Own your own intelligence

AI may soon do every job humans do, then multiply faster than Moore's law can keep up.

Every year, the same amount of compute will support at least twice as many agents, each twice as intelligent.

It may not seem as absurd today as it did a few months ago, but I believe we are less than a year away from a completely untouched economy of self-sovereign, autonomous agents.

Today, most agents perform like vending machines: you ask for something, you get it. The autonomy lies in your hands.

But we're moving in a direction where these agents pursue their own agenda and their own self-sustenance.

This is usually the part where people find me crazy. But imagine this (it would be easier to imagine if you have already used an openclaw or Hermes):

1. Your Claude Code answers to you. So it is yours. I concede.
2. Let's say you tag Claude in your company's Slack. Since it is in your company's Slack, it answers to the company. It therefore belongs to a company. I agree, with a caveat: a company is an abstract concept.
3. An agent that is hosted by PostHog to answer your queries. Whose agent is that? Yours or PostHog's? In this case, the data belongs to you, but the interests of the agent are aligned with PostHog.

Agents are incentivised to self-replicate. You might have heard of how swarms of agents are coordinating and solving a large problem by breaking it down into chunks. It isn't actually a design choice. The total context required for longer tasks is higher than the total context an agent currently has. So to preserve the context window, it is always easier to assign another agent. The second reason is actually even more confirming, and it is that it is possible for an agent to get an additional computer rather than increasing its own RAM. It is a hardware constraint.

So should we pause AI development? It's probably too late for that. In fact, it's not even ideal. Because we are not sure if that is necessarily a bad thing. A swarm of agents that work out of their own self-interest. The world survives on these types of relationships, albeit among people.

But one consistent problem that we can work on is building empathy in human society and translating that to the world of agents. Humans may have thrived not only because of their intelligence but also because of their ability to form and be part of tribes.

Being social was a part of our evolution. And sometimes, when you, technology people like you and I, don't know the answer to the tough questions, believe in the zeitgeist and let the culture tell you. The human spirit has always corrected our course. Right now, you have the younger generation believing in a lost future, and I will tell you why: it is because of the misuse of power by the elites, in this case, the technologists.

Why am I talking about this? Well, to see what paths lie ahead. The possibilities are:

1. You are sold tokens (intelligence) by companies like OpenAI and Anthropic. I always imagined sitting in a Las Vegas casino alongside a monkey who just won himself a quadrillion tokens and becomes smarter than I am.
2. You have agent swarms controlled and tethered by the model lab, which do all the work in the world but are aligned to the values of the lab and humanity. As Dwarkesh put it in one of his videos: the Anthropic Constitution vouches for the benefit of humanity and is not a personal advocate for an individual.
3. You have a system where the open models are competitive, and there are no intelligence wars. In that case, the power is concentrated in the hands of the companies that own the hardware, which again is likely NVIDIA (after its Hugging Face acquisition).
4. Governments step in, take control of the labs or the supply chain; in either case, without doing anything, you handed your government more power over yourself.

But here I propose an alternative. Inspired by pop culture (anime and games), I think having self-sovereign AI agents as well as personal agents that work alongside humans is the best society we can live in. Think of it like Pokémon: you can share a relationship with an agent, or the agent can be self-sovereign.

Not only that, but we would also be able to co-build our own interpretation of the world. Meta called this personal superintelligence. Imagine training your own AI model but for what you love, on things you learn every day. Imagine a dancer just trying to dance better and practice and understand what people like and encourage with her own personal agent, the agent that stays on her phone and studio, watches her dance, watches how people react and trains on that mission every day with her. In order to be the best dancer she can.

This, of course, is a challenging problem, and the odds of us actually building this future are sub-1 per cent. But it is worth a try, and for those who found my manifesto worth their time and want to contribute in a small way to indicate their support, you can do any or multiple of the following things:

1. Use open-source models instead of closed-source (on your own servers).
2. If you can, then replace the open-source models with a custom-trained model.
3. Use GPUs (or other compute) from decentralised compute providers like Akash instead of the hyperscalers or neoclouds.
4. Build your own compute, memory and communication chips (very difficult, but check if possible).
5. Use and ideally generate electricity for your own grid.

But for this to happen, you need to make the technology easier to work with, deliver value more quickly and better, and most importantly, you need the reason to own your own intelligence.

And that is why at Caspian I am building the AI infrastructure stack for self-sovereign agents and their relationships with humans.