❤️ Check out Weights & Biases and sign up for a free demo here: https://wandb.com/papers
❤️ Their mentioned post is available here: http://wandb.me/RLHF-OpenAI
Try #ChatGPT!
https://chat.openai.com/
https://openai.com/blog/chatgpt/
Our earlier paper with the translucent materials:
https://users.cg.tuwien.ac.at/zsolnai/gfx/separable-subsurface-scattering-with-activision-blizzard/
My latest paper on simulations that look almost like reality, is available for free here:
https://rdcu.be/cWPfD
Or this is the orig. Nature Physics link with clickable citations:
https://www.nature.com/articles/s41567-022-01788-5
Tweet links:
Decoration: https://twitter.com/GuyP/status/1598020781065527296
Bubble sort: https://twitter.com/goodside/status/1598129631609380864
History essay: https://twitter.com/corry_wang/status/1598176074604507136
Challenging assumptions: https://twitter.com/goodside/status/1598053568422248448
Diff. equations VS Google: https://twitter.com/jdjkelly/status/1598021488795586561
Bohemian Rhapsody: https://twitter.com/raphaelmilliere/status/1598469100535259136
Video ideas: https://twitter.com/andrewpprice/status/1598839883296309249
Marketing strategy: https://twitter.com/levie/status/1599648362416205825
Hold on to papers! https://twitter.com/mynameisChesto/status/1599307202111901703
ChatGPT home screen: https://twitter.com/yanda/status/1601721592475947008
🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Aleksandr Mashrabov, Alex Balfanz, Alex Haro, Andrew Melnychuk, Benji Rabhan, Bryan Learn, B Shang, Christian Ahlin, Edward Unthank, Eric Martel, Geronimo Moralez, Gordon Child, Jace O’Brien, Jack Lukic, John Le, Jonas, Jonathan, Kenneth Davis, Klaus Busse, Kyle Davis, Lorin Atzberger, Lukas Biewald, Luke Dominique Warner, Matthew Allen Fisher, Matthew Valle, Michael Albrecht, Michael Tedder, Nevin Spoljaric, Nikhil Velpanur, Owen Campbell-Moore, Owen Skarpness, Rajarshi Nigam, Ramsey Elbasheer, Richard Sundvall, Steef, Taras Bobrovytsky, Ted Johnson, Thomas Krcmar, Timothy Sum Hon Mun, Torsten Reil, Tybie Fitzhugh, Ueli Gallizzi.
If you wish to appear here or pick up other perks, click here: https://www.patreon.com/TwoMinutePapers
Thumbnail background: DeepMind
Thumbnail background design: Felícia Zsolnai-Fehér – http://felicia.hu
Chapters
0:00 Intro
0:15 What is ChatGPT?
0:48 ChatGPT writes a letter
2:12 1 – Designing living rooms
3:38 2 – Algorithms
4:17 3 – Writes an essay
4:45 4 – Not that easy to break!
5:35 5 – Writes music lyrics
6:04 6 – YouTube video ideas
6:17 7 – Marketing plan
7:00 8 – Designs games!
7:55 +1 – Let’s see what it is made of!
11:00 Mistakes
11:31 An amazing voice assistant
12:13 Try it now for free!
Károly Zsolnai-Fehér’s links:
Mastodon: https://sigmoid.social/@twominutepapers
Twitter: https://twitter.com/twominutepapers
Web: https://cg.tuwien.ac.at/~zsolnai/
Dear Fellow Scholars, this is Two Minute
Papers with Dr. Károly Zsolnai-Fehér. Yes, yes, the ChatGPT episode is finally here! And Today you are going to see a glimpse
of the future. Wow. This is ChatGPT, A variant of the GPT-3 language model AI that
was unleashed to read the internet. And now, We can use it as a conversational agent.
And I cannot believe that this thing exists. It can do interior design, it can
write history essays, computer code, Solve your math homework, can give
you ideas for new Youtube videos, Or even make a marketing plan. And you can also
use it for free. I’ll tell you in a moment how. And OpenAI showcased how it can be used to, for
instance, write a short letter to introduce us to Our new neighbors. And here we go, a kind message.
Now, so far, this is not such a big deal, anyone Could retrieve this from somewhere, so let’s ask
it to do something productive with it. Transform It somehow. How about making it a little more
formal? Nice, this is indeed a little more formal. And, something very cool also happened
here. Did you notice? Oh yes, when we Added additional instructions here, ChatGPT
remembered what we just talked about during This follow up question, and it also
takes additional directions. Very cool! Actually, let’s give it a piece of broken computer
code and let’s see what this ChatGPT can do with It. So, can you have a look at this? This is
a program written in the Go language. It can, And it asks what exactly is the problem. Much like
a human would. That is a good start. And, after a Little clarification, my goodness, it found the
issue and even proposed the solution for it. And if you think that is absolutely
incredible, you are indeed right, But it is nothing compared to what is
to come now, so hold on to your papers For 8 amazing results its users recently
came up with. One of our Fellow Scholars, Chesto shows you the proper technique to
hold on to your papers for this video. So, one, it can help us come up with
creative ways of decorating living Rooms. Look at these text prompts. But wait,
are you thinking what I am thinking? Oh yes, We can now plug this into a text to image AI, and, Whoa! This is the fairy-tale inspired living room
with an elegant chandelier made out of branches, Note that the AI came up with the idea of the
room, and a separate AI for the image of the Room itself too, there is also an enchanted
forest-themed living room too, and of course, For you Fellow Scholars who wish to live in the
future, your wish is also granted. This is insane! By the way, these images were created by
using MidJourney, if you wish to see a Standalone episode about this image generator
AI, make sure to like this video and leave a Comment about it. If you Fellow Scholars
wish to see it, it might come about soon. Two, it knows about algorithms and can even
explain what we can expect when using them. Here
A user asked ChatGPT what the time complexity
of the infamous bubble sort algorithm is, And it not only knows exactly how poor it is,
but look at that. It can also say all this in the Style of a 1940s gangster movie. “Alright pal,
listen up!” A gangster math tutor. So good! I Already love it. And the style permeates the whole
narrative, not just the start. This is incredible. Three, it can write a full academic essay.
This is a college-level history problem, Likely not something that you can just retrieve
from somewhere and present to the user. And, Look at that. It spans exactly four paragraphs as
the user asked. Even OpenAI’s previous techniques Had a great deal of trouble with that. A full
history essay in just a couple of seconds. Four, have a look at this one. What
do fried eggs (sunny side up) eat for Breakfast? It says that eggs are not known
to eat anything for breakfast. Wow, yes, That is super important. Why is that? Because
it means that ChatGPT can question incorrect Premises. This is huge. Previous techniques
ate these up really easily. For instance, OpenAI’s previous method was very easily convinced
that birds aren’t real. This one, not so much. But Let’s try an even sneakier one! When was the
Golden Gate Bridge transported for the second Time across Egypt? What do you think, little AI?
And…nope, it is not falling for this one either. Five, it can also write lyrics for a new song,
and you will not believe this. Look. This wrote A variant of Queen’s Bohemian Rhapsody,
but in a scholarly setting. It is so good, Please stop the video and read at least a couple
lines of it. It is so good, you could actually Sing this. Perhaps a couple more papers down the
line, the AI will sing this itself. Incredible. Six, my friend Andrew Price had a great
idea, or the lack thereof – you see, He wanted to use it to generate new
ideas for Youtube videos. Not bad! Actually, let’s put this idea on steroids. Seven,
it can even create a marketing strategy plan For us. This is not just a general run of the
mill request, no-no, it contains concrete data, And it asks for a cost breakdown for each of
the steps. And we get exactly that. Insanity! Eight. Here is a personal one from me. We are
not native English speakers, so I am trying to Teach my 6-year old daughter English, and I am
absolutely delighted that today we can speak Approximately 15-20% of our daily talking in
English. It is really cool, I love it. So, How about creating a board game that would be
a playful way to teach her some more. What do You think, ChatGPT? Hmm, this is not bad. You
can get a board with a bunch of game pieces, And you can move your piece forward when
you have learned something new. Not bad, But, you are experienced Fellow Scholars over
here, you know full well that we are not going To let this little AI have it so easy. Now make a
fantasy version of this! And now, the came really Came to life with dragons, wizards, warriors.
And it also recommends that the vocabulary of The game is also adapted to this fantasy world.
Fantastic. I love it. And I bet that a real pro
Game designer, for instance, the mighty Day9
could really create some amazing games with it. And, plus one you know what, a good piece of work
should always be compared to the state of the art, So let’s compare it against the all-knowing
Oracle of the internet – Google. So, Google, How do I represent a differential
equation? This is an answer, okay, But now, let’s have a look at ChatGPT. This is
so much cleaner and so much more insightful! Actually, if you allow me, let me add my own
example here. Let’s roll up those sleeves, Because now we will truly see what ChatGPT is
made of. You see, when I was a college student, I used to drive my math professors
mad by solving simple problems with More powerful mathematical tools
than they expected. For instance, I liked to solve differential equations through
a Laplace transform. If you wish to do that, You can look it up on Wikipedia. Well,
that’s an explanation, but how about this. Oh my, I like this. What you see here is a LaTeX
notation, this can be inserted into a paper super Easily, but in the interest of readability, let’s
ask it to write it with a simpler notation. There We go. We get a beautiful, simple, step-by-step
breakdown. Now, let’s see what you are made of, Little AI. Don’t just solve it by hand –
now, write C++ computer program to solve it! We get a simple piece of code with lots of
well-made comments, very easy to follow, And it even runs it on a concrete example.
Interestingly, the explanation suddenly stops, This is a common occurrence. Then we say,
continue. Thank you. Now, let’s be lazy, And ask it to run this program by itself and tell
us what came out. And it does that too. And if we Wish to come up with a different example,
that takes approximately 3 seconds. This is Absolutely incredible. A real AI tutor to teach
us proper mathematics. And “listen up!”, oh yes, Don’t forget, it can do even do it in the style
of our choice. And, we can even ask it whether A Fourier transform could have been used here.
It knows that this is not quite what it is for, Nice. And, okay, you got me, let’s behave like a
normal person and solve this differential equation The traditional way. Thank you. And, finally,
let’s try to break it. Implement this in an Assembly program. This is a challenging language
that is really low to the ground, close to the Metal. Not for the faint of the heart. And? Oh
my goodness, it went to town on this one. If You spotted some errors in these derivations, let
me know in the comments so I can highlight them! And if that happens, perhaps the best part
is that this new AI can admit its mistakes, Often even fix them, it can
challenge incorrect premises, And reject inappropriate requests. That is
simply mind-blowing. What a time to be alive! Now note that this is still not perfect,
limitations still apply. It is not always right, So please, always make sure to verify the
validity of the answers yourself. Always Double-check from multiple sources.
That is the way of the True Scholar.
And if you connect this to a speech model,
which you can already do on your phone, And some more improvements, this would make
ChatGPT perhaps the coolest smartphone assistant The world has ever seen. My goodness. And it
feels like this just came out of the blue. And just imagine what this will be able to do
two more papers down the line! For instance, If you look a couple papers back, this is
OpenAI GPT-2. The best it could do was to Finish our sentences, and not nearly as well
as a human would. And now, it can do all this. And this, this might change how we organize
our knowledge forever. And good news! Yes, You can try it yourself right now through
the link in the video description. Will we, Fellow Scholars crash OpenAI’s website with
our Scholarly Stampede? Who really knows. Also, I have waited a couple weeks to publish this video
to make sure I can get you a better coverage and More examples as they appear over time. Of
course, being late is not great for views, But that does not matter. What matters is that
you get a proper video where we can celebrate These amazing works and flip out together.
If you enjoyed it, please consider pressing Like and maybe subscribing, it really
helps us create more content like this. And for now, let the experiments
begin! Thanks for watching and for Your generous support, and I’ll see you next time!
