WEBVTT

00:00.000 --> 00:02.617
-: Okay, let's talk about pre warming.

00:02.617 --> 00:03.787
And so this is a pretty naive prompt;

00:03.787 --> 00:05.910
"Can l have a list of product names for a pair of shoes

00:05.910 --> 00:08.400
that can fit any foot size".

00:08.400 --> 00:10.653
Just gonna see what it comes back with.

00:12.960 --> 00:15.813
Okay, we've got some pretty interesting names.

00:19.200 --> 00:21.330
Okay, but we don't know if these names are any good.

00:21.330 --> 00:23.970
And Chatgpt doesn't either, right?

00:23.970 --> 00:26.700
Like it's just following one token after another.

00:26.700 --> 00:28.474
It's not very good at self-reflection here,

00:28.474 --> 00:30.600
and we haven't really given it any guidance.

00:30.600 --> 00:32.610
And you could obviously go in,

00:32.610 --> 00:34.080
give it some tips,

00:34.080 --> 00:38.910
you can say memorable, easy to understand.

00:38.910 --> 00:40.168
And then

00:40.168 --> 00:42.570
(keyboard clicking)

00:42.570 --> 00:43.740
if we put those

00:43.740 --> 00:46.448
into bullet points

00:46.448 --> 00:49.500
and then we'll see what we get back

00:49.500 --> 00:51.060
and we should hopefully get

00:51.060 --> 00:52.710
something a little bit different.

00:56.520 --> 00:57.390
Yeah, so I think these are

00:57.390 --> 00:58.740
definitely a little bit easier, right?

00:58.740 --> 01:01.410
There're a lot less complex, less syllables.

01:01.410 --> 01:02.879
But the problem is

01:02.879 --> 01:04.770
we had to come up with this context

01:04.770 --> 01:05.832
and it's super

01:05.832 --> 01:08.580
painful to have to do that a lot of times.

01:08.580 --> 01:10.204
And so one really quick, easy way

01:10.204 --> 01:13.200
to do it is funny is pre warning.

01:13.200 --> 01:16.140
So what we do instead is you say,

01:16.140 --> 01:18.150
Can you give advice

01:18.150 --> 01:19.833
or let's say five tips.

01:20.790 --> 01:23.430
Five tips on how to

01:23.430 --> 01:24.820
craft a good product name

01:28.313 --> 01:29.530
(keyboard clicking)

01:29.530 --> 01:30.447
and submit.

01:33.960 --> 01:36.120
And it's pretty good at deciding

01:36.120 --> 01:37.980
what is a good product name, right?

01:37.980 --> 01:39.723
You've seen that in the training data

01:39.723 --> 01:41.760
and it can give us some pretty good advice

01:41.760 --> 01:44.425
that it maybe wouldn't even follow itself by default

01:44.425 --> 01:46.740
because you're not giving it any direction.

01:46.740 --> 01:50.280
And in this case, you're asking it to self-direct, right?

01:50.280 --> 01:53.040
We are asking it for tips first.

01:53.040 --> 01:54.510
And then after that,

01:54.510 --> 01:55.740
so the magic happens,

01:55.740 --> 01:56.940
you'd say,

01:56.940 --> 01:59.130
can I have a

01:59.130 --> 02:00.580
list of product names

02:01.560 --> 02:03.130
that for shoes

02:04.157 --> 02:06.783
that fit any shoe size,

02:07.771 --> 02:08.604
(keyboard clicking)

02:08.604 --> 02:09.968
alright,

02:09.968 --> 02:10.990
and now

02:12.990 --> 02:16.380
it should be using this actual advice

02:16.380 --> 02:18.660
and we can make it even more explicit.

02:18.660 --> 02:19.770
This is good in itself.

02:19.770 --> 02:21.404
I find that even just asking

02:21.404 --> 02:22.920
if you have a task,

02:22.920 --> 02:24.420
how would you do this task?

02:24.420 --> 02:26.310
And then asking it to do the task

02:26.310 --> 02:28.320
is pretty good in itself,

02:28.320 --> 02:30.300
but if you wanna make it more explicit,

02:30.300 --> 02:33.884
you can just perpend it with using this advice.

02:33.884 --> 02:37.260
(keyboard clicking)

02:37.260 --> 02:40.200
Can I have a list of shoe sizes

02:40.200 --> 02:41.580
or a list of product names?

02:41.580 --> 02:43.143
So shoes that fit any size.

02:46.380 --> 02:47.310
And if you,

02:47.310 --> 02:49.853
oh sorry, that's my fault.

02:49.853 --> 02:52.800
I left the other message in there.

02:52.800 --> 02:53.943
Forgot to delete it.

02:57.510 --> 02:58.950
Okay, so I think these are a lot better.

02:58.950 --> 03:02.220
They definitely seem more simple, relatable.

03:02.220 --> 03:05.490
They're unique, they are a bit more memorable, I think

03:05.490 --> 03:07.110
OneSizeStep in particular, I like.

03:07.110 --> 03:10.350
Universal Stride, Adaptiwalk,

03:10.350 --> 03:11.370
those are pretty cool,

03:11.370 --> 03:12.900
OmniSize Sneakers.

03:12.900 --> 03:14.190
These are much better names,

03:14.190 --> 03:15.690
I think than we were getting before.

03:15.690 --> 03:18.940
We could also actually, yeah, sorry, let's submit that again

03:20.820 --> 03:24.600
because it was, yeah, MagicSize Shoes, Adaptstep

03:24.600 --> 03:25.433
Cool.

03:25.433 --> 03:27.030
And then we're gonna add this one more thing.

03:27.030 --> 03:30.630
So we've now pre warmed it with the advice

03:30.630 --> 03:32.640
and we've asked it to do the task.

03:32.640 --> 03:35.340
And we can also ask it to rate the task afterwards.

03:35.340 --> 03:40.340
Say rate your names based on the five tips.

03:40.400 --> 03:43.400
(keyboard clicking)

03:45.480 --> 03:48.510
Here we go. And now it's applying its own advice

03:48.510 --> 03:49.654
to its own names

03:49.654 --> 03:51.900
and we've completed the full loop

03:51.900 --> 03:53.340
of quite a complex task

03:53.340 --> 03:55.680
and we have actual justifications

03:55.680 --> 03:57.210
for each name now as well,

03:57.210 --> 03:59.010
which is pretty cool.

03:59.010 --> 04:00.120
This is just a little trick

04:00.120 --> 04:01.560
and it works with Chatgpt,

04:01.560 --> 04:03.240
works with the playground as well.

04:03.240 --> 04:05.370
Don't bother doing your own warmup.

04:05.370 --> 04:07.353
Get it to warm up itself.