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.
334 lines
9.6 KiB
334 lines
9.6 KiB
WEBVTT |
|
|
|
00:00.050 --> 00:05.810 |
|
Well, I'm delighted to welcome you to day three of our eight week journey together. |
|
|
|
00:05.810 --> 00:09.470 |
|
And today we're going to be looking at Frontier Models. |
|
|
|
00:09.470 --> 00:16.760 |
|
The idea that the goal of today is to get deep into these different models so that you can get a true |
|
|
|
00:16.760 --> 00:21.920 |
|
intuition for where are they strong, where are they weak, what are the differences between them? |
|
|
|
00:21.920 --> 00:26.840 |
|
And so that's what I want you to keep in mind throughout today's material, learning about the differences |
|
|
|
00:26.840 --> 00:31.970 |
|
between them and thinking about how you would apply them commercially to your business or to future |
|
|
|
00:31.970 --> 00:35.570 |
|
projects, and understanding when you would pick which model. |
|
|
|
00:35.600 --> 00:37.040 |
|
Let's get to it. |
|
|
|
00:37.100 --> 00:42.320 |
|
So we're going to be talking about six different models today from six different companies, starting |
|
|
|
00:42.350 --> 00:44.750 |
|
of course, with OpenAI's models. |
|
|
|
00:44.780 --> 00:47.270 |
|
OpenAI needs no introduction really. |
|
|
|
00:47.420 --> 00:50.300 |
|
GPT is the most famous model. |
|
|
|
00:50.300 --> 00:53.090 |
|
And we'll also, of course look at 0101 preview. |
|
|
|
00:53.240 --> 01:01.310 |
|
The the newest of their models and ChatGPT is their user interface, the screens where you can interact |
|
|
|
01:01.310 --> 01:02.150 |
|
with it. |
|
|
|
01:02.360 --> 01:05.180 |
|
We'll also look at the models from anthropic. |
|
|
|
01:05.210 --> 01:12.510 |
|
Anthropic is OpenAI's top competitor, based in San Francisco as well, and founded by some people that |
|
|
|
01:12.510 --> 01:16.380 |
|
left OpenAI, and their model is called Claude and Claude. |
|
|
|
01:16.380 --> 01:19.200 |
|
In fact, you may know, comes in sort of three powers. |
|
|
|
01:19.200 --> 01:21.300 |
|
The smallest one is called haiku. |
|
|
|
01:21.330 --> 01:23.310 |
|
Claude haiku, and then the sonnet. |
|
|
|
01:23.310 --> 01:24.750 |
|
And then there's opus. |
|
|
|
01:24.810 --> 01:31.770 |
|
But actually, because sonnet has had much more recent versions, the latest version of sonnet is stronger |
|
|
|
01:31.770 --> 01:35.190 |
|
than the bigger, more expensive opus, as we'll see. |
|
|
|
01:35.190 --> 01:36.930 |
|
That will make more sense later. |
|
|
|
01:37.050 --> 01:42.360 |
|
But Claude Claude 3.5 sonnet is the strongest of Claude's models. |
|
|
|
01:43.050 --> 01:50.580 |
|
Google has Google Gemini, probably latest to the party, and most of us know Gemini most well because |
|
|
|
01:50.580 --> 01:55.530 |
|
nowadays when we do a Google search, very often we see Gemini's responses. |
|
|
|
01:55.650 --> 02:00.720 |
|
Gemini is, of course, the next generation of what was originally called Bard from Google. |
|
|
|
02:01.320 --> 02:04.290 |
|
Cohere is one that you may have heard less about. |
|
|
|
02:04.290 --> 02:11.880 |
|
It's a Canadian AI company, and their model is most well known for being using using a technique called |
|
|
|
02:11.880 --> 02:14.220 |
|
Rag to make sure that it has expertise. |
|
|
|
02:14.220 --> 02:15.630 |
|
So we will see that. |
|
|
|
02:15.720 --> 02:18.930 |
|
And then we know the llama model from meta. |
|
|
|
02:18.930 --> 02:20.940 |
|
We've used it ourselves through llama. |
|
|
|
02:20.970 --> 02:25.710 |
|
This is an open source model, and you may not know that the meta actually also has a website, meta |
|
|
|
02:25.950 --> 02:30.300 |
|
AI, that lets you interact with the llama model. |
|
|
|
02:30.300 --> 02:32.220 |
|
And we will have a look at that. |
|
|
|
02:32.310 --> 02:39.540 |
|
And then perplexity is a bit different, because perplexity is actually a search engine powered by AI, |
|
|
|
02:39.570 --> 02:43.230 |
|
powered by Llms, and it can use some of the other models that we'll talk about. |
|
|
|
02:43.230 --> 02:48.360 |
|
But they do also have their own model too, so it's a slightly different beast, but we'll be looking |
|
|
|
02:48.360 --> 02:50.190 |
|
at perplexity as well. |
|
|
|
02:51.240 --> 02:58.290 |
|
So overall, these llms are astonishing in what they are capable of. |
|
|
|
02:58.320 --> 03:06.330 |
|
They are really very effective indeed at taking a detailed question, a nuanced question, and providing |
|
|
|
03:06.330 --> 03:10.260 |
|
a structured summary that appears well researched. |
|
|
|
03:10.260 --> 03:15.660 |
|
It often has a sort of introduction and a summary, and this is one of the ways that I use it all the |
|
|
|
03:15.660 --> 03:16.470 |
|
time. |
|
|
|
03:16.660 --> 03:22.900 |
|
and I find that across the board, these llms are shocking in how good they are at this. |
|
|
|
03:22.930 --> 03:28.300 |
|
It's something that a couple of years ago, none of us would have imagined that we could get this far |
|
|
|
03:28.300 --> 03:29.290 |
|
this quickly. |
|
|
|
03:30.040 --> 03:36.010 |
|
There are also really good, and I imagine that many of you do this a lot yourselves, and I do it if |
|
|
|
03:36.040 --> 03:41.080 |
|
you put in a few bullets, just a few notes on something and say, hey, can you turn this into an email |
|
|
|
03:41.110 --> 03:44.230 |
|
or can you turn this into a slide? |
|
|
|
03:44.560 --> 03:51.820 |
|
They are really good at fleshing it out and building, say, a blog post, and they're very good at |
|
|
|
03:51.850 --> 03:52.600 |
|
iterating. |
|
|
|
03:52.600 --> 03:55.360 |
|
So they'll do something and are like some of it I won't like others. |
|
|
|
03:55.360 --> 03:58.330 |
|
And you can give feedback and keep going backwards and forwards. |
|
|
|
03:58.330 --> 04:00.340 |
|
And it's a really effective way of working. |
|
|
|
04:00.340 --> 04:05.260 |
|
It's the kind of copilot construct that is so, so, so effective. |
|
|
|
04:06.220 --> 04:09.340 |
|
And then coding, of course. |
|
|
|
04:09.340 --> 04:17.050 |
|
And perhaps this for many of us is is the thing that is most staggering is how very good the llms are |
|
|
|
04:17.050 --> 04:21.010 |
|
at writing code and debugging problems and solving them. |
|
|
|
04:21.190 --> 04:24.320 |
|
It's something which is really remarkable. |
|
|
|
04:24.320 --> 04:28.610 |
|
I've had experiences myself when I've been working on something that's very complex, and it's something |
|
|
|
04:28.610 --> 04:34.040 |
|
that I believe I have deep subject matter expertise in, and I've got a fairly intricate error, and |
|
|
|
04:34.040 --> 04:41.720 |
|
I put the details and the stack trace in into Claude, say, and I get back not only a very precise |
|
|
|
04:41.720 --> 04:46.880 |
|
explanation of what's going wrong, but also the code that will fix it appearing as an artifact on the |
|
|
|
04:46.880 --> 04:47.750 |
|
right in Claude. |
|
|
|
04:47.750 --> 04:50.330 |
|
And it's it's it's amazing. |
|
|
|
04:50.330 --> 04:51.830 |
|
It's absolutely amazing. |
|
|
|
04:52.010 --> 04:56.210 |
|
And in fact, these are often things which I if I try and paste them, if I look for it in Stack Overflow, |
|
|
|
04:56.240 --> 04:57.470 |
|
there's no answer there. |
|
|
|
04:57.470 --> 05:02.840 |
|
Somehow it's it's able to look beyond just, just a regurgitating Stack Overflow answers. |
|
|
|
05:02.840 --> 05:05.990 |
|
And it seems to have real insight into what's going on. |
|
|
|
05:06.020 --> 05:13.280 |
|
And I suppose that's why it's not surprising, really, that Stack Overflow has seen a big falloff in |
|
|
|
05:13.280 --> 05:14.120 |
|
its traffic. |
|
|
|
05:14.150 --> 05:21.650 |
|
You can see that something started to happen in a big way after Q4 2022, which is when ChatGPT was |
|
|
|
05:21.650 --> 05:22.370 |
|
released. |
|
|
|
05:22.610 --> 05:31.800 |
|
So, you know, it's obviously changed the paradigm of how uh, how we how technology people work with |
|
|
|
05:31.800 --> 05:33.930 |
|
with researching our problems. |
|
|
|
05:33.960 --> 05:35.100 |
|
It's very effective. |
|
|
|
05:35.100 --> 05:41.640 |
|
And I encourage you, if you get stuck with some of the things we work on to give Claude or OpenAI GPT |
|
|
|
05:41.670 --> 05:42.570 |
|
a shot. |
|
|
|
05:43.560 --> 05:45.510 |
|
So what about where are they weak? |
|
|
|
05:45.510 --> 05:47.010 |
|
What are the things that they struggle with? |
|
|
|
05:47.010 --> 05:49.230 |
|
Where does humanity still have a chance in all of this? |
|
|
|
05:49.260 --> 05:56.730 |
|
Well, so first of all, they tend to not be as strong with specialized subject matter if it's something |
|
|
|
05:56.730 --> 05:59.190 |
|
that requires detailed knowledge. |
|
|
|
05:59.280 --> 06:02.250 |
|
Most llms are not yet at PhD level. |
|
|
|
06:02.280 --> 06:09.090 |
|
Now, I had to put the word most in there because literally just just just a few weeks ago for me in |
|
|
|
06:09.090 --> 06:16.620 |
|
October, uh, Claude, the newest version of Claude came out, uh, the latest Claude 3.5 sonnet, |
|
|
|
06:16.710 --> 06:23.070 |
|
uh, and it has surpassed PhD level in maths, physics, chemistry. |
|
|
|
06:23.190 --> 06:29.220 |
|
Uh, and so this is something where very quickly we're seeing these models achieving PhD level. |
|
|
|
06:29.220 --> 06:30.360 |
|
So far just Claude. |
|
|
|
06:30.360 --> 06:36.540 |
|
But the others I'm sure are not far behind, but those are in those specific sciences and in a particular |
|
|
|
06:36.540 --> 06:42.180 |
|
domain, like a business domain, they still won't have the specialist knowledge of an expert in that |
|
|
|
06:42.180 --> 06:42.960 |
|
space. |
|
|
|
06:43.530 --> 06:46.260 |
|
And secondly, recent events. |
|
|
|
06:46.260 --> 06:52.170 |
|
So the models have been trained up until a knowledge cutoff, which is for GPT. |
|
|
|
06:52.200 --> 06:53.370 |
|
October of last year. |
|
|
|
06:53.370 --> 06:58.470 |
|
And so they won't be able to answer questions on information that has come since then. |
|
|
|
06:58.500 --> 07:01.740 |
|
And then they have some strange blind spots. |
|
|
|
07:01.740 --> 07:04.260 |
|
There are some questions which they will just get wrong. |
|
|
|
07:04.260 --> 07:08.580 |
|
And when they get them wrong, one of the things that's quite concerning is that they do tend to be |
|
|
|
07:08.580 --> 07:10.440 |
|
confident in their responses. |
|
|
|
07:10.440 --> 07:14.550 |
|
They often don't volunteer the fact that they're uncertain. |
|
|
|
07:14.550 --> 07:19.320 |
|
They just state an answer with the same level of conviction as with something where they do get the |
|
|
|
07:19.320 --> 07:20.160 |
|
answer right. |
|
|
|
07:20.280 --> 07:28.530 |
|
And that that is something which of course causes concern when you see models hallucinate or come up |
|
|
|
07:28.530 --> 07:32.190 |
|
with with new information which it doesn't know and do so with confidence. |
|
|
|
07:32.190 --> 07:38.670 |
|
And we'll see some examples of that and talk about what are the reasons behind those blind spots.
|
|
|