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.
 
 

283 lines
7.6 KiB

WEBVTT
00:00.890 --> 00:03.800
That concludes a mammoth project.
00:03.830 --> 00:05.780
Three weeks in the making.
00:05.780 --> 00:12.440
In the course of those three weeks, starting with the data curation, the working with frontier models,
00:12.440 --> 00:20.360
and then ending with the complete user interface in Gradio with even that very unnecessary but great
00:20.360 --> 00:27.650
plot 3D chart on the bottom right, together with the trace from the agents and the results of the memory.
00:27.770 --> 00:30.110
Um, and all of it came together.
00:30.110 --> 00:32.360
Remember, the user interface was not really the point.
00:32.360 --> 00:34.820
The user interface was a bit extra, so we can monitor it.
00:34.820 --> 00:41.420
The main point is that that just runs and it just keeps running, and as it runs, it will be continually
00:41.420 --> 00:47.090
notifying me every few minutes with a new deal that it hasn't surfaced before.
00:47.150 --> 00:55.070
And that is the conclusion of a very satisfying, great project that we have built end to end.
00:55.910 --> 01:03.310
And so with that, let's take a moment to do a retro of the last eight weeks.
01:03.310 --> 01:04.510
One more time.
01:04.510 --> 01:10.930
I've gone on about this so many times, and so you'll be happy, or maybe a bit sad that this is the
01:10.930 --> 01:14.590
final retrospective you started eight weeks ago.
01:14.590 --> 01:23.260
Over on the left, we wanted to get to being an LM engineer, someone who had mastered LM engineering,
01:23.260 --> 01:27.070
highly proficient, advanced, and this is how we got there.
01:27.070 --> 01:29.530
In the first week we played with lots of models.
01:29.530 --> 01:32.920
We found out how many times the letter A appeared in sentences.
01:32.920 --> 01:35.920
We took a quick look at one preview amongst other things.
01:35.920 --> 01:42.610
In week two we first saw gradient and we played with some multi-modality, which was fun.
01:42.610 --> 01:46.600
We also saw an early version of agent ization.
01:46.600 --> 01:49.540
In week three we got stuck into hugging face.
01:49.540 --> 01:52.750
We had pipelines, we had tokenizers, we had models.
01:52.750 --> 01:55.810
In week four we got deeper into hugging face.
01:55.810 --> 02:02.710
We selected Llms, we generated code and we had that remarkable project, with the 60 000 time performance
02:02.710 --> 02:04.480
improvement in week five.
02:04.510 --> 02:13.930
We built our Rag solution for insert film, which used chroma, and also briefly we used face as well
02:14.170 --> 02:16.390
and created our expert.
02:16.390 --> 02:19.210
And maybe you did the big project associated with that.
02:19.240 --> 02:24.670
Week six we fine tuned a frontier model, although most of which six was spent curating data.
02:24.670 --> 02:26.950
But that is such an important activity.
02:27.160 --> 02:27.910
Week seven.
02:27.910 --> 02:34.600
We did fine tune an open source model that then beat the frontier, and in week eight, we packaged
02:34.600 --> 02:43.480
it together to a genetic AI solution complete with seven agents and a user interface.
02:43.480 --> 02:45.700
And it was fabulous.
02:45.700 --> 02:48.580
So that was the journey.
02:49.600 --> 02:52.270
I need to take a moment to thank you.
02:52.300 --> 02:55.750
Thank you so much for staying through to the end.
02:55.780 --> 03:00.160
You can't, I can't I can't explain how much I appreciate it.
03:00.190 --> 03:06.900
It's really, uh, so, so wonderful to have had people come all the way through the course, gone through
03:06.900 --> 03:12.330
the eight week journey, and take advantage of everything that we've been doing and get to this point.
03:12.330 --> 03:17.910
And I've, I've heard from several of you along the way, and it's been really, really rewarding for
03:17.910 --> 03:19.530
me to experience this.
03:19.560 --> 03:21.330
Uh, super grateful.
03:21.480 --> 03:23.250
I hope you've enjoyed it.
03:23.280 --> 03:25.350
I obviously I've enjoyed it a lot.
03:25.380 --> 03:26.280
Uh, far too much.
03:26.310 --> 03:28.020
I hope you've enjoyed it as well.
03:28.050 --> 03:29.820
I really hope you can stay in touch.
03:29.820 --> 03:32.340
By all means, please do LinkedIn with me if you're open to that.
03:32.340 --> 03:34.470
If you're okay with that, I'm very much welcome.
03:34.470 --> 03:35.880
LinkedIn connections.
03:35.880 --> 03:40.890
And we can have a community, um, and message me if you've got to this point, I definitely want to
03:40.890 --> 03:41.550
hear it.
03:41.700 --> 03:44.160
Uh, and of course, you've got this big challenge.
03:44.160 --> 03:48.150
Now, can you take what you've learned and use it to build your own project?
03:48.180 --> 03:52.260
Maybe that idea I had about using the finance data would be an interesting one.
03:52.290 --> 03:55.230
See if you can build something that could make some money.
03:55.230 --> 03:57.840
If you do, then I expect a lunch out of it.
03:57.870 --> 03:59.340
At the very least, perhaps.
03:59.550 --> 04:01.280
Uh, but that would be a fun challenge.
04:01.520 --> 04:04.010
Whatever you do with it, I want to hear about it.
04:04.040 --> 04:10.130
If you built a great platform that is using some of this learning, then please share it.
04:10.160 --> 04:16.940
I'd love other students to see that too, and it's great to have that kind of output as a tiny little
04:16.940 --> 04:17.360
extra.
04:17.390 --> 04:21.170
I don't know if you remember all the way, way, way, way back in week one, I did mention there was
04:21.170 --> 04:25.100
going to be a little extra juicy nugget at the very end, and this is what it is.
04:25.130 --> 04:30.590
I wanted to tell you that I did a personal project where I fine tuned an LLM.
04:30.590 --> 04:37.880
It was in fact a llama two LLM from the beginning of this year, uh, on all of my text message history.
04:37.910 --> 04:45.050
It turns out that I have 240,000 text messages that have built up over time on my iPhone, uh, since
04:45.050 --> 04:47.300
I had the first iPhone some time ago.
04:47.420 --> 04:53.180
And so I had a lot of text message history, and I was able to use that to train llama two to make a
04:53.180 --> 04:54.800
simulation of me.
04:54.890 --> 04:59.600
Uh, and there's a write up on my website, on my blog, Edward Dot com.
04:59.600 --> 05:03.200
You can take a look and there's instructions for how you can do it too.
05:03.230 --> 05:09.560
And of course, I was using llama two and now llama 3.1 is so much better along with some of the others
05:09.590 --> 05:11.030
like like Kwan and so on.
05:11.210 --> 05:16.640
So you could definitely have a stab at this, and you will probably have even better results than I
05:16.640 --> 05:16.880
had.
05:16.910 --> 05:19.010
And the results I had were spooky.
05:19.310 --> 05:21.230
They really were very good indeed.
05:22.310 --> 05:27.410
And so with that, I have to bring up the final slide.
05:27.470 --> 05:29.300
Congratulations.
05:29.690 --> 05:32.270
I hope that you're proud of what you've accomplished.
05:32.300 --> 05:38.150
I hope you do feel that sense that you have now reached an advanced point in your learning.
05:38.180 --> 05:41.120
You have got to the summit of the mountain.
05:41.360 --> 05:43.610
I'm so, so very happy.
05:43.820 --> 05:51.200
And I hope, I really, really hope that you're able to take this and use it in your day job, in your
05:51.200 --> 05:52.580
career to move forwards.
05:52.580 --> 05:54.860
And I very much want to hear all about it.
05:54.860 --> 05:56.540
So do stay in touch.
05:56.540 --> 05:58.280
Thank you once again.
05:58.280 --> 06:01.250
And a huge congratulations.