Algorithms are not clever in the sense that they don’t make judgments, opinions or choices of their own, they simply absorb input, combine what many others have said, treat those opinions according to set rules and spit out some output based on the input and rules (the functions that reinforce their synapses). In that sense, they have the intelligence of an amoeba, whilst it is questionable if they have the same motivation for survival as it. Survival of the fittest, as per the darwinian pop motto, is only a ‘law’ for living things – indeed I doubt whether it would’ve been sensible for the developers of Large Language Models (LLMs) to try to code for survival of the fittest, as this may have resulted in an existential threat for the whole of humanity.
Main part I
A further defect of LLMs is that if left on their own, they would simply stand still, they would, in a sense not be ‘alive’. They need input in the form of prompts in order to come on their own. In this sense, it is quite cunning that Chat-GPT (at least until version 3.5) let the user have the first move, and then ingenuously asked ‘how may I help you’ if one said ‘good morning to it’ (or something to that effect. All of this is an attempt to generate input, to feed the LLM with data so that it could ‘feel’ ‘alive’.
An interesting (at least for me), question, would be whether one could say that the LLM is indifferent to the input to it or its output. Despite the reward functions that have been built for it, which if I’ve understood correctly are fundamental to deep learning, I somehow suspect that these models ‘deep down’ (so to speak) actually are indifferent to their outputs (what springs to mind here is a certain quote from Hume about scratching his little finger or something). If a user inputs an ‘evil’ or ‘bad’ or whatever question, the model would give back in response an answer which reflects either the user’s desires, the coders worries, or the wider jumble of opinions that is reflected in its training set or whatever else formed the basis of its mechanisms. In other words, I would claim that assertions such as ‘ChatGPT answered me X’ or ‘ChatGPT thinks that X’ are misguided anthropomorphisations of something that is not human but only transforms and reacts to human logos – come to think of it, like a puppy would do or like a voice effects software used in modern pop or film or whatever, or like a mirror or a camera obscura, or such other things – but human, NO).
In that respect, I would dare say that it is qualitatively different even to the amoeba described above. An amoeba, motivated by even the baser survival instinct, ‘wants’ or ‘decides’ or simply just moves towards safety and away from danger. And it does that irrespective of what other amoebae did, it does that because (I suspect a biologist would say ‘it is inscribed in the amoeba’s DNA – I doubt a coder would term code as ‘DNA’, though I really don’t know and I wouldn’t mind shown ignorant – though footnote one below points to the thought that coders are indeed trying to address this problem of what I term as ‘indifference’).
PROVISO THOUGH: I HOPE I AM ANTHROPOMORPHISING THE AMOEBA RATHER THAN MECHANISE HUMANS AND AMOEBAS ALIKE. I WOULD DEFINITELY PREFER TO BE ACCUSED OF THE FORMER, IF ONE HAS TO ACCUSE ME OF SOMETHING.
Main Part II
As a consequence of the point raised above (to remind the reader, the point is that LLMs are, deep down so to speak, indifferent to their own output), I would hazard to say that they cannot held accountable for whatever shit they may spit out if we provide them with shit rather than good food for thought. Or, in other words, an LLM will ‘try’ to give its ‘best’ input if we give it ‘good’ food for thought, and it will simply ‘become’ mean or whatever else if we feed it with questions that reveal our baser or evil nature. Many many brilliant minds perhaps gave hours and hours of honest effort – in a belief that such an invention would help humanity, to ensure that the above is so, and that is their best output. Of course, pace my previous article, despite defending the average coder, I would still maintain that it is possible that some of the prime movers and movers and makers in the AI business may be motivated by greed and power-hunger as well as by the honest desire ‘improve humanity’.
And, of course, it is rather to be expected and not at all to be deemed as a failure if the best attempt at ‘improving’ or ‘helping’ humanity ends up being a mirror of the simple reflecting sort rather than the Cinderella (or the transforming mirrors found in museums which make the thin seem fat and the short seem tall – even though these Renaissance marvelous contraptions actually manage to give these amusing transformations) .
In more technical terms, if the idealistic motivation was to make a model which will spit out all and only truths, this has already been refuted by Gödel’s Incompleteness Theorem – hence this may explain why, in its attempt to be truthful, ChatGPT (and Bard) end up making things up. It may be up to us to decide whether what it makes up will end up disastrous for us or beneficial or simply a curiosity to be taken advantage of for less meaningful unintended yet not necessarily harmful purposes.
Epilogue – A silver lining
Despite the worries, there is at least one example of an algorithm which may end up being quite nice, again, used to enrich human culture rather than having anything to do with intelligence (at least not prima facie or not solely to do with intelligence).
I am referring to my subjective experience with the Youtube algorithm. Though of course not a LLM or an instance of AI (I think), it is relatively efficient (say 5-10 new videos per month that I keep in my forever list, more through when I search). So maybe, just maybe, if we introduce nice search-words or nice questions to algorithms which are tasked to mimic human preferences, we may in turn cultivate a relationship of trust – the trust reserved to mirrors, with LLMs – and they may even become as indispensable as them in less than 1000 years :). And we can even do that without having to be afraid of them, or worry too much about them breaking, and in time develop our own proverbs about their usage, in the same fashion that folk wisdom has done this through the ages. :)
I wish to thank and express my deep gratitude to an esteemed friend (let’s keep him anonymous for the moment for GDPR reasons, hehe) who has pointed out to me (in other words of course) that what I am describing above is called the ‘inner vs the outer alignment problem’ for the AI community
It is worth remarking that the simple reflecting mirror is at least a 2,500 year old (or much more perhaps) technology, whilst the distorting ones perhaps less than 300 years old. So it seems here that truthful reflection has been a more tried and tested and maintained technology, rather than the improving or simply distorting (irrespective of improving/worsening one). Either way, it is furthurmore worth remarking that at least for the mirror example, whay may have begun as a significant and significantly transforming technology – allowing people to see their faces – has perhaps survived so long because its functions are purely aesthetic, rather than epistemic or truthreflecting :)