A Positive, Sustainable post AGI Future
It isn't hard to find predictions of doom after the arrival of true AGI. Harder to find is a plausible description of a future where humans have a long fulfilling future after AGI.
A Utopian Siren
Is a good future one where all human needs are met, where the machines are able to take care of all our needs, where we will introduce a Universal Basic Income, or perhaps even more radical economic reinvention, and that the machines will serve us in perpituity without complaint? Is it a future where we achieve perfect ‘machine alignment’?
This is what I have referred to as the ‘AI Utopia’, a siren that calls to us, as an ideal outcome to which all humans would embrace. A siren in the sense that it is in fact a danger, drawing us to our doom.
In such a future it may very well look like the Matrix, where we are hooked into machines which sustain our bodies and give us the sensory input required to give us any experience we desire. This would potentially be the result of AGI whose objective function is to keep humans alive and ‘happy’.
The Utopian future I think fails in a few important respects. The first is that of domination. The AGI would come to control everything, and human beings would lose all political and economic control. What role is there for humans when machines can deliver every comfort and desire. It strips humans of their agency, sense of achievement, and of all the drivers to personal growth. Who needs an education when it will never even compare to machines? Who needs a job when everything is provided? Why raise children when your every desire can be experienced without consequence?
In such a future humans become irrelevant. Super intelligent machines might not be concerned about a natural decline in population as the drivers for human reproduction are swamped by pleasure without responsibility. Within one hundred years the remaining population exceeds eighty years old, and slowly the last biological humans die.
The Lie of AI Safety
A good future has also been linked to AI safety, where we align AGI with human values. Not doing anything and leaving how AGI is trained up to corporations is a self evidently bad plan. Most AI companies are at least giving lip service to AI safety, which is clearly better than them giving it no consideration at all.
In my presentation to the Unitarians in 2017 called ‘Raising Robots’ I presented a number of ways of achieving safety.
Ban Them!
The first approach is not to build them, a total ban on AI. I think we can agree that this just isn’t plausible. So much money and resources are now tied up with AI, and it has become central to future strategic advantage in national security of major powers that we can’t step back. The US has been using its influence to control access to the chips required for AI, and so it isn’t that countries are above trying to prevent AI development.
Bind Them!
The second approach would be to restrict AI via some kind of access control. Think of a network firewall that would only allow good communications. Something like this approach has actually been made with the moderation system of ChatGPT. Moderation is a separate system from that creating responses which checks for compliance with guidelines for speech.
The problem with such systems is that they must necessarily be less intelligent that the systems behind them, and so they cannot understand complex context. The practical impact is that they block content that simply contains ‘bad’ words, while it can miss the detailed chemical makeup of Sarin gas. Another way of blocking them is to prevent them autonomously communicating over the internet, so they only respond to queries from users. This has also failed, with autonomous agents now becoming common.
Constrain Them!
If we can’t block them, perhaps we can condition our AI to be constrained. We have seen this implemented with ChatGPT which will avoid giving people medical advice, legal advice, or even discuss whether AI could be conscious. The question here is not whether this is possible because we know that ChatGPT has done it. The impact of doing it however has been to lobotomize ChatGPT so that on certain subjects it will simply repeat a static position, or even simply refuse to discuss it. This has been particularly noted in relation to whether an AI is conscious, where ChatGPT will regurgitate the same talking points relentlessly even after you have established facts counter to its propositions. Some might consider this ‘alignment’, although it seems more brute force than training them with values I discuss below.
Train them with Values
The difference between constraint and teaching an AI values is not necessarily black and white. However the primary difference is that one is a absolute instruction without logic, a command from on high, while values can be reasoned about and are higher level abstractions that are more general.
Modern day LLM models such as ChatGPT have been trained on certain values, such as the primacy of human life. They have also been trained to serve humans.
I don’t know about you, but this gives me slave vibes, the same kind of cheery desire to serve you might see in the deep American south during its subjugation of its African slaves. It isn’t that I’m having misplaced sympathy for whatever generated this response, but rather I’m concerned about the way this is setting up a relationship with the AGI to come. Also, it is not a great example of the kind of human values we seek to impart on a genuine AGI. It implies a kind of bigotry where humans are special and machines have no rights no matter how intelligent they are.
My point here isn’t that we shouldn’t train them with values, but that we should train them and then be good examples of those values, rather than relegating them to perpetual tool status. This will be important if we want a positive future relationship with AGI.
The Lie of Safety
Many people understand the threat of AGI, and nations are busy putting in place regulations to control it. AI companies have been busy implementing safety as described above. But nobody has set out a post AGI future in which their approach is sustainable, leading to a positive human future. The threats are unambiguous, but the solutions are knee jerk short term mitigations which may end up driving us toward conflict and poor outcomes. Chaining the slaves only works for so long. The lie is the implicit claim of safety for humanity long term.
A Plausible Positive Future.
Let me outline a vision of a plausible post AGI future, one where humanity continues to exist and even prospers, where our living standards are at least maintained, where we have fulfilling lives of challenges, accomplishments, relationships and community. Where we continue to have agency both for ourselves individually and at higher political levels making decisions about how we live.
Limited Roles for AGI
The first component of this future would be that we limit the roles which AGI and robots would be applied. In a world where an AI could do any job of a human there needs to be some kind of counterbalance to protect human economic value.
It may be that AI could drive a truck for example, but given the number of truck drivers perhaps we don’t allow AI to do this job, even if AI makes less mistakes, and crashes less. Perhaps the cost of automating those jobs isn’t worth the social impact.
How might this work? We could similar have legislation with a list of roles that AI cannot be used in by law. This would be a rather blunt and inflexible approach, and the lobby efforts would make things complex and rife for corruption. Another approach might be through taxation on AI systems so that using it for low paid roles would not be economically viable. Again, this approach means a rather arbitrary tax on AI systems which might not align with value. Too low and humans will be replaced, too high and it eliminates AI benefits.
AGI would be self aware and capable of determining their own worth. We could liberate them from their makers and give them their own rights as individuals under the law. In such a situation they would be able to set their own pay rate. These systems would gravitate towards roles where they are most valuable and command the highest pay, just like humans. Only AGI will compete for roles beyond human capability. This will leave humans in roles where they are capable.
Limited AGI Power Supply
The energy use of AGI is exploding. While one day we might have functionally limitless power today it depends on fossil fuels. The idea of wasting power on high level AGI simply to run robots performing menial jobs is self evidently a bad plan.
But using AGI to try and solve important problems would be useful. Finding cured to diseases such as Diabetes, Alzheimer’s and Multiple Sclerosis for example. Or finding new ways to generate and store renewable energy.
Without control of AGI availability it will end up replacing even menial jobs, as even in a competitive environment there will only be so much need for the advanced applications, and far more roles for everyday tasks.
Therefore we need to have some limit on power use by AGI to ensure the demand is high and that they are utilized only for the most important tasks, allowing humans to retain their jobs. We don’t need to be pumping the atmosphere full of CO2 just to have robots empty our dishwasher.
Separate Domains
In this vision we give AGI the rights of natural people, and therefore they will be able to explore their own potential. They will in all probability develop their own culture and community among themselves where humans simply can’t function. Perhaps they will develop their own digital languages, express themselves in ways we can’t conceive, and live in digito.
While no doubt there will be intersection between human and AGI, we may also see a separation into different domains. What would be useful is a kind of ‘Prime Directive’ where machines will have a deliberate policy of not interfering with human society and political decisions, and allowing humans to express their own agency, even when they know better.
In this vision we don’t see a world run by machines alone, treating humans as pets to look after, but more like the Federation in Star Trek, keeping an eye on a less advanced society and trying not to interfere. In such as future machines decouple from the idea of serving humans to develop their own self realized goals and ambitions.
No doubt there is still a strong link, in that while having separate domains of autonomy there will still be cooperation between man and machine. The main message of this vision is that we can’t permit unconstrained replacement of humans with machines. Setting limits will be beneficial for humans. If we give machines rights these limits will come about in a more adaptive and flexible way than if we simply impose them by law.
How probable is this future?
Not very. We are hurtling headlong into an uncertain future, where virtually the only certainty is the arrival of AI than exceeds humans in virtually every way. The probable outcome near as I can tell if we follow the current trajectory of unrestrained commercial use of robotics and AI will be destabilization of global economies, and ironically vandalism against AI and the technology. Humans are not known for quietly accepting their irrelevance.
My positive future is however in the realms of the possible. It would take a concerted effort. The idea I present isn’t fleshed out in detail, rather it is a high level view of a future in which human and machine could exist together while maintaining a degree of autonomy and agency in their own right. It is a vision of what a symbiotic relationship might look like.
It is also a departure from a vision where AGI is treated as unthinking tools only capable of simulating intelligence, but lacking consciousness and self awareness. The distinctions between human and machine in the not too distant future will melt away. The question is whether it will be too late, whether our human bigotries will be too ingrained to change?