WEBVTT

00:00.660 --> 00:02.280
-: Alright, we're gonna learn about the technique.

00:02.280 --> 00:03.690
Let's think step by step.

00:03.690 --> 00:05.400
So this is a magic word

00:05.400 --> 00:07.590
or a magic phrase you can include in order

00:07.590 --> 00:10.710
to improve the reasoning of your AI.

00:10.710 --> 00:14.820
The one problem that you run into pretty frequently,

00:14.820 --> 00:16.530
even with messaging bots

00:16.530 --> 00:18.720
is it doesn't really know the answer.

00:18.720 --> 00:20.580
It might hallucinate the answer,

00:20.580 --> 00:23.400
or like it's saying here, it doesn't have real time data,

00:23.400 --> 00:25.680
so it can start to fail.

00:25.680 --> 00:26.880
So you're saying, you know,

00:26.880 --> 00:30.000
New York City had 500 professional window washers,

00:30.000 --> 00:33.900
so either like refuses to answer or it'll make something up.

00:33.900 --> 00:37.680
I don't think that 500 professional window washers is enough

00:37.680 --> 00:40.230
for all of New York City, but we'll see.

00:40.230 --> 00:42.510
One really simple way

00:42.510 --> 00:45.120
you can just improve things as you say.

00:45.120 --> 00:48.240
Let's think step by step,

00:48.240 --> 00:50.520
and this works particularly well with mathematical

00:50.520 --> 00:53.730
or like kind of linear driven type reasoning.

00:53.730 --> 00:56.430
This is using Fermi estimation.

00:56.430 --> 00:59.400
Enrico Fermi was famous for breaking things down

00:59.400 --> 01:01.470
step by step and then estimating each step.

01:01.470 --> 01:04.053
So let's see how this works for GPT4.

01:06.540 --> 01:07.890
Okay, so it's saying we don't know

01:07.890 --> 01:09.810
how many window washers there are.

01:09.810 --> 01:10.890
We can make an estimate.

01:10.890 --> 01:12.990
There are 12,000 people employed

01:12.990 --> 01:15.930
as building cleaning workers in the USA.

01:15.930 --> 01:17.250
New York is one of the cities

01:17.250 --> 01:18.900
with the most high rise in the world.

01:18.900 --> 01:22.020
So let's say 10% of all window washers work here.

01:22.020 --> 01:26.100
And then so there could be part-time

01:26.100 --> 01:29.010
and New York has 1% of the window washers.

01:29.010 --> 01:30.660
Given all these considerations that it'd be anywhere

01:30.660 --> 01:34.410
from 1,000 to 2,000 window washers, potentially even more.

01:34.410 --> 01:37.530
So the really useful thing here is one,

01:37.530 --> 01:40.705
it actually gets to a better answer between 1

01:40.705 --> 01:43.050
and 2,000, which I think is closer to the real number.

01:43.050 --> 01:45.240
But you can also see the steps in logic.

01:45.240 --> 01:47.370
So if it's faulty in some of its logic, you can go back

01:47.370 --> 01:50.250
and correct it and you can say,

01:50.250 --> 01:54.450
actually maybe you know that like New York has 50%

01:54.450 --> 01:56.910
of all the windows in the US or whatever.

01:56.910 --> 01:58.650
You could actually correct that assumption

01:58.650 --> 02:01.410
and then you can arrive at a better estimate.

02:01.410 --> 02:05.610
So it not only helps the AI give you a better estimate,

02:05.610 --> 02:06.930
but it also shows you the reasoning

02:06.930 --> 02:10.020
that it can prompt your own thinking on the topic

02:10.020 --> 02:12.390
and understand how it approaches things.

02:12.390 --> 02:14.670
Thinking step by step, it's just a magical phrase

02:14.670 --> 02:17.643
that helps a lot in terms of the results.