Shulou(Shulou.com)06/02 Report--

The tone of talking about Open AI in the industry today is very different from that of a year ago.

Since its last surprise, its team Open AI Five abused human semi-professionals in "Dota2" in August this year. At that time, it even set off a more violent "AI horror theory" than AlphaG, which once became the strength proof of Open AI kicking DeepMind and punching Facebook.

But things have changed recently. In the second half of the year, Open AI is either working on setting a new game record (vengeance of Montezuma surpasses DeepMind), sharing its own game strategies (launching Spinning Up in Deep RL courses to teach people to play games), or researching a new environment that is more suitable for playing games (launching training platform CoinRun).

Although the one-stop technology of these games has also attracted media attention, it has obviously failed to make a big splash.

DeepMind, on the other hand, is not responsible for half of the top papers, or gains an oral Nobel Prize for major breakthroughs such as Alphafold, and there are immortal achievements on the screen from time to time.

All the signs show that Open AI no longer seems to be the best player among the tech giants, and is paying more and more attention to the rhythm of e-sports 's game.

Why are genius geeks addicted to games? Why do tech giants fail so often? Whether it is the distortion of human nature or the loss of morality, today we "approach science" (cross out) to peel off the cocoon to see what went wrong with Open AI.

The trouble of Young Victor: Open AI's Adolescent Syndrome

Before public execution, let's briefly popularize the way Open AI operates:

Open AI, founded in 2015, has always existed in the AI world in the way of the third pole. Funded by Tesla's founder, Elon Musk, and other tech giants, it aims to conduct artificial intelligence research in a non-profit way.

Open AI takes the form of open source projects to open all research results to the public in order to "prevent enterprises and governments from relying on super AI to extract profits or power excessively".

This kind of technological idealism of "AI for the future of mankind" is very appealing to some researchers. Therefore, although the salary is not very competitive, the establishment of Open AI still attracted a lot of top talent to join. Examples include former Stripe CTO Greg Brockman (Greg Brockman), former Google researcher Ilya Sutskova (Ilya Sutskever), talented computer guru Alan Kay (Alan Kay), and many talented young people who have worked in academic institutions, Facebook AI or DeepMind.

Today, Open AI has more than 40 technicians working in an investor-provided office space, while some of the infrastructure used for its work, such as AWS services, is provided by Amazon.

Left: Sam Altman of Y Combinator

Right: Elon Musk of Tesla and Space X

As a non-profit organization, Open AI's contributions since its establishment are mainly reflected in three aspects:

1. Lead the technological breakthrough of deep learning. For example, it held the first reinforcement learning competition for AI, released the deep reinforcement learning education resource Spinning Up, and recently launched a training environment CoinRun to solve the general problems of the AI model.

two。 Promote AI security and sustainable development issues. Because of its own public welfare, Open AI is very concerned about the security of artificial intelligence. In the analysis report released, it repeatedly alerts attention to the security risks and malicious use problems that may be caused by the increase in computing.

3. We will promote the inclusion and democratization of technology. Open AI is one of the main promoters of "democratization of AI". All the fruits of its development will be open to everyone and share information freely with partners. For example, it works with GoogleBrain (GOOG) and the University of California, Berkeley (UC Berkeley) to develop modern machine learning systems.

It seems that Open AI is a "justice alliance" organized by technical elites. But in the last six months, Open AI seems to have experienced a wild and confused rebellious adolescence.

The most obvious is that it is losing its navigational ability in the AI world and becoming an organization focused on playing games. As a result, its appeal to the media has also begun to diminish, so that for more than half a year, people's impression of it is still stuck in the canyon where human players' game records are brutalized.

From being all-powerful to being unable to show a sense of existence, is it that Open AI is too floating, or is it that Google can't use the knife?

Also love playing games, why one became high achiever and the other became an Internet addict?

The confrontation between AI and gamers seems to have become the core focus of Open AI. The reason is also relatively simple:

One is in line with Open AI's AI research and development logic: first develop new reinforcement learning algorithms, then train in the simulation environment to solve difficult problems, and finally apply the system in the real world. The game has natural advantages in training AI, such as clear data support, training results are clear at a glance, and there are a large number of human training data as a frame of reference. Obviously, Open AI focuses on game simulation.

Second, it is easier to make big news. Because of the complexity and unpredictability of e-sports events, onlookers often take the human players and game running scores as the yardstick to measure the value of the algorithm, and then become the standard to evaluate the technical ability of a company. Confrontation with players is easy to attract attention, and it has become the main way for Open AI to brush its sense of existence.

However, the game that once lifted Open AI to the altar is also getting it into a quagmire. The most obvious thing is to play games day after day, forgetting that the main business is born to catalyze change. As a result, many people begin to question Open AI's public welfare model.

Open AI's team.

Some people say that Open AI's initial dream is unrealistic, while others say that too much research on public welfare projects is not strong enough, and it is often difficult to escape the fate of going nowhere.

In contrast, DeepMind, which is next door to Google, teamed up with Blizzard to fight human players in StarCraft Ⅱ, but then contracted half of NIPS's paper indicators. Also playing games, does it prevent people from being high achiever? No!

However, it is obviously not appropriate to simply and rudely sentence the public welfare AI to death. If you want to pull Open AI back from the trend of marginalization, you may have to go through a "bone scraping therapy".

Is the road of AI public welfare really impassable? It's just a congenital deficiency of Open AI.

Is the road of AI public welfare really impassable? Apparently not.

In addition to releasing data sets and training platforms through open source crowdsourcing, Google has also begun to test the public good principle of "democratization of AI".

Last year's Google Cloud Computing Conference put forward the goal of using the power of Google cloud computing to democratize AI technology, and even launched AutoML, which can automatically generate machine learning models.

Also holding high the banner of "AGI" (guiding AI to General Intelligence), Google and Open AI are gradually being squeezed into a narrower and narrower runway. Why in the public welfare field, the root is seedling red Open AI, unexpectedly than the enterprise also lacks stamina? I'm afraid we have to start with the inherent shortcomings of the "non-profit organization" business model:

1. Freedom and confusion: the double-edged sword of public welfare.

Idealism, which is not for profit, is the strength of Open AI and is becoming the weakness of Open AI.

With its grand mission and vision, Open AI can attract talent, win attention and remain competitive. But the lack of business goal drive also makes Open AI lack a sense of purpose, which ultimately shows a lack of clear ideas about what he wants to do in the next few years.

Focus and theme, what is the course of action, and how to meet the needs and expectations of the audience? They are all covered by clouds and mountains, and they can only "dawdle" by playing games.

two。 Excessively loose management system and absent supervision mechanism.

Open AI gives employees a lot of freedom and openness, so you can find a lot of top talent even if the salary is lower than that of the business owner. But it also brings two problems:

First, there is a lack of a clear and effective supervision mechanism, and the output efficiency can only be based on the general observation of investors, which makes it difficult to effectively evaluate the results of the work. it is also easy to make organizational action a mere formality because of inertia, personnel changes, political factors and so on.

Second, the uncertainty of retaining talent is greatly enhanced. "generating electricity with love" is noble, but once it is difficult for senior talents to meet self-realization in their work, they will face the problem of brain drain. Obviously, Open AI does not have the corresponding incentive mechanism to keep employees' initial enthusiasm and fighting spirit consistently.

Problems have begun to emerge, and recently there has been a steady loss of Open AI researchers. Ian Goodfellow, a famous researcher and author of GAN, left Open AI at the end of February and returned to Google brain.

3. Weak public relations and sustainability.

Unlike the marketing and public relations strategies that companies often invest heavily, it is obviously more difficult for non-profit organizations like Open AI to maintain good public relations and visibility.

The artificial intelligence project must invest a lot of manpower and hardware if it is to be carried out smoothly. At present, Open AI is still at the stage of living on its old capital (the $1 billion investment it received in 2015), and how to raise funds in the future is an inevitable problem.

When e-sports strategy began to cause cognitive fatigue, Tesla this tree "can not protect itself" at the moment, Open AI must take the initiative to break the organizational inertia of weak marketing.

In a word, Open AI starts from the public welfare, but also trapped in the public welfare. When outsiders denounce Open AI as "too young to be big", one problem may have been overlooked, that is, the inherent weakness of non-profit organizations.

It is not alone. Paul Allen, the former founder of Microsoft, invested $100m in an advanced technology think-tank and, like many technology public welfare projects, went nowhere. It also reminds us that a weak central organization must first overcome its own difficulties before binding the future of mankind to itself.

Coming out of puberty: where is the future of Open AI?

Compared with the increasingly closed corporate AI and the increasingly commercial academic AI, a rare species like Open AI, whose ideological awareness and business level are online, is really hard to blame.

Therefore, we might as well consider the next question from a more comprehensive perspective-how can Open AI, which takes social responsibility as its responsibility, get out of the confusion?

Here are three tips:

1. Build a long-term strategy. On the basis of organizational consensus, the broad and vague mission of "realizing AI democracy" will be transformed into substantive and measurable action goals to promote the growth and evolution of Open AI itself, even if the game is more targeted.

two。 A strong core of leadership. Even with a clear mission and a lack of strong leaders, it is like a ship with a chart but not a competent captain. In the early days, Open AI was supported by Elon Musk and Sam Altman. Now Musk has left Open AI's board to establish a strong leadership core, which is crucial to Open AI.

3. A sound financial system. At present, people are a little tired of Open AI playing games, "beat xxx in the game" has been very difficult to attract media reports, to change this impression, Open AI still needs to come up with something new. The better the reputation of an organization, the easier it is to get financial assistance from a third party. As Musk said, the general task of a non-profit organization is not urgent, but Open AI is not, and what it has to do is urgent.

All in all, once the honeymoon period of the public and the media and non-profit organizations is over, Open AI, without a filter, will have to face the confusion and growing pains of adolescence. Whether one can complete self-evolution or not depends entirely on one's own adaptability.

Drucker has a famous saying that profit organizations learn their mission from non-profit organizations, and non-profit organizations learn efficiency from for-profit organizations. Even the hegemonic Google has to embark on the broad road of "democratization of AI", and it is time for Open AI to start mapping out a clearer road map of evolution.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.