Technological Directions (Panel 1 of Technology and Society in the Next Generation conference)

[Music]

i would like to thank everyone for

making the trip

uh whether it's a few miles or a few

thousand miles

um to our first panel

okay so um

the conference was organized uh by

myself martine sanchez jankowski

and lynn chancellor who are sitting

there

um

and i myself am a senior researcher

at the institute for the study of

societal issues which is part of the

berkeley campus

uh as martine

was the director of the institute

um

and as you know from last night if you

were there

the conference is sponsored by hunter

uh cuny and the berkeley institute

um and we have funding from the sloan

foundation and the gerald huff fund for

humanity

um and so uh we're going to hear today

about the technological developments

that will form the basis for the rest of

the conference

and so we have four wonderful presenters

to talk to us about some recent

technological innovations

um and we're going to start with

dr pamela silver from harvard

who is a systems biologist and a

bioengineer

dr silver holds the elliott

and oni adams professorship of

biochemistry and systems biology at

harvard medical school

and she's known for her pioneering work

in synthetic biology

so uh

we will uh have her talk for 25 minutes

i think we have time for that yeah

okay thank you

good morning

and good morning people on zoom

i want to think about

how we make

a beautiful sustainable world

by 2050 for the 10 billion people that

will be occupying it

and i don't i also want to mention that

in that time period maybe there'll be

people on mars even and

the technologies we develop

for

going to mars are going to be some of

the same ones that we need

to sustain a beautiful life on earth

now this conference is about technology

and i i want to do a little

framing through the lens that was

mentioned last night of

1930 forward

and thinking about the early part of the

last century where

synthetic organic chemistry was

the technology of the day and that

brought us of course

nylon

plastics remember the the film the

graduate right plastics um

look where that got us but lots of

things that have been um

i think

good for um

for for the growth of of uh

of the world in a better place and that

led to the better living through

chemistry which was a

theme that i grew up with

now mid-last century um i'm a child of

silicon valley and so this is very um

intimate for me

is the development of the microchip

which then led to the technological

revolution

that we've been living through

i want to argue that

the next

thing is the engineering of biology and

that is what is going to

transform

life on as we know it

we understand so much about biology

given the research of the last

many 50 some odd years

that we are in a place to begin to treat

the engineering of biology sort of like

synthetic organic chemistry and so hence

we call what we do synthetic biology

so let me give you some of the impacts

that we have already experienced in

in this century around the

the um precise engineering of biology

within the environment um

i will talk a little bit about some of

our work on carbon sequestration

huge advances in agriculture

of course

health um we're we're living that dream

commodities you are seeing uh advent of

um

so-called green commodities uh the um

many has anyone had an impossible burger

uh that is a product of synth of genetic

engineering it's labeled gmo

um but there are many other um

products of high of both high and low

value commodities

and then i want to interject safety into

this conversation um both safety in

terms of

say keeping us safe

from

what's going on in the food supply for

example and conversely thinking about

how we safely deploy the technology

okay so i think that we are living a

global biology lesson i'd be willing to

um

challenge you i in the audience is a

very diverse age here but

i don't know the age online but

everyone all your parents for those

younger know what mrna is they know what

an antibody is that's amazing

so if nothing else i mean we've made

this great vaccine but we've also

started to put

the the language of biology into

the into the world

um so people can and and this is so hard

to do right and and so this has been a a

really a wonderful case where you can

imagine people at all levels is there a

problem oh at all levels um

using the language of biology

now i want to mention um

the development of the mrna vaccines

this is a slide from

my friend melissa moore who is the

chief at moderna which illustrates the

speed the amazing speed at which they

were able to get to clinical trials

for the vaccine that many of you have

have enjoyed

um why is that and i have a little bit

of a personal vignette on this

now there's a tendency to say oh

the dark state we weren't ready right we

were behind i want to give you a

different point of view

that in fact

the u.s government has been funding

research in this space for at least

probably longer than i can remember but

i was involved in a darpa program

it started i think about 15 years ago

and

the pre-moderna

was part of that program

and the dedicated program manager of

that program believed in them and

supported them and i just have to give a

shout out for dan watendorf he's very

reserved but he made this happen and it

gave me a window into how

it's darpa it can be perceived as

defense based but

that it was amazing now there's another

thing that happens here and that is what

happened in the first couple days so how

was modernity able so they were they

were ready at the right time if it had

been one year earlier maybe not but what

was remarkable was the speed at which

they could synthesize

the messenger rna that became the

foundation for the vaccine so

two days right amazing so so that's

transformative

if you're thinking of a protein based

vaccine it's going to take a lot longer

so this technology is literally

transformative now

we heard some about

open access openness and communication

last night and i want to give you the

ultimate example of where open

communication basically saved the world

so

the sequence

there's there's a lot of details i'm

gonna maybe get wrong here but the

sequence

of sars cop2 is key to that vaccine

right

uh

how

so the the it was sequenced in china

could have waited until they published

it right it goes to nature it could have

taken a few months imagine where we

would have been

but instead through the process of

communication through

the internet

this person in australia

received the sequence from his chinese

colleagues and posted it on twitter

is that amazing or what

and so everybody had access including

moderna

and so they went from a sequence that

was posted on social media to the first

day of synthesizing that mrna now you

might say

how do you believe the sequence how do

you know this is real and it turns out

their technology was so fast and so

cheap it didn't matter they could give

it a shot so that's another underpinning

of an amazing technology

um

and by the way i have no interest no

personal interest in moderna um

but um

i i just was

stunned at the the idea

that you would take a risk from

something you read on twitter

and we're gonna see more and more of

that

you can also argue the negative side of

that but let's go with the positive now

i want to give you a flavor my real

world flavor of of the vaccine situation

every year i like to go off somewhere

remote on a sailboat and

was not possible during covid so i took

my first trip a few months ago

about a month ago or so

down to some remote islands in the

grenadines in the caribbean

and as you move it's a little bit of a

hassle it turns out as you move from

every island which is a separate country

you have to get a new covid test and

that's a pcr test except for coming back

into the us which is so so this is a

covid testing clinic

i think on union island

and it's interesting this is the sign um

notice it says the china coronavirus

so these are the the perspectives you

get outside the u.s

this was a pcr based same-day test this

is the wonderful doctor who adminis

comes to this outdoor clinic and gives

us the test if we want and other islands

they would just come to the boat

and and this is what oops shoot what

happened there

oh my god

this is what the clinic looks like uh

outdoors note the turtle sitting under

there

these are some of my um co-mates on the

boat

after their their pcr test

so for me this was like a magical moment

because

i was just starting in molecular biology

when pcr was

discovered and developed and look now

it's being practiced

in the developing world um so for me

this was like magic

okay so let me just frame a few other

ideas that are on my mind and and what

i'm been a little tiny bit about some

contributions we've made so so thinking

about programming life um i think for me

the key things are

the ability what does life do it grows

it uses energy it can make energy and it

repairs itself and so why

does this build why can buildings be

made of wood but they can't repair

themselves

imagine

what i want to dream about are

why can't your clothes repair themselves

or clean themselves

this is a house made of living matter

but why can't that house do

photosynthesis and regrow

so i see the challenge

the grand challenge in this idea of

using biology

and this has been the underpinning of

synthetic biology from day one is how do

we make the engineering of biology

easier so everyone can do it more

predictable and faster and i go back to

the moderna case it's the perfect

encapsulation of what you want from the

engineering of biology but

if you're going to build the house

you've got to go to a different scale

and complexity and so we're we're still

taking baby steps here but

we are making forward progress towards

some of these goals which i think will

be realized over the next

30 years

now why is that

first of all

dna is the substrate of biology so

everything you do well you could say rna

in the case of modernity but you still

need nucleic acid synthesis

and the cost of

making dna

has dropped and it's dropped faster than

moore's law this is one of the or sorry

the cost of dna sequencing my apologies

has dropped faster than moore's law um

now as you know or may not know

um

we can

in principle sequence every human on the

globe

there are projects to sequence every

organism on the earth

this is just remarkable think about it

we didn't dna the structure of dna was

solved

oh god 60 years ago or so and now we're

talking about knowing everyone's genome

everything's genome

that in turn leads to an enormous amount

of information that can be used for

engineering

which gets me to the red line which

doesn't look as impressive but is is

pretty good and that is the drop in the

cost of synthesizing dna which is the

substrate for

uh engineering biology

so now

and and you've probably heard about this

you can

synthesize a virus

no problem um it's about

depending on who's doing it's about a

penny a base pair so you can do the back

or actually in some cases about

0.1 cents a base pair depending

so you can do the back calculation of

how long till

how cheap would it have to get before i

can synthesize a human chromosome for

example

we can now synthesize a bacterial genome

for under a million dollars

and that price will continue to go down

so we have the

the we have ideas we have a lot of

knowledge so what are some of the

problems that need to be solved

well i think we all agree that energy is

one of the global challenges

so we think about global energy needs

and that growth in population um that

means

increased consumption and need for

energy and also a number of those users

are going to be in the developing world

now we

uh although i think solar is changing

this up but in general have relied on a

centralized form of energy production

and distribution that's been the

the gestalt in the u.s and other

developed countries

the question i think going forward as we

develop new technologies around energy

that we might we are going to think

about more how to have those be

distributed

available to all and actually have a low

cap expenditure

so that's something that i think biology

can help deliver on

so remember

biology can you is the best user of

sunlight to make stuff through the

process of photosynthesis

so i'll just touch on something that

we've done in my laboratory

which was to create a bionic leaf

which is the interface between

chemistry and biology so this is a case

of real

interdisciplinary work it was done

together with my awesome colleague at

harvard

in the chemistry department dan nocera

who invented an electric catalyst

that

in response to light will carry out the

water splitting reaction that's what a

leaf does that's key

now what do you do with that so it's

sort of a storage problem right and so

there are bacteria

that are capable of

taking hydrogen fixing co2 and growing

so we can create an integrated system

that is all in one place

can take in light or enter any kind of

energy so the energy delivery can be

remote

and delivered in the form of electrical

to the leaf which can then

grow

biomass

so this

this is still sort of the gold standard

in this space

these numbers may not look that

impressive to you but if you look at co2

reduction efficiencies plants it's about

one percent

algae

cyanobacteria are the natural winners i

have three percent here but they can get

up to eight percent and the bionic leaf

still wins at about 10

efficiency

now that gets me to another

huge problem that was the great savior

of the last century a technology

to develop fertilizer which of course

brought us

the great revolution in agriculture of

the last century unfortunately

the process for making nitrogen-based

fertilizers now is the culprit right

it's it's it requires high high high

heat it's co2 admitting the so-called

haber bosch process

so we would like we really should find

substitutes for this

so what we can do is take our same

bacteria in the context of the bionic

leaf and they will also fix nitrogen

different bacteria but same idea and i

want to just give you one example these

this is um

radishes which were one of the nasa

favorite plants um grown in the harvard

arboretum and the ones on the

far right are the ones that got our

bionic leaf grown bacteria and they're

clearly bigger and i could go on with a

lot of anecdotes about how these

bacteria can support

agriculture but better yet we were able

to form a new a new company called

coolabio which is just doing amazingly

at trying to develop this technology for

use by farmers

it's been fascinating to talk to farmers

and understand

their their what they have to deal with

and so it's not a simple case to change

up how you do farming

and that's another thing to think about

when you think about technologies

now i want to end with a few

inspirations here

this from some work i did together with

neri oxman at the um

at mit

um at the media lab so neri is a

designer but thinks a lot about science

and so together we thought the the dream

as i said would be living wearables so

imagine if you had clothes that were

photosynthetic now we didn't get there

but nary has designed um

uh

these these closed structures

that can be infused in this case with

photosynthetic bacteria which is shown

on the upper right the lower right is an

example of something that she designed

that was actually on the paris runway

and this

design here

is what we infused the uh photosynthetic

bacteria into

that's one idea

the other one that we really are working

on is imagine

uh do we have enough knowledge to

make freeze-dried organs or freeze-dried

whole organisms

and this has enormous implications

this is an example

of a plant that has been freeze-dried

it's called a resurrection plant you can

actually buy it on amazon and this is a

time lapse

of growing it in our laboratory after

you add water and it will come back to

life

so

this is a technology that nature already

does but imagine

the the transformation if we could

freeze dry organs or

freeze dry eggs so the problem with the

cold chain which you've heard a lot

about could really be addressed with

freeze drying

now

i know that this conversation will lead

to a lot of questions about

misuse and and we can have i'm sure we

will have a lot of discussions about

that but let me give you a framing here

that i

compels me and first of all i'm

i'm not bothered by gmos i'll just say

it up front in fact i want to i want to

open an all gmo restaurant in cambridge

that's my dream but but let me be clear

that

with regard to gmo food

you

especially us coastal elites can make a

choice

we can choose to shop at whole foods pay

more money get our wonderful non-gmo

or we can choose things like the

impossible burger um we can also it's a

little i think there's a social issue

here as well with being anti-gmo because

in the developing world

gmo is going to be critical so i'm often

a little critical of my coastal elitist

colleagues about gmos because i think

there is a social responsibility about

thinking about their deployment in the

developing world but

this is a choice we at least can make

however

let's think about when you go to the

doctor you're in pain let's say you're

in chronic pain you suffer from

crohn's disease or

colitis or

and the doctor says to you i can treat

your chronic pain

you know

you're not going to be asking a lot of

questions about where that drug comes

from i'd be willing to bet you don't ask

any

and unless it doesn't work right

so

i'm gonna offer

that a real key to

socialization around the engineering of

biology really lies in the health realm

um

and and some way to integrate those two

with sort of real world applications i

think has has some opportunities

all right so let me just conclude um so

in case you were

wondering where the money comes from to

do this work

uh there is national investment

um as i mentioned darpa

has been one of the main funders since

day last 20 years i did a darpa isat

report in

2002 i think so by the way we laid out a

road map for synthetic biology then so

that was 20 years ago it's been a little

slow going

however i'm pleased to say that we

reached most of the milestones on that

roadmap

nsf was the supporter of a synbio

research center which is a big deal

when nsf

puts forward that and and dod has

recently made

an initial investment in a manufact

manufacturing is a big deal here

and so

dod has realized that and made at least

an initial investment in

biomanufacturing

and there are a number of private

organizations that fund this work as

well so i want to end with a few other

sociological issues

so why now so there's an enormous

fascination with this field

from students i get letters from high

school students can i please come work

in your lab which resonates for me

because as i said i grew up in silicon

valley and so you know i got to go work

at uh work in people's

labs then so i want to help all these i

get eighth graders it's just amazing um

and and so there's this huge interest

and there's also a huge interest in

building the bioeconomy so there's

there's a huge investment interest and

we we were listening last night about

issues around supply chains and moving

things around and

and of course if you're growing it at

home

and it really is a carbon neutral

process

it begins to solve a lot of those

problems

so we have a generation also i i believe

i'm hopeful that really does want to

solve real world problems and i'm going

to point out climate change and as i

said we also know a huge amount about

how biology works we don't know

everything by any means so we need to

keep funding basic research

we need to learn more

opportunities to combine

disciplines machine learning is key here

and i want to end with

as this grows

and and i think this will resonate for

the aiml people but a need for a trained

workforce we are desperate i mean i work

in the center of the universe in boston

in this area now you guys from

silicon valley can argue with me but i

think i'm at the center i feel like i'm

at the center of the universe and

that said the employment you know as you

have industry growing

remember every startup with a series b

has 200 jobs open layer on top of that

pharma

moderna 800 job openings can and then

plus poor academics what are we going to

do where are the new ideas going to come

from

it we are in we are in trouble here and

i don't have a solution but

i'm thrilled at what hunter is doing i'm

thrilled at any efforts

but we have a national problem so i will

end there thank you

totally fascinating um and a great uh

start

uh to our panel

so we will hear

from uh

dr anita raja who is a cuny

uh professor

um

and she teaches computer science um at

hunter college and she's a member of the

doctoral faculty in computer science at

cuny as well

and so now we'll hear about the ai side

of things

thank you jeremy and

thank you all for being here

and i'm excited to be part of this

conference

so

my talk today is

you know to introduce the definition of

artificial intelligence and what it

covers in

in the current day world and what and

the challenges that we are facing so

that we are prepared for the future

so next slide yeah i should sorry

so i'll start off uh with the definition

uh because there is

i think a little bit of confusion of

what artificial intelligence is and what

machine learning is and i'd like to make

a distinction uh that this is all not

you know the same

[Music]

general definition so let's start off

with ai so what is ai so ai is the study

and design of intelligent agents where

an intelligent agent is capable of

taking perceptions

from the environment

doing some sort of computation or

thinking and then acting on the

environment right so

so that's that entity which is capable

of doing this kind of processing of

information

machine learning on the other hand

a simple definition would be patterned

recognition and

and the ability of a system or a set of

algorithms to improve as you give more

data

and so

an ai system could have several machine

learning algorithms within it because

the ai system will may have to do

understanding of language it has to do

planning it has to cooperate with other

agents and so

at least in my view coming from the ai

side i see machine learning as part of

an integral part of ai but not the

complete definition

of ai

so um

clearly you know just like pamela

discussed we are in an interdisciplinary

area within ai uh

and i think even from the very

foundations of ai um you know we have

cross-cut economics we have to bring in

ideas from philosophy or sociology um

engineering of course and now you you

know some of the projects i will talk

about is how it relates to biology and

uh clearly when we talk about policy

matters you know how does that uh uh

relate to law and so on so mathematics

of course is the foundation of both ai

and

machine learning as well

so

you know very important i think i'm glad

yesterday paul talked about the history

and i wanted to make sure i included the

history of aid so that we can talk about

the

the ups and downs the hype and then not

so great times of or the winter of ai as

it were

so ai basically started around 1940 to

1950

um i think was one of the

new ideas was when mccollum and pitts

tried to model the brain using a boolean

circuit so that was one of the

parallel works that was happening in

addition to

alan turing uh writing about uh com

computing computing machinery and

intelligence and that's where he defined

uh what ai would entail so it's going to

have modules which deal with language

recognition knowledge representation and

learning so at least we were you know he

was able to identify

some of the parts that would be integral

to ai

um and then from 1950 to 1970 i think

there was a

lot of enthusiasm uh in fact this is

where uh you know there was like the

early ai programs

samuels checkers games where we you know

there was this initial uh ai system

beating a human in the checkers game um

and we would officially say the birth of

ai happened

uh at dartmouth in 1956

um and this is a famous quote from john

mccarthy and claude shannon uh from this

dartmouth workshop so it was a group of

scientists who got together and said we

think that a significant advance can be

made

in a.i if we work on it together for a

summer

and they did work on it for summer uh

but here we are uh you know many decades

later uh still trying to understand how

do we move

the needle on ai uh but uh you know at

least we we got

the train started um and so ai became an

industry but as

they started the scientists started

understanding the complexity of solving

ai

we went through ai winter so there was a

darpa funding at that time and

in fact

when people realized that we can't just

imitate human intelligence you know

within a few months

and how far we are from that uh

there was like this dip in in funding um

and

in the 1990s though i think the people

were continuing to work

there was a resurgence and happily you

know people

the scientists were able to move from

what were very strong assumptions uh

about you know linear dependencies and

so on uh to bringing in probabilistic

approaches and in integrating

uncertainty in the reasoning of the

systems that we were building so that's

where we you know started modeling our

systems more to what the real world had

to handle um

so ai became more scientific and uh you

know we would call this a time when

agents were introduced and learning

systems were built as ai spring

and then since 2012 to

present you know i think big data became

the big thing and i think the fact that

we had both uh

computational hardware which could match

the data that was being produced uh and

and

we took the idea of neural networks

which were there between 1950 and 1970

and could make it practical

that's what

really was the uh turning point uh where

where ai became relevant to industry and

that and once we found that killer

application i recall being uh

just finishing up graduate school and

all these discussions in our main triple

ai conference on what is the next killer

application to get ai out there we have

these great ideas

how do we get people to uh to integrate

that into their systems and it was this

notion of big data neural networks and

deep learning

that made that happen

but of course now we are moving through

with production level breakthroughs at

many different angles

which is exciting

so ai is uh

is flourishing you know and as you all

know there's the

we need to have these systems though

that have to have social understanding

and cooperative intelligence so my own

background is from distributed

artificial intelligence which

interestingly was renamed multi-agent

systems during ai winter because

only then could we get funding

and this was pre-grad school for me but

it was a smart move because we could

continue working on systems and then

when ai came into favor we were like we

are back to distributed ai but we were

doing the same researching the same

questions at that time um

so uh the the

yeah so how do we get systems that

you know understand the context that

they are in but also can cooperate with

each other and with humans um to do well

in

problem solving um so some of the

applications you know where these ai

systems need to interact and

rapidly and in a complex way would be

ending poverty and hunger

reducing inequalities

promoting clean energy uh protecting the

planet

offering quality education and in

general leading to better health um so

you know examples would be how do we get

robots to help nurses to deliver stuff

uh to the various rooms in the hospitals

right so um

so my own uh work and i also have done

work both in health and that i'll talk

about today and in traffic how do we get

uh

to you know reduce congestion take

advantage of connected autonomous

vehicles and self-driving vehicles while

humans are still on the road as well um

and do it in the energy in an

environmentally friendly way uh so these

are some of the research questions that

they're working on

and my own north star in research has

been this idea of building cooperative

agent society so how do we get

intelligent software agent systems to

work with humans although my emphasis

has been on the intelligent agent side

shivali will talk about the human and

the intelligent agent connection

which is a nice compliment to uh having

the two talks today today um so

but these agents need to be operating in

environments which are called bounded

rational and here's this notion from

economics right so herb simon uh defines

defined bounded rationality that if you

are going to do if a system is going to

do some computation but does not take

the cost of the computation into account

then the solution that you come up with

is not operational um and so this has

been injected throughout my work so i

bring in this notion of meta-level

reasoning where an agent is or the

software system is able to look at the

problem-solving process from an

end-to-end non-myopic fashion instead of

just the single tasks that you're doing

and you know trade off the resources we

are not only bounded rational we are

also under resource bounds you have to

complete your problem within the set

computational time the computational

processing cost but also other resources

that are needed to solve the problem and

so when we look at this from a single

agent perspective or when we have

multiple agents working together in a

cooperative way to solve problems

it's important to think about these

resource bounds and also uncertainty

[Music]

so one of the projects that i've

been working on for the past decade is

uh is in the area of health

and so it has been how do we

use machine learning algorithms and

eventually builds the decision support

agents to predict and prevent prevent uh

disease and specifically we've been

looking at the pre-term uh births of

babies so babies which are born less

than 37 weeks

so this was funded by nsf and then by

nih most more recently

and what we have found out is that you

know there are great data sets out there

because nih runs all these studies but

how do we leverage these data sets and

um

because there is a huge cost in the

initial part of getting these data sets

which is data cleaning how do you take

this data and massage it

sufficiently enough and extract the

information that you need so that it is

relevant for machine learning right

so we

this is a team with columbia university

hunter college and with cumc columbia

university medical center so we have we

work very closely with the experts in

obgyn so that we have their

input and i think this is a critical

part of any research that you don't want

to be working in your own silos

and just the amount that we learned from

them and vice versa has been tremendous

and it does take time it's been over a

decade um so

we have been making progress in uh you

know the prediction of pre-term birth

but more recently we took our algorithms

which we had developed uh for

improving the prediction which went up

by about 20 percent uh from what the

state of the art was by bringing in our

algorithms to participate in a in a data

challenge that was funded or

uh hosted by the nih um and they again

what they were

trying to get was both industry and in

academia academics to use a

data set that had resulted from a huge

nih study which we were already using as

part of our project fortunately

to to come up with new

solutions so can are there new

methodologies that could

help us in identifying patients who are

at the risk of uh you know morbidity so

mothers who have uh the high risk and

the question that we then took our

algorithms and applied to was

pre-eclampsia ptb pre-term worth itself

is not

a maternal mobility per se but that's

the generalization you know our

understanding of the data and

understanding of the algorithms could

allow us to do

and so we

were one of the seven teams who happily

were able to

win this data challenge and uh

so we came up with new ways of analyzing

the data to identify the mothers who are

at the highest risk but also who are at

the lowest risk of preeclampsia and why

is that important

because if you're at the highest risk

you want to make sure you get the

treatment that is needed

but if you're at the lowest risk you

want to make sure you don't have all

these extra tests or visits or the added

stress uh that could

eventually lead to the disease itself um

so

happily our work was recognized and we

are continuing to do

work on this so i just wanted to give a

brief description of what the data set

is like so you get an understanding of

the scale of this so we had

the data

the study was run from 2010 to 2015 and

data of about 10 000 first-time moms

nulliparous women

um and

and that's one of the most difficult

sets especially for pre-term birth to do

prediction so that's something we

we wanted to focus on because

usually the indicator for instance for

pre-term birth is whether the mother had

a prior preterm birth and that's that

just increases the risk and the care

that would be given

so

we

were happy that we could get this data

set first time mothers and there was a

significant

number of people mothers who had

received

who had resulted in spontaneous peter

but which is another

uh

outcome which is very difficult to

predict so this was a important data set

for us and the data had been collected

over four different visits um so there

were about

i think

yeah so there were three thousand

features from each visit and a feature

would be

information about their blood pressure

or their temperature their ultrasound

information any other surveys

that you know about their socioeconomic

information so we had clinical

socio inform in socioeconomic

information and most importantly we also

had genetic information for the first

time in our study so we're able to

combine this uh this data and uh push

forward the

the prediction results and my interest

specifically is not okay great you're

able to predict who's at high risk so

what what can you do about it and so i

brought my background

um in scheduling and planning to

to to look at prevention right so how do

we take the necessary steps and

specifically i'm interested in doing

early predictions so that you have

enough actions or contingency plans to

handle um what is happening to the

patient

so you know so i won't go into all of

the detail but it took us a year and a

half just to take this data and clean

this up so that again it was amenable to

the kinds of algorithms that we um that

we we had to handle and some of the

issues where there would be

contradicting data

there's missing information you know

because this is real world data so

even that

handling those issues has resulted in

new algorithms for handling missing data

that we are able to share with the

community

so in terms of results so what i'm

showing here is we compared for the

challenge

those who were going to have

preeclampsia versus those who

who did not have like hypertension at

all so preeclampsia is related to uh

blood pressure and

our algorithms were able to show that we

had an in significant uh

improvement in performance when we were

able to use these machine learning

algorithms the other things that we were

able to do and a couple of slides um

images here uh is that i think this one

yeah so this were

can we identify features um you know

when a clinician is looking at a patient

they might have five or six features to

figure out who's at risk but what

machine learning is able to do is give

you more details about the important

features that we're handling

um and also rank order them that this is

what you need to look at um and finally

we also were able to look at thresholds

to give us early indicators so one of

the results that i'm showing here is

that um if the bmi was about 26

for the patient at this particular point

in time of their visit it's actually you

know often you'd be like oh it's 26 and

they're pregnant it's okay but what we

saw from the data was that could be an

indicator that we should take care of

and specifically in our work it was

important to look at subgroups and i

think this is an important issue to

to consider when we come up with machine

learning results that you don't over

generalize what you're looking at

because we had patients from

many different from diverse backgrounds

um in terms of race in terms of

age in terms of socioeconomic

situations and so

right now our analysis is how do we

break this data down so that we are

giving the right prediction and the

options for the right set of patients um

so this is exciting research that is

ongoing um so we want to again that's

part of what i'm saying we want to build

a variety of models so that the same

cohort can be given a chance um

to to both emerge and capture the

different subtypes of uh preeclampsia

another uh

highlight of this work was uh when we

did produce the correct model we wanted

to make sure that it's fair to our uh

to the group that we are studying and

what we noticed was there was inequality

and

i'm not going again these are

uh more technical details but what i

want to emphasize is that we found out

that

there was a high

false positive rate for the uh

african-american

population in our group and that's

because uh there's a smaller

representation of that group in the

training set but when it came to

prediction we were actually giving a

higher rate that of of what was actually

happening among the patients which means

uh a patient was

being could be falsely diagnosed with pe

with preeclampsia and the corresponding

actions could be taken and so we have to

be very careful when we build these

models uh to make sure is it is it fair

and there are you know existing methods

and things that we have developed to

ensure that we take the model and just

don't

implement it immediately but to see that

it's fair and balanced um and so we were

able to then fix the issue uh by by

determining these cert these cut offs

and

and the other thing we found out was the

asian american population was being

under diagnosed so while the

the black race was being over diagnosed

there is under diagnosis of another

group and again just because they're not

fairly represented in the data set it's

not an equal representation um so

so these are i think long-term lessons

to take for all of our as we work on

bringing in machine learning

and eventually decision support systems

and ai into this

work

a second project that i work on is this

idea of connected autonomous vehicles um

so again as i said i come from a

distributed ai background and the key

idea here is that is this notion of

selfish routing so what is selfish

routing it's the

uh you know desire of each driver uh or

each vehicle on the road to optimize

their own path and because everyone is

doing that that is actually bad for the

entire society and we can sort of map

this to a lot of different applications

right so when you are not coordinating

with other agents and there's a limited

resource um and we there are a lot of

interesting paradoxes on

what happens with selfish routing where

everyone for instance for it's for us

it's i-95 or if we take route one um you

know everyone goes to the highway and

instead doesn't take the alternate

routes and hence the highway is super

congest congested right um so some of

the things that we were

trying to bring in is how do we take

this idea of connected

uh nes in vehicles uh especially now as

we move into self-driving cars or

partially self-driven cars um and also

this idea of platooning um

not so great for new york city but let's

think about it you know in the big

highways of texas or something where we

can get groups of vehicles to work

together to sort of coordinate together

and be at very close so there would be a

lead vehicle and it determines the path

for the other vehicles that are part of

this platoon and they are very very

closely uh connected um and we can talk

about long-term

routing plans so that's work that we

have been

doing

so some again a lot of numbers here but

what we're

doing in simulation is looking at

different car following algorithms and

we were able to show that by cooperating

the average speed of the

agents actually increases and the

average travel time drops while we can

also improve

fuel usage and reduce the

emissions

so this was done in this complex uh

simulation called sumo and we are

bringing in multiple agents each car not

knowing

what the other car is going to do but

then they communicate with each other so

that they are able to come up with a

solution which is desirable for the

entire group

so having uh discussed um

some of the projects uh what what what

are issues you know i've already

mentioned this issue of bias and making

sure there's fairness um

we also want to address the issue of

lack of transparency as we do build

these decision support systems and

accountability of the algorithms um

you know

especially when it comes to health or

major finance decisions and you're

making recommendations whether someone

should get a credit uh loan or not um

you want to make sure your decisions are

transparent and so that you're being

accountable

um so yeah i think i'll just skip

through so these are just some examples

you know where we want to bring in

policy uh and uh

and where we have this social network of

of systems which need to work together

in a in a successful fashion

so here are some ideas maybe i should

let me

just pause this and go to the previous

slide

of how of some ideas that i'm currently

looking at to address these issues so as

i mentioned already we are

whenever we build our models we are

checking for fairness but i'm sure there

are

you know

broader ways to uh to consider this in

terms of societal impact

and policy

here's an example i'll just let you

watch this simulation and then i will

describe this

so the goal of this boat is to complete

the circuit uh which is being

described up here right so it's a race

and the maximum maximum utility that the

uh

boat can get is by going through this

circuit but instead what this boat has

learned is by uh

and this the other option is you can

also get some utility by

encountering these turbo

uh pellets as it were um and what the

boat decided is to just skip out of the

race itself and just go through and pick

up all the pellets that are out there

and it was able to maximize its utility

right so

what is the concern here i mean clearly

there is a concern um so

instead of just putting it up there i'll

just

yeah so

what is value alignment um and

this is just a simple example of where

this is you know there's x that you

would like your agent to be

following that the goal for your agent

is x instead it has decided to go up go

and figure out an x prime and maximize

its utility because when it comes down

to how we define

the goal for an agent we are saying

you know here's the problem and what we

want to do is we have defined how you

can get utility and we narrowly say you

have to maximize the expected utility to

win this game or to do well in the cyst

as a system

and that is not sufficient as we build

more complex problems we need to

make sure that the val there's an

alignment of what is considered

useful value uh for the agent and what

the human or the designer would like the

system to do and there are multiple ways

that are suggested so stuart russell um

and you know has one of the ideas that

has come out from his lab is when we do

value as alignment

you don't have to perhaps define the

details of the human preferences just at

a high level

provide the preferences to the agent and

there's this whole area of inverse

reinforcement learning but then the

agent will figure out okay what is it

exactly that the human wants to do and

this is important when we when we direct

an agent to do things which are mission

critical especially right if it's health

or if it's defense

because it could just say oh this is my

goal and at the cost of

ethical decision making or the cost of

the environment or the cost of uh

comfort of of others um so

i think giving a little more freedom

that's one way that one could move

forward with value alignment um to so

that the agent has the freedom to figure

out okay what is it that would help me

get the preference of the human rather

that being too specifically defined

there's also a lot of work as you know

you know the issue of ethics especially

in the past five years

as we have

seen ai being used

in industry there's been a lot of

discussion in our

community

happily it's being integrated into our

education i think having students not

only be aware of what the technical

details are but what are the impact of

their work and this has to stop ground

up

not after they get the job and build

their algorithms which might have uh

challenges uh ethically or you know

societally so um

so

part of these discussions as a research

community have been have been

important and across the board and one

of the other areas that i'm interested

in is building this notion of

cooperative ai and differential

progress there was a

recent paper in nature that

discusses this and i think i put the

link and what are some of the important

uh

characteristics of cooperative ai we

want

systems that have the ability to

understand the consequences of their

actions both on other agents and on

at the human society at large

and all the consequences on the

environment

we also want agents to be

[Music]

communicating

with each other cooperating and

commitment so these are i see these as

almost like research facets how do we do

this better how do we build algorithms

which are transparent unbiased and are

able to be accountable to to follow each

of these research goals

we also want agents to be able to

identify norms and to follow

what are the social norms that are out

there so again when you're put in

different societies

or different uh

groups uh how does the agent identify

what

norms are and some of our past work my

group's past work has been

identifying this type of network that

one is in and doing fast convergence how

do we come up with algorithms so even in

social networks right uh

social network is an example and then we

could have random networks we could have

scale other types of so social network

is a scale-free network um how do we

come up with convergence when networks

have different characteristics so first

identify the type of network and then

have

uh an understanding of identifying what

the norms for that

society would be and uh uh

and achieving that norm

and finally uh integrating this notion

of network thinking has been uh

at least the past decade uh an important

part of my work so how do we

the agents again it it's not too

different from what multi-agent systems

are you just don't think about yourself

but also think about the society at

large

most of my work has been about other

agents but then you know slowly also

integrating the preferences of humans uh

while there are others who are doing

work from the other end like how do we

look at agents and humans and then

move forward with our solution process

so bostrom

and this

is part of the cooperative ai foundation

has written a lot on these

ideas of of differential progress and

this is the notion where

you want to accelerate and i think this

would crosscut any of the technologies

we're talking about today

how do we accelerate the implementation

of beneficial technologies especially

those would that would reduce the

hazards posed by other technologies so

we need we can't be again working in a

silo or in our own narrow box

as other technologies advance how do we

talk to each other and i think

conferences like this are very helpful

in understanding

the impact

of our work

so in conclusion

you know happily ai has made tremendous

progress in the past

decade or more

um

and

i think there's

you know

now instead of just moving from games so

initially when i got i was doing uh my

graduate school we were looking at

machine learning it was mostly just

games and uh

and then it was a game called alphago

which put us on the on

on the radar of industry where

this system called alco beat the best go

player in the world lease it all

and

in a very different fashion from the

chess game that was played in 1997

against kasparov kasperov

so the difference was with kasparov when

casper was beaten uh it was more a

complete search whereas here

the system in alphago was able to

self-learn and come up with algorithms

that no human had even thought about you

know it's moves that looked really bad

to the experts and eventually one uh

beat the experts so um so we are

definitely making progress and and games

are just an abstraction of what we can

do in in real-world situations you can

take that and apply it in traffic you

can apply it for um you know

like farming

and agriculture how do you

get to optimize your resources and get

the best results outcomes that you would

like

but it is critical to consider downsides

i i put a more positive uh

you know radiology often gets beat up in

at ai conferences because

you know vision has gotten so

vision research has gotten so good that

often the you know the discussion is oh

radiologists are going to be replaced by

ai but here's a quote which i completely

agree with you know that i think humans

are important in the loop we are not

replacing the human in the loop

but definitely ai won't replace

radiologists but radiologists who use ai

will probably replace the radiologists

who don't

so it's not just for radiology but in

general i think ai is an assistive

technology

which we should leverage so that we can

as humans be able to focus on the more

complex

problems and ai is able to do the things

which you know which might be easier uh

in some sense but computationally

intense as well so

we

hopefully humans would be

freed to do things which are more

interesting and more complex as a result

of ai so that is the goal that is what i

hope to train my students to do instead

of you know replacing

jobs so that's always a concern one has

to

be aware of and there is the role of

government you know it's important that

government recognize the importance of

ai give it the freedom to do the work

where it's needed but also set rev uh

limitations when things are

are going very right so if uh

so they need i think the government has

to keep people informed about what is

going on in the research and that is

when you fund the researchers the

researchers give the reports and that

information and i think it's already

happening

our reports are becoming more detailed

you know be very detailed of how this is

going to affect society but i also think

that uh researchers should

identify what the potential bad effects

are of their work so let's give that

give give a thought even when we are

publishing papers um you know so what

are the not just the limitations but how

can this you know what could go awry and

what do you need to do about it

um

we couldn't cover the whole space but at

least we start thinking about these

issues

and finally i think there's a role of

the ai community

to share data this is always a

challenging issue

even more so within health but that's

where we need data uh you know 10 000

patients

is with 3 000 features for every visit

sounds a lot but that's not enough for

the kind of work that we are doing we we

actually need like you know five times

more and the right type of data that uh

would help us with our learning process

um also i think the community has to

avoid hype uh unfortunately our field

does go through these cycles of of of

too much hype and it might be from the

outside so

researchers need to

be very specific about what their

results are what the challenges are and

they need to connect to the users um

and finally you know

i think the goal of our research should

be to empower people and not to devalue

uh

humans so that if that is clear and

that's why we are moving forward um i

think there's continued great excitement

for the ai community so

thank you for your time

[Applause]

okay

thank you

for that wonderful talk um so next up

is dr shawali mohan

who is a senior member of the research

staff and principal investigator at the

xerox park

organization in california and she works

also on ai

with a special attention to human

machine collaboration

good morning everyone i'm very excited

to be here i'm shivali

i'm a member of research staff at xerox

park some old-timers may know it used to

be a famous research lab we still

believe we are doing good things but

here i am

i titled my talk humans of ai and

well it's a tip of a hat to the famous

collection of photographs called humans

of new york and we are in new york so i

was excited to title my talk this but

also this is a critical

introspective talk as an ai scientist a

lot of the times as ai scientists we

tend to get super excited about

algorithms and systems and efficiency

and models

but we forget where these that some hum

that humans that these systems are being

designed for humans and those humans

would like to do something interesting

with these systems right so i want to

sort of turn our discussion over its

head and think about humans of ai first

humans who are using ai systems to do

something productive and that's what

i'll be talking about today all right

like with any good ai talk i'll start

with some inspiration from hollywood so

we've been fascinated over past you know

century fascinated with ai systems right

we envisioned these societies where

little robots and humans are living

together doing good things and what's

exciting about these things i think is

that these

entities are intelligent collaborators

they're independent long-living entities

they're goal driven they solve problems

they interact and communicate with

humans empathize with them and they

learn from the experience right that's

what makes them these entities very

exciting but if we look at ai research

now it doesn't look like what we thought

ai would look like right if you are

clued into the hype cycle of we recently

got to know about dali too which can

it's a trillion parameter model

that can generate images that look like

art right so

the focus has been on larger models

larger data sets and

beating the state of art and that sort

of ends where that's the end of ai

research and this doesn't look like what

we you know hollywood taught as ai

should be

right so the question that my research

and other systems

and you know anita is part of that

community they ask is will algorithmic

research by itself lead to intelligent

collaborators that we were excited about

and the answer tends to be no you

actually have to put all of that

algorithmic research machine learning

research and ai systems and there are

communities and people who are exploring

that right so anita showed you what an

ai system would look like

the view right so ai system living in

the world it perceives the world it

represents the current state it knows

certain actions that it can do it thinks

about what's the right action given its

goals and it acts right so and on and on

it goes it's trying to achieve certain

goals in its environment

however

if you deploy this system there is a

critical part of this equation that's

missing there's almost always going to

be a human involved in this loop

somewhere but the the

science doesn't know where to put the

human in right so we don't know how to

put the human into this loop the ai

system is is is operating in and this

has been like recently people have been

starting to bring bring this question up

right so we have articles that are

coming out that say

that you know we in ai science we don't

know how to represent humans how to

model them how to interact with them

and this is where i think social

sciences and ai science computational

sciences can interact because social

sciences have studied human behavior

human learning

human

human lives and we can bring those

models into ai to

you know develop better collaborative ai

systems right so that's where a lot of

this talk and my research

will go in go in the direction office

how do we build intelligent

collaborators that are designed to

support the goals of a human partner so

we are starting from this goal that we

want to

support a human doing a task we want to

model the human partner explicitly we

want to understand how they will react

uh to changes how will they learn how

will they behave and then we are really

focused on figuring out ai systems that

will have effective performance on human

tasks so we are trying to move away from

computation-centric metrics of

efficiency largeness of the model

accuracy to more

uh metrics that focus on effectiveness

human tasks right so we want to look at

safety and health applications we want

to look at

acceptability and transportation and

i'll talk about those a little bit but

what this does is this opens up the

space of applications that we can deploy

ai in and

i'll talk about a few of these problems

so i wish i had a great you know unified

theory to give you like this is how you

would include humans in ai systems or

models of humans and ai systems but i

don't so what i'm going to do is i'm

going to do this case study style so

i'll present three different projects

that i have worked on in the past and

we'll talk in each of those projects

they're from different domains we'll

talk about what does it mean to model

humans and bring theories of human

behavior into ai reasoning and then i'll

conclude with some you know closing

thoughts all right so the first

topic is interactive task learning and

the context of this problem is so right

so rope as if you are from silicon

valley like i am you know robots are

going to be here and they will live with

us in our world right there's already

robots running around in mountain view

delivering pizza to graduate students so

that's already exists and now amazon is

releasing more robots that will follow

us around in our homes right the robots

are here

the problem is is that it's really hard

to program a robot and everyone's home

is different and everyone really has you

know set patterns that they want

you know their their people who live

with them to fall to fall excuse me

now one way of deploying this is that

every time a robot is put into someone's

home there is like a team of engineers

that comes with that robot and then they

program that robot so that now it can do

the things like it could make a chapati

like your mom does right so but but

that's that's just unscalable there's

not going to this solution will just not

work right so the other way to think

about this question is that can we

design robots that can be

that can be trained by the end users of

those robots right so humans

are natural teachers we teach each other

all the time whenever someone comes to

our home we teach them like oh this is

how we have organized our kitchen can

you please help us right so we want to

leverage this natural capacity that

humans have to teach robots and then

have you know robots program them

themselves through these natural

interactions so that's

where the problem of interactive task

learning comes from in 2017 there was

this meeting

a large interdisciplinary meeting with

people from machine learning cognitive

architectures robotics psychology

computational linguistics all these

people got together and talked about

this problem of interactive task

learning how do we design agents

that

you know learn from and teach humans

while they live with them and so it's a

great book that came out of that effort

and i would encourage you to read it if

this is something interesting

right so

interactive task learning is a very

different paradigm

you know if you think about machine

learning and how machine learning

systems are taught these are you give

them an input you know there's a data

set which has the input signal and then

the supervised signal that goes with it

right that's how those systems are

trained but that's not how humans think

about teaching or training and here's an

example of what

put it on your hands too

oh good drawing try again

on your head on your head

put it on your head

what

oh good try again

good try but you're missing your head

you got to put it here on your head

you want to try again

you need help help okay

okay so

this is what like natural interactive

task learning is right so that was my

friend's son ishaan and he's talking to

his nanny satsi

and they're trying to basically said

he's trying to teach isha on how to put

a hat hat to how to wear a hat and you

will note i mean and this is not

surprising but there is no like training

phase testing phase training phase

testing phase it's all incremental

there's experience that ishan is now

analyzing and then extracting useful

information from so that he would be

able to do the task right it's all

online

you know classical machine learning

setup would like okay let's take the

architecture offline retrain it bring it

back online and now it performs

certainly that's not how humans learn or

behave

it's and i think one of the most

interesting things is

this notion that ishan knew when he had

failed at the task right he failed to

orient the cap properly he knew it and

he asked for help and that's very

critical because learners know where

they're failing and why they're failing

and human teachers can really flexibly

adapt to that situation once you realize

that you're you know the person that

you're working with does not know this

component of the task you will really

hyper focus on that right so that that

nature of human learning the active

nature the interactive nature is

critical and that's why we learn so

efficiently right we learn from small

few experiences but these experiences

are highly salient for that task and

then one but the final critical part

about human learning is is this

benevolent teacher right the nanny who

wanted to teach sitsi test so we want to

incorporate all of these notions that

are part of human training into

designing robots right so i the

uh kinds of approaches i've worked with

i'm most excited about an approach

that's built off of cognitive

architectures so you could you've

learned of deep learning architectures

they are not this architecture these are

different class of architectures um

you could think of them as blueprints

for generally intelligent behavior right

so it's an intersection of cognitive

science and artificial intelligence the

research has looked at what is the

computational basis for generally

intelligent behavior in humans because

they are the only exact known examples

of generally intelligent behavior

right and then can we extract something

useful from them and that implement

algorithms and software that operate

like humans right so that's where the

philosophy of the design is

is

basically comes from and they've been in

development for 40 years so there have

been progress there's built systems

architectures that you know do that the

prominent one that i work with is called

soar

and so it's currently hosted at

university of michigan that's where i

went for grad school and we've been

working on this interactive task

learning problem since 2012 right so

this is year 10 of trying to build

systems that can be trained by humans

through natural uh interaction and the

most recent advancement was that we won

the best the the work at michigan won

the best demonstration award at triple

ai so this was an agent called rosie

that could be trained to do certain

kitchen tasks by a human through natural

interactions i'm going to talk a little

bit about two recent advances that we've

made that we've done at park so the

first one was before we design an agent

that can learn from natural human

teaching

it's very useful to understand how do

humans teach and why does that matter

right so

machine learning and a lot of ai people

would have us believe that humans are

only good for giving us label data but

that's not true humans don't think like

that humans don't teach like that right

so we wanted to unpack what human

teaching is and so here's preethi who's

a graduate student at university of

michigan she basically constructed the

study in which she had people come in

and play around in this blocks world on

the top

right corner and use these blocks to

build a wall right and then we analyze

this data and try to get some get out

get extract some general principles of

human teaching so though

and then they were

so

so there's some findings right so the

first one is that people naturally break

down complex tasks into simpler

components right so in in the task of

building a wall people naturally block

broke down you know okay these are red

objects these are green objects you have

to place them next to each other right

so they broke down a very complex task

which was sequencing putting these

objects in a sequence into simpler

components that now the robot can

potentially reason with in a very

limited you know hypothesis space

people use and express the variety of

teaching intentions so it's not just

people are not just giving data to the

robot but they're also trying to

evaluate the boundaries of competence of

the robot so they will ask okay now can

you place the red object next to the

blue object just to evaluate if the

robot actually knows

you know the concept next to people are

flexible and they react to failures in

the robot right so that's a good thing

that means that as

we expose the failures of robot to the

human teachers they would react to it

and would give the robot the right

information and then finally people

organize these concepts into very

distinct what we call curricula

and these curricula were flexible were

adaptive and then they were influenced

by how people you know people's

background so it's a very cool paper um

i encourage you to read it so as now we

know how humans teach we are now also uh

making progress on building

architectures that can learn from that

kind of teaching so the architecture is

called eileen that can demonstrate

human-like learning

and it basically learns in a way that

humans do right so it's an interactive

learning loop where there's a teacher

who's interacting with the agent

using you know a uh situated environment

where you know the teacher is again

trying to teach the robot how to play

with the blocks world and this uh the ai

learning edge as anita said that you

know it has several machine learning

logical inference

uh planning

uh components all put together into a

system so these are not independent

algorithms that usually uh when ai

people talk about they talk about those

algorithms but they are put together in

a system that has interesting behavior

right and what we are seeing is that now

we have architectures that do

multi-level reasoning so at the

bottom-most level we have

inference methods that are reasoning

with the metric space that a robot has

to deal with right the real world the

real perceptions the robotic control and

all of that but then a level up

there is a there's a level of reasoning

that reasons about tasks and goals and

this is like becoming more human-like

where it's thinking about okay what is

the final goal i want to achieve how do

i get there and at the top most level it

is reasoning about concepts in general

right the knowledge that it has general

knowledge it has about the world so we

are we are now beginning to sort of lay

the foundations of an architecture that

can reason at multiple levels just like

humans do and we found that you know

this architecture learns in a way that's

that very similar to how humans learn

right so it can learn various types of

concepts it can ground natural language

onto those concepts it learns very

quickly generalizes rapidly because the

trainer can give it salient examples and

then like you know like humans it's very

opportunistic in its learning it can

figure out where it fails ask for an

answer and then learn from that and so

what i wanted to highlight is that as i

come to the close of this particular

case study is that design of ai systems

is not just algorithms and data right

it's it's also cognitive science because

they're working we are designing these

systems to work with humans so we need

to understand what those humans are

doing

and with the structural guidance from

cognitive science and psychology we are

able to define the desiderata for ai

system design and then implement those

algorithms right so it's it's an

interdisciplinary pursuit

all right so several others of my

colleagues are doing wonderful

fascinating work on human robot

interaction and collaboration all of

which is focused on trying to build

robots into human spaces and make them

safe to to live with humans right so

it's it's a great field there's a

conference that's a yearly conference

called a human robot interaction so if

there's any interest you could go look

up those papers

all right so switching gears a little

bit i'll talk about sustainable

transportation

um so

a lot of you live in new york i come

from california so this scene is very

common where all the highways are choked

up and anita already talked about when

you know how ai can help solve that

problem right congestion wasters waste

several billion hours of time and

several billion gallons of fuel per year

so rpa transnet back in 2013 came up

with this program where they said hey

can we do some can we use technology to

help solve this problem right and we can

like we can come up with energy

efficient routes for everybody who's in

the transportation network but there's a

critical part missing right so you can

come up with the most energy efficient

plan and you can tell it to the human if

the human doesn't accept that as the

recommendation to follow you have done

nothing right so there's a critical

human component to this problem where

you have to influence the human to do

what the system thinks is the best thing

to do so that's what we studied as part

of um the paper here in jer is is what

does it mean to influence a human to

follow a route that a you know an energy

efficient planning system is

recommending so we define this problem

we call the influence problem we imagine

there's dr jane who goes to her office

every day at uh in the morning at 9 00

a.m

and she has a

assistant installed on her phone that's

mo that has access to her calendar that

has access to the transportation network

that dr jane is embedded in and it knows

that you know she can drive she can walk

and she can take this specific bus and

it also knows some personal aspects of

dr jane right so it knows that she's

employed in a regular job she cares

about her environment values her free

time while she travels right so using

all of this information

copter which is the agent that we built

can at a time lead

at a

in a timely fashion at 8 45 a.m come up

with an acceptable option that hey if

you walk to this bus bus stop near your

home there's a direct bus to your office

and that's what you should take and it

makes this message compelling by saying

that by accepting this route you would

reduce emissions to 10 by 10 percent you

would contribute to

reduction in image emissions right so

there's different aspects for

influencing humans

and the paper lays out the mathematical

framework of how we can do that so that

you know ai computations can process

that model

and then

personalize the recommendations to dr

jane so that those recommendations are

actually adopted by dr jane right so

some of us may be familiar familiar with

this theory uh from behavioral economics

choice theory which projects the choices

or the route recommendations that you

may have for dr june into her personal

utility space right and when i'm saying

hey you shouldn't drive but you should

take the bus i'm asking her to pay a

cost in that in terms of that utility

and the probability that dr jane is

going to to accept that recommendation

or adopt

that recommendation is going to be

inversely proportional to the cost so if

the cost of transitioning to bus is

lower than transitioning to you know

biking to her office then she's more

likely to go to the bus right so that's

the main idea and once we define the uh

influence problem that way did we define

acceptability that way we can bring

those notions those notions of human

modeling into ai planning and we can say

that instead of planning

for you know the most energy optimal

route we are trying to figure out routes

that lead us to maximum expected energy

savings right so instead of going from

actual expected

to to like the most optimal energy

expect uh sorry optimization we go to

expected energy optimization

so a few interesting findings first was

that we built we were able to build a

machine learning model that was able to

estimate acceptability of a mode

using a large data set and this model

can um you know predicts with diverse

feature sets and can capture complex

non-non-linear relationships

we found that acceptability does

influence adoption so we validated it

through a choice experiment with

participants from los angeles and we

found that if we set up the system the

way we did

people would actually stop you know

reduce their driving and move to more

sustainable modes of transport and then

we built an agent-based model to our

simulation of los angeles transportation

network and simulated agents or humans

in that model to see if humans were

behaving like our model said they would

would it even lead to energy savings and

we found positive signal there so that

would lead to some significant reduction

in fuel and delay right so that that's

positive news we weren't able to really

deploy this model but i think we have

set up the right sort of evidence to

show how

you know models from behavioral

economics can be combined with ai

planning and together they can solve

this complex problem

all right so very quickly i'm going to

just go through the final

case study that we had

this was

an nsf nih funded

project

and the premise was that

behaviors rooted in sedentary lifestyles

impose major healthcare costs

right so these would be people not

exercising enough not eating well enough

and that lead to severe

health problems that increa uh you know

puts more cost into the healthcare

system

so the project was focused on like how

can we design technology that can

support people

in disrupting non-healthy behaviors and

building healthier behaviors

and one of the most impactful ways of

doing this is through human to human

coaching where a coach sits with a

trainee and helps them you know figure

out what's working in their lives what

isn't and what's the right way they can

change some of their behavior so that

they get healthier

now with the advanced advance of mobile

technology and artificial intelligence

there's lays open the space of what can

an ai coach that lives in your phone do

similar things

right and to design this ai coach

we really need to know what underlies

human behavior and how that behavior

changes right so it's a very cool

problem for ai and ci and cognitive

science and that's why i'm super excited

about it so at park we built two systems

that are designed

um to

solve parts of this problem so new tree

walking

is an exercise prescription

algorithm it's embodied in a mobile

health coach that lives in your phone

and the goal of the system is to get

people walking at the american heart

association recommendation right which

is 30 minutes of moderate intensity

walking five days a week right but if

someone who's super sedentary you can't

really ask them to

walk at that level because first they

might hurt themselves second they will

fail

and then third they will never do it

again because you know they hurt

themselves and then they failed so you

have to put them on this ramp that

slowly progresses them towards this goal

so the the you know our research looked

at what would it mean for an ai system

to have to build out this ramp and have

people you know walk up this ramp slowly

so that they're successfully able to get

to that

right and again it's the similar kind of

an approach we built a model of human

behavior of human walking and how human

aerobic capacity grows and then we tied

it together with

our understanding our models of

motivation so what motivates people to

uh

to pursue certain behaviors how do they

set goals and things like that and then

together that was brought into an ai

heuristic scheduling algorithm

framework to then build an agent that

could actually you know get people

walking more

um so there's several firsts with this

this kind of work first was this was an

actual

longitudinal ecological study of ai so

usually when you think of ai systems

they are very transactional right so you

put in a search query outcomes a page um

you

and usually that's how ai systems are

projected at but this was you know a

long-term ai deployment it worked over

six weeks right so the ai actually had

to have adaptive behavior for six weeks

right it was ecological in that that

people were not brought into labs to

study the behavior of ai people actually

installed this coach on

on their mobile phone and lived their

lives as they would right so we weren't

changing anything about their life we

actually were trying to influence them

as they were living their daily

lifestyles and we found that as we you

know build these

ai systems with human models

they are effective like people are very

able to build healthier behaviors and

what this research also made soup made

it super critical to me

is that ai

usually ai scientists think of

evaluating ai in a very computer

computation-centric manner in which we

look at efficiency

we look at accuracy and

we look at you know the speed of

computation and things like that but

when you bring these ai algorithms into

a context that's critical for humans you

have to revise how you will be measuring

efficacy right so i was working with a

physical therapist on this project and i

told her that you know we you know this

system is very highly accurate and she

said i don't care about accuracy at all

i only care about safety of these

systems because i don't get to see my

patients if the system is deployed which

means that they are likelier to hurt

themselves so i don't care if you get

them you know walking

more at the faster speed i i want them

to be safe right so that exposed to me

that as we're building these systems we

have to rethink how we are evaluating uh

what these systems are doing and again

it's like follows a similar you know

philosophy combining multiple different

fields together with ai to build uh

productive systems

all right so

just ending um

when we think of air trees when ai

scientists usually think of ai we think

of this deployment scenario right

there's an ai system that's interacting

with the world and i delegate some tasks

to ain and they get done right that's

how typically

ai is thought about but actually it's

much more complex than that there's

going to be various different

collaborative interactive

experiences with ai that humans are

going to have and unless we study humans

their behavior their goals and why do we

do certain things we won't be able to

study these really important

ways of creating human

ai pairs

and

that will stop us from studying these

really critical problems right so i

talked about interactive task learning i

talked about sustainable transportation

and health behavior change but then the

spectrum is very big and almost any

deployment of ai that you can think

about they're always going to be

human or a set of humans involved and we

have to think about what they're doing

and what their beliefs are

right so some gratitude to all my

colleagues and funders

and just i'll end with some key

takeaways so modeling humans in ai

systems is necessary for effective

development that's mostly just focused

inwardly focused it's it's a mantra for

ai scientists

um the second thing that's that's maybe

um

interesting to colleagues here is is

that social science frameworks and

theories have really great starting

points that can help ai system

scientists develop better ai systems

right the third thing is um

you know coming up with the right

metrics and evaluation strategies for ai

so if you evaluate ai devoid of context

you will not solve the right problem so

we need to really focus on that

uh we need to build ai systems that

learn and reason about humans like

humans and then i talked about three

different problems but there are several

others that are similar thank you

oh yeah that was fascinating um

so we're going to now uh have a talk

we're going to go back to the natural

world and technology

uh and we will have our final talk by dr

genna wagner

wagner uh and he is a professor now of

finance at the columbia business school

but he has also served as the executive

director of the solar geoengineering

research program at harvard

and he studies geoengineering and

environmental and green tech

thank you

there we go

um okay so

my mission today is to convince you that

anything and everything

related to

technology and society technology and

climate especially

is

at the end of the day

about moral hazard

now just full disclosure

once you have a hammer everything

feels like

the exact same thing so yes there's a

book where this came from

but um

more importantly it is this

interaction between

um

hey there is this newfangled tech

here's the tesla ai guy

and then there is how we'd like to

organize ourselves a society

this is tesla ai guy

going to europe on vacation

and finding that wait

there are ways to organize ourselves

that do not involve having a city with

600 000 people

and the bay area with 20 million people

around it all right san francisco bay

area

and pretending that that has anything to

do with

how we should organize ourselves as a

society sorry

and anyone else living in silicon valley

but right so

um

maybe this is just me in my third floor

lower manhattan walk-up making myself

feel good about

my life

uh but frankly well welcome to my

twitter feed making fun of bourbon

nights jersey heights and anything

anyone in between

about basically

us knowing how to organize ourselves as

society on the one hand and on the other

looking out for techno fix after techno

after technofix

trying to turn the climate ship

around okay

now big picture

this is about climate

emissions have been going up

forever

yes covet

no didn't solve it

we also know the solution

right we basically have known about what

needs to happen

forever

let's say 20 30 years by now

emissions have to come down

eventually that's the solution

and the big question right is it what's

the right mix technology

societal behavioral changes all these

sorts of things well this is the best

week to talk about this of any frankly

monday 3000 pages

2913 pages yet again ipcc the

intergovernmental panel on climate

change comes out and tells us

why it is a problem

what do we need to do

where things currently are going

all right so lots of detail in these

squint at the red line this is where

things are going

emissions over time

look at the other ones that were that's

where we should be

no

graph after graph of the graph

we are not anywhere close to where we

need to

be

right and yes q headlines like

we know what we need to do

it's the politics

and just to hammer home that point

yes it's the politics of course it

always is

the left is the technical summary that

the scientists give us all right there's

the three thousand page report

there's a 140 page summary

provided by the scientists

and then

the past two weeks of this process

the ipcc process to get political buy-in

every country on the planet has a

line-item veto

power

in this two week negotiation to come up

with a summary for policy makers

for the most part that means stuff just

gets cut

well

sometimes

often frankly

it means fairly clear statements

provided by the scientists

we've got to get rid of fossil fuel

infrastructure that's the short summary

of the thing on the left

gets

converted into oh and by the way we can

add

here's the theme a technofix

we don't actually have to get rid of the

fossil fuel plant we can add some carbon

capture and storage that's what the ccs

stands for on top we can just suck it

back out after

and

up to a point yeah we can

the technology

exists

but

just to hammer home the point with

another headline

yeah we are kind of running out of time

we've been running out of time for about

30 years or you know

1965 the very first blur ribbon panel

putting a report

on the desk of the u.s president saying

this is a problem let's act

right and by the way the science is sort

of 19th century science

this is a couple centuries ago at this

point right more co2 higher temperatures

and then by the way

today is sort of an operative day right

if you look at the summary here

uh emissions have to come down by 20 and

25

high

confidence today

april 7th we have exactly 1 000 days

left until

january 1st 2025 right now okay false

precision these deadlines of course you

know yes they are partly political even

though it is a scientific statement

but still

okay so the urgency is clear let's start

with that

now

sometimes these reports do give us new

information and you know in this case

it's

maybe not new information so much as the

new graph the new killer graph the thing

that we'll see over and over again

anyone in the climate world for the next

years

several years

um

there's a whole bunch of different

technologies on the left

and when you squinted these colors here

the blue stuff pays for itself

so let's zoom in a little bit yes it's

wind

yes it's solar of course it's always

been there's a bit of nuclear there are

some other things like reducing methane

emissions and so on and so forth but the

solution is sort of steering us

in the face

right it's

a massive deployment of in this case

existing technologies

yes

and to go back to the theme of

ai guy goes on european vacation

yeah electric vehicles will make a

difference of course

right the bourbonnite stuck

driving

will want an electric vehicle or we

would want them to use an electric

vehicle

and yes that pays for itself so do

hybrids of course they do

but of course well said bourbonite is

still stuck in traffic

right so yeah shifting to bikes and

e-bikes pays for itself

of course it does

it's also a technofix of sorts

right moving into smaller apartments in

cities there's also a techno fix now

more than that right it's a shift in

attitude it's all it's many more things

it's

building new homes in cities

as sort of the ultimate techno fix in

all of this

right noho soho rezoning to put a very

new york specific

um spin on things um but it's absolutely

clear that plan a is to cut co2

emissions

or perhaps more to the point

it's to cut co2 emissions and methane

and other greenhouse gases

and frankly that verdict has been with

us forever

okay um

starting in sort of the 2000s

we basically figured out that look we're

not going to do this

soon enough there is so much hurt

already built in

that we definitely definitely have to

adapt

resiliency um become resilient to

the global average warming already baked

into the equation right new york city

has already moved from the temperate

zone into the subtropical zone a couple

years ago that has already happened

we are not going to stop climate change

here

we know we already have to adapt and by

the way the theme of sort of the moral

hazard right there's a bit of this going

on here already or more to the point

al gore was on the record in the mid

1990s and saying let's not talk about

adaptation quite yet let's solve climate

change first

cutting emissions

and yeah then let's talk about it now

okay fast forward a decade

right al gore himself of course and

frankly every environmental group on the

planet has realized that

actually

talking about adaptation

might actually push us to want to do

more

on the emissions front as well it's not

moral hazard so much as maybe it's

inverse

right sort of the frying pan effect you

whack someone over the head with this

hey this is happening

we have to adapt

maybe we wake up to do more of the

former

as a result

meanwhile

plan a also involves something else

sucking it back out

that's a techno fix

that goes well beyond

wind turbines and solar panels cutting

co2 emissions

right moving to the city and so on and

so forth lots of other things here

and frankly that too we have known for

quite a while

here's

2009 over a decade ago by now well over

a decade by now

um

these are a dozen climate models

climate economy models

asked

to limit

temperature increases concentrations of

co2 in the atmosphere linked to

temperature increases to 2 degrees

centigrade above global average or

pre-industrial

temperatures

that's what this

right set of

columns is these

450 parts per million of co2 for those

in the know

well

look at the outcome

decade ago over a decade ago it was

basically impossible

with some frankly fairly heroic

assumptions

to limit global average temperatures

to the sort of temperatures we thought

are necessary to

have a livable planet we still think

over a decade ago that was basically

gone

now a lot has happened

in that decade

namely

technology

emissions is still going up but frankly

rapid deployment of

solo and

wind low carbon technologies massive

cost decreases

40 years ago when jimmy carter put a

solar panel in the white house and

ronald reagan took it down five years

later

solar panel cost hundred times as much

as it does today

ten years ago

cost ten times as much as it does today

nobody is taking down solar panels from

any roofs anymore we are deploying them

rapidly have come down

capacity has gone up electricity

generated has gone up

massively as a result

things don't look quite as bad anymore

in the sense of

omg catastrophe

but frankly

we are far from

were

most scientists

would say did say monday this week

we ought to be

so in other words

plan a

does in fact include a fourth element

here

right and now we can talk about

inequality and right the rich adapting

the poor suffering and so on the usual

story applies of course

but it is certainly clear that

by now

we more or less know

that suffering is built in

and you know we've known that for

years at this point as well

at the same time

it is clear that there is no plan b

and when i say this definitively

um

i can point to a couple fellow

economists here in shame

um in the

sense of

introduce me by saying i i used to work

on solar engineering i still do um very

much so um

when

people

who are not necessarily initiated in uh

um

finer details of this particular

technological intervention

first discover first hear about

solo geo engineering and i will spend a

couple minutes talking about what it

actually is

their first reaction is very much like

the tesla ai guy

with

um

electric vehicles here's the techno fix

here's the thing that will prevent us

from having to do anything else

in this case of course much more

dramatic right don't have to cut

emissions don't have to adapt don't have

to suck it back out we'll just build

this artificial sun

shield

for the planet it'll cool the entire

planet

everything will be fine

and you know

no surprise the usual suspect i

highlight mr newt gingrich here on the

right right would

pick up on that and for example in his

case literally at the time 10 or so

years ago when on the president obama we

had our last go at trying to pass

sensible climate legislation in this

country

he writes a op-ed saying

ha found solution to climate change

no need to vote for this thing after all

you can solve

climate without actually cutting co2

emissions

now there's a couple interesting things

happening here

in order to be able to say this you

actually have to acknowledge that the

problem exists

which is actually

a good step in the right direction if

you will

but of course then talking about this

as if it were a plan b

you might as well deny the problem exist

right the outcome is the same you're

still voting no

on the legislation

oh in other words

yes i do think we should look into this

particular technology too

we should do the research

solar geoengineering

but no it is no plan b

it's plan a plus

another technology added to the suite of

potential technologies okay so just very

quickly what am i even talking about

it is literally building an artificial

sun shield for the planet

it's basically doing what volcanoes have

been doing forever

so when mount pinatubo erupts in the

philippines in 1991

global average temperatures in 1992

ironically just of the time of the rio

earth summit

june

92

a half a degree centigrade

almost one degree fahrenheit

cooler

than they would have been without the

volcanic eruption

emissions didn't go down

nothing else happened

global average temperatures decreased

didn't solve climate change just to be

clear

oceans are still acidifying

lots of other problems

still way too much co2 in the atmosphere

didn't address the root cause

but

if

temperatures are one of the key metrics

here

well here is a technology and when i say

here's a technology no let's not explode

volcanoes very nearly but maybe

let's do the research to figure out

whether it might be possible

to deliberately

introduce

aerosols tiny reflective

particles

into the lower stratosphere with exactly

this

in mind

okay so

how to think about this

um

it's a very complex

graph

i won't tell you what the time scale is

i won't tell you what climate risks are

but the one thing we know is it doesn't

matter what the time is it doesn't

matter what we how we measure risks if

we burn fossil fuels

those risks will continue to arise

no doubt

here's another definitive statement

if we cut emissions to zero

the sort of thing we know we need to do

well let's assume we get around to it

let's assume we actually do it

in any time scale

relevant to

humans society

decades out

the climate risks

most of them we care about

will stop getting worse they will

and frankly that's the point that's why

we have to cut emissions to zero

they're not getting better and by the

way you see climate risk is largely

climate risks associated linked to

temperatures for example

sea level will rise for centuries after

yes temperatures will stop increasing

yes that's a good thing yes we have to

cut co2 emissions 2-0

on net

but

climate risk isn't going to decrease not

in our lifetimes

so yeah we have to suck it back out

it's the only way to actually decrease

well for the solar geoengineering come

in

taking the edge

off there is plenty of hurt built in

here there's plen there are plenty of

people dying

literally because of unmitigated climate

change and even if we do everything

right

and news flash we won't

but even if we did

there would still be

plenty of hurt built in where solo g

engineering might that's the research

question

actually make a real difference

okay couple more points

yes there are trade-offs

there really are

and when i see heart trade-offs

when one

contemplates does

solo geo engineering

and in this case

highly theoretical in a big way let's do

it to an extent where we stop

temperatures from increasing

well we turn off carbon cycle feedbacks

for example

emitting more co2 naturally because

temperatures are rising

so

actually

even

just looking at

co2 impact co2 burden

if that's the only thing we cared about

solo g engineering itself might still do

a lot of

good

on net

and now we have a hard trade-off

right not that anyone is out there

actually optimizing this in the real

world but if one were to or if we do in

our models

yes

there is a trade-off between spending

the money to cut co2 emissions versus

doing that at a newfangled tank

which of course has its own problems

but then

back to the main theme

there is moral hazard

there's newt gingrich there is my fellow

super freakonomics economist sensoron

and so forth who basically

look at this technology

and say

if we

have this tech available

we

might

will

get away

not doing the hard stuff

not tackling the sort of things that we

actually know need to happen so yes

there are trade-offs

um now

not to put too much of a finer point on

this but

debatable whether this is even

moral hazard in the technical sense or

whether it's closer to a lack of

self-control

essentially right it's us deciding not a

big question is who decides and so on

and so forth but still but

at the very least yes there are in fact

these

these trade-offs

okay

now

does it exist

empirically

when we go out there

and ask people

do you think that the

availability of solar geo engineering

will in fact

detract from the need to cut emissions

in the first place

turns out there are about 30 or so

studies out there that ask just that

question

and frankly

the broad conclusion is that nobody

knows what this technology is

so they'll tell you anything you want

basically and actually that is probably

one of the most important conclusions in

all of this

we went out did one of these surveys

we asked the question two ways

will it detract from or will it lead

toward you wanting to do more

depends on how you ask the question

people just agree with you because why

not right you're a smart scientist

asking them

um

must be a reason why you formulate the

question a certain way

all right the technical

description of this is acquiescence bias

yet yes it's a thing

yes solo g engineering moral hazard is a

thing

but

opinions are so

[Music]

malleable

that you can basically get any question

you want

okay

now

there's a better way of doing this turns

out

we don't just ask a silly question

we observe people

in this case full disclosure by now a

co-author of mine but this study was um

absolutely her own christine america at

all here 600 germans

200 of them are told about solo geo

engineering

in a

lab they're also given money because

they have to show up in this lab or in

this case an online survey

and now they can do with their own money

as they please

including

offset their own emissions

hey wouldn't you want to spend some of

your money to offset your emissions look

you're part of the problem

well turns out

people who are told

about zoology engineering

the 200 germans in this case but still

people

are more likely

to offset more of their emissions

because they've been told about this new

technology

or in other words

the exact

inverse

of what moral hazard would lead us to

believe

right so if basically

technology

nuclear technology geo engineering

doesn't matter what technology is always

about the technology versus behavior

well actually maybe

there's a way to talk about this new

technology

and in that case it doesn't matter which

technology yes geoengineering is one

example

but maybe nuclear as well

or maybe any

technofix

or it is not about

oh

technology might bail us out

absolve of the need to do the hard stuff

but actually remind us that hey

this is a real problem

maybe we actually have to do more than

we thought we did

and

without being too

um

all encompassing here sort of the burden

of economist right we have a hammer and

then we apply it to everything

i'd like to think that applies

much much more broadly

in the sense that

whether it's in bio or gmos or ai or yes

any sort of climate technology from

nuclear on the one hand or anything

that's not quite as

[Music]

kosher to environmentalists as wind and

solar on one end of the spectrum or

solar geoengineering on the other

unless we find a way to have the sort of

conversation

about new

technology that leads to

yes and

we're doing something wrong here right

we are not going to stop

conversation

research into development of

any of these technologies

if they are cheap

if they work

if the risks are socialized and the

benefits are internalized

the usual story

those technologies will come

yes they will

they're not going to solve every

societal problem and yes they have to be

channeled in the right direction and yes

it takes government and yes it takes

lots and lots of other things to channel

these individual wishes and wants and

desires into the right direction

but it is incumbent upon

us

as those working on the technology

talking about the technology trying to

frame the technology trying to figure

out how to channel it in the right

direction more broadly back to our tesla

ai guy

to try to get to this

inverse moral hazard as opposed to

being stuck in this

infinity loop of

any new technology will always be met

with resistance at first because of

course moral hazards dominate

thank you

[Applause]

so we've had uh you know a talk on uh

bioengineering uh

environmental and green tech uh ai and

automation

um and

what do the panelists

think are some of the the sort of

cross-disciplinary themes

that have emerged from these discussions

um and

do any of the panelists

have any curiosities about any of the

other presentations

and do you think that they speak to one

another in any ways

i guess i will open the floor that way

oh that's right

so everyone has their personal link

so yeah so anyone can jump in

oh i have a question from them

so you said that you were

so i don't know enough about gmos to be

angry or excited about it but you said

that you were excited so could you help

us understand

both sides of the debate and then why

you are on the excited side

well that's a big question to understand

both sides of the debate

and i'm probably not the world's expert

but i will just say some things

so first of all i think the first

question is what is a gmo

and that definition actually depends on

what country you're in um

so it makes it even added complexity um

in general it refers to the in the um

uh intro introduction of

foreign dna into an organism

and

let let me say that i think some you

probably know the answer to this some

fraction of the crops in california are

gmos

um of

i think there's corn and soybean

so we are using gmos um

now

uh

i may go off on a tangent here there is

a potential revolution that could happen

in this space

i'm sure you've all heard the words

crispr i

but what crispr allows you to do

in principle is is alter the genome

without introducing foreign dna

and that is one of the potential

upsides because in principle that would

not be a gmo

and in fact i think it's in china they

have declared that crispr engineered

crops are non-gmo

so that's could be a game changer

um

so back to um so i wanted to get the

definition out there

why and then the question is why

is it

what's the danger and

again

there's many answers to that question

the simplest answer is

that you are putting something

non-natural

not made by nature

into the environment and it will have

some adverse effect so the difference

between gmo and medicine is release into

the environment

at least that's my take on it um and so

that

carries a lot of weight

the political side of gmo release

and attitudes about that

are much more complex

and originate actually in europe in the

i believe when monsanto first introduced

uh

the resistant whatever it was

in europe and most of that was

politicized it was political um and the

history and this brought on the history

of the organic food movement

um

but it's been with us now um

so

okay so that so the resist so the

scientific resistance which could also

be in part the em so the emotional

resistance or real is that this could

have some kind of bad effect on me or on

the environment or on things in the

environment it might kill the monarch

butterflies or something like that

okay now the

the biological risks

are

impacts on endogenous species

and and the one that i um

am most

intrigued by and potentially this is the

one that actually underlies the concern

is potential for horizontal gene

transfer and most people don't even know

what that means and that's okay there's

a great movie um oh god i'm gonna block

on it there's science it's it's embodied

in a lot of science fiction movies um

and that's

the actual inherent

scientific thing that people are worried

about now that applies

to corn that is bred at monsanto also

even though it doesn't have foreign dna

in it and if you want to see an amazing

uh

it's not monsanto anymore by the way

isn't it it's buyer i think

but visiting there and watching how they

do plant engineering is a sight to

behold they have this thing called the

the the corn chipper and so they take

thousands of seeds millions of them and

they run them through a machine that

takes a picture of the seed and knows

what the perfect seed should look like

and when it sees that one it chips it

off a piece and sequences his genome so

the corn breeding has

become an amazing state of the art and

so you could say those are

person-bred

organisms being introduced into the

environment so we've been doing that for

millions of years okay so why am i not

worried

first of all i'm i think that um

there's

there's an enormous

awareness of potential dangers and there

are solutions

secondly it's a question of risk versus

reward which is something that it

permeates

all technology and has been

an underpinning of recombinant dna since

day one i was

not very old but my the person i worked

for in

uh hud harvard was one of the star

witnesses in the recombinant dna

tri case before the cambridge city

council and if you want to watch

something an amazing piece of

interaction between

normal people and scientists i recommend

even though they're kind of grainy you

should watch that because it really the

question is what is the definition of

risk and and i think vaccines

capture

the risk versus reward and personally i

think feeding the planet

captures risk versus reward i think

capturing more carbon is a good thing to

do if it requires engineering organisms

and plants so for me personally it's

about

reward i could go on forever but i won't

any any uh

follow-up question i mean i think that's

one of the sort of uh master themes here

um

as well as yeah go ahead maybe just to

put sort of a

another spin on this um

whether it's gmos and bio

ai up to a point although i know very

little about the actual implications

there uh solar g engineering um

our usual shtick

as

economists as policy analysts is a

benefit cost analysis

and frankly

most of these technologies zoology

engineering is sort of vaccine territory

in terms of benefits to costs it's a

thousand ten thousand a hundred thousand

to one

the benefit cost ratio right net

benefits

based on what we know

um are so large

that frankly

to me economist is just the wrong

decision criterion it's not how you can

look at this

and then say oh yes it's a good idea to

do

it is all about

building your point about risk risk

trade-offs

it is all about

not just risk reward in the sense oh

there are these known unknown risks

let's compare them to the known benefits

but it's basically

the risks of unmitigated climate change

compared to the very real risks of a

technological intervention

that attempts to do something about it

it's not about benefit costs it's about

comparing risk comparing instead of don

rumsfeldian right known unknowns and yes

worrying about the unknown unknowns and

both

unmitigated climate change in this case

and zoology engineering yes have both of

them

that's what a research needs to be

that's where the public policy

autofocus

so to follow up you know the question i

had was about incentives with the german

study that you talked about

so

so can you just

put that in the context of so why did

they make the decision so there's a

study population that

that they are going to accept that there

isn't a plan b is that is is that what

was concluded i just want to make sure

okay so this very specific so right the

question is is sology engineering

um

the sort of technology that when you

tell people they will say ha solution to

climate change found don't need to worry

about anything else

and in this case it was the exact

opposite okay why that's your question

um and actually that's my current

research with christina america on

trying to but her hypothesis there are

three hypotheses one is basically it is

this is just such a scary thing

that you want to avoid it at basically

all cost right so if a semi-serious

scientist tells you that this thing

exists or talks about it

um you'll spend your own money

in a frankly vain attempt to try to stop

it right so it's sort of this

this effect of oh my god we gotta stop

this technology

um

the other one is the slightly more

positive spin or very much more positive

spin it's sort of this wake-up effect

right i described this sort of the

frying pan if you whack someone over the

head with this it's kind of like look i

always knew climate was bad but wait if

serious people are talking about this

maybe yeah there is really something

uh to this um and that's sort of the

actually this i think is the

better interpretation the more

appropriate

interpretations of the hypothesis that

seems to hold up fairly well

um which is essentially

a

um

yeah it's sort of the positive version

of this right it's basically hey we

haven't been able to tell people that

climate change is bad based on anything

we've tried over decades

right um maybe this finally does it

which

just to be clear it's a little bit of a

naive

view right it's not like people haven't

tried to shock people interaction before

right

movies exist that talk about new york

disappearing on the you know miles of

seawater or ice right depending on your

perspective um so

um

it's not the first time somebody talks

about this technology clearly not but

frankly

public discourse on zoology engineering

we've only

had it for about a decade 15 years or so

the technology has been around for

decades but there's been a long-standing

taboo not to talk about it because of

the moral hazard because it might

detract from the need to cut emissions

okay and by the way very quickly the

third hypothesis is basically people

just don't know what offsets are right

they think dissolve climate by planting

a tree somewhere and you know those

flash you won't

i'm going to go back to

the engineering of biology because i

think there is

a

not only a perceived but indeed an

inherent risks and

i want to

make sure that

people understand that there is a

huge effort in the synthetic biology

community to define

mitigate and appreciate the risks and

this

this has is really goes hand in hand

with the development of the technology

so for example we have developed

solutions that should protect against

runaway release in the environment but i

want to give one example that is

happening right now that i think is

is interesting and it is

a product of climate change and and that

is um the release of engineered

mosquitoes

into different areas on the earth to to

combat mosquito-borne diseases

dengue

so there's been massive releases in

brazil

there's going to be one in florida

and

this is fascinating technology i'd be

willing to bet a lot of people don't

even understand it but i want to just go

back to you know the 1950s or whenever

when um you know malaria is still

a scourge to the earth right and imagine

if we could cure that with engineered

mosquitoes let's think back to what we

did do about malaria which was to use

ddt

which by the way

worked

india huge that

we just used it at too high a

concentration if they had actually used

it at

maybe 1 100 i don't know the

concentration things might not have gone

so badly for nature

so

and and at the same time had a positive

effect so now

you know ddt is forbidden for the most

part

um

but so i'm thinking about this mosquito

solution

and again if that were to wipe out

local mosquito-borne diseases

i think that would have a huge impact

and a lot of people would not be asking

so what was that mosquito anyway that

they released um they would see the

outcome

and um so i just want to interject that

because it's a modern day thing that's

actually happening

there was another

uh okay to do this i think in florida

which was amazing to me

that's uh interesting and it brings to

mind i mean one of the common themes

here seems to be

sort of how does technology and how can

it train human behavior or retrain human

behavior as you were saying you know the

prospect

of solar geoengineering as a

technological solution might actually

make people believe that these problems

are more tractable right

and i was brought back to your

presentation about training

humans

through artificial intelligence and

intelligent agents um and what's

interesting

there seems to be a sort of common theme

there that human behavior the technology

can aid

humans in retraining themselves right in

some way um and i don't know i was

wondering if you could speak to that and

making the world better for humanity in

some way right

right that's

a great connection i i miss that um

so humans are incredibly adaptable and

flexible so any technology

changes that happen they would adapt to

it like that's the nature of the species

my talk was mostly focused on designing

systems so that we are

supporting that behavior change in the

positive direction right because

behaviors can shift in either directions

and we want at least as an ai scientist

i want i try to be cognizant of what the

eventual outcome is and what is it that

we are building uh

building towards

um

with climate i guess

carnot would say a little bit more about

what he meant when he said

you know that's changing behavior

okay so

i think that you know the

short version

in

i guess i'd focus in on uh and on a

point that isn't doesn't really quite

emerge much but

um

and that has to do with what you said

that humans will really be able

to adapt and are willing to adapt and

i'm not

you know i really kind of challenge you

on that because

i'm not so sure that that's the case

because there's lots of morals involved

uh that people will invoke and whether

we should or shouldn't

that we see this of course in in the

medicine area about whether we should

you know sort of use so i wouldn't quite

go that way in terms of

what people want out of technology which

is something that is you it might be

available but do you want it i think

that's really something that

if you're always dealing with a.i would

be a really serious problem for me

because you'd have to actually whatever

machine you were doing whatever ai we

were going to do

you'd actually have to invoke some sorts

of notions of morals

and in what you were

in in the process of not only it

deciding but providing an answer that

you expected some kind of return to in

in the in the human world

which leads me to the other question

that i would have both for the ai and

for you for automation but i guess i

would say for all of you could you

imagine within the technology field that

yourself

anything that you would say

even if we said knowledge should always

be there that you would be reluctant to

that you would you yourself in the

technology reluctant to go in

i want to address your question in one

and i'll be somewhat brief here about

but in the medical context

and give you a counter example

this alzheimer's drug

that has come on the market

it's

doesn't work

and the fda

even almost showed it doesn't work yet

consumer demand

people want it

so what

what do you do with that and in fact it

may even have negative effects on health

so there's a case where you have extreme

it's almost like climate change

there's

there's no end to it and yet people will

do anything to get that drug

um

and then your second question what was

it what would bad

things yeah i was just gonna say so the

same thing bad drugs

so what is a technology

that can go astray um and that's one

example of many bad drugs

which get

distributed globally

so that's my example

right so i completely agree with you

that

the moral um commitments that we make

drive a lot of our behavior and that's

not where i was going with my answer and

my presentation either i think that's

outside of the realm of

ai and technology that's more about

human societies and how we function as

organizations right that our morals are

defined through our communities through

our experiences through intellectual

exchange with other people and i don't

see a big role of technology in that

specific realm but once you have decided

that you know i want to learn this new

thing or why want to develop this new

behavior that's where i see some of the

work that i've done start playing a role

is to support that

on the second question that you said the

the dangers of

ai technology i think

it's already everywhere so facial

recognition tech for example that's

highlighted

it's heavily talked about

the second thing would be and this is

more recent is using

machine learning ai methods on audio

signals to infer a person's mental

health status right and then

because again these are operating like

black box systems we don't understand

them well we don't unders we can't

inspect them we can't explain their

behavior

the determinations that are being made

which will then lead to consequences for

the person

will be tremendous so that's where we

need more focus more insights on and

there are several examples like that in

ai where you know more critical inquiry

is needed

um yeah

so for a person in technology i'm

actually very skeptical about using

technology myself so i get it you know

coming to

this question or should i use this

technology and what are the moral issues

but also

you know how how was that drug um

authorized i remember reading it in the

news and i'm like this went through too

fast i mean the whole process was weird

in terms of the uh recommendation by the

government uh so it was fda and so on

right

so i think

i do believe that our consumers the

users are going to

have do

do come at it

hopefully doing the right homework right

so when i build my system and i actually

i want to

bring this up when we built the system

for the

recommendation of routes route

recommendation and traffic uh we assumed

that

you know we hoped and assumed that there

would be a certain percent of people

who'll reject the route and that's how

because if everyone took takes the

recommendation by our application

actually it doesn't optimize our our

utility function so we're hoping that

you know there will be these skeptics

and we have to account for that as we

build our system especially when it

comes to like cooperative problem

solving um and in other cases you know

uh maybe there are different ways to

convince again looking at subgroups and

what would be needed to convince them to

um to use that app

for example i i use recently i'm not a

runner but a year ago i got this running

app only because the new york times

after all these years you know i read

this

this article and it talked about oh

couch to 5k i'm like there's no way i'm

going to do this at you know 40 plus

years but i took it up and i can run i'm

not a great runner but i can run for 45

minutes and i'm like okay or more um and

it took several months but it was

because of that incremental um approach

that they took right so um i think

people change over time you and you and

you can there can be different ways of

getting the users to uh to use it and

you don't have to expect that everyone

will use your technology

thank you there's a lot of different

things that i could ask about but i'll

try to focus on the three uh

particularly interesting things that

came to my attention one is when you

talked about the bio leaf how is that

being practically implemented for actual

practical use i'm not a scientist so i'd

like to know how that goes about the

second question that i have deals with

the traffic situation of new york city

one of the solutions to the city of new

york

is to just penalize

drivers from a certain point of entry

into the city of new york or just to

make an increased cost and i was

wondering in terms of your traffic

studies

is there ways to address that i mean

robert moses decided to do that by just

building more roads more roads more

roads and that as we found out they just

kept on increasing the amount of people

that drove on roads and that doesn't

really solve their problems and we also

of course have a problem on the crown

pros expressway and the third thing

deals with the climate change

you talked about geoengineering

it seems to me that that conversation or

that concept doesn't require any quote

cost

associated with in terms of an

individual having to adopt or reduce or

something so it seems to me the germans

might have reacted in a sense to saying

well wait a second i don't really have

to do too much this seems to be a great

you know wonderful solution to

everything so i'll put a little bit of

money to that and that'll solve my

problems if that is the case and you

actually illustrate it on your slide

what is geoengineering what is its cost

and as opposed to introducing it down

the graph

um as a way to reduce the carbon

emissions after you've implemented you

know the other adaptability to reduce

carbon recapturing carbon that kind of

stuff why isn't that at least at the

talking point when a politician goes out

and says like newt gingrich geo

engineering solves everything why can't

that just be whatever form it takes

implemented right at the top and

therefore avoid that discussion about

having to reduce carbons and whatnot and

how do you get around that political

point

okay i guess we should take those in

order um so okay well mine's the easiest

one

the bionic leaf

is a actually a a

closed system it's self-contained so you

can imagine it sitting in a box

and so i think the the comparison to

make

is for example and i'll come back to the

bionic leap but let me make the

comparison to say algae so algae

biofuels big big deal for a while

thing about algae

is it has to be in direct contact with

sunlight

and so that requires huge surface areas

it's like agriculture right

and this has been one of the limiting

factors in terms of algal engineering

for say biofuels

the bionic leaf does not have to be in

direct contact with sunlight it can get

its energy or its electricity from

anywhere so that the box can sit in your

basement you could have a windmill on

the top of your house that is creating

the elect electricity that will go into

the bionic leaf cause the water

splitting reaction

and biomass will grow

so it's it it is not

it's it's a it's a self-contained system

that's one of the the actually

breakthroughs is that you can have the

electrodes in direct contact with the

living system but that system can live

somewhere separate from the incoming

energy

how would that work in dealing with the

problem of food shortages in those

places of the world

where which don't have access is there a

practical application to that the

software don't have access to

food

right so

amazing you bring this up at work so

there is a um

i'm actually in involved i have some

work in this space and also there's a

darpa program and there's even some

startup companies and it's called food

from air

so remember these bacteria

are not they're not using sugar

which would be the normal you'd have to

grow sugar cane give them sugar to grow

to make food

instead they use air they use co2

hydrogen and nitrogen

and so in principle

the bacteria could grow

you could use them directly for food or

you could have them produce

food components

so this is a big deal right now

ironically

if you read the 1970 nasa report

this same idea is in there it just never

was implemented and so um you've hit on

what i think is one of the most exciting

ideas here

i don't know the specific policy that

you were mentioning for transportation

in new york but

i don't think there's only going to be a

solution that's more just technology

right so it's a policy problem it's

incentives and technology would be part

of that ecosystem and there have been

studies that

um

that show that you know uh when there's

uh incentives or sort of artificial

incentives are created so there was a

study out of mit

that said that demonstrated that if we

are paying people to say uh take buses

as soon as we remove that incentive the

behavior goes back to normal right so

they will just revert back to whatever

their baseline preference was whatever

their baseline utility model was and so

that's where the technology part or at

least the

studies that we did start uh making

sense is that there is an underlying

utility that people are using

to make those decision choices and if we

can really leverage that

into our technology systems then the

systems and the route recommendations

are aligned with each how each

individual thinks about their

transportation right so that's where the

technology starts playing a role but i

wouldn't believe that you know that

the technology i talked about

specifically would by itself solve all

the problems that exist and it's going

to be a collaborative policy uh you know

improving our transportation network

changing attitudes so that people do

want to live in cities and you know want

to take public transport so it's a

multi-pronged approach technology would

be only a very small part of it

our next question and let me just agree

with this right so

congestion charging in lower manhattan

right so

you know this is the literally the one

law we have in economics right we run

around pretending that people are like

atoms in a vacuum often and we behave

all like you know

those particular laws not true of course

um the one law we have that holds all

the price up

quantity demanded down right works every

single time or at least a couple exam

exceptions we know co2 isn't one of them

congestion isn't one of them right so

yes

one

the solution if you will for traffic in

new york city is to

make everyone personally pay for the

negative effects onto others when they

choose to drive in new york that's where

congestion charging comes in and just

you know the fraud politics of it all

yes

we're going to get it

but there is currently an environmental

review

run by the state

that will

coincidentally

end conclude

right after the gubernatorial election

why well because it is a state issue for

new york city to be allowed to have

congestion charging and if you're the

governor of new york

um yeah you benefit eight million new

yorkers or in this case only manhattan

sadly because it's just lower manhattan

but of course anyone who drives into the

city right the 12 million

in the suburbs of new york

would not like to pay even though

right

overall societal well-being increases

for everybody

those who then have to drive in a pinch

will be more able to do so

at a cost right and this is sort of the

high cost of free parking right

now okay um

geoengineering right maybe let me end on

a medical analogy um

uh and medical analogies abound here

sort of painkiller chemotherapy and so

on um but it sort of drives home this

point i think of

technology versus behavior um

so

should we tell

the 15 year old who hasn't picked up

smoking yet

hey don't worry

go for it

because if you have stage four lung

cancer

there is chemotherapy we can fix you

or

ask differently should the 15 year old

pick up smoking because chemotherapy is

available the short answer of course is

no

right now if you're the 75 year old

stage 4 cancer patient

chemotherapy is in some sense the only

thing that is going to extend your life

help extend your life right diet and

exercise isn't going to do it anymore

right could or should have

but at that stage

not going to

help well cutting co2 emissions like

diet and exercise

offsetting your emissions is like diet

and exercise in this case um solar

geoengineering is like the technical

intervention

the medicine the surgery the

chemotherapy the painkiller

that might allow you

to

exercise in the first place because it

is the painkiller that allows you to get

up

in the morning at all right so is g uh

sorry uh geo engineering geology

engineering sort of the first thing you

should do

no but of course the real question is is

our planet currently closer to the 15

year old

who should be told to diet and exercise

because that's good for you and will

extend your life or is our planet closer

to the 75 year old cancer patient for

whom

this particular drug this particular

intervention might be the only the best

hope at extending

somebody's life right and

yeah i i won't speculate where i think

we are but

yes let's look into solo geo engineering

because frankly

it is pretty darn late in the game

i i had another question

going back to dr jane when you were

talking about dr jenkins i i found your

example so interesting but i think you

if i if i heard correctly you were

trying to talk about

um

how

how dr j might take the bus right and

to be shown that there would be a lower

cost of that but i was just as i was

listening to i was thinking i mean that

it's so interesting everything you were

saying about bringing in the

the the theme of uh

you know programming humans but but it

would be a very

rational like

you know you're assuming a kind of

rationality on on the part of dr jane

which may or may not you know

work so i was wondering when things seem

to be emotional and not simply rash i

mean if you can sort of count on

rational choice considerations to to be

operative in the example you gave and

then i just wanted to say that i'm

seeing a theme emerging from last

night's paul krugman to a number of you

today and the the sort of important that

technology is doing all these amazing

and

concerning that that balance i think is

going to be a theme but related theme is

that you don't lose the importance of

the human um

the the human balance also like how

human beings are responding seems to

like technology is not overriding that

so how do you make decisions how how do

you you know go in i think and anita was

talking about taking into account a

variety of concerns in some of the

examples you were talking about you know

about transparency etc so it's an

interesting theme and and i just wanted

to ask about

is there a rationalistic assumption here

right so there is i would agree to that

and the reason for that rationalistic

assumption

is that to bring any kind of theory into

ai computations it has to be

mathematical

and the mathematical theories that exist

leverage the rational thinking framework

right so it's just an easy natural fit

so that's

a bias because of the nature of sciences

in the dr jane example we did uh try to

broaden that perspective out so we

did so the influence problem that we

were discussing

does include things like messages that

may be compelling to different people

right so if we know that dr jane cares

about the environment

we can influence her more

oh it's 1 30.

go ahead

so i was reacting to uh your uh push on

that there has to be a more humanistic

component involved and we did think

about it right so the compellingness so

the

the algorithm that i talked about was

talk was mostly concerned with

what should we recommend to dr jane

right the humanistic component would

come into how should we make that

recommendation right because it could be

true that for someone who cares

literally just about how much it cost

telling them that you would cut down

your expenditure by this much by taking

the bus would be the most compelling way

to deliver that message but for dr jain

who cares more about the climate the

better way to think to frame that

message is that you're contributing to

reduction in emissions right so we did

lay out a framework of how that could

happen

integrating those ideas into ai systems

is more challenging because again these

are not mathematical but you're on the

right track with that

just to add to that you know so

i would say that there are models which

take into account that the response can

be different right so just a simple

markov decision process which will take

into account the uncertainty in the

response of the user and then take that

as a sequential process so so because

they have rejected it then these are the

parts possible actions that you can now

take so thinking about an action

strategy not just as a single

uh schedule or a plan but

it's really a tree and you you're

walking through this tree and finding

the right outcome that is occurring and

then adjusting to the current

environment and this can be true not

just about human response but also when

when the agent takes an action the

environment itself can respond

differently you know you think that this

is what's going to happen but that that

did not happen so how do you deal with

that contingency so it has to be part of

the systems that we build um so it's

definitely something that we take into

account

do you do you have an additional

question

i guess it's for the people dealing with

uh you know essentially automation and

ai but i guess

about ai let's assume that you have a

problem that you want have

you want to have solved by

a robot who has been programmed in a

particular kind of way

so you're looking at the program

and so

the outcome of the program is basically

decision one and possibly decision two

given what isn't analyzed by decision

one

but in the process of making the

decision on decision one

it it it is it hasn't

if we talked about the whole human

population

which brings different realities to

a problem that has to be decided then

the then the issue really is taking the

median if you're actually doing this

that is you have these outliers so

you're taking the median and you have to

account for that so even in decision one

you're going to use the median but then

the question really is the interesting

question to me is dcpc is between

decision one and decision two once

that's happening and i'm wondering what

kinds of technologies are actually being

worked on now that actually address

that particular issue

yeah

again a great question um

so i i call this the envelope

effect you know and you can sort of

abstract it down even in a scheduling

problem so if you miss task number one

by a little bit and then that affects

when you start task number two and that

also is missed by a bit

how does how much do you miss your

deadline by and you know i think

some of

the

uh formulations that we would bring in

is like some a thresholding effect like

how far what is the percent by which i

am i mean this is just a simple

um

formulation right so uh how off am i

with

what i expect what am i expecting to be

where i expect to be with this

particular situation and then

re-planning or rescheduling uh what

we're doing right right so uh again it's

a sequential decision-making process uh

which takes into account

all the different ways that you could

land up at least my work takes that

approach but if you keep missing what

you expect where you expect to be and

that adds up beyond a threshold then i

think you should hit the red button and

redo things

go back to that

one question ahead deals with the

radiologists

how do you assure the radiologist

who may just think of his job

as analyzing x-rays

that he himself is not going to really

be out of a job and somehow he's going

to be a radiologist that can incorporate

ai

and the other thing goes back to the

the geoengineering where are we in terms

of geoengineering i mean

you talked about an aerosol kind of a

thing

um what's this status of geoengineering

to deal with that part of the graph

what's the cost uh to society what's the

political framework for that

so i brought up the radiologist quote

so the way i see this i think you know

again this is

uh vision technology and vision

algorithms which have advanced so much

uh

deep learning and so on um and they

are probably doing pretty well

with very specific defined problems and

in the average case you know doing the

right type of prediction so i would say

there are two concerns one is when

things are outliers and i would think

definitely in you know given what we

have seen on the health side of just

preterm babies and clinical data uh on

envision that can be

similar issues right so uh

we can and this this is related to

health and

there are high costs to be wrong with

your results so um i i think that's why

having the expert to handle the outliers

would be one of the questions and the

other is again context

ai systems we are still in the space of

what is called narrow ai you know that

we build systems which are very good at

doing one thing and maybe a couple of

things which are similar but we are

nowhere close to where you know how

humans are able to assess what was this

person's background you know or what

what were they exposed to and what how

does that affect the image screening

that i'm looking at uh so um

i think

as we move towards what is called

general ai um

we have to take this into consideration

but i i think the expertise that

physicians bring in you know sometimes

they are just able to see the patient

and take in thousands of features into

account and make a

decision on where they are in terms of

risk um we are not there yet we are very

specific in that narrow areas ai space

uh where are we with geo engineering so

uh we are at after a couple dozen

peer-reviewed papers

15 years ago

exponential increase in

scientific

interest which means we have a few

hundred peer-reviewed pay 850 or so

much too little to make any kind of

informed decision on

whether it is even a good idea to

contemplate

or put different levy at the stage where

national academies

for the fourth time by now

have come out with a

report last spring in this case

saying

we ought to do research

we ought to have a national research

program

um it ought to be open

transparent and so on right nobody owns

the patent and so on and so forth it

ought to be

jump-starting governance conversations

at the highest levels governance to me

is let's talk about it right let's

make sure we understand what it is

um but frankly and call me biased

founding uh executive director of

harvard's sology engineering research

program here um

we ought to do a lot more fundamental

research into the technology before we

have any of these conversations about

should we deploy sology engineering at a

global scale never mind that nobody

knows who the wii is in this

conversation so if i can

be so bold

to contrast this to the biology solution

we can already

if we

enhance

photosynthesis in plants by twofold

we can already say how much carbon

drawdown there will be

and

so

you know it's it's a technology

that

i would say is ahead

and nature already does it

and

i'm not saying we shouldn't do all

things which you made a very good point

of but i think

you know there are technologies

that they won't complete the solution

but they're good to go in term they need

more research people are working on it

so that's good but um it's not all hope

it there is stuff happening

more should be happening for sure

and to be clear it's not either or right

we need vegas can't be choosers right we

mean we need it all

and we needed it yesterday

that's yeah that's a good point um

so okay

so i i guess i guess that completes our

q a all right

um yeah so i think that's that's a good

statement to end on

beggars can't be choosers um so okay

well you know many many many thanks to

the four of you

for your wonderful talks and q a and and

the audience