First DeepMind AI conquered Go. Now it’s time to stop playing games

where does deepmind go after alphago ai google 1

DeepMind’s AlphaGo artificial intelligence shut out the world’s best Go player, 19-year-old Ke Jie, ending their series at 3-0 in late May. Shortly afterward, the Google-owned company announced that AlphaGo was retiring from active competition, having beaten the very best around.

It makes sense for AlphaGo to bow out while it’s ahead. But by digging deeper into this decision, and looking at historical context, we can see where DeepMind plans to head next. And the future looks very healthy indeed.

There’s No ‘I’ in ‘Team’

On the surface, AlphaGo is retiring because of a lack of competition. But a bigger problem for DeepMind is a more likely culprit: The project has grown bigger than the research group that created it.

Google Trends chart showing Alphago's popularity over Deepmind
Google Trends: Deepmind and Alphago comparison

As the above Google Trends chart demonstrates, there’s more awareness of AlphaGo than there is of DeepMind, at least when one of the system’s high-profile Go matches is taking place. That’s troublesome, because AlphaGo is simply a step along a path towards the goal of creating an adaptable and competent AI, not the goal.

AlphaGo was developed to be good at Go, as its name suggests. Its ability to play the game has no doubt raised the profile of its creators, but it seems that DeepMind is being overlooked. People know AlphaGo, and they’re aware that it’s a project that’s backed by Google. The actual team that’s working on the technology, however, is rather anonymous outside of Silicon Valley.

DeepMind is the brand, and AlphaGo is the product – and right now, the product is eclipsing the brand.

DeepMind is the brand, and AlphaGo is the product – and right now, the product is eclipsing the brand. This could all change when the group launches a new project that, coming from the creators of AlphaGo, will likely be able to hit the ground running in awareness.

“The research team behind AlphaGo will now throw their energy into the next set of grand challenges, developing advanced general algorithms that could one day help scientists as they tackle some of our most complex problems,” wrote Demis Hassabis, Co-Founder & CEO, and David Silver, a Research Scientist with DeepMind.

Stepping away from Go will help DeepMind demonstrate the full breadth of its capabilities when it comes to artificial intelligence. Of course, this isn’t the only reason AlphaGo is retiring. A lengthy reign at the top can only end one other way.

A Fighting Champion

AlphaGo is no longer the underdog. From here on out, its human opponents will enter any series expecting a loss. That puts the impetus on the AI to defend its reputation. Top ranked players will be able to square off with the system, knowing that they can take the loss without losing face, and with the prospect of making a major name for themselves in the unlikely event of a victory.

It’s the perennially relevant story of the prize fighter who climbs to the top and stays there — right up until he’s knocked from his perch. The weight of expectations and the eagerness to see someone slay the giant makes it incredibly difficult to remain a champion, and almost impossible to make a second ascent.

Even holding steady doesn’t do AlphaGo much good. AI is interesting because it’s a new technology that’s constantly evolving. Maintaining a position as the best Go player in the world doesn’t prove DeepMind is moving forward in the same way that its victories up to this point have done.

It’s important to remember that when IBM’s Deep Blue system beat chess grandmaster Gary Kasparov two times in a row in the 1990s, Kasparov was denied the rematch he requested. Projects like AlphaGo and Deep Blue are developed to demonstrate blossoming technology, to drum up awareness among the public, and hopefully garner financial support from entities that can apply the ideas that are in play. Exhibition matches against the best of the best have only a limited amount of utility, no matter what game is played.

If DeepMind can take on a new challenge, however, the slate is wiped clean. The group’s next public-facing project needs to be a similarly insurmountable task as defeating the world’s greatest Go player. The challenge is to find an activity that doesn’t just measure up to AlphaGo’s recent accomplishments, but exceeds them entirely.

The house always wins

DeepMind isn’t the only group pushing the limits of AI, and various other teams have ideas about how to demonstrate the capabilities of their technology. For instance, one project saw a neural network take on top Super Smash Bros. players. DeepMind itself is collaborating with Blizzard to teach an AI how to play Starcraft II.

Video games are more complex than board games, but there’s something less impressive about seeing a machine excel at digital entertainment. These games are made up of ones and zeroes; it’s easy to assume that a computer would be better than a human player. Whether true or not, exhibition matches that pit man against machine are all about public perception.

Close up of hand playing Go

For the same reason, DeepMind probably won’t teach a machine to play Arimaa, a board game developed with the specific purpose of being difficult for machines to play. It would be a great demonstration of how far this technology has come, but since it’s an incredibly niche game, it wouldn’t appeal to the masses in the same way an AI playing Go could.

Looking to traditional board games in the same vein as Go, there are a couple of options. The Chinese xiangqi and the Japanese shogi both stand out, each being complex and competitive. But good progress has already been made in teaching computers to play both, so it’s doubtful DeepMind’s name would be thrown into that hat. Mastering another one-on-one strategy game wouldn’t be too much of a diversion from the AlphaGo project.

Instead, expect DeepMind to pursue a challenge with clear relevance to all onlookers. In January 2017, an AI called Libratus took on some pro poker players and won, making headlines around the world. It wouldn’t be surprising to see DeepMind field a similar project.

Even though it’s already been done, poker could be a potent choice, with bigger stakes and better competition — and providing more spectacle. Libratus took on players at the Rivers Casino in Pittsburgh, PA but it wouldn’t be surprising to see DeepMind head to Vegas or Monte Carlo to add some pomp and ceremony.

Poker is an interesting challenge for AI because it represents a marked increase in the volume of human competition. From Deep Blue facing Kasparov, to AlphaGo squaring up to Ke Jie, there have always been detractors who have claimed that the computer players have been programmed with a specific opponent in mind. With several players around the table, DeepMind would be able to demonstrate an ability to take on a selection of different playstyles at once.

There’s another possibility: it may turn out that DeepMind is done playing games. It might be time for AlphaGo — or at least the technology that underpins it — to grow up and get a real job.

An end to the exhibition era

Everybody knows that Deep Blue took down Kasparov at his own game, and anyone that keeps an eye on tech news knows AlphaGo has demonstrated a similar mastery of Go. Over the past twenty years, we’ve been given plenty of evidence that AI is more than just science fiction.

That said, most of us are yet to feel any palpable benefit from AI in our day-to-day lives. Sure, various companies wheel out terms related to the technology as a marketing buzzword — but the results are seldom the quantum leap we’re looking for.

While the AlphaGo project was making headlines, the company was also working on several less showy projects.

This is the real challenge DeepMind must face. Few doubt the potential AI holds to change the world for the better, and AlphaGo has quite capably demonstrated what the group’s specific strain of the technology can accomplish. The time for exhibition matches is over. Now, we need to see what AI can do in terms of practical applications.

DeepMind is aware of this. While the AlphaGo project was making headlines, the company was also working on several less showy projects. These efforts are set to take center stage going forward

In DeepMind’s blog post officially announcing AlphaGo’s retirement from competitive play, Hassabis and Silver noted that the team behind the technology is moving on to algorithms that could help with tasks like “finding new cures for diseases, dramatically reducing energy consumption, or inventing revolutionary new materials.”

We’ve seen how the company’s AI can improve energy consumption in the past. In July 2016, Google detailed how the technology was being used, and it seems the next step is to roll out on a larger scale. This project isn’t going to grab headlines, but it also won’t rub anyone up the wrong way.

Nurse checking temperate with iPhone app

The same can’t be said for DeepMind’s efforts to branch out into medicine. The idea of using AI to detect early warning signs of disease sounds like a PR win, but the execution of the project has already caused controversy.

In May 2017, DeepMind was criticized after it was given access to over 1.6 million medical records by the U.K.’s NHS. The issue was whether the company was providing direct care, given that the records it was supplied with were personally identifiable.

That’s the kind of problem that Deepmind didn’t have to worry about while conquering Go. AI in the real world, performing tasks that affect real people will cause real consequences — whether from the information being analyzed or the aftermath of something going wrong.

AlphaGo was AI with training wheels; losing the game was the biggest risk. While it might’ve been embarrassing for Deepmind, failure was an option. The next stage of the company’s lifespan is going to see more consequential hurdles. Let’s see how high AI can jump.