The revenue doubled, but it all depends on Google.
Editor’s note: This article is from WeChat public account “qubits” (ID: QbitAI), author of the chestnut Annie.
The talent is still very expensive, and the revenue has improved, but the huge losses are still expanding.
This is not just an AI company. This is the DeepMind that created AlphaGo, and it may be a true portrayal of the entire AI industry.
We started with the latest year’s financial data from DeepMind, which was just exposed.
burn 10 million per day, with an annual salary of 4 million per person
The report shows that DeepMind continued to expand its losses in 2018, reaching 470 million pounds, about 4 billion yuan, and burning more than 10 million on average every day.
What is the core reason why DeepMind burns money like this?
People are too expensive. According to DeepMind’s financial report, most of its expenditures are concentrated on personnel. The related expenses increased to 398 million pounds in 2018, almost doubled compared with 2017.
By LinkedIn, DeepMind employeesThe number has reached 839 employees. On average, the cost per employee is close to 470,000 pounds, equivalent to 4 million yuan – the real number of millions of annual salary.
Under large expenditures, DeepMind still owes a debt. By the end of this year, it will have more than 1 billion pounds of debt due, of which 880 million are from Google.
But the good news is that although the burning rate is amazingly fast, Google still can stand up and stay in the end.
DeepMind is looking forward to the report. In the next 12 months, Google will continue to provide them with adequate financial support and write a written guarantee.
But don’t worry about DeepMind, it doesn’t mean that every AI company can have a good father like Google.
Investing in Google, making money also depends on Google
And the support provided by Google is also reflected in the source of revenue.
The financial report shows that in 2018, DeepMind’s revenue was 102.8 million, an increase of 88.9% compared with 54.4 million in the previous year.
Maybe you will ask, what is the DeepMind revenue model?
DeepMind Disclosure: At present, the main revenue business is cutting-edge machine learning research and application, including providing technical services to other companies.
And this “other company” is not someone else, it is a good father Google.
DeepMind sold some software to Google. Google has used its AI system to make cooling units in the data center more energy efficient and increase the battery life of Android devices.
However, even if revenue doubles, DeepMind’s losses in 2018 will still reach 470 million pounds, an increase of 55.6% from last year’s 302 million pounds. Among them, Operating Loss is 465 million pounds, 201In 7 years it was 27.9 billion pounds.
As mentioned above, the core reason for the loss is that the staff costs are too high and the wages are not low, totaling 398 million pounds, accounting for 84.6% of operating losses.
In recent years, DeepMind has been actively recruiting machine learning researchers and data scientists.
At the end of 2016, the number of DeepMind employees exceeded 350. At the end of 2017, the number reached 700. Today, LinkedIn shows that DeepMind employees have at least 839 people.
The financial report also stated that the increase in technical infrastructure is also an important expense.
In addition, DeepMind also disclosed that last year’s donation and sponsorship to the academic community reached 13.5 million pounds, or about 116 million yuan – possibly also for talent robbing.
Finally, there is a total of about 13 million pounds, about 110 million yuan, for construction, renovation and furniture purchases.
According to the plan, DeepMind will move into the new headquarters in London in 2020.
So the British are indeed elegant, even if they are hugely losing money, making money and investing, they have to rely on Google’s father, but it is not a problem, and they can’t work hard, and they can’t afford a poor salary.
burning money, keep burning money
Loss, the ever-expanding loss is the normal state of DeepMind. Since its inception in 2010, this AI company has not been profitable.
The loss has surged from £293,000 in 2011 to 470 million pounds last year, an increase of more than 1,600 times in seven years.
Since 2014, after being acquired by Google, DeepMind began to disclose employee costs in its earnings report, which increased from £26.04 million to last year’s £398 million, a five-fold increase in four years.
But Google has money.
January 2014,Google acquired it for £400 million (approximately $519 million) and provided solid financial support.
So, even with huge losses, DeepMind can continue to donate and build a new headquarters.
In contrast, OpenAI, another AI research institution that marks it, is less moisturized.
The huge research and development costs have forced this non-profit organization to become a for-profit organization, and the research results have also been clouded, and it was once accused of being CloseAI.
Even the latest $1 billion in financing from Microsoft was also questioned by The New York Times, which was thought to be redeemed with cloud computing resources, and at the cost of signing Microsoft’s exclusive cloud computing resource provider. To commercialize Microsoft agent technology.
So nowadays, AI is also going to be a fight.
Where are you going?
DeepMind also responded to the current financial situation.
In a statement to Forbes, DeepMind stated that its long-term mission is to advance AI research and achieve positive social benefits.
They said they believe that AI has great potential to promote scientific discovery and is proud of the impact of work in areas such as protein folding.
DeepMind explained that the DeepMind for Google team will make great progress and will further apply its experience and knowledge to Google’s research and apply it to real-world challenges. Because of this, the income of the past year has almost doubled. In the future, we will continue to invest in basic research and world-class interdisciplinary teams, and look forward to future breakthroughs.
Before the release of this earnings report, foreign media Wired published an exclusive interview with CEO of DeepMind, which involved a lot of thinking about the future direction of DeepMind.
Kazabis said that the most interesting field in science is competingDifferences between disciplines. When this happens, a miracle happens.
So, the current research focus of DeepMind is to link AI to protein folding research in biology.
If you understand the process of making proteins in nature, can humans make their own proteins? As a result, there are no solutions for Parkinson’s syndrome, Alzheimer’s disease, and diabetes due to protein folding errors.
Deep learning happens to be a key step in this research. Although the amount of genomic data is large and complex, machine learning is just a part of research on predicting protein structure from amino acid sequences.
Kazabis also revealed that in the past few months, they have completed predictions of 90 protein structures. On average, the results of the internal AlphaFold system team outperform other studies:
The AlphaFold team’s prediction of 25 proteins is the most accurate for modeling protein sequences from scratch.
One of the protein skeleton structures predicted by △AlphaFold
In such an academic competition, DeepMind also faces challenges from teams from top universities.
Kazabis feels that DeepMind’s engineering engineering examples should not be underestimated, and execution will exceed the research department, because academic groups have always been too secretive in research.
One more point, DeepMind is self-sufficient in terms of computing resources. Hasabis did not lie, and before training “the strongest GAN in history” BigGAN, DeepMind used 512 TPUs.
Please refer to the previous article:
The best GAN in training history used 512 TPUs, one for self-report: this is not an algorithmic progress, it is an improvement in computing power
Kassabis also talked about project management cheats: the most important thing is to speed up the update.
He said that every six months, senior management will review priorities, reorganize some projects, and encourage teams, especially engineers, to work harder or adjust tasks.
Because the intersection of disciplines will increase the difficulty of research, many project studies take more than six months, usually a large cycle of two to four years.
Hassabis believes that although he is a CEO, he hopes to spend 50% of his time directly on research. He has repeatedly mentioned that despite small progress, there is still a long way to go.
The ultimate goal of DeepMind, past, present and future, is still to implement universal artificial intelligence – AGI.
Interactive one
If you want to get the complete file of this DeepMind annual report, the qubit public number dialog box can reply “DeepMind2018”.
Reference article:
https://www.forbes.com/sites/samshead/2019/08/07/deepmind-losses-soared-to-570-million-in-2018/#7b98a36e3504
https://www.wired.co.uk/article/deepmind-protein-folding