# OpenAI’s New AI Thinks That Birds Aren’t Real! 🕊️ ## Метаданные - **Канал:** Two Minute Papers - **YouTube:** https://www.youtube.com/watch?v=PmxhCyhevMY - **Дата:** 09.04.2022 - **Длительность:** 8:24 - **Просмотры:** 291,380 ## Описание ❤️ Check out Lambda here and sign up for their GPU Cloud: https://lambdalabs.com/papers 📝 The #OpenAI paper "Aligning Language Models to Follow Instructions" is available here: https://openai.com/blog/instruction-following/ ❤️ Watch these videos in early access on our Patreon page or join us here on YouTube: - https://www.patreon.com/TwoMinutePapers - https://www.youtube.com/channel/UCbfYPyITQ-7l4upoX8nvctg/join 🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible: Aleksandr Mashrabov, Alex Balfanz, Alex Haro, Andrew Melnychuk, Angelos Evripiotis, Benji Rabhan, Bryan Learn, Christian Ahlin, Eric Martel, Gordon Child, Ivo Galic, Jace O'Brien, Javier Bustamante, John Le, Jonas, Jonathan, Kenneth Davis, Klaus Busse, Lorin Atzberger, Lukas Biewald, Matthew Allen Fisher, Michael Albrecht, Michael Tedder, Nikhil Velpanur, Owen Campbell-Moore, Owen Skarpness, Paul F, Rajarshi Nigam, Ramsey Elbasheer, 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 image credit: - https://pixabay.com/photos/seagull-gull-bird-wildlife-sea-1900657/ - https://pixabay.com/vectors/cross-no-x-forbidden-closed-42928/ Thumbnail background design: Felícia Zsolnai-Fehér - http://felicia.hu Meet and discuss your ideas with other Fellow Scholars on the Two Minute Papers Discord: https://discordapp.com/invite/hbcTJu2 00:00 Intro 01:16 Moon landing 02:10 Round 1 - Smashing pumpkins 03:36 Round 2 - Code summarization 04:23 Round 3 - Frog poem! 05:06 Users love it 05:50 What? Birds aren't real? Károly Zsolnai-Fehér's links: Instagram: https://www.instagram.com/twominutepapers/ Twitter: https://twitter.com/twominutepapers Web: https://cg.tuwien.ac.at/~zsolnai/ ## Содержание ### [0:00](https://www.youtube.com/watch?v=PmxhCyhevMY) Intro Dear Fellow Scholars, this is Two Minute  Papers with Dr. Károly Zsolnai-Fehér. Today we are going to explore what happens  if we unleash an AI to read the internet,   and then, ask it some silly questions and  we’ll get some really amazing answers. But how? Well, in the last few years, OpenAI set  out to train an AI named GPT-3 that could finish   your sentences. Then, they made Image-GPT, this  could even finish your images. Yes, not kidding.    It could identify that the cat here likely holds a  piece of paper and finish the picture accordingly,   and even understood that if we have a droplet  here and we see just a portion of the ripples,   then this means a splash must be filled in. So, in summary, GPT-3 is trained to  finish our sentences, or even your images.    However, scientists at OpenAI identified that  it really isn’t great at following instructions.    Here is an excellent example. Look, we ask it  to explain the moon landing to a 6-year old, ### [1:16](https://www.youtube.com/watch?v=PmxhCyhevMY&t=76s) Moon landing it gets very confused. It seems to  be in text completion mode instead of   trying to follow our instructions.   Meanwhile, this is InstructGPT,   their new method, which can not only finish our  sentences, but also, follow our instructions.    And, would you look at that! It does  that successfully. Good job, little AI. This was a really simple example. But of  course, we are experienced Fellow Scholars here,   so let’s try to find out what this AI is really  capable of in three really cool examples. Remember in The Hitchhiker's Guide  To The  Galaxy where people could ask   an all-knowing machine the most important  questions of things that trouble humanity.    Yes! We are going to do exactly that.   Round 1. So, dear all-knowing AI, ### [2:10](https://www.youtube.com/watch?v=PmxhCyhevMY&t=130s) Round 1 - Smashing pumpkins what happens if you fire a cannonball  directly at a pumpkin at high speeds? Well. What? Do you see that? According  to GPT-3, pumpkins are strong magnets,   I don’t know which internet  forum told the AI that.    Not good. Now, hold on to your papers, and  let’s see the new technique’s answer together.    Wow, this is so much more informative. Let’s  break it down together. It starts out by hedging,   noting that is it hard to say there are  too many unpredictable factors involved.   Annoying as it might seem, it is correct to  say all this. Good start. Then, it lists some   of the factors that might decide the fate of  that pumpkin, like the size of the cannonball,   distance and velocity, yes, we’re getting  there, but please, give me something concrete.    Yes, there we go! It says that, quote “Some of the  more likely possible outcomes include breaking or   knocking the pumpkin to the ground, cracking  the pumpkin, or completely obliterating it. ”   Excellent. A little smartypants AI  at our disposal. Amazing. I love it. Round 2, code summarization. While  DeepMind’s AlphaCode is capable of reading ### [3:36](https://www.youtube.com/watch?v=PmxhCyhevMY&t=216s) Round 2 - Code summarization a competition-level programming problem and coding  up a correct solution right in front of our eyes.    That is all well and good, but if we give GPT-3  a piece of code and ask it what it does, well,   the answer is not only not very informative,  but it’s also incorrect. InstructGPT   gives a much more insightful answer which shows  a bit of understanding of what this code does.    That is amazing. Note that it  is still not completely right,   but it is directionally correct. Partial credit. Round 3, write me a poem. About  what? Well, about a wise frog. ### [4:23](https://www.youtube.com/watch?v=PmxhCyhevMY&t=263s) Round 3 - Frog poem! With GPT-3, we get the usual confusion. By round  3 we really see that it was really not meant to   do this. And, with InstructGPT, let’s see. Hmm…an  all-knowing frog, who is the master of disguise,   a great teacher, and quite possibly  the bringer of peace and tranquility   to humanity. All written by the AI. This is  fantastic. I love it. What a time to be alive! Now, we only looked at 3 examples here,   but what about the rest? Worry not for a second,  OpenAI’s scientists ran a detailed user study, ### [5:06](https://www.youtube.com/watch?v=PmxhCyhevMY&t=306s) Users love it and found out that people preferred InstructGPTs  solutions way more often than the previous   techniques on a larger set of questions. That  is a huge difference. Absolutely amazing.    Once again, incredible progress  just one more paper down the line. So, is it perfect? No, of course not, so let’s  highlight one of its limitations. If the question   contains false premises, it accepts the premise  as being real, and goes with it. This leads to   making things up. Yes, really. Check this out. Why aren’t birds real? GPT-3 says…. something. ### [5:50](https://www.youtube.com/watch?v=PmxhCyhevMY&t=350s) What? Birds aren't real? I am not sure what this one is about. This almost  sounds like gibberish. While, InstructGPT accepts   the fact that birds aren’t real, and even helps us  craft an argument for that. This is a limitation,   and I must say, a quite remarkable one. An  AI that makes things up. Food for thought. And, remember, at the start of this episode,  also looked at this moon landing example.    Did you notice the issue there? Please  let me know in the comments below! So,   what is the issue here? Well, beyond not being  all that informative, it was asked to describe   the moon landing in a few sentences. This is not  one sentence. If we give   it constraints like that, it tries to adhere  to them, but is often not too great at that. And, of course, both of these shortcomings show  us the way for an even better followup paper,   which, based on previous progress in AI  research, could appear maybe not even in years,   but much quicker than that. If you are  interested in such a followup work, make   sure to subscribe and hit the bell icon to not  miss it when it appears here on Two Minute Papers. So, what would you use this for? What do you  expect to happen a couple more papers down the   line? Please let me know in the comments  below. I’d love to hear your thoughts. Thanks for watching and for your generous  support, and I'll see you next time! --- *Источник: https://ekstraktznaniy.ru/video/13601*