This article is reproduced from the Silicon Star, author Juny
While people were marveling at DeepMind's successive technological breakthroughs, DeepMind handed over a loss bill of up to 649 million US dollars (about 4.2 billion yuan). And even worse than DeepMind's huge losses is E lem ent AI, the Canadian star unicorn backed by the world's largest deep learning community, which was recently sold at a low price under the difficult situation of making ends meet. The same commercialization challenge actually exists in Open AI, an artificial intelligence research organization in Silicon Valley.
At a glance, cutting-edge artificial intelligence technologies such as AlphaFold, MuZero, and GPT are shiningly standing on the mountain peaks stacked with banknotes, but it seems that few people can actually climb up to shake hands with them.
So, why has artificial intelligence technology passed through the barriers and developed to the fourth generation, but the road to commercialization is still full of difficulties?
It's not only DeepMind who loses money
In fact, DeepMind is one of the largest artificial intelligence research institutions in the world. Since its official establishment in 2010, DeepMind has never achieved profitability. Instead, the "burning money" index is still rising year by year. If it were not for Google in 2013, it had spent 6 DeepMind was acquired by billions of dollars, and the laboratory may have been bankrupt and dissolved.
Fortunately, DeepMind did come across a "gold master dad" who spoiled it enough. Although it has suffered huge losses year after year after being acquired, Google recently expressed that it is still willing to continue to provide DeepMind with expensive AI research funds, saying the company's recent research Very satisfied with the progress, he also forgave his 1.1 billion pounds of debt.
Figure 2014-2019 DeepMind's loss and its growth, graphed by Silicon Star
And like DeepMind, the artificial intelligence research institution "spoofed" by big companies is Open AI. Open AI is an artificial intelligence non-profit organization founded in December 2015 by Musk, former YC president Altman, and many Silicon Valley entrepreneurs. It aims to prevent the disastrous effects of artificial intelligence and open patents to the public. And research results, by 2017, about 1 billion US dollars in financing had been raised.
After parting ways with Musk in 2018 due to Tesla's development and conflict of ideas, Open AI announced its transformation from a non-profit laboratory to a "capped profit" company, and in July 2019 it received up to $1 billion in investment from Microsoft. The direct reason for Open AI's commercial transformation is that its existing capital and income situation cannot meet the continuous investment demand for computing and talents.
Compared with DeepMind and Open AI that can focus on technological research with the support of large companies, Element AI, an artificial intelligence unicorn that was once regarded as a "star of tomorrow" by Canada, seems to be less fortunate.
Regarding the development of Element AI, the Silicon Stars also introduced in a series of articles in Canada: an "AIaaS" wind is blowing: the unicorn Element AI has turned into an artificial intelligence "big Mac", which is the winner of the Turing Award, Yoshua Bengio, one of the "Big Three of Deep Learning", was founded in October 2016. The Montreal MILA laboratory led by Bengio is currently the world's largest deep learning academic community. The products developed by Element AI are based on the research on artificial intelligence technology of a large number of scholars in MILA over the past few decades. Element AI has also developed in the world in recent years. One of the fastest artificial intelligence startups, once considered the leader of Canada's artificial intelligence industry in the future.
However, this once high-spirited star company that aims to be the pioneer of the global artificial intelligence consulting industry was revealed last month that it will be acquired by the US cloud computing platform service provider ServiceNow, and the purchase price is only 230 million US dollars. Not only is it far below its valuation of more than 1 billion Canadian dollars in its last round of financing, it is even less than the total financing of US$257 million in 4 years.
According to a document previously disclosed by The Globe and Mail, on the eve of the acquisition, Element AI's capital chain was almost exhausted, a large number of employees were fired, and the annual income was only about 10 million Canadian dollars. After deducting costs, the final purchase price of Element AI may be less than $195 million.
"Burning money" machine and application dilemma
Although the destinies of these artificial intelligence laboratories are different, they have one thing in common, that is, they all carry out artificial intelligence technology research around deep learning, reinforcement learning and other cutting-edge fields. DeepMind and Open AI are all aimed at the ultimate realization Artificial General Intelligence (AGI). The so-called general artificial intelligence is to allow machines to have general human intelligence and can perform any intellectual task that humans can perform.
In recent years, these artificial intelligence research institutions have indeed been moving in the direction of general artificial intelligence, and have achieved good results, but technological breakthroughs have not brought commercial returns, and input and output are far apart. And the two huge rocks on the road to AI commercialization, one is funding and the other is application.
First of all, if you want to make a breakthrough in the field of machine learning, it is impossible not to "burn money", not only "burning" or "burning". Since the deep reinforcement learning exhibition is based on massive data processing and complex knowledge reasoning, it is difficult to support the conventional stand-alone computing mode, so the demand for computer resources when training the model is extremely high.
For example, in May 2020, Microsoft launched a supercomputer specially built for Open AI for AI model training. It cost hundreds of millions of dollars. Google's TPU has been in a state of being leased to DeepMind. In addition to hardware, in terms of training, take Open AI’s well-known text generation algorithm GPT as an example. A model with 1.5 billion parameters costs US$2,048 per hour to train, while the AlphaGo algorithm similar to DeepMind requires at least completion of data before success. A million times of self-game, the training fee alone will cost 35 million US dollars.
In addition to technical infrastructure, the salary of researchers is another big one. Although there are more and more people studying computers, the person who can engage in deep reinforcement learning research is water chestnut. The annual salary of top researchers in the United States is at least one million US dollars. According to the financial report data disclosed by DeepMind last month, its personnel expenditure in 2019 reached 460 million pounds (about 4 billion yuan), an increase of 17.6% over the previous year.
Figure DeepMind's 2019 fee structure, the picture is cut from DeepMind's public disclosure of financial reports
Therefore, Element AI's $260 million financing in less than 4 years may be a huge sum of money for other startups, but it is just a drop in the money for Element AI, which has nearly 500 employees. If it does not have the likes of Google, Musk, and Microsoft. With the support of such a financially large funder, it is no wonder that it eventually came into the dilemma of exhaustion of the capital chain and unsustainable.
On the other hand, when “throttling” is impossible, “increasing revenue” is not easy.
Even if the technology is deified or imagined more powerfully, it has to be admitted that the large-scale commercial application scenarios of deep reinforcement learning have not yet appeared. It has been 5 or 6 years since AlphaGo defeated humans for the first time in 2015, but the disruptive impact it had expected to bring to humans at the time has not yet appeared too much.
The main reason is that in the real world, many things are not restricted environments like chess and games, and there are not enough conditions and data to allow machines to train and learn. In the development process of the laboratory, most models are trained for specific tasks rather than generalization. Therefore, although artificial intelligence can show excellent abilities in playing various electronic games and board games, it has May become very fragile in the real world.
In short, the current reliability of these technologies in real-world applications is not enough.
Although DeepMind's operating income seems to be increasing year by year in recent years, from the disclosed financial data, Google is actually paying for its technology products. Although the first AI commercial product that accurately diagnoses eye diseases within 30 seconds was launched last year, it did not cause much splash due to high-risk areas.
Figure DeepMind's eye disease diagnosis AI, the image comes from the Internet
Although Open AI released the first commercial text generator based on GPT-3 in 2020, in the actual application process, statements with discriminatory, biased or low-level errors appeared, but it still requires the purchaser to invest manpower to check Correction is therefore only used as an auxiliary tool.
Element AI began to introduce several AI products to the market in 2019, including providing cybersecurity services for financial institutions, helping port operators to predict the waiting time of truck drivers, etc., but they were all ineffective. According to the disclosure, the company and Several customers encountered difficulties in the process of cooperation, and finally ended up dimly with an annual income of only 10 million Canadian dollars.
Everything may take time to settle
But for a market where cutting-edge artificial intelligence technology is unsatisfactory, we may be able to say that the time is not yet ripe.
At present, in the commercial application scenarios of artificial intelligence, the most widely used are computer vision, intelligent speech, natural language processing and other major technical directions, but you should know that these mainstream AI technologies began in the 7th and 80s of the last century It has been continuously revised and demonstrated, and after decades of development, it has not been widely used until now. The general artificial intelligence technology pursued by DeepMind and Open AI is currently only in the opening stage.
As for the commercialization of AI, it is not that the existing technology is immature, or that there is no application scenario at all, but the running-in period of technology and scenario is not long enough, and there is no purposeful optimization for the scenario.
The transformative power brought about by artificial intelligence is difficult to predict. In fact, global technology giants are investing heavily. In China, Baidu has been almost "All in AI" in recent years. Since the establishment of China's first deep learning research institute in 2013, Baidu has spent 15% of its annual revenue on artificial intelligence research and development and invested about 10 billion yuan. It has accumulated nearly 100 billion yuan. At the Baidu World 2020 conference, Robin Li still stated that he must "persist in and firmly believe in artificial intelligence for a long time."
Undoubtedly, everyone is betting on a future. After all, if DeepMind can bring breakthroughs to Google’s Waymo and other projects, or Open AI can help Microsoft develop powerful text tools, then, compared with the huge benefits, the current investment is only insignificant.
At the level of corporate strategy, DeepMind has also paid back a portion of the fare for Google. At least DeepMind has recruited and reserved a large number of top talents for Google in the AI competition of global technology companies, and research results have been successively ambushed by Google in more than ten projects including health diagnosis, wind power generation, and voice assistants.
It's just that the road to pursue breakthroughs in artificial intelligence technology is destined to be difficult and long. Some people have fallen on the road, such as Element AI, and some people are still running, such as DeepMind and Open AI. We no longer have to argue about whether these research institutions are working for the well-being of mankind or being trapped by capital. To a certain extent, capital and technology are inherently complementary to each other.
It is precisely because some people are willing to continue to provide support for theoretical research that can create a safe environment for more scientific researchers to move forward, and the wheels of science and technology can roll forward. Perhaps, soon we will see the accumulation of AI cutting-edge technologies.