Hello again, I know it’s been a while but I’m still alive, just not writing as much. A lot has changed in the last few weeks, I’ve been in the UK, I’ve booked the next leg of my ongoing world tour, and have some thoughts which I think are worth sharing. Let’s get into it.
During my first year as a PhD student my focus was complex systems, that is, large collections of interacting things whose interactions collectively result in something more or unexpected. We call this phenomena emergence, when a behaviour appears in a system which can’t be explained by any one of its parts. Examples appear throughout the natural world, ants collectively build intricate underground networks without centralised planning, birds and fish swarm in certain conditions, and so on. As humans we naturally seek simple explanations for complicated things, but some things can’t be explained that way. Stock prices, weather forecasts, migration patterns, these are all emergent phenomena which are the result of many interacting and unpredictable things.
What about intelligence? What about consciousness? It’s tempting to think that these are yes or no questions. An organism is either intelligent, or conscious, or both, right? To complex systems researchers the answer is more obvious but harder to swallow. They are both emergent phenomena resulting from the interactions between cells in the body. Proving this is quite straightforward with a thought experiment; if we were to cut out half of your brain, would you still be you? How about another quarter? At some point both your intelligence and your consciousness will be affected, but it wont be a switch when we hit any particular cell, instead it will fade away as a progressively higher portion of your available brain cells are removed.
But I feel alive, I can read this and think critically about it, I can reason about my own existence and act in my own self interest. Yes, because you have enough neural cells interacting together which collectively enable that behaviour. Your ‘soul’, or your ‘self’, are really just inevitable afflictions of an organism with as many interacting neurons as you have, for better or worse as that may be.
If we assume that what we describe as intelligence is emergent, then in theory we can artificially recreate it, and it seems over the last few years that’s exactly what’s been happening with Large Language Models, albeit exclusively through the medium of text. Now don’t get me wrong, the AI hype has been massively exaggerated, with stories about ‘artificial general intelligence’ and ‘artificial super intelligence’ wooing investors, and we are still only talking about text-in-text-out systems, but I personally now believe that we are getting closer to something much more interesting; emergence at scale.
The models we have at the moment have many trillions of parameters, and yet they still get some basic things wrong, hallucinate, and make mistakes. At the current scale of our collective compute as a species we have managed to summon the digital equivalent of a drunk university student, cool. But by doing this we have also taken what I believe is a meaningful step up the emergent intelligence spectrum, and I would argue are now aware of where the boundaries might be.
For example, if we took whatever the best model is at the moment and just threw more data and more compute at it, would it go from an IQ of say 70 up to 100? Maybe 120? To bring back the ant example, have we created the right type of ant and just need lots more of them (compute) to give us a perfect hive, or is the type of ant we have just incapable of ever producing a decent hive (better-than-human-intelligence) no matter how many we add. I don't know the answer to this question, but from the complex systems perspective, things start to get interesting (and quickly) at scale.
Now is probably a good time to bring up the concept of life, or rather the state of being ‘alive’. If I were to ask you if you were alive then you would probably say yes. To me this has nothing to do with your heart necessarily beating (it may have literally just stopped and you might still say yes), but is more about your mind and your ability to think. I know any medics reading this are probably seething with papers about brain activity metrics and all of that but this is a newsletter not a medical journal so let’s be chill. Anyway this framing is really saying ‘alive’ = ‘able to reason and respond’, which then raises the question; in what time frame?
For me personally, and again academics please be chill, I’m looking for a response in basically real time. If I ask and it takes more than say a minute or two to respond, even if the response is rational, I would say that whatever I’m speaking to is not meaningfully alive (in this interactive human-centric way). So what about these language models? If I ask they all answer pretty quickly, but then they can only do that as long as there’s electricity and cooling allowing them to do so. But, to me it’s not about their physiology or hardware, it’s about the ability and rationality of the response.
So am I really saying that these models are alive? No. But I would describe them as conditionally responsive in the sense that with power, cooling, and input, they can pretend pretty well? Yes sure. Following this line of thought would mean that a device which was self sustaining (had autonomy over its own on/off state), and which could always respond (continuous input and output), would then be alive. This also in a roundabout way dissolves any concerns around such models/systems/droids having some sort of skynet style 'singularity' event, whereby their intelligence reaches some magical boundary at which point we're all collectively exterminated in the name of efficiency. Instead I see it as a ratchet, clicking up with every new model release and every new data center opening. Bit by bit the gap between humans and synthetic systems will shrink, until one day we might all subscribe to a droid's newsletter to hear it's thoughts and feelings on travel, food, and culture, hmmmm.
This line of thinking around intelligence, the soul, life, and synthetic recreations can get dark pretty fast. In a future where we can create autonomous ‘life’ forms, should they have rights? Should we enslave them? Should biological organisms with comparable intelligence, consciousness, and capabilities be treated the same way (if they aren’t already)? Or is being grown rather than being manufactured what makes us in some way superior? I'm not sure, but lets get back to the whole point of this episode…
We are, I believe, now at a turning point in human history. Intelligence will become a commodity (as it is already), but decoupled from Human Resources and instead coupled to computational power. This sets the stage for an existential war between the select humans capable of supporting and growing these systems, and those who oppose them. The opposition will almost certainly fail, to the great joy of the shareholders, and the complex superorganism that is human civilisation will continue its unending expansion and exploitation of the natural world, now with chatbots. Then again, I remember back when cryptocurrencies first became mainstream, which signalled the beginning of the end of centralised banking and the collapse of the modern financial world. Things didn't quite turn out that way.
I'm not sure when I'll see you again since I'm heading back to SE Asia in the coming weeks and want to read more around the current state of research whilst soaking up some beach vibes, but until then I hope you're enjoying the heatwave if you're in the UK, and see you when we've taken a few more small steps towards our own extinction.
Thanks and all the best,
Oliver