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

but uh tin flew in from silic Valley

he's going to be here in th as well for

B part but he's going to talk about

agents so please en Jo yep hi everyone

so I just going to quickly jump into it

right away imagine you want to research

a topic of the future of AI you want to

understand what's going to happen soon

in terms of AI space and so on and you

don't want to see it on on a 1 hour and

30 minutes lecture um you don't want to

click on a thousands of links you don't

want to research and spend long time

doing that you want to have a concise

report right away right there this is

possible now with autonomous agents it's

also possible now to basically send a

promt to chpt and get your favorite

pizza delivered from the favorite

restaurant um furthermore it is possible

now to execute on any almost any online

task like say for instance doing um a

California driving test online and you

can look at the Sinister face of this

guy you can trust he did it himself so

yeah uh my name is Art I'm uh originally

I'm software engineer I was born and

raised in Ukraine I have been developing

AI products for the past more than a

decade I guess as every second Ukrainian

software engineer right now because of

the geopolitical situation I'm also an

ethical hacker and work a lot in cyber

security space um and I'm a Serial

entrepreneur I came to MIT to do my MBA

degree but to many people ask me why

software engineer do an MBA degree and I

dropped out well not exactly because of

that I started my own company which is

called Kraken AGI and we are building

agents with the mission to drastically

Elevate Global digital reliability and

security we basically build in

autonomous AI agents for cyber security

and software development

space now today I want to quickly touch

on the topic of um terminology in this

field it's extremely confusion it's

extremely confusing and convoluted

unfortunately right now as in any new

field I also want to explore autonomous

AI Agents from the perspective of

specifically AGI artificial general

intelligence and I want to expose you to

some um techniques and mechanics that

are being used in the industry to build

to build such an agents right now so

that you maybe can go offline and

research more on these Topics in terms

of

terminology GPT is a model gpt1 gpt2

gpt3 GPT 4 chat GPT is a SAS

product gpts are now agents or co-pilots

or assistants so the the terminology in

the industry is extremely extremely

confusing different things mean the same

um although named differently different

things that are named the same mean

different stuff so it's it's like super

confusing right right now and this is

fine you need to understand that this is

fine when you search something on the

topic of autonomous a agents you will

see well in some paper neuros symbolic

linking in another retrieval augmented

generation Google will tell you this is

called grounding some other folks would

use emotional grounding concept so this

is fine but let's touch on a couple of

key terms here so the agents have been

developing for quite some time um since

the Advent of AI basically in during

this old um AI deep learning um hype um

this architecture emerged so it it it

was quite some time ago and basically

the main components of it is we have a

system that can autonomously perform it

has sensors it it has actuators it

interacts with an

environment post generative AI hype so

like a couple of years ago we got a more

simplified architecture and a more

simplified approach now actuators and

sensors are kind of called tools now in

this system we have either an llm as a

core or a family of AI models basically

um providing reasoning capability for

the whole agent itself we give out a

task to an agent um that it needs to

achieve on a

goal another thing I want to Define here

today as well is what is Agi what is

artificial general intelligence and I

think um Google has taken an approach to

that and and actually had very good

definition AGI is basically something

that can accomplish any task that human

can accomplish or can do or kind of can

interact and react in any environment

human can interact and

react now what's today's AI is missing

to to to be AI well except of this lame

joke about letter G um today's AI is

like a monk in a cave meditating so it

like the llm or Char GPD when you

interact with it it's a

snapshot of time and space and knowledge

it's very wise it has a lot of general

knowledge but it exists outside of time

outside of the

environment and basically you come up to

char GPT you bring a letter with the

task written on it CHP gets this letter

reads it as a monk writes down the

result gives you back and keeps

meditating it also has quite a lot of

constraints specific scalability

constraints one of the scalability

constraints is the fact that well if you

bring like a bunch of books to this mon

um it will just truncate half of the

information that has been provided and

we only work with information that

allows that its context window

allows another thing to look on the

current AI to kind of put in an

additional analogy what rard mentioned

on today's lecture about the foundation

models that the um current llms

Foundation models are components of the

brain but not the whole brain our whole

brains are much more complicated it's a

family of AI intertwined together and

they're not the whole body the body has

the sensors interfaces to act to feel to

see to have vision and so on current llm

models on only have limited interfaces

to do

that now still considering this can we

achieve AGI like capabilities

today and the answer is we kind of can

or we at least can push to a

capabilities and there are a couple of

approaches to that one of the approach

is what Yan leun he's a chief scientist

at mattera proposes it's basically he

basically says Hey llms are dumb they

can't execute on a bunch of tasks like

say planning they still hallucinate

so we need to build a completely new

architecture basically inherit our own

knowledge and build a completely new

approach completely new AI model

probably based on Transformers or

something else and um or give the new AI

new architecture of this new

ability but I'm software engineer I like

love to problem solve and in software

engineering industry we say

composability over inheritance so uh we

potentially still can try and use

something

that's already exist in the field and

try to apply try to work around

limitation that exist there try to uh

create new techniques around it and

actually still try to push um the

existing llms towards agla capabilities

only using what we have and that's

another approach that currently the

industry and most of the Silicon volley

startups and companies are are taking

right now how to do that specifically

how can we build how can we push current

llms those monks sitting in the caves to

be more like AGI first we can give it we

can give the llms um ability to

self-reflect we as hum humans we are

thinking iteratively we don't have like

when we produce an idea our next thought

is being fed of the S thought that

happened a second ago so we iteratively

think about something until we come up

with the result we don't just go ahead

right away and produce the idea or or

suggestion or something this technique

of like thinking associatively in

iterations is called The Chain of

Thought and it's being applied in prompt

engineering you can apply it on your own