While some of the other books I recently reviewed here warn against the emergence of unaligned super-human artificial intelligence, Erik Larson criticizes such predictions as misplaced and argues that we have in fact no clue how to build an AI because we do not have much of an idea what human intelligence really is. His arguments eloquently address many aspects of AI and he cites many concrete examples of seeming success stories that brought us in fact hardly closer towards general AI, like Deep Blue (Chess), Alpha Go (Go), Watson (Jeopardy!), Goostman (Turing test), Google Translate (language translation). I would like to highlight three main directions of critique: the implausibility of an imminent intelligence explosion, the incompleteness of our understanding of inference, the folly of big data AI.
I.J. Good was the first to describe the intelligence explosion in a 1965 paper on “Speculations Concerning the First Ultraintelligent Machine”:
Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an “intelligence explosion,” and the intelligence of man would be left far behind. Thus the first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control.
(cited from p 33).
Larson argues that we have not only little idea how to build a machine with general intelligence equaling our own but we also have made little progress over the last decades. As he illustrates with many examples all our attempts since the 1950s have ended up in creating narrow AI systems that excel in one narrow task but are clueless in anything else. Deep Blue cannot play Go, Alpha Go cannot play Jeopardy! and Watson cannot play chess.
Even if we could design a machine with general intelligence tomorrow, it would not know how to improve itself, at least not quickly. After all humans have been around for hundred thousand years and we still do not know how to improve ourselves. In Erik Larson’s words,
we have no evidence in the biological world of anything intelligent ever designing a more intellgent version of itself. Humans are intelligent, but in the span of human history we have never constructed more intelligent versions of ourselves.
He cetainly has a point, but I do not think it invalidates the central argument of I.J. Good. Nick Bostrom, Toby Ord and others who have warned about the potentially dire consequence of an intelligence explosion do admit that there is uncertainty about both, when the intelligence explosion will occur and how fast it unravels. Larson is correct that we still do not know how to create a machine with general intelligence and he may be quite right that we still require major innovations, and we do not know when, or if at all, these break throughs will occur. However, he admits himself that break through innovations are by nature unpredictable and there are glaring examples of innovations that have been declared impossible at the eve of their occurrence (e.g. Rutherford declared that the production of energy from splitting of atoms within 20-30 years is “moonshine”, 24 hours before Leo Szillard devised an effective mechanism how to do it.)
So if a general AI is developed, which Larson cannot rule out, would it be in the same difficult position to improve upon its own design as humans have been the last 70 thousand years? Obviously it would be in a much better position because their human designers would provide it with all the knowledge that we have accumulated over the millennia, about the laws of nature, how to build computers and how to design machines with general intelligence. So it could start improving itself on the very first day of its existence. It is hard to predict how fast such self-improving might progress, but given the fact that an AI does not face the same limitations in terms of size and power consumption that constrain human brains, it could advance fast, and with accelerating speed.
While Larson highlights the uncertainties around a possible intelligence explosion, when it might occur and how it develops, he does not invalidate the main reasons for caution that Bostrom and others have put forward.
According to Erik Larson there exist three types of inference: deduction, induction and abduction. Deduction has been formalized since Aristotle and forms the basis of logic reasoning since then. If we know that the street gets wet when it rains, and it is raining now, we know with certainty that the street is wet now. If we know that A implies B and we know that A is true, then we also know with certainty that B is true. Deductive reasoning is formal, general and leads to knowledge that is certain.
Unfortunately, deductive reasoning is useful only in few real-life situations. If we have seen hundred sheep that are all white and make baa, deduction does not tell us what to expect from the next sheep that we encounter. There is no deductive rule that tells us how to deduce anything based on 100 observations. And there cannot be because the next observation may or may not comply with previous observations. The 101st sheep may be white or it may be black. Induction helps here because it extracts regularities and patterns from observation data, which then can be used to predict properties of future observations. Induction is the basic principle behind all modern machine learning algorithms. A typical machine learning algorithm is trained with a large number of pictures depicting cats and dogs, that are used to extract common features in dog images and cat images. When the algorithm is confronted with a new image, it will with fairly high accuracy identify dogs and cats. These kind of algorithms have been phenomenally successful in thousands of applications, and are now central to autonomous driving, cyber security protection, advertisement in internet services like search engines and social networks, and many more applications that we all use on a daily basis.
Their overwhelming success notwithstanding, induction based algorithms do not understand their observations on a deeper level. They cannot predict future events based on an understanding of causal dependencies. In a certain sense, they are like Pavlov’s dogs. If they have observed bell ringing followed by the serving of lunch, they will predict the appearance of lunch upon hearing the sound of a bell. They would not understand that the lunch does not causally depend on the bell ringing, and they would not become suspicious about this observed connection, and they would not muse about this curious coincidence.
Abduction helps here because it is about building of hypotheses that explain observations. Larson argues that we do not have a theory about abduction and we do not know how to build good hypotheses. This missing element in our understanding of inference is a deplorable omission and a major reason why current attempts at AI will not lead to general intelligence. They will only result in narrow AI, custom tools for solving specific well defined problems, like distinguishing the image of a cat from the image of a dog. Human thinking is all about hypothesis building. In Larson’s words,
When we seek to understand particular facts … we are inevitably forced into a kind of conjuring, the selection or invention of a hypothesis that might explain the fact.
Larson considers it central to what it means to be intelligent, and he traces this observation back to the the American mathematician C.S. Peirce.
The origin of intelligence, then, is conjectural or abductive, and of paramount importance is having a powerful conceptual framework within which to view facts or data. Once an intelligent agent (person or machine) generates a conjecture, Peirce explains, downstream inference like deduction and induction make clear the implications of the conjecture (deduction) and provide a means of testing it against experience (induction).
Indeed, Charles Sanders Peirce was a very influential American philosopher, logician and mathematician at the end o the nineteenth century. For instance Bertrand Russell wrote “Beyond doubt […] he was one of the most original minds of the later nineteenth century and certainly the greatest American thinker ever”. (Russell, Bertrand (1959), Wisdom of the West, p. 276). In his 1887 paper Logical Machines, Peirce wrote
Every reasoning machine […] has two inherent impotences. In the first place, it is destute of al originality, of all initiative. It cannot find its own problems; […] This however, is no defect in a machine; we do not want it to do its own business, but ours.
(From Logical Machines, p 233)
130 years later this impotence is still a feature of all machines we have built, even those that are named AI. But Peirce considered this defect to be a feature, not a bug, a position that has recently been expressed by Stuart Russel (Human Compatible, Stuart Russel, 2019), among others.
So even though abductive reasoning, with its foundations laid by Peirce and a recent renewed interest in the 1990s, has a long history, we still lack a formal framework of abduction that could serve as the basis to include it in artificial intelligent agents.
With the growth of the internet and ever increasing computing power, large data sets have become available. Billions of people interact online providing endless sources of text, images, video, and audio data. Countless organizations make large amounts of data systematically and easily accessible to many what previously was cumbersome to retrieve by few. The Canadian state offers many official documents in both English and French, which makes them a welcome training resource for automatic language translation. The EU does the same for many European languages. Worldwide researchers distribute data sets of images, videos, biomedical signals (EEGs, ECGs, …), environmental data (climate, maritime observations, fauna and flora, …) to their peers and the general public that have been collected and augmented with great efforts. All that data gave rise to the era of Big Data Analysis, and modern AI has greatly benefited.
In particular modern machine learning based on Deep Neural Networks require huge amount of training data. Since 2012, when a DNN called Alex Net blew away the competitors in a competition on object identification in images, called the Imagenet competetion, this approach has led to a tsunami of research in DNNs and applications of DNNs, ranging from games, autonomous driving, language translation, medical diagnostics, predictive maintenance, financial trading, robotics, social media, etc. etc.
While the success of DNNs have been truly amazing, not a single DNN has developed a deeper understanding of its data, as Erik Larson correctly argues. He describes many telling examples from the domain of language translation, a subject he has worked on personally for many years. The use of human languages is always context dependent and refers somehow to the world outside the text. But even the most sophisticated language translation systems only deal with the sequence of characters without any understanding of the world the text refers to. This gives rise to countless subtle deficiencies in translated text. One set of problems arises from words with multiple meanings, and the usage of those meanings occurs at very different frequencies. E.g. “pen” in English usually denotes a writing device, but sometimes it refers to the cage of an animal. If you ask Google Translate for an English - German translation, this is what you get:
|The box is in the pen||->||Die Box ist im Stift|
This translation is unlikely to be correct because boxes tend to be bigger than writing pens. Considering the real world objects a better translation would be “Die Box ist im Gehege”, but Google Translate does not propose this solution because in the overwhelming majority of occurrences in its training texts, pen is better translated to Stift.
Google Translate is not unaware of this dual meaning, because:
|The dog carries the box into its pen||->||Der Hund trägt die Kiste in seinen Pferch|
So Google Translate knows that sometimes pen translates to Stift and sometimes it translates to Gehege (or Pferch). But it cannot use the knowledge of relative sizes of boxes and pens because it lacks that knowledge. Google Translate is also surprisingly unaware of the context, even if that context is strikingly obvious. For example
|The dog carries the box into its pen. Now the box is in the pen.||->||Der Hund trägt die Kiste in seinen Pferch. Jetzt ist die Box im Stift.|
is positively weird.
Erik Larson convincingly argues that big data analysis, as it is practiced today in machine learning and AI is overrated because of its exclusive focus on the syntax of the data, overlooking the semantics and pragmatics. Because current algorithms do not understand the meaning of the data in the real world, they are ultimately bound to fail.
In summary, Erik Larson offers an insightful, enlightening analysis of the shortcomings of current AI, even though it does not invalidate the warnings of Bostrom, Ord, Tegmark and others that are worried about the emergence of true artificial general superintelligence.
(AJ January 2022)