With the aim of developing Artificial Intelligence that can greatly and surely benefit humanity, OpenAI was founded in 2015. This company is located in San Francisco and is owned by Elon Musk and Sam Altman. The contributions to the creation would amount to one billion US dollars.
Table of contents
OpenAI and its human resources
Elon Musk and Sam Altman set out to create OpenAI with other investors. The goal was of course to form an endowment ofone billion dollars. The company declared to be open for collaborations with other institutions. In this optic, the patents as well as the researches have been opened to the public within the limits of the respect of the security.
Nine researchers were involved in the startup in 2016. Two years later, one of the two main founders, Elon Muskdecided to resign because of a conflict of interest with Tesla, which is his workplace. That said, he still remains a donor. The key people at OpenAI are currently Sam Altman CEO, Ilya Sutskever the research director, and Greg Brockman CTO. The strength of the OpenAI group lies in its human capital. Indeed, the group has excellent researchers in its ranks.
What are the company’s objectives?
OpenAI’s main vocation is to work in the field of artificial intelligence. In fact, the suffix “AI” indicates Artificial Intelligence. The founders as well as the collaborators considered that artificial intelligences had risks. Indeed, a too extensive use of AIs made that anomalies can occur. The main goal of OpenAI is therefore to overcome these problems. The ideal for the founders is the creation of an artificial intelligence endowed with a greater safety and being able to be more beneficial to the human being. In order to achieve this, the company is open to any collaboration, from any organization or individual.
Focus on OpenAI Achievements
In the area of artificial intelligence, OpenAI has been very quick to implement programs. Indeed, as early as 2016, a still beta version of OpenAI Gym is available to the general public. The latter is a manifestation of a research base on learning. In 2020, OpenAI makes an announcement. It is about GTP-3 which consists in answering questions in a more natural human language. This one is formed with a lot of words taken from the internet. Moreover, it is also capable of translating languages since an API is associated with it. Remember that the strength of GTP-3 lies in its ability to produce coherent texts in more than one language.
Note that while OpenAI was originally a non-profit entity, it is now a capped for-profit entity. It should even be noted that a Microsoft – OpenAI association.
OpenAI has never stopped innovating for optimal development. Its products are diverse, and focus mainly on reinforcement learning.
Gym is a documentation site in order to provide the user with a simple interface. Indeed, it offers a general intelligence reference that is easy to configure with a wide variety of environments. In fact, its goal is to standardize the definition of environments in AI research publications. This will make the research easier to replicate. Gym is like a larger version ofImageNetThe challenge of large-scale visual recognition.
Since June 2017, it can only be used with Python. Then in September 2017, the site was not maintained and active work focused instead on its GitHub page.
“RoboSumo” consists in teaching virtual humanoid robots ” metalearning” to be move by pushing the opposing agent out of the ring. With this learning process, the agents learn to adapt to changing conditions. When moving to a more violent virtual environment, for example, the agent prepares to stand. We consider that he had learned to balance in a generalized way.
According to Igor Mordatch of OpenAIcompetition between agents can create a “arms race” intelligence. Thus even out of competition, the agent’s ability to function could increase.
Debate Game was launched in 2018. Its purpose is in fact to Research an approach that can help audit AI decisions and develop explainable AI.. In this debate game, machines are introduced to debate toy problems before a human judge.
Using machine learning, Dactyl forms a robot Shadow Hand at from scratch. However, he uses the same algorithm code thatOpenAI Five. The robot hand is fully trained in a physically inaccurate simulation.
Generative pre-training or Generative Pre-trained Transformer (GPT) testifies to its abilities to gain worldly knowledge and deal with long-term dependencies. Then come its successors GPT-2 and GPT-3of the language models of transformer unsupervised. GPTs are like general purpose learners.
MuseNet is a deep neural network. It can generate songs with ten different instruments in fifteen different styles.
API consists in particular in “accessing new AI models. ” It can be called for “any AI task in English.”
DALL-E and CLIP
DALL-E is a Transformer model that creates images from textual descriptions. On the contrary, CLIP creates a description for a given image.
OpenAI Microscope was created to easily analyze features that form inside neural networks.
OpenAI Codex is a descendant of GPT-3. It is considered the AI powering the autocomplete code GitHub Copilot tool.
Video game robots and other benchmarks
OpenAI Five or five robots capable of playing against humans in video games. This particular team of players is essentially guided by trial and error algorithms. It plays on Dota 2 The famous five-on-five fighting video game.
The first steps of OpenAI Five
It was during The international 2017 – the first tournament of the game’s annual championship – that a professional player suffered his first defeat against a bot. Dendi, the player in question had fought live against a bot and lost. Greg Brockmanthe CTO, explained the situation: this victorious bot had been training against itself for two weeks. It was learning to play through a reinforcement learning mechanism. When it succeeded in reaching one or more objectives in the game, this robot was rewarded. Intensive training had been taking place for several months for these particular players. The learning software was aimed at the advancement in programming software competent to solve complex tasks.
OpenAI Five between defeats and achievements
A year later, bots could already unite in teams of five to play against amateurs and semi-professionals. OpenAI Five was able to participate in The international 2018 : two matches against professional players however they came out losers.
But in April 2019 in San FranciscoBy the end of the month, the bots were still playing during a four-day online gaming competition. Their victory amounted to 99,4 % out of a total of 42,179 games played.
Platform-based reinforcement learning is a popular area of research today. Gym Retro is one of them. Generally, it is used to conduct research on RL algorithms and to study their generalization. Long before, research in RL was based on the optimization of agents in solving single tasks. Gym Retro allows in the same way to make generalization between games but differentiating on the appearances.