You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 

112 lines
3.6 KiB

WEBVTT
00:00.020 --> 00:03.800
Let me talk about some other phenomena that have happened over the last few years.
00:03.800 --> 00:10.910
One of them has been the rise and fall, perhaps of a new type of job called the Prompt engineer, someone
00:10.910 --> 00:17.780
who specializes in knowing how to craft the right kind of prompts to get the best outcomes from llms.
00:17.810 --> 00:23.780
At one point, this role commanded a $500,000 salary and was hot in demand.
00:23.810 --> 00:26.480
It does seem to have fallen downwards a bit now.
00:26.480 --> 00:31.100
The demand is less, partly because knowing how to prompt well has become ubiquitous.
00:31.100 --> 00:35.690
There's so much content now about the right ways to go about prompting, and also because there are
00:35.690 --> 00:39.050
now tools that will actually create a prompt for you.
00:39.080 --> 00:41.450
Anthropic, in fact, has has one of those tools.
00:41.450 --> 00:45.770
So it is now something that has become relatively common.
00:46.010 --> 00:50.240
Another another phenomenon was the custom gpts.
00:50.270 --> 00:55.160
OpenAI has a GPT store that was incredibly popular for a while.
00:55.190 --> 00:58.010
It's become a little bit saturated at this point.
00:58.010 --> 01:01.960
I think people became fatigued of building custom GPT, gpts, but it's still there.
01:01.990 --> 01:07.210
The GPT store is still reasonably popular, and you can go there to experiment with different kinds
01:07.210 --> 01:09.220
of tuned gpts.
01:10.090 --> 01:20.290
Then of course, still very important was the emergence of co-pilots ways in which a human and an LLM
01:20.320 --> 01:22.180
could collaborate together.
01:22.270 --> 01:27.160
Famously, Microsoft Copilot, I think, was perhaps the first one that really took the world by storm.
01:27.190 --> 01:34.810
GitHub copilot of course, and there are many more co-pilots that are being embedded into a lot of applications.
01:34.810 --> 01:38.890
And in a way, we sort of saw that with canvas a moment ago.
01:39.430 --> 01:48.070
And then the new hot trend right now is all about agent ization, about using Agentic AI, which is
01:48.070 --> 01:52.150
where multiple llms collaborate to solve a problem.
01:52.180 --> 01:59.440
A more complex problem is broken down into smaller steps or smaller tasks, and then particularly tuned
01:59.470 --> 02:06.610
llms are used to tackle each of those steps, perhaps also with an LM responsible for planning and deciding
02:06.610 --> 02:08.410
which LM is doing what.
02:08.440 --> 02:13.900
Also, with a concept of memory, that there's some kind of persistent information that lasts, that
02:13.900 --> 02:20.830
can be exchanged between the LMS and a sense of autonomy, that the LMS don't just exist for the purposes
02:20.830 --> 02:29.080
of a chat interface with a human, but they have a sort of time horizon that that spans multiple chats,
02:29.080 --> 02:29.830
potentially.
02:29.830 --> 02:37.660
So that sense of autonomy and memory and being able to plan tasks and divide tasks down, those are
02:37.660 --> 02:41.320
all some of the core tenants of Agentic AI.
02:41.350 --> 02:45.880
And we'll be coming back to this a few times during the course, but in particular at the end of the
02:45.880 --> 02:51.850
course in week eight, we will build a full Agentic AI solution, as I might have said, a time or two,
02:51.850 --> 02:55.270
but it's really great and it will have eight, seven, seven.
02:55.300 --> 02:56.680
I don't get overexcited.
02:56.710 --> 03:02.050
There are seven agents that will collaborate as part of what we will build at the end.