MIT 6.S087: Foundation Models & Generative AI. IMAGE GENERATION

okay welcome to the fourth lecture uh on

the course called fation Model intive

AI uh today we will do a very brief uh

managing of data

uh data is one of the key components in

this new type of AI and really deserves

its own course but we're going to talk a

little bit about it and then we're going

to cover stable diffusion which is a uh

text to image uh generation

ai dtive ai um okay so what do we have

left on this course so next week we are

going to have a lecture on emerging

Foundation models and and their

applications so this is foundation

models in the wild in the in the market

and in commercial settings especially

we'll have two uh guest speakers as well

uh so manoles will talk about AI in

genomics and Applied Biology and art

time will talk about autonomous agents

um and then the last lecture will be on

AI ethics and Regulation and then we'll

have a panel at the end of it then

discuss uh the ethical aspects of

AI okay so to summarize a little bit

right first lecture was an introduction

a very quick intuitive answer to what is

uh Foundation models and generative

AI we went on a little bit of a

philosophical digression and and the

history of AI the second lecture went

through all the different algorithms uh

from a high level perspective and then

we went into depth the last lecture on

chpt and right so what do we do in in uh

what's the key behind Foundation models

it's be able to learn from observations

you don't need human beings in a loop

and you can scale up as as much as you

want and that you get from this is a

very kind of contextual relational

understanding of meaning that mean is

defined by the company it keeps it's

self

referential right so dog is something

that's walk by owner with a leash it's

something that has anistic relationship

with cats it's something with that

chases F chases fris PES when those are

thrown this is how we understand what a

dog is right not actually your parent

labeling it or really you optimizing

certain goals like reinforcement

learning it's you observing dogs in

different context and correlating dogs

with other Concepts that's how you

understand what a dog is that's your

know that's the main

trick uh again uh

uh something that I think is you know

true when it comes to AI is that you

know there a very most understanding is

intuitive and is relies on all the

different relational uh edges so you

don't never really fully understand

something you can always become better

and it's a lot about just familiar

familiarizing yourself so if you don't

understand everything in lecture that's

fine just try to get some intuition and

familiarize familiarize yourself with it

and then kind of keep going

okay so uh data why is it so important

well I think a kind of uh complimentary

perspective on all the breakthroughs in

AI that we've been seeing is thinking

about in terms of data and actually how

the new AI just looks at data the old

type of data but looks at data in a new

type of way so it becomes more powerful

and can use more of it and that's

basically a very big part of what's

happening right now and the data is is

really really key

um and in some some ways I think data is

the you know if you want to apply AI in

actual settings you might care about

looking at the data understanding data

is going to be key for you and so that's

a very interdependent concept like Ai

and data go goes hand to hand it's very

very hard to develop better models if

you don't understand data and vice versa

so if you start applying AI to your own

settings understanding how the data uh

you know what data you have and also how

AI

how AI leverag data and what kind of

data he wants is going to be very very

important for you and if we take this

picture we had before about kind of you

look at the this new AI development as

some kind of Iceberg right the tip of

the iceberg are

uh chat GPT and stable diffusion for

example right hype stuff that that uh

people are talking about and then course

below that it's about understanding soft

Su press learning the training

methodology to get the these AIS right

foundation models and generative Ai and

then a really big you know significant

chunk below that the people don't talk

about as much is the data right that's

what feeds this whole

Revolution

and right and I

think

so let's look at open AI for example the

build chipt right maybe it's like 10 to

you know 10 engineers working for a year

or six months actually developing the

chbt version that we're using right now

right that's not a lot I mean that's why

you have a lot of startups now that try

to replicate CHP chbt just for a lot of

money in computer train on on data with

a small team and they can reproduce it

so that's like in itself I mean it's of

course impressive but it kind of pales

in comparison to how the internet was

created right so they they're able to

leverage all of internet download it and

and train on it but the inter was you

know we spent 20 years you know billions

of people putting a lot of data online

that that's a huge effort that you Cann

replicate right nobody can replicate

that uh process of creating the internet

from scratch no company can it would be

too too expensive and

so somehow the internet makes this all

this available and is basically the

greatest data collection effort in human

history and that's I think is even more

vital for CHP than any of the technology

right right so data is really really

key and again um I mean if I had to

choose between and for you as well if

you had to choose between having chtp or

having the internet you would rather

have the data and the internet because

you can retrain CHP you can create

better versions so the data is really

key and having access to it as well it's

also going to be problematic now with

copyright Etc where people want to say

like well stack Overflow but we have

this data now you you kind of out

competing us by using our data and

people are going to Value data more and

more now as well and it becomes very

interesting how it's going to affect the

development of AI because the internet

has been so easy to use and and people

haven't really

cared and again when you work for a

company whatever you do and you have

your own problems start looking at the

data the data has all the secrets so

it's a it's a very very important to

think about and not Overlook

um all right so

the small piece of philosophizing in

this class will be about data and so I

mean equally that that for chtp the

interesting thing is not the technology

right it's not the brain of CHP but it's

the internet maybe it's a similar

perspective of also human beings like

maybe we're not that impressive as

intelligence but actually the data that

we've created and the whole like get

biosphere has created that's what really

matters so maybe we're just more data

creators for the gene for example if you

know about this theory about the selfish

Gene trying to reproduce itself like

maybe we just collecting data for for

another purpose basically so right let's

say that an alien came to uh Earth and

discover earth and we be like oh they

probably want to kidnap us and look at

our brains and dissect us understand us

maybe they'll be like well no we don't

care about you guys we want your data we

just want like all the history of the

earth and what's going on here collect

that and then can they can derive their