Bob Friday Talks: What place does AI have in education and film?

0:00 hello everyone thanks for joining us on

0:02 another episode of Bob Friday talks

0:04 today I'm joined by Oscar winter James

0:07 O'Brien professor of computer science at

0:09 UC Berkeley James maybe we can start

0:12 with giving the audience a little bit of

0:13 your journey how did you go from

0:15 Professor to Oscar winter where did the

0:17 journey start well my my journey in

0:19 visual effects and I guess visual

0:21 Computing in general started when I was

0:22 a grad student in Georgia Tech um I

0:24 became very interested in the idea of

0:26 using computer simulations to uh create

0:29 visual IM

0:31 um and so that's what I focused my PhD

0:33 work on when I was at Georgia Tech after

0:34 I graduated I came to Professor position

0:37 here at UC Berkeley um and still focus a

0:40 lot on visual effects but also um sort

0:43 of branched out to using uh simulation

0:45 optimization for a whole wide variety of

0:47 problems again mostly mostly related to

0:49 imagery um so what I would call Visual

0:52 Computing now you you've been a

0:53 professor since at Berkeley since 2000

0:56 or yeah since 2000 about 25 years okay

1:00 and so we're here today to talk about AI

1:01 so I thought maybe we could start the

1:03 discussion a little bit about how you

1:04 see AI affecting our education

1:07 University you know there's a lot of

1:09 talk about chat GPT you know what are

1:12 you seeing inside of the Berkeley right

1:13 now well I've had some firsthand

1:15 experience with this right um in in a

1:18 couple of ways uh you know last sem this

1:20 semester I'm teaching a graduate course

1:22 but last semester I was teaching a large

1:23 undergraduate course and we I was going

1:25 to give them a take-home final um and I

1:27 wanted to run the questions through gp4

1:29 just just to see what would happen and

1:31 it got them all right and so I said okay

1:33 well let's make the questions a little

1:34 trickier a little harder and it still

1:36 got them right a little bit trickier

1:38 harder and then finally it was getting

1:39 them wrong but before I could give this

1:41 the test to my students I want to you

1:42 know give it sort of a dry run a test

1:44 out so I asked my grad student Tas would

1:46 you you know to take the test and by the

1:48 time I made it hard enough that gp4

1:50 didn't get it right the grad students

1:52 didn't get right either um which is uh

1:55 you know it's a it's it's sort of a

1:56 little concerning like you know what you

1:57 know what is it that we're we're to be

2:00 teaching in our classes um and the other

2:03 as the other experience I had with uh

2:05 with with AI in this class was that we

2:08 had at one point one my taas came to me

2:09 and said Hey listen I've got some

2:10 cheaters and we've got several students

2:13 who all turned in almost identical code

2:15 and so I sat down I talk with these

2:16 students one individually one by one and

2:18 they all had the same story they didn't

2:20 they didn't think they had done anything

2:21 wrong they were told we were told we

2:23 were told to use Visual Studio to work

2:25 on our assignment Visual Studio includes

2:27 co-pilot we use co-pilot it wrote some

2:30 code for us and what was wrong with that

2:32 yeah you know for me Copa has become you

2:34 know for summarizing meetings now it is

2:37 totally changed my way I work now right

2:39 I don't have to take notes anymore I

2:41 just have copi help summarize it you

2:43 know but when you look inside of our

2:44 educational system you really see that

2:46 this AI is a positive negative you know

2:49 for me I kind of look at it as like

2:50 calculators you know when when side

2:52 rules calculators everyone I try to ban

2:55 calculators you know how do we adapt to

2:58 this new AI Tools in fact fting all of

3:00 our Lives well I mean this is this is

3:02 why I didn't penalize the students in

3:04 any way right if those students were out

3:06 working in the world and they had a job

3:07 to do then it would be totally

3:08 reasonable for them use co-pilot just

3:10 like it' be totally reasonable to use a

3:11 calculator you know it doesn't make

3:12 sense to tell them not to use it um it

3:15 does create the problem because we do

3:16 want to teach Basics still we still want

3:17 to teach the fundamentals and now it's

3:19 kind of a little bit of a problem

3:21 because if the students can use the AI

3:23 tools to do the fundamental work to do

3:26 these Basics how do you test them on

3:27 this how do you how do you teach the

3:29 students

3:30 not to not use the calculator but to

3:32 understand what the calculator is doing

3:34 so that when they use it they know how

3:35 to use it in the right way well I'm with

3:37 you because you know when I was growing

3:38 up it's like you still have to learn

3:39 your multiplication tables exactly you

3:41 know you have calculators so what do you

3:43 think the answer is now you know we're

3:45 in this new environment people have to

3:47 learn the basics somehow you know when I

3:49 look at the educational system and I

3:51 think we discussed earlier you know 340

3:53 years ago I had to buy books you I was

3:55 trying to learn to build Wireless

3:56 networking radios and everything you

3:58 know nowadays I don't buy books to learn

4:00 stuff I mean I go to the internet in the

4:02 cloud right now if I want to learn some

4:05 new AI topic you know I learn it on

4:07 YouTube right I mean this I mean even

4:10 before we had the AI tools to write

4:11 software it was also the case that you

4:13 know you could find a lot of if you had

4:15 a problem to solve you could look on

4:16 GitHub and find some example code that

4:18 somebody else had written and you know

4:21 you know I guess there are copyright

4:22 issues when you're looking at somebody

4:23 else's code but generally speaking if

4:25 you have a job to do you want to use the

4:26 best tools to do the job whether it's

4:29 finding open source code or using an AI

4:31 tool to generate new code you want to

4:33 use the best tools to do the best job

4:35 you can but that does make challenging

4:37 to try to teach the students right

4:39 because we need to now adjust how we

4:41 teach them just like we don't you know

4:43 have to memorize lots of stuff anymore

4:45 because there is not necessary they can

4:46 look it up how do we evaluate them how

4:49 do we teach them in a useful way and you

4:51 know when we talk about grades how do we

4:53 give them grades that make sense right

4:55 if it's if everybody's able to do the

4:57 assignment and do it perfectly because

4:58 they have the tools available ble to do

5:00 it perfectly do I give them all A's do I

5:02 find a way to somehow differentiate them

5:04 and not give them all A's what what do I

5:06 do it's a it's a challenging question

5:08 yeah I mean so now Juniper missed right

5:10 we're in the networking business so I'm

5:12 curious now like when you're giving your

5:13 test in universities now you know do you

5:17 let the inter do you let the students

5:18 use the network and access these tools

5:21 while they're taking the test or you

5:23 have kind of a no network policy going

5:25 on well if they're taking a test inside

5:26 the classroom then that's easy we can

5:28 ask we can they can put their phones

5:29 away and we we notice if you have your

5:30 phone on the desk um and so that that's

5:33 a little bit easier of a context it's

5:34 really the take-home stuff a take-home

5:36 final a homework assignment a project

5:38 that they're doing on their own um I

5:40 mean luckily the class I'm teaching they

5:42 tend to be like I teacher sorry I teach

5:45 upper division classes and graduate

5:47 classes and those they have Project work

5:49 and for the projects we say whatever

5:51 resources are available use those

5:52 resources right it's just like if you

5:54 hired someone to do a job use the

5:55 resources available to you um for the

5:58 lower division classes where there

5:59 learning the basics where they're

6:00 learning the fundamentals I think that

6:02 the problem the question is a lot more

6:04 challenging because you do still want to

6:05 learn those fundamentals and you know

6:08 the university requires us to give

6:09 students grades so we want to test them

6:12 on the fundamentals we want them to

6:13 learn the fundamentals but we also don't

6:15 you know we don't want to create a

6:17 situation where the honest students do

6:19 the work and then get a grade that might

6:21 not be an A and the dishonest students

6:23 use some other system you use a

6:25 generative AI to do the work and they

6:26 all get A's right we don't want a system

6:27 that's unfair um it's very I don't think

6:30 there's a clear answer yet well maybe

6:32 the other topic I want to try to cover a

6:33 little bit in this session today is

6:35 really around your career in film

6:37 because I know you were the inventor or

6:39 one of the co-inventors of digital

6:41 molecular matter algorithm did I get

6:43 that right yeah okay you know and for

6:46 those who don't know have you ever seen

6:47 the first Avatar

6:49 movie James technology was in that movie

6:52 was uh some of the X-Men movies as well

6:54 it's actually been used in over 200

6:55 films um and what this what the software

6:58 did is it's like it it kind what I

7:01 described earlier if you want to have a

7:03 a scene where something is a building is

7:04 being blown up or spaceship's crashing

7:07 into something instead of trying to

7:08 build a little model and film it and try

7:10 to make it look right you set up a

7:12 simulation you want a building to be Sat

7:14 On by Godzilla and claps well you you

7:16 take a model of the building you build

7:18 the building virtually and you set up a

7:20 big finite element simulation that's

7:21 going to compute you know as you know

7:23 the monster pushes on it or whatever is

7:25 happening what are the forces in the

7:26 building how's the concrete going to

7:28 crack where you know how the will the

7:30 windows shatter and it actually computes

7:33 all this and generates the raw data that

7:35 thing goes into what we call a renderer

7:36 that creates the images and those are

7:38 the effects that go into your in the

7:40 film that's what you see you're not

7:41 looking at something real you're looking

7:42 at something that was completely virtual

7:45 so I I have to ask you know of all the

7:47 scenes all these movie scenes I think

7:49 you're in like Technologies in 90 or

7:51 more feature films what scene stands out

7:55 you remember the most you go hey that

7:56 was me you know that was my technology

7:58 making that happen so actually the first

8:00 movie that really used a lot of our

8:02 software was the movie Sucker Punch um

8:05 and it I guess it didn't do super well

8:07 in the box office but it has some

8:08 beautiful special effects and to me that

8:11 that that's the first place that our our

8:13 software was really used in a very major

8:15 way it's a you know there I I forget

8:17 there lots and lots of scenes that used

8:19 it in the movie um very heavily and so

8:21 for me that's a very special one there's

8:23 also you know in addition to the movie

8:25 industry we've also used the software

8:26 for video games so Star Wars the Force

8:28 Unleashed was the first game that came

8:30 out that made extensive use of our

8:32 destruction modeling software now I I

8:34 know you started this adventure back in

8:36 2000 and I think when we talked about it

8:38 you started with finite element these

8:41 DMM algorithms you know maybe you give

8:43 the audience a little bit of the journey

8:45 from 2000 to we are today you know when

8:49 we look what's going on in the film

8:50 industry and we have actors and writers

8:52 going on strike being worried about AI

8:54 when did this AI really start to become

8:57 a thing inside the film industry so so

9:00 so AI um machine learning um deep deep

9:03 learning you know have different terms

9:05 for these things this has been you know

9:08 as a te as the techniques you know back

9:09 back in like the 90s for example the the

9:12 AI Tech AI the capabilities of AI

9:14 systems were very limited and so they

9:16 weren't used for a lot um at least not

9:18 in the movie industry um but by the time

9:20 you got to about like 2017 or so the

9:23 system started to be um much more

9:25 powerful much more useful and so you'd

9:27 see them being used um as a component

9:29 right so if I want to animate a

9:31 character I hire an an like in the movie

9:33 How to Train Your Dragon or something

9:35 like that I hire an animator who

9:36 animates Toth list by hand but now maybe

9:39 I can create a tool that an AI tool or

9:42 machine learning tool that makes it a

9:44 little bit easier for them to do it so

9:45 rather than very painstakingly moving

9:47 something and then waiting for an update

9:49 it will be more smooth because the AI

9:51 system is able to you know more quickly

9:53 compute the answer but you still have a

9:54 human anime doing the work um and what

9:57 we're seeing now of course is the the

9:59 generative AI systems that can actually

10:01 create the special effects themselves or

10:04 and uh you know that that's that's a

10:06 it's a sort of a phase transition it's

10:07 very different than using the the AI

10:10 system in a traditional pipeline you're

10:14 now getting sort of a new pipeline

10:15 that's built around the AI systems and

10:17 when did this gen AI I mean when did

10:19 that happen I mean it sounds like you

10:21 know somewhere from 2000 you know

10:24 somewhere sounds like in 2017 we started

10:26 getting machine learning be having

10:28 helped speed up these

10:30 action scenes you know when did these

10:32 actors and writers really have something

10:34 to worry about well I I think I think in

10:36 the last couple you know literally last

10:38 two two years basically um it's become

10:40 very clear that you know uh we now have

10:43 generative AI systems that are capable

10:45 of producing images that are almost

10:46 indistinguishable from reality um they

10:49 still make mistakes sometimes and give

10:50 you extra fingers and stuff like that

10:52 but the systems are getting better and

10:54 uh these problems are becoming fewer and

10:57 fewer um we also have now syst are able

10:59 to do video um so animated sequences not

11:02 just still images and you know there

11:05 there's there's a good and bad side to

11:06 this right the good side is it's very

11:08 democratizing right if if you and I

11:10 wanted to to you know we had some great

11:12 idea we wanted to make a film about the

11:13 story we wanted to tell

11:16 well today or maybe a few yesterday we

11:19 would have needed you know a big team of

11:21 people to do it like in this even right

11:23 now in the studio there's a team of

11:24 people running cameras and so on it's

11:26 kind of a big deal um but when if you

11:29 have an AI system where you can simply

11:31 describe the scene you want and it'll

11:32 generate the imagery for you then now it

11:35 becomes possible for individuals to now

11:38 produce high quality um video um the

11:41 downside is that you know there are a

11:43 lot of people whose profession is to

11:45 make that highquality video or make some

11:47 aspect some piece of it and you know

11:49 it's concerning those jobs might go away

11:52 well you know so I'm in the networking

11:53 space here you know what I always tell

11:55 people and I get the same question is

11:57 like Bob what's different about AI

11:59 versus what we were doing 30 years ago

12:01 and I usually tell them that you know

12:03 we're kind of moving AI is really this

12:05 next step in the evolution of automation

12:07 right you know the difference here is in

12:09 the past we were building very

12:10 deterministic repeatable things robots

12:12 that built your car a script that did

12:14 the same thing you know this AI

12:17 automation is really starting to do

12:19 things on par with humans some sort of

12:21 cognitive reing skill right and that's

12:23 kind of what we're seeing with chat gbt

12:26 you know so in that sense we're starting

12:27 to build automation that is replacing

12:29 humans which kind of leads into our next

12:32 kind of question you know in this next

12:34 thing you know most Technologies

12:36 typically create they eliminate jobs but

12:39 they typically create more jobs than

12:40 they eliminated you know so when you

12:42 look at you know what's happening with

12:44 AI automation movies even maybe in

12:47 education professors any concern that

12:49 you know you're going to get replaced

12:51 professors are going to get replaced by

12:52 AI assistants I think everybody should I

12:54 you know as these machine as you said

12:56 they're they're they're they're getting

12:57 basic or essentially the maches are able

12:59 to sort of emulate cognitive skills

13:01 right they can make decisions they can

13:03 sort of look at General General

13:04 situations and make general purpose

13:06 decisions and I think that what that

13:09 means is that no one's job whether it's

13:11 my job as a professor um an artist's job

13:15 creating something uh any profession you

13:18 have um Can potentially and probably

13:20 will get to the point where our

13:23 automated systems our machines are able

13:25 to perform that task and I think that

13:27 this is a little different than what we

13:29 we've seen in the past in the past we

13:30 would see technologies that made people

13:32 more efficient so I you know I have a

13:34 factory it makes a certain number of

13:35 cars and now I get the assembly line and

13:38 now it can make the same number of

13:39 people can make more cars but if I want

13:41 to make twice as many cars I still need

13:43 to hire twice as many people right or

13:45 then I have robotic arms that might help

13:47 build the cars but I still need people

13:48 to oversee that so now the same Factory

13:50 can make you know a hundred times as

13:52 many cars but I still need people

13:53 running the factory at some point we'll

13:55 get to the point where the number of

13:57 people you need in the factory is zero

13:59 and that's now very different because if

14:01 I want to scale up today I need to hire

14:03 more people if I want to me expand I

14:05 need to hire more people there are more

14:06 jobs get created but if I need zero

14:09 people to run the factory and I want to

14:11 now scale up my factory to make twice as

14:12 many cars well I still need zero people

14:15 and so that's a big difference zero is

14:16 different than every other number yeah

14:18 so me I kind of get you know hey we're

14:20 starting to automate human tasks you

14:22 know I here we're going to have robots

14:23 walking around the hospital cleaning

14:25 things here shortly you know when you

14:26 look at the Berkeley and what's

14:28 happening on the University there the

14:29 campus there any examples where you

14:31 start to see AI starting to automate

14:33 things at the school or the campus

14:35 itself oh I mean I I'm not exactly sure

14:37 what the administration is doing in

14:39 terms of deploying AI systems I I

14:40 suspect there may be uh you know

14:42 universities tend to be a little slow in

14:43 terms of how they move um so they Pro I

14:46 I suspect you see you know out in the

14:48 industry World you'll probably see more

14:51 automation being done with AI systems I

14:53 do think that there's a question about

14:54 teaching that's very relevant so I

14:56 already mentioned how AI the students

14:57 might use Ai and we have to evolve how

15:00 we teach them to now Encompass this just

15:02 like we changed changed how we taught

15:04 when calculators became a thing um but

15:07 we also have the question about actual

15:08 instruction right if I'm teaching a one

15:10 a grad class to a small group of

15:11 students that's great I sit down with

15:13 the students we basically talk

15:14 one-on-one it's a you know it's a

15:16 conversation um but if I'm teaching a

15:18 large undergraduate class and I've got

15:20 300 people in my class and I'm lecturing

15:22 to them I think it's a real question

15:24 maybe they'd be better off having a

15:25 one-on-one tutor an AI tutor teaching

15:28 them the subject material me I I'm an

15:29 expert in that material and I think I do

15:31 a good job teaching the class at least I

15:32 hope so but I have to ask the question

15:35 may maybe the student would be better

15:36 off if they had an AI tutor so we we

15:39 look at the University right now I'm

15:40 trying to think you know with the AI

15:42 coming if you have a student you know

15:44 someone's coming into you know headed to

15:47 college you know any guidance or

15:49 recommendations you know you look

15:50 forward in the future where AI is going

15:52 to be impacting all the industries it's

15:55 going to be impacting how we do our work

15:58 you know to our point we talked about

16:00 before you know it may be you know being

16:02 a plumber maybe the safest job in the

16:04 future it it is true that any job where

16:07 you basically are putting your hands on

16:08 a keyboard that's a little bit easier to

16:10 to automate using AI because uh you know

16:13 if you're putting your hands on the

16:14 keyboard you're working digitally and

16:15 and that's the domain of these uh these

16:17 these computer systems now if we talk

16:19 about a job such as you know you said a

16:21 plumber or electrician somebody who has

16:22 to use their hands to actually you know

16:24 run some wires in a

16:26 building even if we had an AI system

16:28 that's capable of understanding and and

16:30 you deciding what should be done there's

16:32 still the matter of dexterity you know

16:33 being able to actually you know push the

16:35 wires and run them or you know whatever

16:37 it is that needs to be done you need a

16:39 dextrous human to do it right now and

16:41 our robotic systems aren't there so I I

16:43 do think that as we look at jobs getting

16:44 replac by automated AI systems I think

16:47 the jobs that will take longer to be

16:49 replaced I mean eventually they will

16:50 because the robots will get there but

16:52 the jobs that will take longer be

16:53 replaced where are the ones that

16:54 actually require physically manipulating

16:56 something okay and you maybe wrap you

16:59 know I'm Jer mest I'm in networking you

17:01 know Wireless Mobility any thoughts on

17:05 where you see Mobility impacting the

17:07 future of either

17:10 film professorships campuses well I mean

17:14 one of so one of the big limitations on

17:16 AI systems right now is that they

17:18 consume in a monstrous amount of

17:19 electricity um both for training but

17:22 that can happen in in your server Farms

17:24 um but the then there's also the real

17:27 the actual real-time use of it when you

17:28 actually use the AI to answer a question

17:30 or draw a picture or whatever and right

17:32 now they use just tremendous amounts of

17:34 electricity and the idea that you can

17:37 now and more than capable you know

17:40 practical car around in your pocket

17:41 right it will drain your phone battery

17:42 like that and so you know people are

17:45 looking at ways to make the the

17:46 electrical demand smaller but one of the

17:48 big important ways that you can address

17:50 this problem is you move the computation

17:52 into the cloud and now I have my device

17:54 that I'm carrying with me whether it's

17:55 an apple headset or a phone or whatever

17:58 and

17:59 the computation that I'm using

18:01 personally as I move around in the world

18:04 is uh being relayed from the cloud okay

18:06 so this is like moving computation to

18:08 the edge this is like AI at the edge you

18:10 know do I want to do this on my phone or

18:11 back in the edge you know the other

18:13 thing I'd be curious about is you know

18:15 in your classes now or in the film

18:17 industry you know are you moving to gpus

18:20 are your classes actually using these

18:22 big GPU AI clusters well yeah yes they

18:25 are I mean we we've been using them for

18:27 a while because we also did computer

18:28 graphics on the gpus Once Upon a Time we

18:30 actually use the gra GPU Graphics

18:32 processing unit to do Graphics once um

18:35 but now now more and more it's been used

18:36 to do a do AI type stuff um if I if you

18:40 don't mind plugging I'm involved with a

18:41 startup called juice juice um juice

18:43 Technologies and you know their their

18:46 whole product is a way of virtualizing

18:48 GPU so that you've got a machine here

18:50 your GPU is off in the cloud and you can

18:52 access that GPU as if it was on your own

18:55 local device and that makes the it

18:57 mainly makes your programing development

18:58 a lot easier um so so yeah we we you

19:02 know gpus are used all ubiquitously all

19:04 over the place Juniper's big on uh

19:07 networking for AI and that seems to be

19:09 one of the big transitions happening you

19:11 know moving from this x86 data center to

19:14 Bly a GPU data center you know they

19:17 almost call front of the house you know

19:18 we have x86 in the front of the house

19:20 and we have gpus in the back of the

19:21 house and it seems like you know Nvidia

19:23 is now the third most valuable company

19:25 in the world you started it with the

19:28 graphics right all that GPU stuff

19:29 started with digital Graphics yeah in

19:31 the in gen AI now it seems just be

19:34 driving an exponential growth well you

19:36 know uh when when there's a gold rush

19:38 you make your money selling shovels

19:40 James thank you for joining us it's been

19:41 my pleasure and look forward to seeing

19:43 you on the next episode of Bob fry talks