The Biopolitics of Intelligence

Here I frame a macro-level biopolitics of AI by arguing that a) not all machine intelligence is just artificial, b) the nature of access over topline machine intelligence will shape or constrain critical societal intelligence functions, c) the ability to blockade intelligence, enhance or degrade the quality of its delivery, represents a major threat to societies and organisations, and d) the only reliable and long-term solution to the problem of intelligence blockade is sovereign local deployment.

  
a) Not all machine intelligence is artificial.

I go back to Herbert Simon here - artificial are those designs which embody a human purpose. The garden is artificial, the woods are natural. Most machine intelligence we encounter does embody a human purpose. However, it is not necessary, theoretically, that the goals an AI pursues draw upon human purposes. There is more to this. Consider a single biological neuron. It is a simple electrochemical switch. It either fires, or does not, based on its immediate synaptic inputs. It does not "know" what language is, what memory feels like, or how to solve a math problem, but trillions of these dumb switches self-organize to exhibit the human consciousness.

The key here is very complex systems which emerge due to simple agents interacting using very simple rules. A single node in a deep artificial neural network performs simple linear combination following an activation function and holds no intrinsic knowledge. Yet, when billions of these nodes iteratively adjust their connections and weights, high-level abstraction spontaneously emerges. We don't know how (that is where the challenge of explainability comes), it just does and more so as you keep scaling. If you can have a functionally smarter, functionally creative system, arguing over whether it has "thought" or "simulation" is pedantic - you've got bigger problems. 

This self-organised coherence is why it is possible that an AI may not always embody human purposes, despite working on a substrate that was built by humans. This means that machine intelligence may instead become increasingly shaped (like human intelligence) by the adaptive requirements of the socio-technical networks within which they are embedded. This throws your align-in-the-lab-deploy-in-the-wild out of the window. The real alignment will happen in the woods, not in the garden.     


b) Available machine intelligence will shape societal intelligence   

An acquaintance of mine recently began exploring AI models for live financial/stock analysis. His complaint is that the same model has become stupid now (earlier it was smarter). His problem is that he never had the model, he only has whatever artificial intelligence trickled down to him through free, or nominally paid, APIs. It shapes his own performance in relation to those who could have access to better models. Trading is a zero-sum game, so he loses to those who have better infra and better AI. Now extrapolate his experience to organisations, governments, populations, vulnerable sections of society. Differential AI capacity will lead to some humans, probably clustered in specific geographies, absolutely dominating over the rest simply by being able to know, predict, and do more, better, and faster than others.  

It is highly unlikely that the Big AI labs will ever release the absolute best of their models to the public, it is more likely that those will be monetised on a B2B/B2G basis and the consumer grade models will offer a degraded performance over their limited access counterparts. All this is just economics and has nothing to do with AI risks or any other pontifications about saving humanity. The only hope is open-source catching up and converging with some more technical breakthroughs. 

c) The coming of intelligence blockades 

The claim that biology self-organizes intrinsically while AI is organized extrinsically (by human "loss functions") is drawing an arbitrary line where the system ends and the environment begins. Once you take the whole and start removing system boundaries, the living world basically consists of agents capable of intelligence functions who shape their environment, form coalitions and try to influence others (using soft or hard power). This implies that we will have hybrid coalitions (of humans and AI, we already have had humans and animals) which will try to optimise their resource utilisation and develop some unique or differential capacities which can be sustained in relation to other hybrid coalitions. This is where and how the real alignment happens.

It is increasingly likely then that as human dependence of AI increases (which is likely because of prisoner's dilemma, you adopt or perish), coalitions will be incentivised to degrade the intelligence functions of competing coalitions or turn them into dependents. This is sort of like cognitive colonialism, a more resilient one as dependence will be maintained through convenience, not coercion. One may define intelligence blockade as the possible restriction, degradation, or conditional provision of machine intelligence necessary for maintaining competitive cognitive performance, whether at individual level or at the level of broader networks that form coalitions and nation-states. 

Because the locus of agency shifts from individual agents to networks of agents (humans and machines) that increasingly function as integrated cognitive systems, the age-old logic of blockades appears at a higher level of abstraction where the objective is no longer merely to deny a nation access to material resources, but to constrain its capacity to perceive, learn, and decide at a level comparable to its political or economic adversary. Just as states depended on foreign energy supplies or maritime trade routes, they will depend on foreign compute, technical standards and information products.     

d) We have to explore pathways to sovereign local deployments.  

The only reliable and long-term solution to the problem of intelligence blockade is Sovereign Local Deployment (SLD). But, SLD exists on a spectrum and is therefore also aspirational, i.e. a goal to continuously move towards. At the most optimal level of SLD, it is working at individual level where everyone has a private AI infrastructure. They control their AI, the hardware, curate data, rely on local inferences and self-hosted deployment. There are economic, skill and knowledge barriers to a truly functioning SLD. But it is the star to aim for if you want to address the risk of intelligence blockades. SLD is more complex at society/nation level.

Consider the global system in terms of relations and dependencies nations have with others, and just map those relations as arrows in your mind: A truly sovereign system is one where all essential paths remain internal. A dependent system will contain some critical external/outgoing connections. Now assuming the global system contains everything required for topline intelligence production, it is necessary for any truly sovereign system to contain enough structure to continue producing intelligence even if the external connections disappear. Basically, the local system should be a subset of the global system for SLD at that level to work - it needs all the institutional structures, supply chains, talent, energy, security, and governance capacity which exists at the global level. 

Unless that happens, blockades will continue to risk reducing the set of connections essential to achieve intelligence functions in local environment - whether this be at individual, village, city, state, or national/population levels.