I’ve started listening to Max Tegmark’s book from 2017 about the Age of Artificial Intelligence. The world has changed quite a lot since 2017, especially the world of AI. But Tegmark’s books is still one of the best, and in my taste, most readable books on AI.
For me, as a philosopher, I’m primarily interested in the underlying assumptions, or, you could say, I am interested in what is not said in the text and yet operative and determinative for the content. Which is why I rarely write about technicalities, such as how someone got some descriptions wrong and how something is slightly inaccurate. I think what matters more and what’s also more interesting are these underlying foundations of what somebody is saying. In a book on AI, these will of course concern the core concepts of AI Research, – in Tegmark’s book primarily: intelligence and information.
Information and Intelligence, The Underlying Basic Assumptions of Tegmark’s Book
Tegmark is giving a definition of these terms: “intelligence = ability to accomplish complex goals”.
This concept or definition of “intelligence” rests on Tegmark’s understanding of what information is, and, as we will see, on the assumption, that information is some kind of underlying fundamental reality. The connection is established with the following questions (p.30): “The conventional wisdom among artificial researchers is that intelligence is ultimately all about information and computation […] But what are information really […]? How can something intangible and ehereal as information and computation be embodied by tangible physical stuff?”
What Tegmark calls a conventional wisdom, we could say, is a taken-for-granted and unquestioned assumption about the nature of intelligence. Tegmark is telling us that intelligence is all about information processing and computation. If that were the case, the conclusion that computers, robots, machines – which are all about computation – will one day be as intelligent or more intelligent obviously follows from this basic assumption.
Assumption: Intelligence = Computation
Conclusion: Things that can compute = Intelligent
Tegmark therefore has to give us a prove that intelligence is all about information processing. And he does that indirectly by demonstrating that the world itself and all things in it, consist of nothing else than information.
Notice how Tegmark is challenging the metaphysical assumption about reality first: “physics has taught us … everything is matter” ~ his questions indicate that he will propose that this might turn out to be wrong; and that there is something that – even though it behaves according to the laws of physics – is not at all “matter” in the traditional sense, or in Tegmark’s words: is “substrate independent”.
As I will show in the following, information thereby becomes something like a “more fundamental reality”; this we could call the “informationalist turn”, because it is a common pattern with many of the leading AI researchers, or in technology and science in general. Tegmark mentions that he tried to prove that basic assumption in one of his previous books but does not further elaborate (and I haven’t read his previous books). But bottom line, what Tegmark says about information must be characterized as a metaphysical or ontological assumption about reality that takes information to be the primary building block of the universe. But not in a simple materialist sense according to which the “atom” as the smallest unit is now replaced by an even smaller unit called “information”.
The informationalists’ move is more sophisticated: it assumes that information (in the form of data) insofar as we can compute it, can substitute the objects. Or the other way around: that objects are “in reality” information, bits, data.
Tegmark’s first description of information takes it to be a relation between propositions (written in a book) and states in the world (p. 31):
This is the common sense and traditional understanding of how knowledge, truth, etc. works. Propositions are the containers of “information”, in this case. What’s important is, however, that the essential property of “information” is that it is computatable because it can be translated, or it basically consists of “bits”:
Tegmark operates on the background assumption that everything can be reduced to or translated to these atoms of information = bits. What he calls “substrate independence” is more like a “substrate indifferentness”: bits still have to take on a physical form (hence they are subordinate to the laws of physics), but they are indifferent regarding what form that might be. The philosopher G. Günther already in book on Cybernetics (1957) interprets this as an initial step by which informationalists aim to overcome the traditional ontology and its dependence on matter: in this case, humans are not restricted by “nature” and by what is given to us, but they can “programme” their own reality because they are able speakers of the language of information. In that case, computing is understood as a “construction of reality”; and the examples for that would be what is nowadays called “simulation” or “virtual reality.” Notice the subtle implications of these basic assumptions of the “power of information” when reading what Tegmark himself writes:
Tegmark takes this aspect to be an indication that information is something special, special in the sense of: of higher dignity, higher importance. And we will come back to this later. But what Tegmark calls independence is important since it can function as a challenge to the materialists’ ontology of matter. “Independence” sounds as if information does not need matter. The word evokes the meaning that “information” is a distinct class that is not further “grounded” in matter.
—
Also note that the justification for this new kind of ontology that in my opinion belongs to the new paradigm of informationalism is functionality in terms of translatability into information. The fact that things and processes in the world can be translated or reduced or respresented by information in the form of bits is taken as a prove of the ontological claim that information underlies reality (qua universal building block of the universe).
Two Problems with the Information Paradigm
There are at least two problems with this approach: 1. it does not question the translatability and takes it to be universal, assuming that the representation of X in terms of information (for example, a model of X) adequately represents what X is. If the collected information about X is sufficient to predict the behaviour of X, this information can then replace the original X.
From the Translation of X into Information,
to the Computation of these Information
to the Replacement of X.
2. The second problem follows from this process of substitution but not with necessity. Because now what has taken the place of the original X, the computable information of X, is taken to be the ultimate reality of X. This is an “ontologization” that follows after the translation, computation, and replacement of X. The claim is that “information of X” is what X really is.
This ontologization has many drastic implications that can generally described with drifting away from first-hand experience (our lifeworld) to the world of computable information. This drift could be described with Husserl as an substraction, but as the underlying ultimate reality the calculations of mathematical physics are now replaced by the stochastics of data science.
One of the main problems could be described as the “tyranny of the statistical,” taking into account the limitations of the translation process (problem 1): The problem is that the representation of X through information is considered to be an adequate representation based on the functionality of the models; and that this form of representation is considered to be universally applicable.
One of the many problems that follow from this can be gathered from a anecdote Gadamer liked to tell, namely that it can happen that the statistics used for the evaluation of the health of a person can make a healthy person sick, if the healthy person does not display what the computer takes as the correct parameters for “health.” The point is: The data gathered from 1.000.000.000 persons still does not justify universal applicability. In medicine as well as in jurisprudence, human judgement cannot and should not be replaced by computers. The application of (universal) rules requires more than machines can do, since it is neither following-rules, not translatable into machine-learning based on stochastic gradient descent.
Memory, Computation, Information
But back to Tegmark to prove that these problems are actually inherent in his book and in his understanding of “intelligence.” Tegmark is trying to demonstrate the translatability of everything into information in small subchapters called: memory and computation. He is referring to examples of how computation works, to the computronium; he is referring to S. Wolfram and A. Turing, – their work is taken as prove for the claim that everything is translatable into information (that can then be computed) (p. 30-40).
I’d argue that what these examples prove is up for debate. And that this would lead into a discussion of functionalism vs. … let’s say … non-functionalism. The assumption is always that if I can build something that functions in the same way as X, then what I have built is identical to X. X is identified with this function. What I would not deny is that was has been built is actually working, functioning, — nor would I deny that the function of what has been built is resembling the function of the original; but the next step, which claims that what has been built is identical to X … this is what is debatable, in my opinion.
And this functionalism becomes super tricky if we move to higher organisms such as animals (or the brains of animals). There is an essential distinction between a human moving his finger and a robot moving his robot-finger. Even if the movement is identical in all its intricate details (the “information” can be identical), the “function” of moving a finger in the human context is determined holistically by what/who/how the person is. The external outside perspective and the translation of the movement of the finger into information that can be processed and computed and then copied by the robot is never an adequate representation of this movement.
Another thing that I found curious is Tegmark’s analogy between the “substrate independence” of computation (by which he wants to support the claim of the universal applicability of information+computation) and “beautiful examples in physics” such as waves (soundwaves). Tegmark’s point is that we can study waves independent to the particular substrate in which they are manifested. But is that a correct and just analogy? the analogy between information and waves? After all, waves and the property of waves are still observable facts. They are “real” in a way that information is not. Tegmark is taking an example from the world of physics and from within the context and paradigm of mathematical physics and conflates it with the example of the computation of information. Tegmark wants to deduce three important points from this analogy (37), but he is only using metaphors to make sense of this analogy; I think it shows an attempt to show nonexisting similarities between the old and already established paradigm of mathematics and physics and the new paradigm of informationalism. I think it would be more helpful to point out the differences than to try to blur the lines by using metaphors.
The distinction is then pointed out neatly by Tegmark on p. 38:
“computation is a pattern in the spacetime arrangement of particles, and it’s not the particles but the pattern that really matters! Matter doesn’t matter.
In other words, the hardware is the matter and the software is the pattern. This substrate independence of computation implies that AI is possible: intelligence doesn’t require flesh, blood or carbon atoms.”
What “substrate independence” means in the case of waves and then in the case of information seems to be qualitatively different. Even what “pattern” means in both cases is different. A mathematical wave function is a formalization of reality, abstracted from the particular substrate of the wave. But how is information “abstracted” from reality? By a process of formalization? Laws of Nature? Rules that are inherent in nature that are then represented by mathematical functions? The patterns represented by wave functions are laws. The patterns by which informationalists aim to replace reality with computable information are probabilities. Something different is going on in the world of information-processing. And these distinctions matter.
Consciousness
To show why all of this matters, I want to jump into the 8th chapter of Tegmark’s book: a chapter called Consciousness (p. 130ff). This is where things become tricky and perhaps even problematic, because we’ll see an application of the basic assumptions that Tegmark has laid out in the beginning chapters.