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.
190 lines
5.4 KiB
190 lines
5.4 KiB
WEBVTT |
|
|
|
00:00.110 --> 00:05.750 |
|
And we will conclude our expedition into the world of frontier models through their chat interface by |
|
|
|
00:05.780 --> 00:08.570 |
|
looking at meta AI and perplexity. |
|
|
|
00:08.600 --> 00:15.770 |
|
Meta AI is, of course, the front end version to llama that's running behind the scenes. |
|
|
|
00:15.770 --> 00:21.680 |
|
We can ask it the same question about how does it compare itself to other models, and we'll get back |
|
|
|
00:21.680 --> 00:25.190 |
|
something that has some strengths and weaknesses. |
|
|
|
00:25.190 --> 00:27.440 |
|
It doesn't do a great job. |
|
|
|
00:27.440 --> 00:33.530 |
|
It gives some rather old fashioned complimentary llms, but it's it's okay. |
|
|
|
00:33.530 --> 00:38.000 |
|
And generally speaking, I think you'd find asking various questions that you'll get answers that are |
|
|
|
00:38.000 --> 00:42.200 |
|
okay, but not the same power as some of the others. |
|
|
|
00:42.590 --> 00:44.060 |
|
Let's ask the same question. |
|
|
|
00:44.060 --> 00:52.370 |
|
You guys are going to be fed up with me for doing this, but how many times does the letter A appear |
|
|
|
00:52.370 --> 00:54.650 |
|
in this sentence? |
|
|
|
00:54.920 --> 00:58.190 |
|
And let's see what we get from meta AI. |
|
|
|
00:58.340 --> 01:04.160 |
|
It appears five times, so meta is also not able to handle that particular question. |
|
|
|
01:04.190 --> 01:08.210 |
|
Now, one thing that the meta is able to handle is image generation. |
|
|
|
01:08.210 --> 01:17.320 |
|
And we can say please generate an image of a rainbow of rainbows |
|
|
|
01:19.270 --> 01:23.830 |
|
leaping from Hawaii to 17. |
|
|
|
01:24.100 --> 01:31.450 |
|
Uh, and you'll find that this is the kind of, of challenge that, uh, lama is up for. |
|
|
|
01:31.450 --> 01:35.980 |
|
And it very nicely does it with these four possibilities. |
|
|
|
01:36.070 --> 01:38.350 |
|
And you get this, uh, this kind of effect. |
|
|
|
01:38.350 --> 01:44.260 |
|
And it seems to have, uh, well, Hawaii has appeared, uh, but but some of these are very respectable, |
|
|
|
01:44.260 --> 01:49.840 |
|
particularly for the open source model that is, uh, Lama sitting behind the scenes. |
|
|
|
01:49.960 --> 01:55.510 |
|
All right, let's flip over to perplexity, which is, of course, a search engine, not an LLM. |
|
|
|
01:55.510 --> 01:56.080 |
|
So it doesn't. |
|
|
|
01:56.110 --> 02:01.660 |
|
It's the odd one out in this group, although actually, uh, OpenAI also is now in the search space, |
|
|
|
02:01.660 --> 02:02.320 |
|
too. |
|
|
|
02:02.440 --> 02:08.020 |
|
Um, so really it's looking for, uh, factual questions that it can then research and provide an answer |
|
|
|
02:08.020 --> 02:08.380 |
|
for. |
|
|
|
02:08.380 --> 02:13.240 |
|
And I'm recording this on November the 6th, the day after the elections in the US. |
|
|
|
02:13.240 --> 02:22.940 |
|
So I can say something like, who is the president elect of the United States, and it will do some |
|
|
|
02:22.940 --> 02:23.570 |
|
thinking. |
|
|
|
02:23.570 --> 02:29.960 |
|
And I would not be surprised at all to see that it's able to summarize back the outcome and give key |
|
|
|
02:29.960 --> 02:33.110 |
|
points, reactions and the like. |
|
|
|
02:33.170 --> 02:40.280 |
|
Uh, and so it's able to do this and provide a nuanced, well crafted response to current events. |
|
|
|
02:40.310 --> 02:46.160 |
|
Actually, if you ask a question like that to, uh, OpenAI to GPT right now, it will also give you |
|
|
|
02:46.160 --> 02:50.990 |
|
a good answer, uh, based on current events, despite its knowledge cutoff being last year. |
|
|
|
02:51.110 --> 02:53.210 |
|
But if you ask Claude, it won't be able to do that. |
|
|
|
02:53.210 --> 02:55.610 |
|
And it will say it will say to to be direct. |
|
|
|
02:55.610 --> 03:00.440 |
|
My knowledge cutoff is, uh, you get that very specific answer. |
|
|
|
03:00.890 --> 03:04.550 |
|
Uh, let's ask the question, uh, how many? |
|
|
|
03:04.580 --> 03:15.350 |
|
Let's start a new, fresh chat and say, how many times does the letter A appear in this sentence? |
|
|
|
03:16.130 --> 03:17.960 |
|
It's the last time you have to see this. |
|
|
|
03:18.170 --> 03:20.510 |
|
Uh, and it says four times. |
|
|
|
03:20.510 --> 03:23.030 |
|
So it is able to count. |
|
|
|
03:23.060 --> 03:24.800 |
|
Uh, so it's impressive. |
|
|
|
03:24.830 --> 03:28.310 |
|
Uh, don't know whether it's a coincidence, whether it's because other people have written articles |
|
|
|
03:28.310 --> 03:32.300 |
|
about this that it's found, but it is able to count for times. |
|
|
|
03:32.300 --> 03:39.080 |
|
So perplexity is with the oh one preview version and being able to get this right. |
|
|
|
03:39.110 --> 03:41.420 |
|
Congratulations to perplexity. |
|
|
|
03:41.630 --> 03:48.290 |
|
Um, and now ask a slightly curious question, which is, ah, uh, question about comparing to other |
|
|
|
03:48.290 --> 03:52.910 |
|
models, and you'll see a, um, here's the response. |
|
|
|
03:53.000 --> 03:55.340 |
|
Uh, the area is indicated by perplexity. |
|
|
|
03:55.340 --> 03:57.260 |
|
I don't have the capabilities. |
|
|
|
03:57.260 --> 04:00.440 |
|
So it definitely pushes back firmly on that. |
|
|
|
04:00.680 --> 04:04.460 |
|
And that is a wrap on our exploration of frontier models. |
|
|
|
04:04.460 --> 04:07.160 |
|
But now I encourage you to do the same. |
|
|
|
04:07.160 --> 04:12.560 |
|
Come up with interesting questions, particularly try and find questions which are able to bring to |
|
|
|
04:12.590 --> 04:17.960 |
|
the surface the differences between the models, their characters, what they're good at, where they're |
|
|
|
04:17.960 --> 04:18.590 |
|
weak. |
|
|
|
04:18.620 --> 04:23.420 |
|
And if you find something good, then please share it with me or post it in messages. |
|
|
|
04:23.450 --> 04:29.810 |
|
Uh, this is it's really great to find the kinds of prompts that help to surface these differences, |
|
|
|
04:29.810 --> 04:32.660 |
|
and also that help to highlight where they are so strong. |
|
|
|
04:32.750 --> 04:36.050 |
|
And I will see you in the next video to wrap this up.
|
|
|