Equipped with one of the the most brilliant brains in human history, Albert Einstein was nonetheless notoriously absent-minded, and struggled with simple tasks like finding his way home. We can draw a parallel with today’s “narrow artificial intelligence” like AlphaGo — which can beat top human Go players, but remains woefully unable to distinguish, for example, a fish from a bicycle.
This is not what AI should be, according to Marek Rosa. The 38 year-old Czech-born tech entrepreneur is a big fan of science fiction movies like Gravity, Interstellar, and his favorite: Chappie — a 2015 film centred on an AI robot which is able to learn and think.
Rosa’s dream is to create a machine that improves itself recursively to achieve an intelligence level humans cannot imagine, allowing it to tackle problems of any scale. In contrast to narrow AI, General AI should know how to reason, judge, communicate with humans in natural language, and integrate problem-solving skills toward common goals.
To make it happen, Rosa needs to do two things: first, teach AI how to learn; and then teach AI the skills in demand.
“Children are not expected to gain knowledge on their own, they attend school in order to learn new things and build on their existing skills. In a similar manner, we expect our General AI to accumulate skills in a step-by-step self-improving fashion,” says Rosa.
In July 2015, Rosa founded the R&D startup GoodAI in Prague, Czech Republic. The company now has an international team of 20 research scientists, with CEO/CTO Rosa at the helm. GoodAI’s goal is to develop General AI as fast as possible, in order to help humanity and potentially unlock the secrets of the universe.
Rosa took US$10 million from his own pocket to fund GoodAI, and plans to hand another $5 million to the winner of his General AI Challenge — a global competition that encourages participants to tackle multiple crucial research problems in human-level AI development. More than 500 people have joined the first round, and the winner will be announced this fall.
Rosa owes his fortune to remarkable success in video game programming. He is the CEO and founder of Keen Software House, an independent game development studio best known for their best-seller Space Engineers (SE), which has sold over two million copies.
Space Engineers is a sandbox game about engineering, constructing, and maintaining creations in space.
Space Engineers is similar to Minecraft, wherein players can construct gameplay elements using rudimentary building blocks. The difference is that Space Engineers sets the universe as the background and adds features such as programming.
Says Rosa, “Twenty years ago, AI was still in the realm of science fiction, so I started with game development. I wanted to build great games, build up the skills I need for AI research, and earn enough money to be able to launch my own AI company and work in it full-time.” While Rosa did not suggest any working connection between SE and GoodAI, SE is considered to be a good all-purpose training arena for General AI.
At present, GoodAI’s research team is making progress on gradual learning — the ability of an AI agent to acquire new skills and knowledge incrementally, to reuse learned skills in order to learn new skills more efficiently, to recursively self-improve itself, and more. This is considered the foremost challenge at this stage.
In a table titled “High-level overview of milestones for the development and education of General AI” there are five stages, spanning from stage 0, where AI has no hardcoded skills, to stage 4 — the super-human-level general purpose AI. The GoodAI research team is currently navigating between stage 0 and stage 1.
Rosa is building a virtual school for his futuristic General AI “students,” where an optimized set of learning tasks will be designed to teach the AI useful skills and abilities. Researchers will apply the AI’s performance in the curriculum’s learning tasks as feedback, in order to improve both the curriculum and the hard-coded AI skills. The communication will be based on reward and punishment, which are “rudimentary signals on how to steer the agent’s behavior,” according to Rosa.
The agent can move to its next task only after it solves the previous task with an acceptable performance. The ultimate goal of the school for AI is to teach the agent how to learn better on additional tasks.
The GoodAI team is also creating an entity for overall planning — the AI Roadmap Institute — which will collect, compare and study different pathways to General AI. This is expected to encourage discussion and progress within the General AI research community.
Building a General AI is a long-term process, and Rosa believes any short-term narrow AI design will not help in pursuit of the goal. At GoodAI, narrow AI research has been split into a separate entity called GoodAI Applied. “Instead of designing AI that can solve specific and narrow problems, we want to invest our effort into AI that can design other AI, which may be specialized for narrower domains. So instead of just building AI, we want to build meta-AI.”
While that may seem ambitious enough, Rosa envisions even more. Supported by strong computing power and human cognitive abilities, General AI is expected to deliver advanced skillsets, allowing it to break through the limits of human abilities. “Problems such as understanding the universe or analyzing all the underlying interactions in a human body are nearly impossible to solve with the human brain’s capacity,” says Rosa.
Rosa stressed this does not mean his advanced AI will exceed human control. General AI is considered a perfect self-improving tool to augment human abilities and capacities, not to override human value systems. Think of it like a car, which helps humans move faster, but does not conflict with anyone’s will or natural behaviour.
GoodAI is confident that with help of General AI, humans might potentially study and understand the universe like Einstein did. This is unlikely to be realized in the immediate future — but when it is, it could emerge as the last thing humans ever need to invent.
Author: Tony Peng | Editor: Michael Sarazen