MIT 6.S087: Foundation Models & Generative AI. PANEL

um I'm going to ask you each one um kind

of a more targeted question at first and

you guys also think about what what you

want to ask our poist today so um

Professor first question to you um rard

touch on this topic and you are an

expert in computational biology probably

exposed a lot to Evolution and and

mechanics how it worked I want to um get

your opinion your perspective on the

Dilemma that exists right now um in

terms of centralization versus

decentralization in terms of alignment

versus um more risk and diversity

so let me pass this to

you in terms of perfect yeah in terms of

alignment versus more risks and

diversity specifically meaning that well

we as humans are very diverse uh we have

diverse cultures you know

um lived in in Greece and France rard

lived in Sweden I lived in Ukraine um we

are BR in different environments and the

evolution might have been pushed by our

differences um but now we have a very

defragmented um defragmented AI

defragmented organization defragmented

Society in terms of who's pushing AI

forward they are implementing their own

AI alignment systems um they're reducing

the diversity but potentially also

reducing biases and stereotypes that

have already existed in society so we

kind of have a dilemma between high

risks um but more opportunity for

Innovation or lower risks and lower

opportunity for Innovation what's your

perspective on that coming from biology

Evolution and things around that

beautiful uh fantastic question

extremely rich extremely uh deep broad

reaching Etc so um let me start with

Biology a little bit so basically uh as

you mentioned sort of humans are forced

to be diverse we don't have a choice we

basically have genetic variation that

modifies every aspects of our brain and

of our body and of our behavior and of

our inclinations and so so forth I have

three children uh you know they are

completely different from each other and

and and they were completely different

when when they first came out and

they're still completely different now

and um as much as we would love to as

parents think that nurture matters a lot

it's only about 50% and another 50% is

just like nature and and there's very

little you can do about that and um

that's I think part of the beauty of

humanity the fact that whether we like

it or not we're all programmed to

actually think differently to interpret

things differently to uh Etc and that

that's just the nurture component the

nature component Al sorry that's just

the nature component the nurture

component also gives us extraordinary

diversity in sort of where we grew up

the things that we saw as cultural

references at different points in our

lives as you mentioned different

cultures different families even in the

same sort of street block you can have

kids growing up with completely

different perspectives on life and I

think that's what makes MIT work that's

what makes any team work the fact that

we think differently and we can bounce

ideas with each other with mutual

respect but also uh completely different

perspectives and that shapes the ideas

very interestingly so I think one way to

achieve that with AI even with a single

underlying large language models is to

instill different personalities in a set

of agents that are interacting together

in the same system so that forces the

agents to actually process ideas in

different ways so if you want to have

the most Creative Solutions you don't

want a single AI That's going to give

some average you want a lot of different

AI that are going to be bouncing off

each other each with own personality and

you can encode that you can give them

personalities you can basically say you

know you are a professor who grew up in

Iran and who has you know these kinds of

backgrounds you are a waiter who grew up

in I don't know Scandinavia and has this

background Etc and then based on these

personalities you can sort of build a

life story and a set of attributes for

each of the agents and then push them

towards uh more creativity um in terms

of bias we all worry so much that AI

will be biased but I have to say that

humans are you know have a terrible

track record on bias we are horrible

when it comes to bias and icii as a hope

for being able to not just debias but

anti-bias uh our thoughts to be able to

sort of

artificially tag on different biases

with different attributes and push us

off our comfort zone in terms of

expectations and have the AI push itself

off its comfort zone so you can

basically create create again these

personalities with very different

stereotypes and with mismatch of these

stereotypes and sort of have the AI

interact with those and actually learn

how to uncouple uh those biases so

that's on the bias a little bit on the

diversity and in terms of the

centralization I you know again I think

the scenario of Skynet in Terminator is

exactly one of centralization it's

basically US versus the AI and I think

that the for of the market are such that

as centralization happens in One

Direction you will have forces pushing

against it in the other direction and

there is uh there are laws against

Monopoly there are antitrust laws that

are sort of go going to kick in if we

see that in fact centralization is

pushing too far and I think that's

healthy I think that the forces of the

market are are healthy I think that the

best way to combat the Skynet scenario

of the AI apocalypse is not to pause AI

it's to double down and to sort of you

know expand out and to democratize and

to sort of you know provide

opportunities for many others to build

on the same architectures on the same

Hardware on the same software and

sometimes even on open AI to basically

create diverse agents on top of it and

that's what we saw a few months ago with

the Chachi pts the fact that everyone

can program their own Ai and even if

there's an underlying architecture you

can still have diversity in the

utilizations and in the

outcomes okay so thank you that's very

interesting I do think that saying that

you have one single big big AI that you

incorporate different personalities into

it sounds like if you take the biology

and evolution similarity like well all

of humankind would share a single brain

you know that would be prompted

differently and that sounds like w why

don't Humanity have a single brain

because it's very fragile if it screws

up we're all screwed so I I you see I

mean I think it's and also I think if

you have that such a big thing it's

gonna even if it's just less biased it's

going to be biased systemically in

exactly the same way for all of those

users right well a human being exactly

is very biased but differently so which

is I think much more in line with nature

and evolution which I think is great

guiding Stars uh so like since you since

you work with this I feel also the last

thing you point as well that let's just

push through right but like what's and

what's the what can we learn in terms of

innovation and change from nature well

most of change is bad and how you know

we understand that is by the passing of

time like if you push things very very

quickly what systems like Evolution will

understand what's bad Innovation is

going to kill us and what's not if you

don't give it enough time to see the

effects does that make sense uh no

absolutely these are a great idea so so

basically on the first comment of the

single brand many many personalities

even if you have a

single giant llm it has 5 billion

parameters if you look at the human

brain the way that thought happens it's