Cybernethics / Cybernéthique

So these halfwits say that rational people ought to be bayesian. Sure. But bayesian about just what? By imposing not the right a priori probability distribution (that will eventually be fixed by experience) but the right descriptional paradigm, the right "meta-context", memetic parasites can have even "bayesian" people believe a lot of nonsense; just propose as "independent" facts on which to be bayesians things that are not so independent after all -- by bombarding readers with systematically biased data to evaluate, you will easily tip their opinions toward the direction of your choice; and they don't even have to choose "right" rather than "wrong" -- they only have to accept your problematics, your choice of data to evaluate. That's where Solomonoff induction comes into play.

La méthode experimentale peut établir des faits expérimentaux. Mais elle ne peut jamais établir de loi ou d'explication rationnelle. Pire, valider une expérimentation en tant que telle dépend soi-même d'un bagage théorique qui fait appel à des théorisations qui n'ont rien d'expérimentales. Pour dire les choses sans jargon, les faits seuls sont muets; pour les faire parler, il faut une explication qui est de l'ordre non pas des faits, mais des lois qui relient les faits. (Sans parler que tout cela suppose qu'il y ait déjà un accord sur l'établissement des faits eux-mêmes.) Bref, affirmer que la science s'appuie sur la seule méthode expérimentale, c'est une imposture épistémologique. Certaines sciences utilisent voire requièrent la méthode expérimentale pour étudier leur sujet; mais cette méthode expérimentale ne saurait résumer à elle seule l'approche épistémique d'aucune science, pas plus qu'elle ne saurait caractériser aucunement la science en général.

La science, c'est le savoir partagé et partageable. La méthode épistémique pour relier les faits en lois explicatives, c'est l'induction. Affirmations de l'ordre de la définition, indiscutables. Ensuite, pour ce qui est d'étudier la nature et les propriétés du savoir partageable, on pourra voir un épistémologue des sciences (suggestions bienvenues: Popper? Carnap? Quine? Dennett?). Quant à la nature de l'induction, la chose a été remarquablement formalisée par Solomonoff.

When discussing epistemology, and particularly when disparaging Karl Popper as a half-wit (which still makes him waaaay ahead of the usual crowd of witless zeroes who are paraded as philosophers), I often like to tell about Solomonoff Induction, this compelling solution to the problem of induction. Popper believed such solution was impossible (and redefined "induction" so as to prove his point with Hume's argument), even though all humans (nay, mammals) had been showing him wrong for millenia, and though a formal solution had been hinted as Occam's Razor (tell me about Diogenes the Cynic's response to Zeno's Paradox).

Well, it so happens that Ray Solomonoff has written a nice historical account of his discovery (kudos to CiteSeer). Lots of reflective epistemological insight there -- and a lot of other great heroes who played a role in this story, including (surprise!?) a key contribution by the founders of LISP.

Of course, Solomonoff Induction is not meant to formalize exactly the way we either think or ought to think. Most importantly, it doesn't take into account the cost of thought, which limits the depth and precision of our probability computations. But that's not the point. The point is that we can formalize the way that induction is being done, and capture in an understandable way the essence of understanding -- and its limits, as Chaitin would no doubt remark.

PS: of course, there exist or can be invented infinitely many variants of Solomonoff Induction, that will take into account constraints on computation such as costs, reversibility, etc. These variants can help serve a variety of purposes, from modelling human learning processes to building machines that learn and otherwise mine data. But the details of these variants is secondary and comes naturally; the great discovery was in Solomonoff Induction.

I've once written an article about Microkernels, where I denounce them as being mostly an intellectual fraud.

Someone from the internet (possibly a random student doing homework) sent a mail to me and Jonathan Shapiro asking whether a Kernel was faster than a Microkernel. Here is what I replied.

[Darn, I didn't insist enough on the whole kernel/microkernel issue being irrelevant in the face of the overwhelming issue of the structuration of the higher-level system components. And about the solution being of course no kernel and expressive contracting between components instead. Reminds me of the metaphor of system as society and kernel as government: no one needs a government; everyone needs an expressive market of property rights.]