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.
 
 

211 lines
6.4 KiB

WEBVTT
00:00.950 --> 00:08.210
Now I want to take a quick moment to give you a flyby of five different ways that llms are used commercially,
00:08.210 --> 00:15.710
and you probably know of 50 more, but it's good to take a moment to look at them and to be food for
00:15.710 --> 00:20.840
thought for you to consider other commercial applications of llms.
00:20.840 --> 00:25.850
And while we're doing this, also be thinking about how you would assess the right model for these different
00:25.850 --> 00:31.700
problems, using the kinds of leaderboards and the arena that we've already talked about.
00:31.730 --> 00:37.970
The first one I wanted to show you is a company called Harvey, which uses llms in the field of law,
00:37.970 --> 00:43.610
and you can read through their site, but it gives products for lawyers that will do things like answering
00:43.610 --> 00:47.390
questions on on law, what what is a claim of disloyalty?
00:47.660 --> 00:53.630
And I believe it also does things like looking through legal documents to find key terms and the like.
00:53.630 --> 01:01.370
It makes so much sense to be applying llms to the field of law that it's a no brainer.
01:01.370 --> 01:09.820
It's a classic example of needing to use language and nuance and apply it to difficult business Challenges.
01:09.970 --> 01:11.140
Here's another.
01:11.140 --> 01:14.200
And this is near and dear to my heart because this is my day job.
01:14.200 --> 01:15.730
This is Nebula.
01:15.790 --> 01:24.160
Uh, Nebula io, which is using llms to apply to the world of talent and recruitment, helping managers
01:24.160 --> 01:27.160
to hire and engage with great candidates.
01:27.160 --> 01:33.310
But it's also helping people to explore and understand where they can be most satisfied and successful.
01:33.340 --> 01:37.210
Using Llms to understand the content of people's careers.
01:37.240 --> 01:41.230
Again, it's the kind of use case that makes so much sense.
01:42.610 --> 01:46.420
This one is one that I find particularly annoying.
01:46.420 --> 01:50.950
I am, uh, upset by this particular company.
01:50.950 --> 01:58.480
Blooper AI um bloop, which is a platform that ports.
01:58.510 --> 02:00.790
Legacy code into Java.
02:00.790 --> 02:06.220
And the reason that I find this, uh, galling is because I wish I had thought of it.
02:06.250 --> 02:07.870
It's such a great idea.
02:07.870 --> 02:09.310
It's such an obvious idea.
02:09.340 --> 02:09.700
Obviously.
02:09.700 --> 02:10.450
Brilliant.
02:10.480 --> 02:14.260
As soon as I heard it, I thought, oh, that's a great idea.
02:14.410 --> 02:21.180
Uh, so of course there is tons of COBOL code out there and other legacy code.
02:21.180 --> 02:27.150
And it's it's a huge challenge for large corporations to figure out how to maintain this code, as people
02:27.150 --> 02:29.700
increasingly don't want to work with languages like COBOL.
02:29.700 --> 02:33.600
And there's a lot of legacy code that people can't read and understand.
02:33.720 --> 02:42.270
Um, and it's such a perfect use case for coding models that can learn different programming languages
02:42.270 --> 02:48.480
and then can use that to port from one language to another, and presumably can do things like adding
02:48.480 --> 02:52.290
in comments and test cases and everything else.
02:52.440 --> 02:56.700
Um, so, uh, yes, it's uh, dang it.
02:56.730 --> 03:03.270
They've they've obviously, uh, really struck on a fantastic idea.
03:03.270 --> 03:09.420
Uh, I think this is a great looking product from a Y Combinator backed company.
03:09.420 --> 03:12.780
And, uh, yes, congratulations to to bloop.
03:13.530 --> 03:15.210
And I love the name bloop.
03:15.420 --> 03:24.570
Uh, on on less uh, memorable name would be, uh, this Salesforce product Salesforce's Einstein copilot
03:24.600 --> 03:31.110
health actions Quite a mouthful, but I will say that aside from the slightly clunky name, the product
03:31.110 --> 03:34.680
itself again is one of those ones that's like, oh yes, that makes sense.
03:34.830 --> 03:42.810
Uh, a sort of, uh, dashboard that can be used by healthcare practitioners, for example, to do things
03:42.810 --> 03:51.420
like summarize for a care coordinator, the outcomes of a medical appointment, saving presumably tons
03:51.420 --> 03:58.320
of time, and giving this kind of very compelling summary of what happened, uh, which for all we know,
03:58.350 --> 04:00.600
could be a Gradio app because that easy.
04:00.720 --> 04:08.190
Uh, but it's probably something Salesforce, uh, but it's such a good use case makes a lot of sense.
04:08.190 --> 04:12.990
And obviously, uh, no doubt Salesforce is going to do really well with that product.
04:13.650 --> 04:16.140
And then I came across Khan.
04:16.140 --> 04:17.550
Mingo Khan.
04:17.580 --> 04:17.940
Yes.
04:17.940 --> 04:20.070
I think it's probably how it's pronounced conmigo.
04:20.310 --> 04:24.810
Um, uh, which is, uh, the Khan Academy.
04:24.810 --> 04:28.710
Uh, presumably, uh, with a mingo at the end of it.
04:28.770 --> 04:35.780
Uh, and this is an LLM based platform to be a companion for teachers, learners and parents.
04:35.780 --> 04:37.760
And what a great idea again.
04:37.760 --> 04:44.450
And the Khan Academy is a obviously a fabulous resource, and I'm sure that this is something that this
04:44.450 --> 04:47.750
solution is something that will be immensely valuable.
04:47.750 --> 04:55.610
And the application of LMS to education is something which is again a no brainer, absolute no brainer.
04:55.610 --> 05:05.540
So across these different examples, law, talent, coding, the medical field and education, you can
05:05.540 --> 05:10.100
see the impact that LMS will be able to make in each of these products.
05:10.100 --> 05:15.140
And I'm sure you have many other examples yourself, but hopefully, as I say, this has given you food
05:15.140 --> 05:21.350
for thought and also interesting to think about how you would be assessing different models that you
05:21.350 --> 05:23.210
would pick to solve any of these problems.
05:23.210 --> 05:29.750
To state the obvious, of course, for this one, no doubt we'd be looking at the medical LMS leaderboard
05:29.750 --> 05:35.660
and hugging face, and here we'd be looking at many of the coding metrics, not just human eval for
05:35.660 --> 05:39.650
Python, but the coding metrics for other languages too.
05:40.610 --> 05:43.850
Okay, with that we will go over to wrap up.