WEBVTT

00:39:42.520 --> 00:39:45.800
We hear about agents
and even autonomous agents.

00:39:45.800 --> 00:39:47.960
How different is this from a program?

00:39:52.440 --> 00:39:53.520
By autonomous agents,

00:39:53.520 --> 00:39:59.240
we understand systems that can
iteratively perform certain actions.

00:39:59.240 --> 00:40:00.680
They can sense the environment,

00:40:00.680 --> 00:40:03.800
they can set up plans,
and then they can execute and act

00:40:03.800 --> 00:40:07.600
on the environment, collect back
feedback and do all this iteratively.

00:40:07.600 --> 00:40:11.120
So, that is what makes it different
from a typical program.

00:40:11.120 --> 00:40:14.120
The concept has existed
since the seventies.

00:40:14.120 --> 00:40:17.920
But what's interesting
is that only recently have

00:40:17.920 --> 00:40:24.040
we found a way to make them really
autonomous by providing LLMs.

00:40:24.040 --> 00:40:27.200
So, in this case, LLMs can be used as a brain

00:40:27.200 --> 00:40:31.560
for this type of agents, can actually
plan for them, can collect

00:40:31.560 --> 00:40:35.960
and analyze the feedback that agents
receive from the environment.

00:40:35.960 --> 00:40:38.680
Can you give us situations, examples,

00:40:38.680 --> 00:40:43.240
where you've seen, we've seen
those autonomous agents in action?

00:40:43.240 --> 00:40:46.080
So again, this is a very,
very new development.

00:40:46.080 --> 00:40:51.240
But you can think of using autonomous
agents to control other software.

00:40:51.240 --> 00:40:54.520
For example, an autonomous agent can
create configuration files,

00:40:54.520 --> 00:40:58.320
can launch software,
can receive results from those

00:40:58.320 --> 00:41:02.360
computations done by other software,
can analyze them.

00:41:02.360 --> 00:41:06.720
And if doesn't like it, change
again the configuration and run it again.

00:41:06.720 --> 00:41:09.520
So, in some sense,
it can replace a lot of human work

00:41:09.520 --> 00:41:11.960
and become really an agent.

00:41:11.960 --> 00:41:15.360
It's both fascinating,
but also a bit scary.

00:41:15.360 --> 00:41:17.680
Do you think there's a line
we should draw somewhere

00:41:17.680 --> 00:41:20.600
to protect society, mankind?

00:41:20.600 --> 00:41:22.840
One is we need to prevent
an autonomous agent

00:41:22.840 --> 00:41:25.240
from being able to modify itself,

00:41:25.240 --> 00:41:30.320
to change its own code,
change its own program.

00:41:30.320 --> 00:41:33.240
And the second point is, we want to limit

00:41:33.240 --> 00:41:35.960
the ability of the agents
to do goal settings.

00:41:35.960 --> 00:41:38.960
We still want humans to control overarching goals.

00:41:38.960 --> 00:41:41.680
Any other technology that has been created so far

00:41:41.680 --> 00:41:45.960
has been capable of spreading
intelligence, transmitting intelligence.

00:41:45.960 --> 00:41:47.440
This is the first time we’ve created

00:41:47.440 --> 00:41:50.600
technology that can create intelligence.
And this can be a little scary.