What are AI agents? Agent swarms? Autonomous swarms? What's all the buzz about?

https://www.youtube.com/watch?v=IqovGsaTro8

Transcript:
(00:00) all right level check hello level check cool what is an AI agent and what is an agent
swarm and what makes it autonomous there's been a lot of hype over the last week uh
driven in part by myself but not just me uh for a little bit of background open AI released their
assistance API which a lot of people are just calling agent because it's easier to say it's two
syllables instead of three and so now all the hype is around how do we create a agents so
before we dive in let's just talk about what is
(00:35) this at a high level so if you're watching this video you might be a developer who's
just coming to the idea or business leader or maybe you're just curious and you want to
know what all the buzz is about so the easiest and simplest way to think about this is that it
is just the next evolution of Automation and so you've probably seen the videos of you know
the little swarms of little floor robots driving around Amazon work uh wearing warehouses
you know they can go fetch boxes from shelves and they can you know put it in
(01:04) a larger box to be shipped off that is an example of a physical swarm where you
have uh pretty much interchangeable robots that any one of them can jump in or out of the
work cues and get work done in the physical world they are interacting with the environment
they're receiving instructions from some kind of controller and they're also able to make
some autonomous decisions on their own such as avoiding collisions and that sort of thing
uh that is kind of the the physical manifestation and so why we're
(01:35) all excited about these digital agents is because now they're powered by large
language models and soon they'll be powered by more multimodal models but also they
have the ability to interact with a digital environment and so what I mean by that is they
operate in cyberspace they operate by connecting to apis and databases and that sort of
thing so an individual AI agent has uh it's basically defined by a few things it has
autonomous function meaning you give it a task or a list of tasks and it's able to prosecute
that
(02:08) list of tasks and rather than in in interacting with a warehouse It's interacting with
whatever digital space you give it so you might give it access to certain apis or functions to
call you might give it access to databases or other coding abilities it's able to make some
decisions on its own such as deciding what API to talk to or how to talk to that API uh and
because it it's able to make some of those decisions on its own it has to judge those
decisions based on either uh specific goals or other principles that you can bake into
(02:42) it with instruction sets and this might sound really simple but this is like this is the
kind of automaton that was only dreamt of as of just a few years ago and one thing that I
want to point out is that a year ago this time people were still mocked for talking about AG
AGI now mainstream media is not only talking about AGI on a regular basis they've just
started talking about super Intelligence on a regular bases so we are ramping up very
quickly the Overton window is Shifting and it's due in no small part because people like Sam
Alman
(03:16) Mustafa sulaman and Sachi nadela are actively talking about AGI and so once the
CEO of a big tech company talks about AGI now it is okay to talk about this stuff so it is
ramping up quickly and it is coming but this is what one single AI agent is what is an agent
swarm so an agent swarm is what we realized was was the next logical step once you can
instantiate a single agent very easily which there are several platforms and projects out there
that allow you to do this it is easy to spin up an arbitrary number of agents and so
(03:50) if you're in software development or architecture this is very similar to spinning up
containers so you know for those not in the know uh basically the way that software runs
today at least large scalable deployments of software is that it's all run in tiny virtual
machines called containers which is a self-contained environment that can run some bit of
code so imagine that you have one agent or several agents per container and now you can
instantiate literally millions or billions of Agents on the fly all day every day again it's
(04:24) not that much different this is basically just the merging the the next logical step of
large language models and container technology we've already got the underlying
technology now we're just putting it together and so whether you have a fleet of containers or
an agent swarm the idea is now your containers are smarter because they are enabled by
large language models and generative AI technology again this is a brain dead simple next
logical Evolution but with any uh leap like this there's going to be new problems to solve
such
(04:54) as how do you orchestrate these things when they're this intelligent how do you tell
them what to do how do they know that they're succeeding we have several projects that
we're working on that I'll talk about closer to the end of the video if you want to learn more
about how we're actually implementing this code and you want to get involved it's all open
source and so then finally because we have agents that are able to make decisions and
prosecute missions we have the ability to have agents that
(05:20) instantiate other agents which means that you can have a fully hypothetically you
can have a fully autonomous agent swarm meaning that you have agents that are
responsible for controlling puppeting and instantiating other agents uh and then the idea
here is that it is self-organizing principal driven and self-improving so this is the this is the
dream we don't know if it's possible yet the the current models might not be smart enough
but it is eminently possible meaning if we can't do it today certainly by this time next year we
will
(05:54) have fully autonomous agent swarms able to do pretty much anything if it's got an
API it'll be able to talk to it which that means that uh apis that connect to data stores to
codebases uh to model training you know AI based uh apis pretty much anything with an API
will be automatable uh by these agent swarms this time next year which is why I'm still
confident saying that any definition that you have of AGI will almost certainly be satisfied by
this time next year if not sooner I I ran a poll yesterday and I think
(06:27) something around 40% of people 30 to 40% of people think that AGI has already
been achieved in secret and then another 40 to 50% think that AGI uh is eminently going to
be achieved and I suspect that agent swarms are going to be one of the most practical ways
to implement it because rather than putting all of your eggs in one basket what you're doing
with agent swarms is that you're putting you're basically dividing and conquering uh so that's
kind of the the high level now real quick plug for my patreon I've
(06:57) recently moved to doing webinars for the pre premium uh tier subscribers um so the
next webinar is December 8th at 11:00 a.m. us Eastern um I just sent out the invite to all of
my existing patreon subscribers um but yeah so this is this is what I'm calling my AI master
class so in this in these sessions which I already had the first one it's recorded So if you sign
up you can go watch the uh the recording from November um but it's a monthly webinar
session where I will give you the latest and greatest
(07:26) news um from everything that I'm seeing and all the people that I talk to so this is
technical news as well as more uh political and financial news so I talked to Business
Leaders I talked to researchers I talked to lawyers I talk to all kinds of people um that are
plugged into not just AI but the high-tech and and fast moving space so I can give you kind
of a little bit of behind the scenes uh word on the street and then we also talk about systems
thinking because systems thinking is how I do everything that I do and I've got
(07:56) an upand cominging book on systems thinking uh and then finally we move to a Q&A
session uh which the since I just upgraded to zoom webinars it'll allow actually all of you to to
to have sidebar chats and vote on the questions that you want me to answer so uh go ahead
and jump in on patreon I have two tiers there's the basic Discord tier which will get you
access to the uh my private Discord Community as well as some exclusive content and then
there's the premium tier which gets you even more exclusive content on top of
(08:27) everything else such as these webinars okay so as promised there are two
open-source projects that I am running that have to do with autonomous AI so the first one is
the Ace framework the autonomous cognitive Entity framework so this is the culmination of
more than four years of research that I've been doing on cognitive architectures so what is a
cognitive architecture a cognitive architecture is just a software blueprint or a software plan
to take these AI models these latest and greatest AI models and organize them in
(09:00) such a way that you can create basically a single agent so the ace framework has
taken a while to find its footing but we finally made an executive decision uh particularly in
light of open ai's progress towards cloud-based agents so we have decided that the ace
framework is going to be 100% open source so that means we're only going to be using
open source models like llama and orca and Falcon and we're also going to focus
exclusively on local hardware so the reason for this is because we want to
(09:31) ensure that autonomous agents are democratically accessible we don't want to we
don't we want to avoid vendor lockin we don't want you to be 100% dependent on Microsoft
and open aai and Amazon and meta and Google Now that being said we do have a second
project that will use those and we'll get into that in just a second but the two overarching
principles the two primary constraints for the ace framework are that it was open source and
local only and so obviously that will constrain what we're capable of doing right now
(10:02) but these models are improving fast the hardware is improving fast as well and so uh
the the link for both of these is on my GitHub which is in the description as well as patreon
by the way all the links are in the description of the video so that's the ace framework so if
you want to jump in you can uh hop on over there uh the GitHub is uh public and then the
hos project so the hierarchical autonomous agent swarm so the hos project is an evolution of
the ace framework but rather than being a little bit more monolithic this is
(10:33) cloud-based and it's uh also going to be probably container based but bit this the
idea here is that we'll use the latest and greatest models such as the assistant API by open
AI to rapidly instantiate uh dozens if not hundreds or thousands of agents and then not only
that but we will also create a class of agents that is designed to control and orchestrate the
rest of the Swarm um so then you have supervisors and Watchers and then you have uh you
know uh agents that are exclusively tasked with monitoring you know things like ethics
(11:07) and legality and resources and that sort of thing so these are both open source
projects now obviously some people will say well open AI is an open source which is true but
the code that we're using to instantiate these agents is open source whereas the ace
framework even the models that we're using were going to be open source entirely open
source so those are the two projects that I'm running you've probably seen on my channel I
kind of post daily updates with these um I have appointed uh some
(11:36) Community leaders so I probably won't use quite as many daily updates on the
YouTube channel um but uh they are there and we also have a private Discord or not private
a public Discord called the autonomous AI lab so longtime viewers of my channel will
remember that uh almost two years ago I I I created the cognitive AI lab which um it kind of
was directionless I didn't really know where what it was for but so we're revisiting this we're
creating a brand new Discord server uh link is on the GitHub Repose
(12:09) it is the autonomous AI lab and it is the Discord server for both the ace framework
and the hos project um and there's also a robotic platform one called the open Murphy
project so Murphy stands for multi-use robotic platform humanoid intelligent entity so
basically we're trying to create the Nester class 5 eventually um so those are the three open
source projects that I'm running um and the again one one uh Discord to rule them all and
there's three repos one for each uh we've already got some Community
(12:39) managers uh set up so these are these are the folks that are kind of leading the
charge on the coding efforts it's completely open collaboration I do have a contributing um
uh document as well as code of conduct so make sure you check those out uh but yeah go
ahead and jump on in and start contributing and start collaborating the one thing that I will
say is that because it is full fully open um the number one signal that I'm optimizing is
reducing the signal to noise ratio or I guess increasing the
(13:06) signal to noise ratio so uh don't add don't add Superfluous noise and everything will
be fine so thanks for watching I hope you got a lot out of this I hope this uh helps Orient you
to the direction that things are going and um yeah I'm not going to say whether or not agent
swarms are AGI uh I don't really think that that's a useful description but objectively speaking
what we're building right now is Agent SW and we'll take it as far as we can get so uh look
forward to seeing all of
(13:33) you jumping in on Ace hos and Murphy cheers