Many articles on Seeking Alpha deal with Alphabet Inc.'s (NASDAQ:GOOG, NASDAQ:GOOGL) finances, general business, and prospects. This is not one of those articles. Instead, it's all about DeepMind: a division of Alphabet listed in the reports alongside Waymo and other business models under "other bets." I intend to provide an overview of what DeepMind already contributes to Alphabet and what potential there is.
AI is probably the most important thing humanity has ever worked on. I think of it as something more profound than electricity or fire.
Sundar Pichai (Google CEO)
A little bit of history
DeepMind was a British startup founded in 2010 by Demis Hassabis and others. He is still the most famous face of the company today. Among the first investors were Elon Musk and Peter Thiel. In 2014 DeepMind was acquired by Google (Facebook (FB) was also interested) and since then it has had the goal to "solve intelligence."
Past achievements: From computer games to protein folding
Even before the takeover by Google, DeepMind was mainly engaged in solving older computer games. The idea was to let an artificial intelligence learn the game from scratch, playing against itself and learning from its defeats and victories. In this way, natural learning behavior should be imitated as best as possible.
After the acquisition, DeepMind took on more and more complex games. For example, in chess, DeepMind's AlphaZero defeated the best chess engine to date after only 4 hours of training. Engines have beaten humans in chess for many years, but most were fed by massive databases and calculated further moves with powerful computers. The fact that DeepMind's AI beats the other engines with this independent learning is an impressive success.
AlphaGo caused even more of a stir when it defeated the then-world champion Lee Sedol in a best-of-five game of Go. No AI had managed that before, partly because Go is not calculable. The number of possible moves is too high: Go is considered one of the most complex games ever. There is a documentary about this duel that is well worth watching.
A system like AlphaFold that is able to accurately predict the structure of proteins could accelerate progress in many areas of research that are important for society. AlphaFold is already being used by our partners. For instance, the Drugs for Neglected Diseases Initiative (DNDi) has advanced their research into life-saving cures for diseases that disproportionately affect the poorer parts of the world, and the Centre for Enzyme Innovation at the University of Portsmouth is using AlphaFold's predictions to help engineer faster enzymes for recycling some of our most polluting single-use plastics.
The long-term goal is the development of general-purpose AI. However, all experts currently assume that this will not be achievable until 2030. A general-purpose AI would theoretically be able to combine all available information in all areas and apply it creatively, similar to the way humans can. For example, it would no longer be possible to determine whether one is talking to a human or a machine when used in a speech AI. Such an AI would probably offer us creative solution suggestions for all our problems and would make human research pointless since it would be inferior from then on anyway.
DeepMind is of the opinion that reinforcement learning is sufficient to achieve a general AI in the long term. Until then, however, AI will be applied to smaller subareas in order to optimize processes and solve problems. Similar to what has already been successfully tested in games.
Is DeepMind making money yet?
There are two ways DeepMind can contribute to Google's bottom line. Either to generate revenue, or to reduce costs. So far, it's the latter that's happening, which usually causes less of a stir than big revenue numbers but is just as valuable. DeepMind generates "revenue" by applying its technology to other Alphabet projects. And these more than tripled (!) to £826 million in 2020, according to the latest financial statement. Expenses rose from £717M to £780M.
In 2019, the company still made a loss of 477M pounds. So from 2019 to 2020, there was a tripling in sales and profit for the first time. As an independent company, DeepMind would undoubtedly have a valuation in the billions, given these numbers and the intellectual assets it has accumulated.
Of course, you have to be careful with all of this. Since it is revenue within the same company, it could be that Alphabet is artificially inflating this revenue a bit. But it's more likely that DeepMind's developments have tangible benefits that add income or reduce costs elsewhere at Alphabet. We can also find examples of this on DeepMind's website.
One more observation: DeepMind is part of Alphabet's "other" bets, but I'm not sure how that revenue is booked. This is because the total revenue of "other bets" in 2020 was $657M, which is less than DeepMind's reported revenue.
Where DeepMind is already operational
Here are some examples of what DeepMind is already doing right now.
- Identifying eye diseases: DeepMind's algorithm can detect over 50 sight-threatening eye diseases as well as leading experts in the field. What's more, the system can even predict who is more likely to develop age-related eye diseases.
- Save electricity: Google's data centers have to be cooled constantly, consuming a tremendous amount of energy. DeepMind has helped optimize this cooling, saving 30% (!) of the power.
- Increase energy output: It was new to me, but Alphabet operates 700MW of wind power. DeepMind has developed algorithms that use past performance and the weather forecast "to make optimal hourly delivery commitments to the power grid a full day in advance. This is important because energy sources that can be scheduled (i.e. can deliver a set amount of electricity at a set time) are often more valuable to the grid." (source). As a result, the value of the generated energy has increased by about 20%.
- More natural-sounding AI voices: WaveNet is used in Google Assistant and other systems to make the artificial voices sound more human and improve the user experience. Among other things, it mimics natural pauses in breathing.
- Ancient texts completing: Archaeologists usually find only fragments of ancient stone tablets and other relics. Putting these together into a meaningful whole is enormously complex. DeepMind has developed an AI that helps archaeologists to classify and complete historical texts.
- Traffic prediction: DeepMind has helped improve the estimated time of arrival accuracy by up to 20% to 50% by making better predictions of surrounding traffic.
Many additional opportunities
- Applied to sports, big data can lead to better decisions: Traditional sports is an area where big data has not been used much, even though a lot of money is at stake. However, if algorithms can help teams or individual players make better decisions, it will probably only be a matter of time before they are used more and more. For example, DeepMind describes a situation in a penalty shootout in soccer. An analysis showed that evaluating various data (past victories, which is the strong foot) leads to much better predictions of where the shooter will shoot: precious information for the goalkeeper.
- Breast cancer screening: In medicine, there are many possible applications. One of them is the more accurate detection of cancer. In one study, this reduced both false positive detections of breast cancer and identified cancer cases undetected by humans.
- Game balancing: In the past, it took a lot of work for game developers to create balanced games. An excellent example is the real-time strategy genre. Usually, the player can choose several races or countries, and none of them should be stronger than the other. Humans would need thousands of hours to create a balanced system. However, an algorithm could do that within hours, and DeepMind is very well positioned here due to its experience in the gaming field.
- Controlling nuclear fusion plasma: Nuclear fusion could bring a revolution in energy production, but there are still many challenges to solve.
A control system needs to coordinate the tokamak's many magnetic coils and adjust the voltage on them thousands of times per second to ensure the plasma never touches the walls of the vessel, which would result in heat loss and possibly damage. To help solve this problem and as part of DeepMind’s mission to advance science, we collaborated with the Swiss Plasma Center at EPFL to develop the first deep reinforcement learning system to autonomously discover how to control these coils and successfully contain the plasma in a tokamak, opening new avenues to advance nuclear fusion research.
Article: Accelerating fusion science through learned plasma control.
These lists are far from complete. The company's blog is full of descriptions of research, including links that could keep you busy for days. By this point, it should be clear that DeepMind has enormous potential.
I think it's fair to say that DeepMind is still in its infancy. However, many of the previous researches are slowly starting to be integrated into social aspects and become monetarily worthwhile. The fact that in 2020 alone the stated revenue tripled is quite remarkable.
In addition, one must also consider that many of the application areas are only used by Google itself. This means that Google has not yet decided to resell these solutions to other companies. An example of this potential is how the sales of Google's wind turbines increased by 20%. With the many gigawatts of installed capacity and the amount that will follow, even a minimal optimization would have an enormous value for the operators.
As said at the beginning, this is only about DeepMind. The current revenue drivers for Alphabet are, of course, other areas. However, a strong and promising DeepMind is one more reason to continue investing or staying invested in Alphabet.