The robot Sophia

I was reminded about Sophia in the course of posting reference 1 and was moved to make further enquiries, with the results which follow.

Sophia was built and is operated by Hanson Robotics (reference 2), a Hong Kong based company led by a Texan (reference 3). It is not clear exactly where or how they operate, but the headquarters have moved from Texas to a science park in Hong Kong. And given that, according to an article in Forbes, the company employs less than 50 people, maybe the headquarters is all there is and maybe collaborations with lots of others is the name of the game.

Sophia is the latest, or at least one of the latest, of a line of humanoid robots built by Hanson, quite different from the sort of robots which one finds in factories and warehouses. She is top left in the snap above, lifted from the Hanson website, and Bina48 – the Rothblatt flavoured robot mentioned at reference 1 – is top right. I think it is oldest bottom right and youngest top left.

Sophia has a mobile face with expression. She can see with her eyes, hear with her ears, talk with her mouth and grasp & feel with her hands. She can get about, although rolling is a lot easier than walking. All of this under the control of sophisticated AI engines. A lot of effort has been put into working the arms and hands – using something called inverse kinematics – and to giving the fingers sensation. She makes media appearances and appears to be popular in Saudi Arabia – popular to the point of being made a citizen there – although I don’t suppose she conforms to the official faith.

Follows current theory in these matters to the effect that feelings, emotions and the sense of self are, certainly in the first instance, the product of data about the body, data coming in from the body and its periphery, as opposed to the data which is mostly about the world outside that comes in through the ears and eyes. So Sophia has some of the machinery, even if she does not have the real feelings which come with subjectivity, with consciousness.

Activated in 2016, she seems to have gone a bit quiet since 2020 or so, but she is now the subject of a decentralized autonomous organization (DAO) and there is a merchandising spin-off at reference 5.

I was particularly interested in their measuring of how conscious Sophia was, using a measure based on Tononi’s integrated information theory with its measure Phi (or Φ), which took me to references 6 and then to reference 7, which provides some support. A theory which I first came across just about ten years ago, noticed at reference 8, and in which I have dabbled since. Dabbled being the operative word: I do not understand the technical ramifications and I am not a true believer, this last in the sense that I do not think what we mean by ‘consciousness’ is going to be captured by an information-only theory of this sort. But that is not to say that it is without interest.

Indeed, the authors of reference 6 are agnostic about the wider merits of Φ, and are interested in how it performs as a measure of information integration, as one among a number of more or less elaborate techniques available for quantifying the consciousness of a system, be that system a man or a machine. Important in other contexts, for example, in testing whether someone is in a vegetative state, that is to say near dead, or just locked-in – fortunately, a reasonably rare problem. 

While I don’t pretend to understand the detail, the idea seems to be that if you have a lot of time series coming out of a system and that there is no neat way of dividing that set into two very roughly equal parts, then that system is well integrated, has a high value of Φ. While, contrariwise, if it can be divided into two more or less independent parts, then the value of Φ is zero.

Some of the systems used by Sonia are built on an AGI platform (AGI for artificial general intelligence) from OpenCog of reference 9. The core of this platform is a general purpose store called AtomSpace, introduced at reference 10, where the atoms are the nodes and links of a graph structure. Some of these atoms which might look like the words of a natural language. In some ways, a generalised version of an SQL database, in others a version of XML. Programmers or other systems can talk to an AtomSpace using a language called Atomese. Supporting data, possibly large amounts of data, possibly dynamic, can be stored with those nodes, where they are called values. One of Sophia’s data items is called STI for short term importance, which fluctuates as the atom in question moves into and out of something called the Attentional Focus, a something which is computationally expensive and has to be rationed, rather in the way that consciousness is very selective, with most data not making the cut. These fluctuations gives us the time series which are needed to have a stab at computing Φ.

Conclusions

It will be interesting to see how things go. Will Sonia move forward or will she be traded in for the new model?

As regards Φ, the impression given is that a great deal of machinery has been deployed on what is very much work in progress. I had been hoping for something a bit less complicated! But very much worth a look for all that.

References

Reference 1: https://psmv5.blogspot.com/2023/04/pigs-and-printers.html. 

Reference 2: https://www.hansonrobotics.com/. 

Reference 3: https://en.wikipedia.org/wiki/David_Hanson_(robotics_designer). 

Reference 4: https://en.wikipedia.org/wiki/Sophia_(robot). 

Reference 5: https://www.play-asia.com/. 

Reference 6: Using Tononi Phi to Measure Consciousness of a Cognitive System While Reading and Conversing – Matthew Iklé and others – 2020?

Reference 7: Improved measures of integrated information – Tegmark, M. – 2016.

Reference 8: http://psmv2.blogspot.com/2013/04/phi.html. 

Reference 9: https://opencog.org/. 

Reference 10: https://wiki.opencog.org/w/AtomSpace. 

Reference 11: https://drive.google.com/file/d/14yAkLqXIPckMI4MqLxJ_l4qMP23BeoSe/view. A poster for AAAS 21: another version of the story.