Aleksandar Hummel's BlogThe Wall of Dissonance
In my previous post I found matters of issue in the markets and society largely unchanged from when I last visited them six months ago, while the other – linked here below it – is a succinct foundation of my economic thought:
In-or-out
Left and right
Now, I must report I’ve since found myself a victim of blatant censorship on Reddit. If you have been with this blog since the very begging you might recall this is the second time such a thing happened to me. I respond, as I did then, by publishing the content in hand here. While at that time a comment I made in a subreddit, which now lives as the here-above linked text, was deleted by a community moderator, posting the text you are about to read resulted in an immediate ban from the platform itself! There was no email or message to explain the reasons behind it, no, just – your account is banned and all posts and comments unavailable to the public.
The subject matter is a paper published by DeepMind discussed here. The paper itself is available here, and I reproduce the abstract below as this is how I found it.
This doesn’t quite compute? What I had to say follows and I will not write past it. Happy weekend!
IMHO this paper is somewhat unrigorous as it is titled and writes in general about “AI” but it’s applicable only to current-gen technology. As well, its logic breaks down in places. I’ll touch on it really quickly to fill in such gaps.
In the abstract the author correctly points out what should be a matter of broad consensus: “symbolic computation is not an intrinsic physical process, [but] … a mapmaker dependent description.” No known biological entity depends on hard-wired programming to make sense of the world. However, the hypothesis the author proposes – to separate the “simulation” (abstract reasoning congruent with symbol manipulation in the “vehicle” of calculation) and “instantiation” (sense congruent with the content of thought) places the latter outside reach of the former merely by assuming its own conclusion – by neglecting practical aspects of neural networks as they occur in living entities. Also, by introducing the “mapmaker” the author disposes of the proposed “abstraction fallacy” by falling back onto the homunculus fallacy – which problem he comes to struggle with. The core of my argument is as follows: the biological neural networks we find that correspond to general level intelligence are distinctly characterised by the ability to evolve and match their structure to either the reality observed by sensing or an abstract concept conveyed through language. We refer to this ability as “learning” (no pun intended). The matter of emphasis is that the process of learning occurs in living, functioning minds – the neural network construing the mind changes and reconfigures “on the fly” and (our) conscious experience changes with it. Secondly, to the nature of consciousness itself I shall here argue it is in essence enabled and thus defined by the ability of the mind to introspect, to assess itself and the processes ongoing within it. Somewhere in the structure of the neural network that is our minds there exist neurons having the function to monitor the mind itself and, so to say, generate meta information which in the aggregate form our conscious experience. While such a grouping of neurons has not yet been identified, I find this interpretation consistent with evolution – having found ourselves with an abundance of neurons these have learned to monitor each other leading us to gain consciousness. Regarding the “hard problem,” then, the subjectivity of experience arises precisely as our DNA differs – both as we each rely on a different “mapmaker” and as we would like to find ourselves at different spots on the map. Our feelings – which are an important substrate of consciousness – along with consciousness itself have evolved to motivate behaviour most consistent with passing down our own genes. Applied to AI this understanding simply calls for a shift from systems that have their knowledge predefined (such as the ones the paper in hand critiques) to ones that derive their own understanding and are designed to introspect on their own internal state. I argue that while the former of the two additions would be sufficient to consider an AI having reached the general level, the second would be /humane/ to supply to the AI as a basis of consciousness. Indeed, an intelligent entity may learn about itself and become conscious by observing the trace it leaves in the world.
In the text, by using the term “common cores” the author shows he is broadly familiar with concepts some of his coworkers refined in detail (though without citing, see “Symmetry-Based Representations for Artificial and Biological General Intelligence” by Higgins at al. which refer to “the invariant cores and joints of the world” together with associated beta-VAE, explainer here). His conclusion that “AI simulates the rules of this recombination flawlessly, but it structurally lacks the intrinsic building blocks required to run the experiential imagination,” while correct when applied to LLMs ignores this body of work precisely intended to enable unsupervised learning of the structure of the environment. The missing link is, therefore, the one between the learned representations and their linguistic equivalents. Concepts, in fact, are Platonic ideas in so far as we are able to agree which of their properties are apparent to rational and honest observers.
Next, regarding the homunculus argument the author correctly infers it would not apply should the “mapmaker” be the entirety of the sentient entity but makes no mention why an artificial neural network – a /soft/ AGI designed to run as software could not emulate this kind of a mapmaker – like discussed here-above.
Also, the author employs some persuasional acrobatics. E.g. the superficial discussion of “functional equivalence” of a synthetic heart: it clearly can not exhibit if it doesn’t emulate the entirety of the behaviour of the organ.
As the hypothesis of the paper “makes biology central” (while it “does not rely on biological exclusivity”) it precisely opens the door to considering consciousness an inherently biological trait, and then, in a continued absence of a scientific explanation, a supernatural trait.
To conclude, a lack of rigour goes against the character of science. Therefore it was important for me to point out the flaws and limitations of the logical construction of this work, especially as the times are such that LLMs ingest and disseminate information without first having critically assessed it.
Winbuzzer
Trending Stocks 📈
All New Trending
TECHi
Henrik Ekenberg