# DeepMind’s Take on How To Create a Benign AI

## Метаданные

- **Канал:** Two Minute Papers
- **YouTube:** https://www.youtube.com/watch?v=pc_k-sgUYmY
- **Дата:** 02.01.2019
- **Длительность:** 4:15
- **Просмотры:** 54,901

## Описание

The paper "Scalable agent alignment via reward modeling: a research direction" is available here:
1. https://arxiv.org/abs/1811.07871
2. https://medium.com/@deepmindsafetyresearch/scalable-agent-alignment-via-reward-modeling-bf4ab06dfd84

Pick up cool perks on our Patreon page:
› https://www.patreon.com/TwoMinutePapers

We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
313V, Alex Haro, Andrew Melnychuk, Angelos Evripiotis, Anthony Vdovitchenko, Brian Gilman, Christian Ahlin, Christoph Jadanowski, Dennis Abts, Emmanuel, Eric Haddad, Eric Martel, Evan Breznyik, Geronimo Moralez, Jason Rollins, Javier Bustamante, John De Witt, Kaiesh Vohra, Kjartan Olason, Lorin Atzberger, Marcin Dukaczewski, Marcin Dukaczewski, Marten Rauschenberg, Maurits van Mastrigt, Michael Albrecht, Michael Jensen, Morten Punnerud Engelstad, Nader Shakerin, Owen Skarpness, Raul Araújo da Silva, Richard Reis, Rob Rowe, Robin Graham, Ryan Monsurate, Shawn Azman, Steef, Steve Messina, Sunil Kim, Thomas Krcmar, Torsten Reil, Zach Boldyga, Zach Doty.
https://www.patreon.com/TwoMinutePapers

Crypto and PayPal links are available below. Thank you very much for your generous support!
› PayPal: https://www.paypal.me/TwoMinutePapers
› Bitcoin: 1a5ttKiVQiDcr9j8JT2DoHGzLG7XTJccX
› Ethereum: 0xbBD767C0e14be1886c6610bf3F592A91D866d380
› LTC: LM8AUh5bGcNgzq6HaV1jeaJrFvmKxxgiXg

Thumbnail background image credit: https://pixabay.com/photo-3706562/
Splash screen/thumbnail design: Felícia Fehér - http://felicia.hu

Károly Zsolnai-Fehér's links:
Facebook: https://www.facebook.com/TwoMinutePapers/
Twitter: https://twitter.com/karoly_zsolnai
Web: https://cg.tuwien.ac.at/~zsolnai/

## Содержание

### [0:00](https://www.youtube.com/watch?v=pc_k-sgUYmY) Segment 1 (00:00 - 04:00)

de fellows caelers dit is toen meneer papers wit cardoso live share this app is al death note heb die u sowieso fireworks but a really wanted to cover this paper picasa tassen story dds itenc zij important for all of a story about when creating a new life of a password het esc hoe heftig somehow god is een jij wat we consider to be desirable solution is en richting dons wel dit we find out the best way to accomplish het dit is easy one kleding simpele video games week als we can just radio 3fm to maximize de score zien in de game for instance de mowgli geert in atari breakout de clausule get to finishing the level however in real life de hand hebt anyone the wing kassen score totale foutloos wie hardware object of what you were some times me heftig make decisions that seem bedden time but we service wel in de future try to save money for studying for a few years younger artikel live decisions that payoff in de langue wrong but may seem en design robot avatar the opposite is of social media is dat may sound right at the time mee immédiate lee backfire mijn in a car chase dan tess de korea to unload allah necessary words to go faster more if you to prepare to be from play is actief van de kar zo gauw kenway possibly kreet en eaj that somehow anderstein sar intentions and acts inline lead them that challenge en question and this often referred to as the agents alignment probleem dit has to be aligned we daar varios what can we do but as well soort of herring en my reading the wise week en medicontrol te begeven of the high flow iets rewards systeem scientists and deep mind has published a paper on this topic while i started their proces van toe assumptions and samsung number one korting the authors for many guests to want to surf the violation of outcomes is easier than to using de correct behavior in short it is easier to reality tv them to be calm and every size reasonable price nota van complexe d-serie memo dat is tas met olie school waar dit is in die troep voor een large number of de fakkel problems and samsung number to user intentions can be learned met high accuracy in other words the van in fda dat sam van riel it store intentions de eaj should be able to learn dat leaning on this to some sense week and change de basics formule shaun of reinforcement learning in the following great normale we hebben agents de choosers zelf actions in een bar meant to maximize score for instance this came in moving de peddel around to get any blocks as possible and finish the labor day standard is familie show nu weet dat de user can periodically provide feedback en houden score should be calculated na de in your website to maximize dus 0 score en twee hoop dat this will be more inline die intentions for in a car chase e-semble weekend moni fire uw woord to make sure would remain in the car and not getting het wraps de most remarkable property of the formulation is dat dit 1000e van require a stool for instance slee de dmr all to demonstrate argent entrance to the young gifted de formules een faal als er principles and not arexons we can just eat in a favourite armchair ben die een yard war l by changing your function every now and then and let the guide to the ruling work this a slide jolink apathie 4x hebt dat dit xlii works loving the idea of via de loketten peper usa time or the tiles and how to do this efficiently and een casestudy we de viewer tarly games als al sinds dit huis en love implications pertaining to life safety and houten creëert online de agents and increasingly important op idee is use respect voor deep mind voor in westing more and more of their time and money in the air lexar walking and for your generous support and i'll see you next time

---
*Источник: https://ekstraktznaniy.ru/video/14376*