In the race to develop artificial intelligence, one player is fielding an uncommon strategy. OpenAI, an initiative launched last December, sets itself apart from the highly competitive and staunchly secret efforts of major corporations like Google, Facebook, and Baidu. It is unique in that it is a non-profit, and its goal is “to advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return.”
Quote on the wall at the entrance of OpenAI offices.
New Approach to AI
OpenAI was founded by a heavy hitting group of technology executives such as SpaceX and Tesla founder Elon Musk, LinkedIn founder Reid Hoffman, Y Combinator chairman Sam Altman and Palantir co-founder Peter Thiel. Together, the group has committed $1 billion over a number of years to ensure the initiative can progress without being constrained by meeting quarterly financial benchmarks.
Additionally, it is structured to create artificial intelligence that improves humanity in safe and beneficial ways, that are broadly and evenly distributed rather than hoarding talent, research findings and tools.
Eventually, the company hopes to produce “artificial general intelligence” – a machine with an intellect comparable to humans. In the more immediate future however, it is focused on a popular and rapidly developing subset of AI research known as “deep learning“.
One element of this is reinforcement learning, where machines learn how to complete tasks through trial and error. Another is unsupervised learning, where machines learn from unlabeled data without being corrected by humans, thus developing their own understandings of concepts contained within.
Example of a simulated robot using OpenAI to learn how to walk through reinforcement training.
Goals are Set
Earlier this year, OpenAI published four technical goals it has set for itself. The first is to define a metric for intelligence, to be able to measure their progress towards getting machines to reach set goals in a range of settings.
The second goal is to build a physical robot that does housework. They would not manufacture it, but instead share the plans and technology so others can make general purpose robots that can perform basic tasks right off the shelf.
A third goal is to make a chatbot that can carry on a conversation, fully understand text, follow complicated instructions, and even know when to ask for clarification when something is too vague. Eventually, this agent should be able to actually perform complex tasks.
Along the same lines, the fourth goal is to build a single AI that can win a wide variety of games. There are already AIs that can win at specific games such as chess and go, but not all of the above.
Google subsidiary DeepMind’s AI learning how to play Breakout for Atari through reinforcement training.
Anticipating Problems
Part of reaching these goals is thinking of the bad things that can happen and solving them first. OpenAI recently listed four special projects whose goal is to make sure artificial intelligence doesn’t hurt humanity.
The first project is to “detect if someone is using a covert breakthrough AI system in the world”. Silicon Valley isn’t the only site of this research, and having an adversary crack the nut first could be disastrous. OpenAI imagines they can figure out if someone has done this by looking at news, financial markets, online games, and more.
The second is to “build an agent to win online programming competitions.” OpenAI is not just looking to top leaderboards – as it succinctly says, a program that can write other programs would be, for obvious reasons, very powerful.
The third project is “cyber-security defense”, because someone will eventually create an AI cyber-security offense and we should fight fire with fire. Fourth is “a complex simulation with many long-lived agents”, where a variety of AI agents will interact with each other to discover language and accomplish a rich variety of goals.
Reddit Power User
The latest addition to the OpenAI offices is a cutting edge computer. Donated by graphics chip manufacturer NVIDIA, its very first DGX-1 was delivered in August. It took 3,000 people working over three years to produce.
Elon Musk and NVIDIA CEO Jen-Hsun Huang with the DGX-1 at OpenAI offices.
The computer’s value is that it brings 170 teraflops of computing power – equal to 1 billion arithmetic calculations per second – to solve AI problems.
What will OpenAI do with the world’s fastest AI system? Make it read Reddit, of course. Researchers are feeding it many years worth of conversations from the content aggregation and discussion site to explore how AI can develop language skills to talk with us.
Safety in Numbers
Ultimately, OpenAI’s greatest contribution may be that it democratizes access to AI. As Musk says, “the best defense against the misuse of AI is to empower as many people as possible to have AI. If everyone has AI powers, then there’s not any one person or a small set of individuals who can have AI superpower.”
Let’s just hope OpenAI’s agents come out a little more well rounded than Microsoft’s Twitter-trained Tay.
Microsoft’s Tay chatbot has bad role models.