From the uDemy course on LLM engineering.
https://www.udemy.com/course/llm-engineering-master-ai-and-large-language-models
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.
199 lines
6.1 KiB
199 lines
6.1 KiB
WEBVTT |
|
|
|
00:00.530 --> 00:06.770 |
|
I realized my enthusiasm for this project is a bit insane, but I have to tell you that I am very satisfied |
|
|
|
00:06.770 --> 00:07.670 |
|
with the outcome. |
|
|
|
00:07.700 --> 00:12.620 |
|
It's rare that you come up with a project like this that has so many moving parts and is technically |
|
|
|
00:12.620 --> 00:20.240 |
|
quite complex and has a lot of substance to it, but have such a simple and tangible outcome as our |
|
|
|
00:20.240 --> 00:22.340 |
|
push notification for a good deal. |
|
|
|
00:22.520 --> 00:29.750 |
|
There is something beyond the king of the notification that really makes this very satisfying indeed. |
|
|
|
00:30.290 --> 00:35.660 |
|
But one point that I wanted to make about it, perhaps a more, more seriously, is to point out that |
|
|
|
00:35.660 --> 00:41.120 |
|
this sort of template that we've got here for solving this problem is something that you can apply to |
|
|
|
00:41.150 --> 00:46.940 |
|
a lot of other similar kinds of business problems and startup ideas and commercial problems that you |
|
|
|
00:46.940 --> 00:48.680 |
|
might even have been thinking about already. |
|
|
|
00:48.710 --> 00:53.570 |
|
I mean, to give you something that comes to mind right away is that when I was looking on hugging face |
|
|
|
00:53.570 --> 01:00.020 |
|
for the data for this this project, for the Amazon pricing data, I also came across a bunch of data |
|
|
|
01:00.020 --> 01:07.670 |
|
that that has company reports, annual reports from companies and shows how that relates to the change |
|
|
|
01:07.670 --> 01:14.610 |
|
in share price on the day and the few days after the report comes out, which is a very interesting |
|
|
|
01:14.610 --> 01:15.420 |
|
data set. |
|
|
|
01:15.420 --> 01:22.110 |
|
And it occurs to me that whilst this is more of a stretch, it is also obviously a bigger opportunity |
|
|
|
01:22.110 --> 01:23.580 |
|
if you can find a way to crack it. |
|
|
|
01:23.610 --> 01:30.810 |
|
If there's a way to build a proprietary model that's able to look at that and make some prediction about |
|
|
|
01:30.810 --> 01:35.130 |
|
what will happen to the share price as a result of a company's report being published. |
|
|
|
01:35.160 --> 01:41.910 |
|
And you could imagine then having a scanner agent that scans the new company reports being being being |
|
|
|
01:41.910 --> 01:42.630 |
|
released. |
|
|
|
01:42.660 --> 01:49.170 |
|
And so you can really see the very similar agent workflow that you would adapt for that kind of of of |
|
|
|
01:49.170 --> 01:49.710 |
|
problem. |
|
|
|
01:49.710 --> 01:54.660 |
|
Now, of course, I will tell you that any trading decisions that you make are entirely at your own |
|
|
|
01:54.660 --> 01:56.880 |
|
risk, and I have nothing to do with them at all. |
|
|
|
01:57.060 --> 02:03.120 |
|
And obviously no one has yet cracked that problem of being able to build an automated AI system that |
|
|
|
02:03.120 --> 02:09.000 |
|
can do things like read company reports and make, uh, trading decisions robust trading decisions as |
|
|
|
02:09.000 --> 02:09.510 |
|
a result. |
|
|
|
02:09.510 --> 02:15.810 |
|
But it seems like a serious, a very, um, important challenge to take on. |
|
|
|
02:15.810 --> 02:21.060 |
|
And I wanted to point that out because it shows you how similar that is in terms of the the bits and |
|
|
|
02:21.060 --> 02:24.360 |
|
pieces of the puzzle that you would look at if you wanted to do that. |
|
|
|
02:24.360 --> 02:33.160 |
|
So that could certainly be a big challenge, if nothing else, just for the interest of seeing what |
|
|
|
02:33.160 --> 02:38.200 |
|
happens and seeing whether you can find signal in that kind of data and using it as an opportunity to |
|
|
|
02:38.230 --> 02:40.600 |
|
build a similar agent workflow. |
|
|
|
02:41.410 --> 02:42.040 |
|
Okay. |
|
|
|
02:42.070 --> 02:49.930 |
|
Anyway, with that, let's move to the last slide of this week, which is of this sorry, not of this |
|
|
|
02:49.930 --> 02:54.430 |
|
week, of this day, which is to tell you that there is now one day left. |
|
|
|
02:54.430 --> 02:59.890 |
|
There is one final day, week eight, day five, which will be the conclusion of this course. |
|
|
|
02:59.890 --> 03:01.930 |
|
And you don't want to miss it. |
|
|
|
03:02.050 --> 03:07.210 |
|
Uh, before I tell you what you'll do, then just to remind you, at this point you can generate code |
|
|
|
03:07.210 --> 03:13.510 |
|
and text with frontier models, with open source models, using hugging face, using APIs with tools, |
|
|
|
03:13.510 --> 03:18.760 |
|
structured outputs, rag with assistance from from Gradio. |
|
|
|
03:18.790 --> 03:25.180 |
|
You can follow that strategy we went through to solve problems including curating your data, which |
|
|
|
03:25.180 --> 03:30.280 |
|
is hard and grueling, but perhaps the single most important step. |
|
|
|
03:30.460 --> 03:34.330 |
|
Um, making your baseline model and fine tuning a frontier model. |
|
|
|
03:34.330 --> 03:38.380 |
|
And then you can Select and train an open source model. |
|
|
|
03:38.380 --> 03:39.370 |
|
And we did it. |
|
|
|
03:39.400 --> 03:42.010 |
|
We beat the frontier when we did that. |
|
|
|
03:42.040 --> 03:44.770 |
|
You can deploy models to production. |
|
|
|
03:44.770 --> 03:48.730 |
|
Now you can use modal to to put something up there on the cloud. |
|
|
|
03:48.760 --> 03:53.980 |
|
You can also productionize your code in other ways and prepare it into something that can be run in |
|
|
|
03:53.980 --> 03:55.750 |
|
production environments. |
|
|
|
03:56.140 --> 04:03.400 |
|
And now you can build agentic workflows with planning memory, databases and multiple agents working |
|
|
|
04:03.430 --> 04:07.030 |
|
together to take a complex task and break it down. |
|
|
|
04:07.360 --> 04:11.470 |
|
And so that brings us to preparing for tomorrow. |
|
|
|
04:11.470 --> 04:16.780 |
|
Tomorrow, of course, we're going to be building a user interface, which is going to be glorious because |
|
|
|
04:16.780 --> 04:17.920 |
|
it's in Gradio. |
|
|
|
04:18.040 --> 04:23.980 |
|
Uh, we'll also be looking at the fact that today when we ran that workflow, you'll notice that it |
|
|
|
04:23.980 --> 04:25.270 |
|
ran and then it completed. |
|
|
|
04:25.270 --> 04:30.730 |
|
So it's not really autonomous and it doesn't keep running and have it still relies on the human to press |
|
|
|
04:30.730 --> 04:31.330 |
|
it each time. |
|
|
|
04:31.330 --> 04:32.950 |
|
So we need to fix that of course. |
|
|
|
04:32.950 --> 04:35.320 |
|
So we will do both of those things tomorrow. |
|
|
|
04:35.320 --> 04:43.270 |
|
And once they are done, that will leave you in a position where you will have mastered AI and LM engineering. |
|
|
|
04:43.330 --> 04:45.100 |
|
And I can't wait for that point. |
|
|
|
04:45.100 --> 04:49.300 |
|
I hope that you are both excited and proud and I will see you there.
|
|
|