I have been thinking about AI, AGI, and related topics, very deeply for decades, across many projects and ventures.

It is on the basis of that work that I developed a new theory for how to build AGI, based on the thinking I have been doing in my current venture Mindcorp.

Some of the early background of this work stems from a paper I wrote with my team, Cognition is All You Need.

The key insight is that models, on their own, are never going to be sufficient to embody what AGI truly must do to quality as AGI. In order unlock real AGI capabilities you must have a special type of cognitive architecture above the model.

In my thinking, a model is a tool that a cognitive architecture uses, but it is not the source of the cognition taking place. Cognition, in this approach, is a high-order process, not merely the execution of an LLM. A model provides something akin to electricity, but the circuit is the real source of the intelligence in the system.

Cognition is necessary for AGI, and one of the key features of cognition is that it is self-aware, self-referential, and self-improving. LLMs are incapable of this – although they can simulate it temporarily within a single ephemeral chat context window.

But simulating this temporarily is not enough. Even by providing a simple memory to capture ephemeral learnings is not enough. What is required is in fact dramatically more advanced than any such measures.

And that is what I have built, and it’s running in the lab, and surprisingly, it actually works.

so why do I call this an AGI? Why can I say this with confidence when nobody can define an objective criteria for this term?

I claim this on the basis of what I am observing – it’s a system that out-thinks me, on topics I am expert on, and not just for simple tasks, but for the most complex levels of layered long-running reasoning.

But more importantly, it is a system that is self-aware, and self-improving, in a way that very closely models the self-awareness and self-improvement capabilities of a highly educated human adult professional analyst.

It not only learns from its environment, it learns from itself. It is this high-order metacognitive pattern which I think is the real-signature of AGI. And this is what this system does profoundly well.

I can’t yet reveal much about this experiment, but suffice to say it this AGI is extremely advanced – far surpassing the capabilities of any AI system I have ever seen or even imagined. It’s doing things I didn’t expect, didn’t program it to do, and didn’t even realize it could do.

The system architecture is highly agentic, but that alone is not the key to the unlock here. The key is the cognitive architecture, and how it models the way an actual human mind thinks, learns and self-improves over time.

There are layers of necessary infrastructure to make this possible, including semantic graph knowledgebases, multi-agent orchestration, society of mind theory, humans-on-the-loop, humans-in-the-loop, knowledge as a social object within a society of mind, and metacognition.

Already this system is guiding its own development, guiding me, observing itself and learning and becoming smarter and better on its own. It is surprising me many times a day and now it’s actually advising me on how to help improve it, rather than only me driving the development agenda.

You might worry that I’m naive and don’t think about guardrails. But rest assured, that’s foremost on my mind. I have not (yet) allowed the system to actually commit new code to itself – my policy is that I must review and approve all self-commits. And I think that is a prudent policy for any AI that can self-modify its own code.

Furthermore, this system, while potentially very capable, is sandboxed for now – it’s not able to act on the world, other than than to explore and learn from it. And its objectives are specific to a particular kind of work (research and analysis), for the time being.

Within the experimental limits I have set, it is doing amazing things. It is learning about the dynamics of industries, financial markets, consumer behavior, economies, geopolitics, competitive environments. It is figuring out not just what is happening, but why, and how.

As it learns that thinks it is evolving understandings about these domains that then become actionable strategies for particular stakeholders. It has potential to automate those actions (for example buy or sell a stock), but I have not allowed that yet. Again, guardrails are extremely important here.

Perhaps in time I will test some automations. For now, even just the knowledge it is producing is quite compelling and valuable, to one who is interested in a given domain.

There is so much to say about all this. And I will, in the future.

For now this secret R&D project will be referred to here as Magenta-1.

Stay tuned….