[00:00:00] Okay, in this video we're gonna be looking
at a form of autonomous AI or an autonomous AI agent. And this is actually from a paper
that was released online called Task-Driven Autonomous Agent using,GPT-4, Pinecone, LangChain
for diverse applications. it was actually announced on Twitter. I saw this,last week, by Johei Nakajima. and
he had a nice sort of, tweet post announcing it and if we go through it, we can see that,
okay. The idea here is that we are using a large language model to generate ideas, do
some sort of critique on those ideas and thenideally execute tools So this has
elements of throwbacks to Toolformer, some of the other key papers around this, that
have been here. one of these things shows, it's got some nice diagrams in here, showing
like how this works and some of the key points into it. So the user basically provides an objective
and a task and then once that is set into a task queue, [00:01:00] the large language
model basically decides how to execute that, and then as it, develops stuff, it saves it
to a memory. and that memory then can be accessed through different things. it then, goes back
a, around these loops. it's got a prioritization agent to decide
what. Task comes next and what you know is priority, et cetera. And it basically goes,
through this, over time. So if we look in here, there, there's, they've got another
diagram in here, just going through it as well, of basically you've got this task queue. we've got the execution agent. And each of
these things,for these parts is actually just the same language. you're just using different
prompts. You're using different ways to, manipulate the output. And then you've got the memory,
which in this case they're using Pinecone, which is a vector store database, so that
they can store things in there and that you can basically do lookups on them and stuff
like that as well. So the. The idea here I think is really good.
the paper talks a lot about using, GPT-4 Pine Cone [00:02:00] and the LangChain Framework,to
basically do this. along with this, he also released, some, code. So the code, has got
the nickname, baby AGI. I don't think this is approaching AGI in any way, but it's cute
name. And you know what, this is supposedly a very
paired down version of the original one. So we don't actually know how good the original
one was. As far as I understand, we hasn't been released or we haven't, got, videos of
trying it and stuff like that. looking at the baby g i, that's here. this is de this code has definitely been released.
We can play around with it. In fact, I've set. A CoLab so that you can also have a play
with it. And so we can just have a look at, how it works, some of the ideas behind it,
and what you could do with an agent like this, in the future. So you need quite a number
of API keys to get this going. you'll need an opening AI key. of course you
can basically set it up to use either GPT-4, or in this case I'm using GPT 3.5 turbo. [00:03:00]
It's nice enough to have a print statement that does say if you're using GPT-4, that,
you know, this could get expensive. we also need Pinecone API key and you need to set
up the Pinecone environment. So this will change based on your, where you're
setting it up. I just left this in so that you could get an idea of what you should be
putting in there. Cause I think it's not always totally clear. you need to put in a table
name. you can't use underscores or anything in this table name, here. And then you need to set up an objective and
initial task. So here I've basically said, okay, plan a romantic dinner for my wife this
Friday night in Central Singapore. and the initial task is make a list of the tasks,
and you'll see That while this particular, one is not set up with any tools or anything,
it does a nice job at going through, the thinking processes that it would need to do. Now in here it does seem like they're planning
on adding tools. to, to this, looking at the, at the code base there are, tools being added.
it's also interesting that looks like they're planning to add, or they've added,the [00:04:00]
llama, I'm guessing this is the four bit version of llama that runs locally, to do this thing. So that, that would be interesting to. See
how well that does. if we go through, we can look and see, okay, after it's got a lot of
setup code, for doing this and for Pinecone setup code,the main logic behind these things,
is not, overly complex. So we can see that we've got the task creation agent. So this is basically got its own prompt here
and we've just got nice sring basically substituting, doing kind of what LangChain does, with its
prompts. and it's then able to, basically call that it's got a prioritization, agent
again, same concept, but with a different prompt. So you are using a different prompt
to do the same thing. you can see here you're a task prioritization,
AI tasked with cleaning, the formatting and reprioritizing the following tasks. And the
tasks in tasks. Consider the ultimate objective of your team passes in that as well. So this
is very similar to some of the agents that we see in LangChain and that we've,looked
at as well.[00:05:00] We've then got at the execution agent, again,
same concept, different prompt. You are an AI who performs one task based on the following
objective. Take into account these previously completed tasks, passing into context, your
task, and then the response. so it's interesting here that they're setting the temperature
quite high, for. whereas normally in Chen you would actually
set the temperature pretty low, like close to zero or zero, for this kind of thing. so
that may have also of affected its output, for doing this. alright, so then you basically
just go,you've just got this huge leap. It goes through it. Does it, let's look at some of the output
that we are getting from. So first off, make a list of tasks. so you can see that it does
a pretty nice job of choose a romantic restaurant in central Singapore. Make a reservation for
two at the chosen restaurant. select a bouquet of flowers to surprise your wife with. This is definitely not something I asked for,
but okay. Maybe it's something that it, decided. you could imagine in the [00:06:00] future
though, you would want the agent to actually come back to you with suggestions and then
you would say yes or no to these suggestions. choose a romantic gift for your wife. purchased a selected gift from the store in
central Singapore. it's really going all out on this dinner. And then, finally confirm
the dinner, et cetera. Okay. So, research and choose a romantic activity to compliment
the dinner experience. if anything, I would say that the agent is very verbose. and again, this would all be down to the manipulation
of the prompt, that you would want for something like this. And you could imagine that this
prompt might be really good for one task, but not great for another task. All right.
It goes through, it comes up with su suggesting a private Sunset yacht cruise, along the Marina
Bay. Now, this, it is, it's very good in that it's
getting,locations, right? And it's getting things like that. Again, this is to be expected,
because we're using, one of the large, open AI models, for doing this. it's [00:07:00]
quite funny how it's, choose a romantic outfit, hire, rent a luxury car,a lot of things that
it, it's making suggestions, but they may not be, Ideal sort of suggestions for, a romantic
date in Singapore. one of the things I did find out was interesting
was, and it's funny how it goes on to say, please note that you may need to provide a
valid driver's license and stuff like that. I, you could imagine in the future that these
things will have a variety of information on you. And then be able to use that, like
if it's got a knowledge base on you of your driver's license, your credit card number,
all those sorts of things. I certainly wouldn't give this one my credit
card number cuz it seems to want to spend a lot of money. so you can see here it's picked
out, a jewelry store. It actually picks out, three real jewelry stores and it get, seems
to get their location correct. Uh, pretty impressive. it then also picks out a nice,
restaurant. and it's interesting that, the restaurant
that it picks is a luxury restaurant. I think it's a Michelin, I'm pretty sure it's a three
star Michelin restaurant, in Singapore. and [00:08:00] so again, this is all coming from
the open AI model. there, there's nothing, unique about this that's coming from Baby
agi. It's just nice manipulation of the open AI
APIs, in this, it goes on and on. It takes, a bit of time to run, through these. in the
end I just of stopped it. cuz it certainly, can ping the API quite a bit and get a lot
of things back. it is interesting to, you know, I tried another one. Planning a party. And that also, did the basic
stuff quite well. what we are lacking here, is the ability for it to come back to you
and to know what it should come back to you about. and this is gonna be one of the key
things, I think for a lot of these things going forward. it claims that it's contacted
the restaurant and it's made,a booking. It's very strange that, about their policy
on bringing outside candles. Again, this is I would say, dying in the details, o of this
kind of thing. Anyway, it's here. You can have a play with it yourself. I ended up stopping
it, just cuz it seemed to be going on and on. [00:09:00] Alright. The idea, I think is the key thing here. the
idea of developing these, agents that have the ability to run a variety of different
tasks and are incorporated with a variety of different tools. That's what we're gonna
see a lot of in the future. we're gonna see this with the, chatGPT plugins or the open
AI plugins format that's coming along. We're already seeing this with some of the
things in LangChain, so this sort of just. Is a nice way of wrapping up some of these
ideas and giving you some idea of how they could be in the future. As always, if you've
got, any questions, put them in the comments below. if you found this useful, please click, like
can subscribe. I will see you in the next video.
