If you ask researchers, what are the strongest AI laboratories in the world? Everyone must be “radish and cabbage, each has their own love”, but it is certain that the top three will not waver: DeepMind, OpenAI and FAIR. However, some people also questioned that the reputation of these three laboratories is not good, and there are also elements of “forced public relations”.

This article is from WeChat official account:Xinzhiyuan (ID: AI_era) ,Source: cnbc, edit: Q, Xiaoyun, Original title: “Who is the world’s top AI laboratory? DeepMind, OpenAI and FAIR dominate the top three”, the head picture comes from: Visual China

The large American technology companies such as Google, Facebook, Amazon, Apple, and Microsoft have established specialized artificial intelligence laboratories in the past decade.

If you ask AI researchers, which is the top AI laboratory in the world?

Many people can’t give an answer, but when it comes to the top three, everyone can almost give a consensus: DeepMind, OpenAI and FAIR.

The three top AI laboratories are backed by Google, Microsoft, and Facebook respectively. At the same time, these three laboratories are pure AI research laboratories. Alphabet allocates hundreds of millions of dollars to DeepMind every year. Microsoft’s founding investor in OpenAI is $1 billion. It also invested 1 billion US dollars on the basis. Facebook did not classify FAIR’s investment funds, but it was also expensive.

“In terms of reputation, DeepMind, OpenAI and FAIR are the top three.” Associate Professor Mark · Georgia Institute of Technology School of Interactive ComputingRiddle said so.

An anonymous expert said that DeepMind, OpenAI, and FAIR may be the three pure artificial intelligence research laboratories with the largest known funding. At the same time, he also mentioned the experiments of Baidu and Tencent when referring to Chinese technology giants. The condition of the room is not yet known.


DeepMind is a British artificial intelligence company. The company was founded in 2010, originally named DeepMind Technology (DeepMind Technologies Limited), and was acquired by Google in 2014.

DeepMind is best known as AlphaGo, which challenged and defeated the best human players in the world in the game of Go, and even has a Netflix documentary about AlphaGo’s victory over South Korean Go legend Lee Sedol.

The company is currently committed to using artificial intelligence to solve the biggest scientific problem of mankind. At the end of last year, the company made a breakthrough in the biological field called protein folding: in the international protein structure prediction called the “Protein Olympiad” In the competition (CASP), AlphaFold defeated the rest of the participants and was able to accurately predict the 3D structure of the protein based on the amino acid sequence.

DeepMind began to develop the artificial intelligence Go software AlphaGo in 2014.

In October 2015, the distributed version of AlphaGo first defeated the European Go champion and Chinese-French professional chess player Fan Hui 2nd Dan by 5:0. This is the first time that a computer Go program has defeated a professional Go player on a nineteen-way board and split first.

In March 2016, AlphaGo challenged the world champion Korean professional chess player Lee Sedol 9th ​​dan. The result of the game was that AlphaGo defeated Lee Sedol 4:1.

On January 25, 2019, DeepMind’s artificial intelligence AlphaStar defeated the human profession by 10:1 in StarCraft IIPlayers.

On December 23, 2020, DeepMind announced its AI algorithm MuZero.


OpenAI was established at the end of 2015 and is headquartered in San Francisco. Its founders Elon Musk and Sam Altman originally founded it out of concerns about the potential risks of strong artificial intelligence. After just five years of establishment, it has become one of the world’s leading artificial intelligence research laboratories. Most importantly, it is highly regarded by the public for its mission: the goal is to be the first to create AGI-one A machine with the ability of human thinking to learn and reason. And the laboratory hopes to evenly distribute its benefits to the world.

OpenAI has also developed game artificial intelligence software that can defeat humans in games such as Dota II. However, it is more famous for GPT-3 and artificial intelligence image generator DALL-E.

DALL-E is a Transformer-based language model that uses the 12 billion parameter version of GPT-3. It simultaneously receives text and images as a single data stream, which contains up to 1280 tokens, and uses maximum likelihood estimation for training to generate all tokens one by one. This training process not only allows DALL-E to generate images from scratch, but alsoIn order to regenerate any rectangular area of ​​the existing image, it is basically consistent with the content of the text prompt.


FAIR itself does not have as well-known models and applications as AlphaGo and GPT-3, but its team has published academic papers in areas of interest to Facebook itself, including computer vision, natural language processing, and conversational AI.

The idea of ​​establishing FAIR began in 2013. Facebook’s founder Zuckerberg, CTO Mike Schroepfer and other company’s stock-holding leaders are all looking for ways to keep the company competitive in the next 10 to 20 years. technology.

Facebook has previously used machine learning technology to determine what kind of message flow users will see on their social networks, but compared to the most cutting-edge neural network model, this is a relatively simple matter.

At that time, some Facebook engineers were also experimenting with convolutional neural networks (CNNs), which is a powerful method in the field of machine learning. Now it is commonly used in the image field. Zuckerberg was extremely impressed with the potential of artificial intelligence, even in the early stages, so he hired an engineer from Google Brain, Marc’Aurelio Ranzato. Then, he found the creator of convolutional neural networks: Yann LeCun.

During his tenure as the head of Facebook’s artificial intelligence laboratory, he talked about Facebook’s artificial intelligence layout and defined the organizational structure as the following:

AI at Facebook = FAIR + Applied Machine Learning + Product Groups

It includes:

1. FAIR, Facebook Artificial Intelligence Lab

2. Applied Machine Learning, Applied Machine Learning Department

3. Product Groups, product deployment team

However, he later stepped down as the dean of the Facebook Artificial Intelligence Research Institute (FAIR), and then became the chief AI scientist, focusing on research work.

One way to measure the impact of an artificial intelligence laboratory is to see how many academic papers it has published in two artificial intelligence conferences: NeurIPS and ICML.

In 2020, Google has 178 papers accepted and published by NeurIPS, Microsft has 95 papers, DeepMind has 59 papers, Facebook has 58 papers, IBM has 38 papers, and Amazon has less than 30 papers.

On ICML in the same year, 114 papers were accepted and published by Google, 51 papers for DeepMind, 49 papers for Microsoft, 34 papers for Facebook, 19 papers for IBM, and 18 papers for Amazon.