OpenAI

from Wikipedia, the free encyclopedia
OpenAI LP

logo
legal form Limited partnership
founding December 11, 2015
Seat San Francisco , California, USA
management Sam Altman , and others
Website www.openai.com

OpenAI LP is a company which, controlled by the non-profit organization OpenAI Inc, deals with the research of artificial intelligence ( AI ). The organization's central donors are the investor and entrepreneur Elon Musk and Microsoft . The goal of OpenAI is to develop and commercialize artificial intelligence on an open source basis in a way that it is beneficial to society, not harmful. The organization enables "free collaboration" with other institutions and researchers by making their patents and research results available to the public. The company is funded with over $ 1 billion in donations, with only a small portion of these funds expected to be consumed in the early years. OpenAI deals with the question of the "existential threat from artificial intelligence" - that is, the possible surpassing and replacement of human by artificial intelligence. The organization so far only serves to research artificial intelligence and will likely be linked to Neuralink later . In February 2018 it was announced that Elon Musk is leaving the management of Open AI in order to avoid conflicts of interest with Tesla's developments in the field of artificial intelligence. He remains a donor and advisor to the organization.

Motifs

Scientists Stephen Hawking and Stuart Jonathan Russell expressed fears that if artificial intelligence could one day gain the ability to improve itself, it could lead to an "intelligence explosion." Due to the associated superiority of AI, a displacement of the human species by super-intelligent artificial intelligence , which is not subject to any human values, would be conceivable.

Even Elon Musk, co-founder of OpenAI, describes artificial intelligence as the greatest existential threat to humanity. To counter this risk in the development of artificial intelligence, OpenAI was designed as an open-source non-profit organization. This should ensure that the organization is independent of obligations towards donors and their interests. This should give research the freedom to have a positive long-term impact on society.

The company OpenAI itself suggests that it is difficult to estimate how much society could benefit from high-quality artificial intelligence. In return, it is just as complex to determine what damage an improper use of this technology could bring. From today's perspective, it is also not foreseeable when artificial intelligence will be on the level to compete with human intelligence. In any case, the board of OpenAI is convinced that artificial intelligence is an extension of the individual human will. This technology should therefore also be available nationwide and be accessible to everyone to the same extent. A surpassing of human intelligence through the OpenAI project, which was conceived over several decades, is at least conceivable for the company's co-chairman, Sam Altman .

For Infosys CEO Vishal Sikka , disclosing the development of OpenAI was instrumental in helping him. It was central that the aim of OpenAI lies in the fact that the research always had "the greater good" - i.e. a holistic positive intention for humanity - in focus. Wired's Cade Metz said that companies such as Amazon are interested in using open source software and open source data to compete against companies such as Google or Facebook, which have a strong proprietary database. Y Combinator companies also share their data with OpenAI, Altman said in an interview.

strategy

Elon Musk asks what should be done so that the future would be good. According to Musk, there are opportunities to just watch or to participate in the development. The design must be done in a responsible and safe manner in order to have a positive effect on society. Musk does not deny that there is a risk associated with the development and advancement of friendly artificial intelligence. He believes it is important to make artificial intelligence available to as many people as possible so that a small part cannot claim the advantages of AI for themselves. Musk was convinced in an interview that this is the best defense against the improper use of AI systems.

This strategy by Musk and Altman to reduce the risk of damage caused by artificial intelligence through open source AI for everyone is controversial in the community. The philosopher Nick Bostrom is one of the skeptics of this approach and makes it clear with the following metaphor: "If you have a button that can cause harm in the world, you wouldn't make it available to everyone."

Products

OpenAI Gym

Starting in 2016, OpenAI has developed its platform " OpenAI Gym ", which deals with reinforcement learning . The current version of the source code is available on GitHub. The aim is to offer a basic system that is easy to set up and supports a wide range of different development environments. OpenAI Gym tries to offer a standardization for the publication of results in the research of artificial intelligence in order to be able to compare and reproduce publications more easily.

OpenAI Five

With OpenAI Five, a system of cooperating algorithms was developed. AI systems usually act independently of one another. For the first time, AI systems are working together in a team towards a common goal. In the complex strategic computer game Dota 2 , teams of several players compete against each other. In a competition, these cooperating AI systems were superior to the human players, but it should be noted that a greatly simplified version of the game was used in this competition, in which only 17 of the 115 heroes can be played and some objects and illusion runes are not were available. Later, players from all over the world were allowed to play online against the AI ​​for 3 days as part of the simplified version, whereby OpenAI had a win rate of 99.4% in 7257 games, but was also defeated several times in a row by many teams - by a team even 10 times in a row. A competition with people with no restrictions in the game is not planned.

GPT-2

Another project is GPT-2 (Generative Pretrained Transformer 2), which is an artificial intelligence that can independently complete English-language texts and whose texts can sometimes not be distinguished from texts written by humans. According to the researchers themselves, they thought their own software worked so well that they themselves became scared of it.

It differs from some other automatic text generators in that the software does not use predefined text blocks for text generation. It is also not restricted to a subject area, but can generate some suitable sentences for any text. The language model is based on an artificial neural network with 1.5 billion parameters . It is a statistical language model that describes how likely a certain sentence is in a context. The data basis for GPT-2 is approximately 45 million pages of text processed. The technology behind GPT-2 is similar to the technology of older software, such as was developed by Google , among others . Only the ways in which OpenAI assembles the data and how the software models interlock is new.

Scientists criticized the breakthroughs reported by OpenAI as clickbaiting , the lurid headlines distracted from the fact that the breakthrough achieved by OpenAI was being exaggerated. As with other similar systems, experts have judged that GPT-2 clearly shows that it has difficulties understanding content and that it tends to cover up these difficulties with the help of statistical analyzes and does not understand the content. For example, a text about unicorns turned into an imaginative text about unicorns with four horns. If the topics are more unusual and special, the generated texts seem to lose quality. For many researchers, it remains controversial whether artificial intelligence can even be achieved with more data. Deep learning needs too much data and, in contrast to humans, needs a lot of examples in order to be able to learn. According to researchers, an algorithmic possibility of simulating human learning remains a mystery to research in artificial intelligence and cognitive science, even with a view to GPT-2 . According to a scientist from the German Research Center for Artificial Intelligence , with a view to GPT-2, the polarizing question of a competition between man and machine, which is often raised in the present, does not arise, but rather the question of how journalists will write their texts in the future be written with the help of artificial intelligence.

In November 2019, OpenAI released GPT-2 on a website where the text generator can be tried out. GPT-2 has already been copied by students and surpassed by more extensive models.

GPT-3

GPT-3 represents a further development of GPT-2 and was first presented in the form of a scientific publication on March 28, 2020. The OpenAI API based on it is the first commercial product from OpenAI. With over 175 billion parameters, it by far exceeds the previously largest T-NLG model from Microsoft, which was comparable with 17 billion parameters.

Other developments

Other products such as B. a robot simulation that can learn a new task by observing it once. The source code can be viewed on GitHub.

Participant in the project

The chairman of the project:

until February 2018:

Other supporters of the project:

High-ranking employees:

The development group started at the beginning of January 2016 with a team of nine researchers. Microsoft's Peter Lee mentioned that the cost of a top researcher in artificial intelligence research exceeded the cost of an NFL quarterback. Although OpenAI pays salaries that are geared towards profit organizations and not those of non-profit organizations, the company cannot keep up with salary schemes from, for example, Google or Facebook. Despite the differences in pay, Sutskever said he was willing to turn his back on Google for OpenAI. The reason for this lies in both the good group of researchers at OpenAI and the company's mission. Brockman stated that it was key to him to bring society closer to the availability of real, safe artificial intelligence. OpenAI researcher Wojciech Zaremba turned down offers from companies that offered him two to three times his market value as a salary in order to join OpenAI instead.

Web links

Individual evidence

  1. ^ John Markoff: Artificial-Intelligence Research Center Is Founded by Silicon Valley Investors . The New York Times . December 11, 2015. Accessed December 11, 2015.
  2. ↑ Dave Gershgorn: New 'OpenAI' Artificial Intelligence Group Formed By Elon Musk, Peter Thiel, And More . Popular Science . December 11, 2015. Accessed December 12, 2015.
  3. ↑ Open AI: Effort to democratize artificial intelligence research? . In: The Christian Science Monitor , December 14, 2015. Retrieved December 19, 2015. 
  4. ↑ a b c d Tech giants pledge $ 1bn for 'altruistic AI' venture, OpenAI , BBC News . December 12, 2015. Accessed December 19, 2015. 
  5. ↑ Tech leaders launch nonprofit to save the world from killer robots . In: The Christian Science Monitor . 
  6. ↑ a b Elon Musk's Billion-Dollar AI Plan Is About Far More Than Saving the World . In: Wired , December 15, 2015. Retrieved December 19, 2015. "Altman said they expect this decades-long project to surpass human intelligence." 
  7. ^ Vishal Sikka : OpenAI: AI for All . Infosys . December 14, 2015. Archived from the original on December 22, 2015. Info: The archive link was automatically inserted and has not yet been checked. Please check the original and archive link according to the instructions and then remove this notice. Retrieved December 22, 2015. @1@ 2Template: Webachiv / IABot / www.infosysblogs.com
  8. ↑ a b c d Silicon Valley investors to bankroll artificial-intelligence center . In: The Seattle Times , December 13, 2015. Retrieved December 19, 2015. 
  9. ↑ a b Cade Metz: Inside OpenAI, Elon Musk's Wild Plan to Set Artificial Intelligence Free (en-US) . In: Wired , April 27, 2016. Retrieved April 28, 2016. 
  10. ^ Gym: A toolkit for developing and comparing reinforcement learning algorithms. OpenAI, accessed on June 28, 2018 .
  11. ↑ GitHub : openai / gym. Accessed June 28, 2018 .
  12. ↑ Dave Gershgorn: Elon Musk's Artificial Intelligence Group Opens A 'Gym' To Train AI . In: Popular Science , April 27, 2016. Retrieved April 29, 2016. 
  13. ^ Greg Brockman: OpenAI Gym Beta ( en-us ) OpenAI. April 27, 2016. Retrieved April 29, 2016.
  14. ↑ OpenAI rules . Retrieved April 24, 2019.
  15. ↑ OpenAI results . Retrieved April 24, 2019.
  16. ↑ a b Matthias Kremp: "Talk to Transformer". This artificial intelligence writes frighteningly well. In: Spiegel Online . November 7, 2019, accessed January 4, 2020 .
  17. ↑ a b Manuela Lenzen: Super trolls at the start. In: FAZ.NET . February 19, 2019, accessed January 4, 2020 .
  18. ↑ Christian Alt: Too dangerous? The AI ​​tool from OpenAi. In: BR television . February 21, 2019, accessed January 4, 2020 .
  19. ↑ Olivera Stajić: Robot journalism: This text is (still) AI-free. In: The Standard . October 24, 2019, accessed January 5, 2020 .
  20. ↑ Matthias Bastian: OpenAI: "Dangerous" Text-KI GPT-2 test now as full version. In: mixed.de. November 6, 2019, accessed January 5, 2020 .
  21. ↑ arxiv : 2005.14165
  22. ↑ OpenAI API. June 11, 2020, accessed on July 20, 2020 .
  23. ^ Robots that Learn. In: OpenAI Blog. May 16, 2017, accessed June 28, 2018 .
  24. ↑ GitHub : openai. Accessed June 28, 2018 .
  25. ↑ Elon Musk Invests in $ 1B effort to thwart the dangers of AI , CBS News . December 14, 2015. Retrieved December 19, 2015. 
  26. ↑ a b c Elon Musk, Peter Thiel, Reid Hoffman, others back $ 1 billion OpenAI research center . In: San Jose Mercury News . Retrieved December 19, 2015.