Transcript video
File created by Sun Dec 03 2023 10:39:59 GMT+0100 (Central European Standard Time)
Video URL: https://www.youtube.com/watch?v=7VSWyghVZIg
Title: (2) OpenAI GPT-4 - The Future Is Here! - YouTube
Total: 111


#####

Time: 00:00

Dear Fellow Scholars, this is Two Minute Papers
with Dr. Károly Zsolnai-Fehér.

#####


#####

Time: 00:05

Finally, OpenAI’s legendary language model
AI, GPT-4 which also now powers their ChatGPT

#####


#####

Time: 00:13

assistant, is here.

#####


#####

Time: 00:15

And my goodness, they promise a lot of goodies.

#####


#####

Time: 00:19

Advanced reasoning.

#####


#####

Time: 00:20

Complex instructions.

#####


#####

Time: 00:22

More creativity.

#####


#####

Time: 00:23

This is going to be an incredible journey,
so hold on to your papers right now, Fellow

#####


#####

Time: 00:29

Scholars.

#####


#####

Time: 00:30

And, get this, it can also accept image inputs.

#####


#####

Time: 00:34

Let’s put this one to work immediately.

#####


#####

Time: 00:38

Dear GPT-4, imagine that we have these ingredients.

#####


#####

Time: 00:43

What can we make from them?

#####


#####

Time: 00:44

Note that the input is not a text anymore,
but a photo.

#####


#####

Time: 00:48

That is excellent, and don’t forget, if
one of these caught your eye, you can also

#####


#####

Time: 00:55

ask follow up questions here on how to exactly
make them.

#####


#####

Time: 00:59

Now, earlier we looked at some preliminary
results of an IQ-test, where it scored 147.

#####


#####

Time: 01:08

However, I noted immediately that we should
be good Fellow Scholars and take this with

#####


#####

Time: 01:15

a little bit of skepticism, and wait for a
more formal examination.

#####


#####

Time: 01:20

And, yes - here it is!

#####


#####

Time: 01:23

It finally happened.

#####


#####

Time: 01:25

GPT-4 give us human-level performance on a
variety of academic benchmarks.

#####


#####

Time: 01:31

And they report something absolutely incredible.

#####


#####

Time: 01:35

Look.

#####


#####

Time: 01:36

This is what its previous incarnation in ChatGPT
was capable of doing.

#####


#####

Time: 01:42

10th percentile on the bar exam.

#####


#####

Time: 01:45

And now, maybe 30?

#####


#####

Time: 01:47

50?

#####


#####

Time: 01:48

Nope, 90th percentile.

#####


#####

Time: 01:50

This means that the new one is within the
top 10 percent of test takers.

#####


#####

Time: 01:56

That kind of improvement in just one paper
is outstanding.

#####


#####

Time: 02:00

And, I know you are looking, and the answer
is of course, we are going to talk about this

#####


#####

Time: 02:06

too in a moment.

#####


#####

Time: 02:08

And this is where being able to process visual
information means an incredible breakthrough.

#####


#####

Time: 02:14

It doesn’t just mean that it can help us
with the ingredients, or can explain memes.

#####


#####

Time: 02:19

Although that is quite helpful.

#####


#####

Time: 02:21

But, it gets better.

#####


#####

Time: 02:23

For instance, look at this one.

#####


#####

Time: 02:26

Yummy.

#####


#####

Time: 02:27

On the AMC mathematical exam, GPT-4 marked
with light green performs so much better than

#####


#####

Time: 02:35

the previous version.

#####


#####

Time: 02:37

However, here is where the magic happens.

#####


#####

Time: 02:40

Look, this exam contains some light visual
information, and when we add the capability

#####


#####

Time: 02:46

to process this to GPT-4, it does so much
better.

#####


#####

Time: 02:50

It is rapidly improving to match human-level
results.

#####


#####

Time: 02:55

We are one step closer to human-level intelligence.

#####


#####

Time: 03:00

This work is history in the making.

#####


#####

Time: 03:02

And, I was also very excited about this.

#####


#####

Time: 03:06

The USA Biology Olympiad exam.

#####


#####

Time: 03:09

This needs next-level capabilities, because
it requires visually inspecting and evaluating

#####


#####

Time: 03:16

electrocardiograms, and in general, taking
in a ton of visual information and giving

#####


#####

Time: 03:22

correct answers.

#####


#####

Time: 03:23

So how correct was it?

#####


#####

Time: 03:27

What?

#####


#####

Time: 03:28

Are you kidding?

#####


#####

Time: 03:29

99th percentile.

#####


#####

Time: 03:31

Better than almost all humans.

#####


#####

Time: 03:34

Wow.

#####


#####

Time: 03:35

However, wait a minute.

#####


#####

Time: 03:36

I see light green, not dark green.

#####


#####

Time: 03:39

What does that mean?

#####


#####

Time: 03:40

I mean that is the non-vision version of GPT-4.

#####


#####

Time: 03:45

But this exam has plenty of images.

#####


#####

Time: 03:48

So, what happened?

#####


#####

Time: 03:50

Well, the AI got a little help.

#####


#####

Time: 03:53

The images have been transcribed by a human.

#####


#####

Time: 03:57

I would love to see those transcriptions,
I bet it was not an easy job.

#####


#####

Time: 04:02

So, it seemingly cannot perform this level
of visual inspection yet, but I am certain

#####


#####

Time: 04:09

that we will see something like this one or
two more papers down the line.

#####


#####

Time: 04:14

Perhaps with GPT-5.

#####


#####

Time: 04:17

If that sounds interesting, consider subscribing
and hitting the bell icon to not miss it when

#####


#####

Time: 04:22

the next paper drops.

#####


#####

Time: 04:24

Here is also an example that I absolutely
loved.

#####


#####

Time: 04:28

Little AI, explain the plot of Cinderella
and each word has to begin with the next letter

#####


#####

Time: 04:35

in the alphabet from A to Z.

#####


#####

Time: 04:37

Wow.

#####


#####

Time: 04:39

This one is tough.

#####


#####

Time: 04:40

I have to admit, I am not sure if most people
would be able to solve this in a satisfactory

#####


#####

Time: 04:47

manner.

#####


#####

Time: 04:48

And…look at that.

#####


#####

Time: 04:51

Flying colors.

#####


#####

Time: 04:52

My goodness.

#####


#####

Time: 04:54

And this is one more really cool example from
their official demo.

#####


#####

Time: 04:59

What?

#####


#####

Time: 05:00

Are you saying that this crude drawing for
a website comes in, with some jokes and buttons.

#####


#####

Time: 05:07

And it not only creates this website with
proper javascript, but it also understands

#####


#####

Time: 05:13

that it needs to fill in the jokes, and it
does that too.

#####


#####

Time: 05:17

Loving it.

#####


#####

Time: 05:18

So, when can we try it?

#####


#####

Time: 05:20

Well, OpenAI deployed GPT-4 to their ChatGPT
Plus subscribers, and they also made a waitlist

#####


#####

Time: 05:28

for API access.

#####


#####

Time: 05:29

Now, I was lucky enough to have access to
it, thank you so much, and it already helped

#####


#####

Time: 05:35

me remember my discussion with a geologist
professor from a few years ago.

#####


#####

Time: 05:41

There were some details I did not remember,
and now I do.

#####


#####

Time: 05:45

Thank you!

#####


#####

Time: 05:46

So, who is using it?

#####


#####

Time: 05:48

The answer is that everyone is using it.

#####


#####

Time: 05:51

It can help us organize a huge knowledge base,
or it can help us learn a new language, and

#####


#####

Time: 05:57

so much more.

#####


#####

Time: 05:59

And this is tech transfer, so in other words,
from paper to product, not in decades, but

#####


#####

Time: 06:06

in a matter of days, and in some cases, even
a matter of hours.

#####


#####

Time: 06:12

What a time to be alive!

#####


#####

Time: 06:14

And if you have been watching Two Minute Papers,
you see the whole history of these works,

#####


#####

Time: 06:19

from GPT-2, to GPT-3 and now, GPT-4.

#####


#####

Time: 06:24

You saw history unfold right before your eyes.

#####


#####

Time: 06:28

Thank you so much for being on this incredible
journey with me.

#####


#####

Time: 06:32

Now, I will also note that the paper is really
detailed in terms of the evaluation of the

#####


#####

Time: 06:37

results, but a little light on the inner workings
of the algorithm itself.

#####


#####

Time: 06:42

I hope we get to know more soon.

#####


#####

Time: 06:49

Thanks for watching and 
for your generous support, and I'll see 

#####


#####

Time: 07:59

you next time!

#####
