looked so kind of drunk. They were falling all over the place. They were stu­pid, very very slow. And I spoke to Dr. Gill Pratt who ran that, and he’s of the opin­ion as well that we’re going to get a huge increase in robot­ics capa­bil­i­ty from the appli­ca­tion of sens­ing sys­tems from neur­al net­works.

Pieters: But con­nect­ed to that, what would you rec­om­mend, as a grad­u­ate stu­dent in machine learn­ing? Stay in acad­e­mia? Join one of the estab­lished cor­po­rate research labs? Or start your own start­up?

Clark: I think what we’re going to get deep learn­ing artists at some point. Now that is the worst idea in the world if you would like to have mon­ey or a career. But it is some­thing that will hap­pen. We’re going to get some artists who are using this gen­er­a­tive stuff. I’ve seen work that you’ve done—I’ve seen a lot of it. I think that the field of new aes­thet­ics we’re see­ing from this could become very very inter­est­ing once we have a bet­ter under­stand­ing on how to use the mod­els.

To answer your actu­al ques­tion, though, there are two areas of pos­si­bil­i­ty. One is cus­tom accel­er­a­tors. If you can fig­ure out how to do good FPGAs to do learn­ing, then you can start to put clas­si­fiers on low-cost drones and things like that. That would be the area I’d rec­om­mend. But FPGAs are incred­i­bly dif­fi­cult. So you know, good luck.

Winiger: So, recent­ly we’ve seen lots of friends of ours that are work­ing on deep learn­ing libraries or key pieces of the puz­zle get­ting hired imme­di­ate­ly by Facebook Google, etc. And so I guess as a stu­dent espe­cial­ly, you’re con­front­ed with this real­i­ty right now. You can either get imme­di­ate­ly hired by one of the big play­ers, or go and do a PhD by an insti­tu­tion that is spon­sored by one of the big play­ers, or start a start­up. To refor­mu­late that ques­tion, which one of these is that less evil poi­son, in a sense?

Clark: Well, here’s the prob­lem. This is a poi­son ques­tion, right. Because from a soci­ety view, every­one should stay in acad­e­mia because it begets the largest quan­ti­ty of open research and the best teach­ing for the next gen­er­a­tion. I don’t know about Europe so much, but in America with the way fund­ing is and com­pet­i­tive tenure sit­u­a­tions, and just the mis­ery of being a post-doc at say, if you’re not a top-tier insti­tu­tion? That’s such a hard life that the ratio­nal deci­sion for many is going to be to go and work at Google or Facebook. Because you will get excel­lent train­ing, you will get some of the best data you can access in the world, and you will be paid giant amounts of mon­ey. So, go for the com­pa­ny. But keep in mind that every­one in this com­mu­ni­ty has a respon­si­bil­i­ty to do the best research they can in the most open way.

Pieters: Yeah. Connected to this is the ques­tion of biopol­i­tics, you know, in a post-structuralist sense, where it’s about con­trol of the land­scape. Where you have all these peo­ple in their free time cre­at­ing libraries. But I’m get­ting hired by one of the big com­pa­nies. It’s a land grab for tal­ent, that’s clear. But at the same time it’s also a land grab for con­trol of the resources on the soft­ware lev­el. So that’s a trend, at least. Do you see this trend?

Clark: Yes, def­i­nite­ly. Partly it’s that AI is a rel­a­tive­ly small com­mu­ni­ty. You know, Yann LeCun and his friends at NYU all work on Torch, where­as DeepMind has done work on oth­er libraries. Even the lan­guages. Some peo­ple real­ly like Lua, oth­ers are more doing stuff with Python. Some peo­ple who are very intel­li­gent are just writ­ing things in C. But I am afraid of those peo­ple because I don’t know how you can do that. You know, there’s a diver­si­ty hap­pen­ing here.

Now, I don’t know how it gets fixed. Someone needs to grow up and com­mit the com­mu­ni­ty to one or two of them. If we look at the his­to­ry of soft­ware, that’s not gonna hap­pen for a few years. 

Winiger: Maybe piv­ot­ing from the cor­po­rate dis­cus­sion, in the pre­vi­ous episode where we had Mark Riedl, we had a long chat about gen­er­a­tive tech in sto­ry gen­er­a­tion, that kind of thing. And we touched briefly on gen­er­a­tive jour­nal­ism. And he men­tioned that was one of the few areas where gen­er­a­tive text cre­ation real­ly is being deployed in indus­try in a large way. Which was to me eye-opening. And I knew at the bor­ders this was hap­pen­ing, but to hear it first­hand was very inter­est­ing. I mean, what you think? Is it going to be a rad­i­cal­ly new jour­nal­ism quite soon, or what do you think about that?

Clark: …Yes. [laughs]

Winiger: Right.

Clark: So num­ber one, I work at Bloomberg. We obvi­ous­ly do very com­pet­i­tive sto­ries. When you do an earn­ings sto­ry, we try and have a first ver­sion out with­in three or four min­utes. We do that by writ­ing incred­i­bly detailed tem­plates. We talk to peo­ple in the days run­ning up. We have a whole team of peo­ple stand­ing around with num­bers, check­ing every one when you push it live. 

Obviously this is some­thing that is going to be increas­ing­ly auto­mat­ed. Because this is a job where I am try­ing to be like a com­put­er. And when­ev­er you’re doing that job, you real­ize at some point a com­put­er will do this.” The Associated Press already uses tech­nol­o­gy from Narrative Sciences to gen­er­ate earn­ings reports for com­pa­nies that they don’t cov­er.

The only prob­lem that these tools have is that they can’t do con­text. So it will look at all of the indi­ca­tors in the earn­ings release. It will look at the ana­lysts’ rec­om­men­da­tions. And it will make sen­ti­ments deci­sion based on whether the com­pa­ny beat or didn’t beat. The prob­lem is that the mar­ket isn’t ratio­nal, so some­times a com­pa­ny can beat on all of the ana­lyst esti­mates but go way down because buried some­where in the release is a ref­er­ence to how they’re chang­ing account­ing, or they’re writ­ing down some­thing or what­ev­er.

Generative jour­nal­ism will be a real­i­ty. We’re not quite there…yet, but it’s very clear to me that it’s going to fos­ter a sort of winner-take-all sit­u­a­tion where if we at Bloomberg devel­op some small tools, I will be able to do sto­ries much more effi­cient­ly and that leaves more time for inves­ti­ga­tion.

The New York Times, in a pro­to­type ver­sion of their new CMS, they’re using recur­rent neur­al net­works to sug­gest tags based on the sto­ry. So they’re already kind of aug­ment­ing arti­cles with some of this machine intel­li­gence. Which is a great idea. 

Pieters: Well, and the oth­er way around there’s this recent work on ques­tion answer­ing sys­tems, where some news­pa­pers, very very spe­cif­ic in their style by using both the sum­ma­ry and the actu­al arti­cle to train a long-short term mem­o­ry but a ques­tion answer­ing sys­tem. Which basi­cal­ly is to say that if it can go one way it should also be able to go the oth­er way.

Clark: It should be able to. You may have been aware DeepMind did some of the work there. And took in all of the Daily Mail, which is a large tabloid web site. And what they found is that if you put the phrase into their learn sys­tem you know, Does cof­fee cause…” or Does eat­ing lob­ster cause…” every time “…” be can­cer, because the Daily Mail loves writ­ing arti­cles about how everything’s going to give you can­cer. So that shows you how even with a very large data set there could be some prob­lems that are very unpre­dictable.

Winiger: Yeah, I mean this is in my expe­ri­ence as well work­ing with gen­er­a­tive sys­tems. You real­ly have to rethink the design process as one of choos­ing inputs and out­puts. Choosing the Daily Mail seems like an exer­cise in com­e­dy more than any­thing else.

Clark: It’s fun­ny, but one of the things that Google has been talk­ing about a lot is that… And you may know more about this. The European Commission pub­lish­es huge doc­u­ments it pub­lish­es them con­cur­rent­ly in twenty-seven dif­fer­ent lan­guages. So there’s an idea of not only can we use that as a very good store of text to learn con­cepts, but we can learn con­cepts as they cross from one lan­guage to anoth­er because we have that map­ping. And because it’s not just a French-German dic­tio­nary it’s French-German-Italian-English all on one thing, you can learn a very com­plex, rich rep­re­sen­ta­tion across the dif­fer­ent cul­tures. So that seems fruit­ful to me.

Winiger: It just brings back a dis­cus­sion I had a few days ago with [Guy Acosta?] one of the devel­op­ers in deep learn­ing. And he brought up this inter­est­ing idea that the trained weights of these nets can be set in con­nec­tion with what pre­vi­ous­ly was the data­base. So the next ora­cle in a sense will be the one that is wield­ing the most pre-trained weights and that there will be a whole set of law cas­es unfold­ing soon where peo­ple will get sued for the spe­cial­ized train­ing on top of pre-trained nets and things like that. Which I found inter­est­ing. It just came to mind when you were talk­ing just there.

Clark: Well, if you think about it what we’re doing is we’re turn­ing very high-dimensional math­e­mat­ic rep­re­sen­ta­tions of a sort of large knowl­edge space into intel­lec­tu­al prop­er­ty. Which should be the most fright­en­ing idea in the world to any­one. This is from most abstract thing you could pos­si­bly try and turn into a cap­i­tal­ist object, and we’re head­ing down that direc­tion. I don’t think that can work. I think that if you look at the way that you encode infor­ma­tion from a trained net, the legal cas­es will be huge­ly com­plex. But Google has been acquir­ing many patents and so has IBM and so has Microsoft. So we might get a Cold War sce­nario where there is no law suits because all of them have enough patents to threat­en each oth­er with you know, nuclear bomb law suits. Who knows?

Winiger: I mean, it’s a hor­ri­ble sce­nario. Nobody wants to see that because I sup­pose it would sti­fle inno­va­tion, real­ly.

Clark: How do we avoid it? You know, what are things that you guys think could be done to stop it hap­pen­ing?

Winiger: I think step one is to sup­port the cur­rent open­ness. Because we see this mar­ket­ing mon­ey rush­ing in and obvi­ous­ly they’re smart mar­ket­ing peo­ple. They try to set a cul­tur­al sen­ti­ment. And I think that’s one dial that as a soci­ety we can start to turn in the oth­er direc­tion. Upgrading the now some­what old-sounding notion of the pub­lic domain. Especially in the US it sounds com­plete­ly out of date but it could quite eas­i­ly actu­al­ly be dialed back into fash­ion. So that’s one approach, I sup­pose. It’s a real­ly hard prob­lem, isn’t it?

Pieters: There’s Creative Commons licens­es, right. I mean, why does Google for instance bring out a patent for all these dif­fer­ent algo­rithms they devel­op. Why not launch it with a Creative Commons license, where there may be [an] attri­bu­tion clause. Okay, you have to attribute it to Google, so Google is still pro­tect­ed on the attri­bu­tion prin­ci­ple but it still would be actu­al­ly open.

Clark: Well, it real­ly is unfor­tu­nate­ly this hor­ri­ble sort of mutu­al­ly assured destruc­tion game the­o­ry sce­nario, where Google may not have want­ed to patent this, but what it may have done (which com­pa­nies do reg­u­lar­ly) is looked at all of the patents IBM has on AI, said like, Holy moly. If we don’t have some AI patents, we can be in a legal­ly weak posi­tion with respect to IBM should there be a law suit.” So it cre­ates this sce­nario where even if it’s not a good idea, you’re going to amass these patents because as a cor­po­ra­tion you have to do ratio­nal things for your investors. And any investor would kind of right­ly say, Hey, Google. By not patent­ing any of this stuff, you’re putting your­self at a dis­ad­van­tage to your com­peti­tors in the mar­ket­place who are amass­ing the tools nec­es­sary to mount a legal attack.” It’s a bit depress­ing but I think that is the ratio­nal cor­po­rate response.

Pieters: One of the big sto­ries is also the more future con­cerns about the devel­op­ment of AI, lead­ing mass employ­ments, or to Terminators, ets. But let’s stick to the mass employ­ment sce­nario for now. It’s being argued if there is a mass employ­ment there needs to be a solu­tion for the peo­ple who are unem­ployed, which might be some­thing like a min­i­mum income. So where do you stand on this issue?

Clark: It’s a huge prob­lem. I’ve read a lot of research by David Autor, who is a great econ­o­mist I believe at MIT. He has writ­ten a lot about this. His analy­sis is that we don’t have the data to be able to project that AI could lead to large-scale unem­ploy­ment. But we also don’t have the data to say that that won’t hap­pen. And then if you look at peo­ple like Andrew Ning, who does AI at Baidu you know, he said to me, When the US went from 50% of peo­ple work­ing in farm­ing to 2% over fifty years, that was fine because the farmer knew that their son or daugh­ter should go to col­lege because farm­ing would be mech­a­nized and there wouldn’t be a job.” 

The speed of today’s econ­o­my means that this same tran­si­tion is hap­pen­ing with­in a sin­gle gen­er­a­tion. And that’s where the prob­lems come in, is that we have no sys­tem in soci­ety for retrain­ing peo­ple when they’re mid­way through their lives to take on a new type of employ­ment or job. And that will be the issue that AI brings to the table. Because if you’re a truck dri­ver, if you’re a lawyer work­ing in e-discovery or data stuff, you’re a jour­nal­ist doing a lot of jour­nal­ism that just requires sort analy­sis of num­bers which are out there, there is huge evi­dence that AI is com­ing for you and is mov­ing very rapid­ly.

And just from anoth­er slight­ly more basic eco­nom­ic point of view, the thing that AI does is it makes your exist­ing cap­i­tal expenditures—you know, your ware­hous­es, your factories—it increas­es the effi­cien­cy of them and it low­ers the depre­ci­a­tion of them. So as a busi­ness oper­a­tor like Amazon, you have the hugest incen­tive in the world to roll out Kiva Systems robots to as many of your ware­hous­es as fast as you can. Because when­ev­er you look at the num­bers, the effi­cien­cy is so much greater than with a staffed mod­el. This is going to be a defin­ing issue, maybe. I expect with­in the next five to ten years we see the big effects. If self-driving cars come on sched­ule and receive the kind of uptake that peo­ple at J.P. Morgan, peo­ple at the big banks are object­ing, it’s com­ing, you know.

Winiger: It’s super inter­est­ing. So when it gets to the kind of hard ques­tion of how soci­ety should frame automa­tion, etc., in the West espe­cial­ly self-worth and these more philo­soph­i­cal con­structs are real­ly based on full employ­ment and so forth, right. I mean, the whole psy­chic in the West is built on these notions. And so in a sense we are say­ing it’s all going to col­lapse soon­er or lat­er. It’s already now an elec­tion cycle top­ic.

Clark: It’s going to be chal­leng­ing. I have a friend who’s actu­al­ly also English. They work in New York in finance. So they’re aware of tech­nol­o­gy and what hap­pens. I speak to them about this issue and they say to me, Well, but Jack, what will peo­ple do if they don’t have to work? People have to work. It’s nat­ur­al.” And I talked to a lot of peo­ple who have that view. So as you say, we have such a deep-set psy­cho­log­i­cal asso­ci­a­tion between work and self-worth that watch­ing that change is going to be dif­fi­cult. Maybe this is an area where the Europeans can take lead­er­ship because we’ve always had an appre­ci­a­tion for hol­i­day and not work­ing. Maybe that will help us, you know.

Pieters: Maybe you could argue that Google and Facebook and [inaudi­ble], it would be in their inter­est to push this new phi­los­o­phy of there will be peo­ple unem­ployed and it’s fine in the sense that they will have the [?] peo­ple work­ing against this trend. In polit­i­cal lead­er­ship of this kind of trend, this a good thing. It can be framed in the nar­ra­tive of this is good for busi­ness, good for the world, then they are, at least on the cor­po­rate lev­el the ones who are most to gain from this, right? 

Clark: Yeah. And from a pub­lic rela­tions stand­point, as a com­pa­ny you nev­er want to be asso­ci­at­ed with the destruc­tion of jobs and increas­ing inequal­i­ty. And unfor­tu­nate­ly for these AI com­pa­nies like Facebook and Google, they’re already being tagged with that. Because they have a very com­pet­i­tive mar­ket and they give engi­neers free food and mas­sages and bus­es. And we’re in San Francisco where you have huge home­less prob­lem and huge inequal­i­ty as well. But this is an issue that they’re going to need to take a lead­er­ship role on because oth­er­wise they’ll risk dis­con­tent from soci­ety becom­ing direct­ed at them because they’ve become a sym­bol.

Winiger: Right. That actu­al­ly ties it real­ly beau­ti­ful­ly back to begin­ning of the dis­cus­sion of the greater need of explain­ing this real­ly com­plex set of issues to the pub­lic much bet­ter. Because oth­er­wise that whole debate is going to actu­al­ly break down. I mean, that might end in a real­ly nasty sce­nario, I guess.

Clark: Yeah. The oth­er issue this is bound up with is why can’t I pay for Facebook? Why can’t I pay for Twitter? Why can’t I do some some sit­u­a­tion where either I pay them and they don’t get my data, or they pay me a very small amount of mon­ey and I give them my data? Because if you taught peo­ple that their data has some val­ue, or that they have the option of cap­i­tal­iz­ing on that val­ue by buy­ing a ser­vice instead of get­ting it for free, they would under­stand what AI means.

Because all AI is in the way that it all affects a lot of soci­ety, it’s the out­come of us trans­fer­ring loads of very well-annotated, clean data to cor­po­ra­tions. And that issue will have to become one because if you’re Google and you say, Well, you don’t pay for Gmail because we sub­si­dize it with adverts and you get a lot of val­ue from it even though we take your data,” that is a very rea­son­able argu­ment, but at some point peo­ple are going to ask well why can’t I have the oth­er option. 

And then the com­pa­nies have to say, Well actu­al­ly, your data is so valu­able when we com­bine it with every­one else’s that we have no incen­tive to do this, from an AI devel­op­ment stand­point.” I believe that would start a con­ver­sa­tion among peo­ple about this.

Winiger: And for the gen­er­a­tive work I’ve been doing, we copy­right­ed it. I mean, what hap­pens if you train a net with copy­right­ed images and you gen­er­ate out­puts with that, you get into a real­ly inter­est­ing sit­u­a­tion with copy­right law very quick­ly, real­ly. And that’s just the begin­ning.

Clark: But I can think of real­ly puz­zling sce­nar­ios, like if I train a gen­er­a­tive music sys­tem based on a CD I buy of the New York Philharmonic play­ing Bach, a what point am I still using copy­right­ed per­for­mances from the New York Philharmonic ver­sus at what point is it just Bach played by a gen­er­a­tive sys­tem? It’s very very hard to dis­cern that bor­der­line. Because we know that what hap­pens as you do this gen­er­a­tive stuff is you dis­tort the under­ly­ing mate­r­i­al to the point that maybe it is fair use, that maybe it isn’t the orig­i­nal IP any­more. There’s no way to answer this stuff sim­ply. It’s going to be a very hor­ri­bly com­pli­cat­ed time, I think.

You can imag­ine you train a movie sys­tem on every sin­gle action movie in his­to­ry. And then you cre­ate a gen­er­a­tive sys­tem which will go from frame to frame or scene to scene and sort of inter­po­late a new movie out of this. At what point are you infring­ing copy­right? Like how do you even judge that any­more? It’s crazy. And as I said ear­li­er, we’re going to get artists who do this. We’re going to get peo­ple who want to con­tribute to the cul­tur­al dis­cus­sion and the aes­thet­ic dis­cus­sion, that do this stuff, and the legal sys­tem and right­sh­old­ers will have no clear path for how to react. It’s new ter­ri­to­ry.

Pieters: Well, even more prob­lem­at­ic, once you start charg­ing mon­ey for the gen­er­at­ed mate­r­i­al or you want to copy­right that, then it becomes inter­est­ing for all those copy­right hold­ers from the orig­i­nal inputs. Because they might want to also cash out that. 

Clark: You know, do you end up licens­ing the object itself? So say the val­ue of a pho­to­graph of my cus­tomized vehi­cle is quite low. Would it be greater for me to share a data set which is a set of sev­er­al hun­dred pho­tos of the car from every sin­gle angle so you can train a sys­tem to have a rep­re­sen­ta­tion of it? And should I charge my data on the rich­ness of the AI rep­re­sen­ta­tion you can derive from it? Again, I don’t know but this feels like con­ver­sa­tions which cre­ative peo­ple are going to have to start hav­ing.

And then you get to the real­ly out­landish sce­nar­ios when you start to com­bine all of this which we’re talk­ing about with a dis­trib­uted trust-based blockchain sys­tem, or the run­ning of code and val­i­da­tion of it. Again, you start to get autonomous pro­grams that will mine the Internet for con­tent and then will sell gen­er­a­tive art for bit­coin, anony­mous­ly. Well, what do we do when that has hap­pened in the world?

Pieters: So there’s a prece­dent in Holland of some­one cre­at­ing lan­guage with nat­ur­al lan­guage pro­cess­ing (I think it was not a neur­al net­work, but I mean machine learn­ing in any case), which was argued that it was hate speech or threat­en­ing tweets. And he got his door kicked in by the police. And in the end the ques­tion was who was respon­si­ble for this con­tent on Twitter? Was it the machine learn­ing algo­rithms or the per­son behind it, who argued it was—

Clark: Put the serv­er in prison.

Pieters: Yeah. At least accord­ing to the Dutch leg­is­la­tion, in the end it was the per­son who cre­at­ed the algo­rithm who was held respon­si­ble for this.

Clark: You know, Google when they launched Google Photos had a huge prob­lem, which was that the sys­tem was iden­ti­fy­ing peo­ple of col­or as goril­las. Which is lit­er­al­ly the most offen­sive thing your sys­tem could do, pret­ty much. Again, is the Google per­son respon­si­ble for not test­ing all of the cor­ner cas­es? Is Google the cor­po­ra­tion respon­si­ble for not doing QA? Valid and com­pli­cat­ed ques­tions, because it cer­tain­ly offend­ed and hurt some peo­ple. But then, they weren’t hurt by a per­son, they were hurt by gen­er­a­tive deci­sions of an algo­rithm that emerged out of a data set whose prove­nance we as the pub­lic aren’t told because it comes from a pri­vate com­pa­ny. Where is account­abil­i­ty in this uni­verse?

Winiger: It’s tricky. I mean, I sup­pose on the one hand yeah sure, gen­er­a­tive sys­tems do man­i­fest a lot of auton­o­my, in a sense. On the oth­er hand it’s the per­fect black box to hide nefar­i­ous human action behind. And you know, you just stand there and you raise your hand and say, Well you know, it was the black box. Excuse my—or it’s—behav­ior. Don’t sue me.” I mean, it’s a bit of both real­ly, isn’t it?

Clark: I had a con­ver­sa­tion with a hedge fund recent­ly, who I can’t name, but it was about what they thought of deep lead­ing and deep learn­ing sys­tems and how it applies to trad­ing. These peo­ple are very con­cerned by deep learn­ing because we can’t inspect the mod­els very eas­i­ly. We can’t get very good teleme­try. And the whole con­cept of tak­ing an emer­gent gen­er­a­tive sys­tem and plug­ging it into a trad­ing envi­ron­ment gives these peo­ple night­mares. It is like the worst thing they could imag­ine.

And yet, there is going to be a huge incen­tive to use deep learn­ing to pick up on some sig­nals which are not yet being processed by tech­ni­cal trad­ing firms. So we’re going to see some very inter­est­ing arms race there. You know, we already see it with satel­lite imagery being run through deep learn­ing sys­tems to look at the height of oil and gas tow­ers to infer sup­ply. As we get the roll­out of low-cost drones, mak­ing it pos­si­ble to sur­veil far out com­modi­ties, we’re going to get a whole range of learn­ing sys­tems plug­ging into the mar­ket. Again, it’ll be a good thing for effi­cien­cy but it will also open us up to hor­ri­ble prob­lems that we can­not even imag­ine. Which is excit­ing and kind of unnerv­ing as well.

Winiger: If you made it this far, thanks for lis­ten­ing.

Pieters: And also we would real­ly love to hear your com­ments and any kind of feed­back. So drop us a line at info@​ethicalmachines.​com

Winiger: See you next time.

Pieters: Adios.

Winiger: Bye bye.

Help Support Open Transcripts

If you found this useful or interesting, please consider supporting the project monthly at Patreon or once via Square Cash, or even just sharing the link. Thanks.