A Walking Skeleton in a Beautiful Face: Ben Goertzel on What Sophia Really Is

Goertzel begins by saying that to understand Sophia you first have to understand David Hanson, whom he calls the brilliant mind behind the robot and still more interesting than his creation. They met around 2007 at a futurist conference where both were speaking, and quickly saw how much they shared: a passion for AGI and the singularity, and a love of the same science fiction.

Both wanted benevolent AGI that would improve life for all humans and all sentient beings, but they differed on method. Goertzel, who describes himself as very math-oriented, thinks in terms of the abstract cognitive algorithm that could internalize the patterns of human values. Hanson wanted "computational compassion": machines that look people in the face, are loved by them, and love them back.

Goertzel's path then took him to Hong Kong in 2011, partly because he fell in love with a woman he later married. There he worked on machine learning for stock and futures prediction and discovered the hardware ecosystem across the border in Shenzhen, which he calls astounding. He introduced Hanson to local investors, who helped fund the move of Hanson Robotics to Hong Kong.

At first Goertzel was too busy with AGI research and trading work to get closely involved, but as the two spent time together he began thinking about how to make Hanson's robots smarter. For a few years he served as chief scientist and head of software at the company. He later stepped back as he moved deeper into blockchain and co-founded SingularityNET, with Hanson as one of the co-founders.

The original idea was that many copies of the Sophia robot would be user interfaces to the globally distributed SingularityNET cloud mind. Goertzel says Hanson always saw Sophia as a platform with well-designed APIs, much like SoftBank's Pepper, on which anyone could try their own AI algorithms. SingularityNET, being an API marketplace, fits naturally into that picture, whereas OpenCog is more Goertzel's own project than Hanson's.

For Goertzel, Sophia was also a way to put AI algorithms in front of many people in an emotionally compelling way, with a link to AGI ethics. He says he is an optimist by nature but must intellectually assign a nonzero probability to nasty outcomes, because no one can know what something ten times smarter than them will do. Just as his dog cannot predict what he will do tomorrow, there is irreducible uncertainty.

His proposed response is to bias early AGI toward benevolence by infusing it with love and compassion, as we do with our children. If AIs interact with people in the context of compassionate and beneficial tasks, they learn by doing and absorb an intuitive sense of what those words mean, which they can later abstract in their own way. Hanson saw it the same way, hence the name Sophia, meaning wisdom.

The grand vision was to roll out early-stage AGI as loving robots in home service, education and nursing, where they would learn human values and ethics from people rather than only facts from databases. Goertzel then contrasts the vision with the reality of robotics, citing the first principle that the robot is always broken. He recalls soldering robots together himself in the 1990s while experimenting with neural networks and reinforcement learning.

Faced with making Sophia actually work, Goertzel says, many of his robotic AGI aspirations were set aside. Computer vision suffers from glare, and the speech-to-text systems of the time could not handle Hong Kong people's English accents, so the robot often received the wrong text and sounded stupid. This led him to think of the system as a "walking skeleton", a software engineering approach.

In a walking skeleton you build a simple version of every part the whole system will need, and then improve each part in the context of the integrated whole. Sophia has something that sees, hears, moves, remembers and learns, each in a simple form and all connected. Goertzel is clear that there is a lot of AI in there but not any AGI: face recognition, detection of people entering or leaving, and a dialog system that mixes hand-coded rules with deep neural nets.

The dialog system also tries to give conversations a narrative structure with a beginning, middle and end. Lex Fridman notes that the systems that beat the Turing test in contests like the Loebner Prize are heavily rule-based, because neural networks currently struggle with story arcs, even something like Google's Meena in full-scale conversations. Goertzel agrees and describes an experiment of his own.

For some weeks he had Sophia's chatbot talk to Facebook's open-sourced transformer chatbot, generating training data on the server. The Facebook bot produced a wider variety of fluent sentences but rambled, while Sophia's was more repetitive yet kept a conversation arc much longer. He thinks architectures like the Reformer might close the gap, but neither bot understands what is going on.

If he were doing a robotics project aimed purely at AGI, he says, he would build a robo-toddler learning about what it sees, so that language is grounded in experience. Sophia must instead talk about sports, weather or whatever conference she is at, with no grounding for any of it. That requirement pushes short-term work toward heuristic, hard-coded, Watson-like solutions and away from AGI.

Fridman notes that Sophia is partly an art creation: Hanson is a great sculptor, and people project intelligence onto a beautiful face even if nothing is going on inside, hence the long selfie lines. Many people attribute AGI-level capabilities to her, and researchers such as Yann LeCun get frustrated and criticize those, like Goertzel, who stay quiet and let the world stay captivated.

Goertzel concedes that if he were not involved with Hanson Robotics he would probably have been annoyed too, since people trying to raise money for real AGI could be hurt if the public thinks it already exists. But once he dug into Hanson's intentions, he says, he stopped being angry, whereas LeCun remained so.

He finds this ironic because LeCun works for Facebook, which in his words uses machine learning to program people's brains toward vapid consumerism and political extremism. If one's ethics permit that, he asks, why would they forbid a lovable theatrical robot that draws some people into its illusion? He adds that criticism from Yoshua Bengio would have humbled him more, since Bengio is not using AI for blatant evil and does not go around trashing others' work.

Goertzel also says he has not agreed with everything Hanson has said or done about Sophia. He calls Hanson an artistic wild man whose off-the-cuff style is part of his genius. Goertzel pushed the company to be more open about how Sophia works, and Hanson did not oppose that.

Hanson's response to calls for openness, Goertzel recalls, was that you can tell people exactly how it works and they will not care, because they want to be drawn into the illusion. He saw it confirmed at an event in Austin, Texas, with the Philip K. Dick robot, operated by a friend typing from a back room. Afterward, shown the operator, participants still insisted they had talked to Phil, as if the writer's spirit animated the typist's mind.

Goertzel half-seriously says a small part of him thinks they may be right, given how little we understand the universe, and he mentions a cosmic mind field. More concretely, he says our imagination about Sophia, and LeCun's frustration with it, are part of a beautiful dance of creating AI. He draws a parallel with Boston Dynamics' Spot, which is far from intelligent yet makes us feel bad when it is kicked and falls.

On transparency he points to what he has published: an article describing three software systems used in Sophia. These are a timeline editor, a rule-based authoring tool where humans script the output; a chatbot with rule-based and neural parts; and sometimes OpenCog, with more learning and reasoning. He admits he often cannot tell which one is running unless he interacts for several minutes.

He sees two ethical questions: whether to be transparent, which he says they have been, and whether to keep fooling people who want to be fooled once transparency fails to change their beliefs. He notes that the whole media industry fools people the way they want to be fooled, and argues Hanson Robotics does so toward a good end, using empathy to deliver a good user experience.

The good end he has in mind is medical and care robotics, where he has been discussing collaborations with Hanson Robotics for elder care, work with children and autism therapy. In the Covid era, a nursing-assistant robot does not spread infection and can give more attention than human nurses can. If a sick patient attributes more understanding to a human-looking, loving robot than it really has, he asks, is that really bad, and says the key is how it is used.