The following are the notes on the talk I gave with the same title, during the 2021 summer edition of “Tech Conference Europe”. The ideas presented are mostly based on a 2021 book entitled “Framers” by Kenneth Cukier, Viktor Mayer-Schonberger, and Francis de Vericourt.
“AI might show us that the human qualities we most value are disappointingly simple to mechanize”.Melanie Mitchell paraphrasing Douglas Hofstadter (“Artificial Intelligence A Guide For Thinking Humans”)
“Technology learns to do more of what we do better than we can do it”ohn Green (“The Anthropocene Reviewed”)
Insights into how the human mind works
“Halicin”, a new chemical compound, can be used to destroy antibiotic resistant bacteria and it was created by AI. However, it was a human, Moldavian-born dr. Regina Barzilay, who was able to “frame” the entire problem in a different way and then apply a neural network to her hypothesis.
An AI system is an idiot savant. It can do a single task very well, given very specific circumstances. But it will continue to play chess even if the building is on fire. By contrast, humans think using very complex mental models, manipulation of concepts, abstractions, and framing counterfactuals. The human mind is the only thing capable of envisioning something that isn’t there. Even more, our cognitive abilities are capable of reframing problems and applying different strategies, which ML algorithms are incapable of doing.
When the human mind works with symbols, it assigns them meaning using incredibly complex semantics based on the real world and intersubjective experience of communities of people. When computers manipulate symbols, they simply refer to operations and no ‘meaning’.
Above all these there is also an added layer of emotional living, making all of us aware that not everything can be reduced to logical formulas. A note should be made that while much of the cognitive science literature also focuses on this, there are thinkers such as Steven Pinker, Yuval Harari or Sam Harris who believe strictly in the “computational theory of the mind” and that all of intelligence can be reduced to some sort of information processing. However, the consensus within the psychology literature is that intelligence is multi-dimentional and emotional intelligence is highly important.
AI traits and achievements in a larger perspective
For many questions, algorithms can give much better answers than humans. We must also consider that, by using their imagination and creativity, humans can pose never-before-asked questions. Appreciating and applying the lessons learnt by AI is something only we can do.
AI is incredible, thanks to advances in processing power, when it comes to pattern recognition in very large amounts of data. Humans use pattern recognition to be able to generalize from the specific, but algorithms do this in an entirely inefficient way: you don’t show a child a thousand pictures of cats until it can reliably identify the animal. And what happens when we have scarce data on a particular problem? (such as modeling of rare events)
Our technology has become much too complex for it to have been the product of the few geniuses mentioned in science books. We understand now, with the advantage of hindsight, that it was created by the aggregation of generations of people who were able to communicate (even, or especially, after their own deaths). Most of AI systems today work in isolation.
Will AI be able to create better and better algorithms by itself? In the current state of affairs: no. The example of a system of deep reinforcement learning is perfect in this case. We are mesmerized by how better algorithms can get at chess, Go or Dota 2, however in reinforcement learning (which is how these types of AI were trained), what constitutes a “reward” is defined explicitly by humans, and is not learnt by the system itself: human ex machina.
Music composed by AI can now no longer be distinguished from that of major composers. It’s true. But the algorithms calculate that music, they don’t imagine it. They are trained on the structure of melodic lines and rich harmonies of human composers and then randomly remove notes and try to guess (based on statistical distributions) what would best fit next.
“Not until a machine can write a sonnet or compose a concerto because of thoughts and emotions felt and not by the chance fall of symbols could we agree that machine equals brain. That is, not only write it but know that it had written it. No mechanism could feel (and not merely artificially signal, an easy contrivance) pleasure at its success, grief when its valves fuse, be warmed by flattery, be made miserable by its mistakes, be charmed by sex, and be angry or depressed when it cannot get what it wants.”Geoffrey Jefferson
Other aspects to consider
Experts have no insights into General Artificial Intelligence.
According to Francois Chollet, humans have “an ability to adapt to novel, never-before-experienced situations using little data or even no new data at all” and that “is the defining characteristic of human cognition”.
Brute force training versus ability to generalize by working with concepts leads us into the final point: difference of degree and difference of nature. No matter how much better artificial intelligence gets, it will never be human intelligence.