I just read a very interesting true story about an android built in the early 2000s based on the renowned sci-fi author, Philip K Dick, both in personality and physical appearance. It was displayed in public at a few prominent events where it interacted with the public in 2005, then was lost on a flight between Dallas and Las Vegas in 2006, and has never been seen since. The book is called Lost In Transit; The Strange Story of the Philip K Dick Android by David F Duffy.
You have to read the back cover to know it’s non-fiction published by Melbourne University Press in 2011, so surprisingly a local publication. I bought it from my local bookstore at a 30% discount price as they were closing down for good. They were planning to close by Good Friday but the COVID-19 pandemic forced them to close a good 2 weeks earlier and I acquired it at the 11th hour, looking for anything I might find interesting.
To quote the back cover:
David F Duffy was a postdoctoral fellow at the University of Memphis at the time the android was being developed... David completed a psychology degree with honours at the University of Newcastle [Australia] and a PhD in psychology at Macquarie University, before his fellowship at the University of Memphis, Tennessee. He returned to Australia in 2007 and lives in Canberra with his wife and son.
The book is written chronologically and is based on extensive interviews with the team of scientists involved, as well as Duffy’s own personal interaction with the android. He had an insider’s perspective as a cognitive psychologist who had access to members of the team while the project was active. Like everyone else involved, he is a bit of a sci-fi nerd with a particular affinity and knowledge of the works of Philip K Dick.
My specific interest is in the technical development of the android and how its creators attempted to simulate human intelligence. As a cognitive psychologist, with professionally respected access to the team, Duffy is well placed to provide some esoteric knowledge to an interested bystander like myself.
There were effectively 2 people responsible (or 2 team leaders), David Hanson and Andrew Olney, who were brought together by Professor Art Greasser, head of the Institute of Intelligent Systems, a research lab in the psychology building at the University of Memphis (hence the connection with the author).
Hanson is actually an artist, and his specialty was building ‘heads’ with humanlike features and humanlike abilities to express facial emotions. His heads included mini-motors that pulled on a ‘skin’, which could mimic a range of facial movements, including talking.
Olney developed the ‘brains’ of the android that actually resided on a laptop and was connected by wires going into the back of the android’s head. Hanson’s objective was to make an android head that was so humanlike that people would interact with it on an emotional and intellectual level. For him, the goal was to achieve ‘empathy’. He had made at least 2 heads before the Philip K Dick project.
Even though the project got the ‘blessing’ of Dick’s daughters, Laura and Isa, and access to an inordinate amount of material, including transcripts of extensive interviews, they had mixed feelings about the end result, and, tellingly, they were ‘relieved’ when the head disappeared. It suggests that it’s not the way they wanted him to be remembered.
In a chapter called Life Inside a Laptop, Duffy gives a potted history of AI, specifically in relation to the Turing test, which challenges someone to distinguish an AI from a human. He also explains the 3 levels of processing that were used to create the android’s ‘brain’. The first level was what Olney called ‘canned’ answers, which were pre-recorded answers to obvious questions and interactions, like ‘Hi’, ‘What’s your name?’, ‘What are you?’ and so on. Another level was ‘Latent Semantic Analysis’ (LSA), which was originally developed in a lab in Colorado, with close ties to Graesser’s lab in Memphis, and was the basis of Grasser’s pet project, ‘AutoTutor’ with Olney as its ‘chief programmer’. AutoTutor was an AI designed to answer technical questions as a ‘tutor’ for students in subjects like physics.
To create the Philip K Dick database, Olney downloaded all of Dick’s opus, plus a vast collection of transcribed interviews from later in his life. The Author conjectures that ‘There is probably more dialogue in print of interviews with Philip K Dick than any other person, alive or dead.’
The third layer ‘broke the input (the interlocutor’s side of the dialogue) into sections and looked for fragments in the dialogue database that seemed relevant’ (to paraphrase Duffy). Duffy gives a cursory explanation of how LSA works – a mathematical matrix using vector algebra – that’s probably a little too esoteric for the content of this post.
In practice, this search and synthesise approach could create a self-referencing loop, where the android would endlessly riff on a subject, going off on tangents, that sounded cogent but never stopped. To overcome this, Olney developed a ‘kill switch’ that removed the ‘buffer’ he could see building up on his laptop. At one display at ComicCon (July 2005) as part of the promotion for A Scanner Darkly (a rotoscope movie by Richard Linklater, starring Keanu Reeves), Hanson had to present the android without Olney, and he couldn’t get the kill switch to work, so Hanson stopped the audio with the mouth still working and asked for the next question. The android simply continued with its monolithic monologue which had no relevance to any question at all. I think it was its last public appearance before it was lost. Dick’s daughters, Laura and Isa, were in the audience and they were not impressed.
It’s a very informative and insightful book, presented like a documentary without video, capturing a very quirky, unique and intellectually curious project. There is a lot of discussion about whether we can produce an AI that can truly mimic human intelligence. For me, the pertinent word in that phrase is ‘mimic’, because I believe that’s the best we can do, as opposed to having an AI that actually ‘thinks’ like a human.
In many parts of the book, Duffy compares what Graesser’s team is trying to do with LSA with how we learn language as children, where we create a memory store of words, phrases and stock responses, based on our interaction with others and the world at large. It’s a personal prejudice of mine, but I think that words and phrases have a ‘meaning’ to us that an AI can never capture.
I’ve contended before that language for humans is like ‘software’ in that it is ‘downloaded’ from generation to generation. I believe that this is unique to the human species and it goes further than communication, which is its obvious genesis. It’s what we literally think in. The human brain can connect and manipulate concepts in all sorts of contexts that go far beyond the simple need to tell someone what they want them to do in a given situation, or ask what they did with their time the day before or last year or whenever. We can relate concepts that have a spiritual connection or are mathematical or are stories. In other words, we can converse in topics that relate not just to physical objects, but are products of pure imagination.
Any android follows a set of algorithms that are designed to respond to human generated dialogue, but, despite appearances, the android has no idea what it’s talking about. Some of the sample dialogue that Duffy presented in his book, drifted into gibberish as far as I could tell, and that didn’t surprise me.
I’ve explored the idea of a very advanced AI in my own fiction, where ‘he’ became a prominent character in the narrative. But he (yes, I gave him a gender) was often restrained by rules. He can converse on virtually any topic because he has a Google-like database and he makes logical sense of someone’s vocalisations. If they are not logical, he’s quick to point it out. I play cognitive games with him and his main interlocutor because they have a symbiotic relationship. They spend so much time together that they develop a psychological interdependence that’s central to the narrative. It’s fiction, but even in my fiction I see a subtle difference: he thinks and talks so well, he almost passes for human, but he is a piece of software that can make logical deductions based on inputs and past experiences. Of course, we do that as well, and we do it so well it separates us from other species. But we also have empathy, not only with other humans, but other species. Even in my fiction, the AI doesn’t display empathy, though he’s been programmed to be ‘loyal’.
Duffy also talks about the ‘uncanny valley’, which I’ve discussed before. Apparently, Hanson believed it was a ‘myth’ and that there was no scientific data to support it. Duffy appears to agree. But according to a New Scientist article I read in Jan 2013 (by Joe Kloc, a New York correspondent), MRI studies tell another story. Neuroscientists believe the symptom is real and is caused by a cognitive dissonance between 3 types of empathy: cognitive, motor and emotional. Apparently, it’s emotional empathy that breaks the spell of suspended disbelief.
Hanson claims that he never saw evidence of the ‘uncanny valley’ with any of his androids. On YouTube you can watch a celebrity android called Sophie and I didn’t see any evidence of the phenomenon with her either. But I think the reason is that none of these androids appear human enough to evoke the response. The uncanny valley is a sense of unease and disassociation we would feel because it’s unnatural; similar to seeing a ghost - a human in all respects except actually being flesh and blood.
I expect, as androids, like the Philip K Dick simulation and Sophie, become more commonplace, the sense of ‘unnaturalness’ would dissipate - a natural consequence of habituation. Androids in movies don’t have this effect, but then a story is a medium of suspended disbelief already.
No comments:
Post a Comment