I have experience out performing the best in the world for 1000x less and I’m confident I’ll be able to do the same building a Microsoft for Droids with The Droid Flywheel Coldstart.

All that matters is the derivative rate of improvement. The smallest possible system with the highest maximum derivative rate of improvement (acceleration) will eventually dominate. It is not expensive to build these systems and they are hyper efficient to scale.

The way you maximize acceleration is by maximizing:

• speed of feedback
• signal of feedback

The higher the speed and signal, the greater the rate of compounding improvement.

The Opportunity

Both of these dimensions are incredibly weak (relative to potential) in modern AI. Cycles are long and signal is fuzzy because data is force fed to models. See 😖The Fallacies of AI .

The cycle I’m talking about is going from error in performance to improved performance.

Ideally, this cycle is measured in seconds. In modern systems that do mass training, for example with LLMs, you have a big long multi-month or longer cycle where the model grows stale. You have to crawl the internet and collect chat sessions to get data. You have to hire many, many people to manually give feedback to the system.

The real problem is that it is hard to even identify a real cycle because it’s hard to even define improvement in many systems. The best we can do is look at their “versions”. In the last half decade models like ChatGPT have had about 5 major evolutions.

You can look at charts like this analysis from Stanford and rough out an annual improvement on the order of 10% or so each year.

This sounds impressive but it’s actually quite low (relative to potential). In order to thwap these giant companies we will increase the cycle time by 1000x — to less than half a day.

Sound impossible? How could all these smart people with billions not do this already? It’s not a question of logistics, it’s a question of framing.

Our Approach

So, how do you maximize differential of improvement?

First, I’ll tell you how it won’t be done:

• hiring people to provide feedback to your models
• using the same systems everyone else is
• force feeding your models data that you sourced
• ungrounded models
• spraying billions at the problem
• using language or arbitrary environments

How will it be done? Our system will be auto-didactic by doing the following:

• self-assess — identify where it is weak
• theorize — self-generate curriculums and test them synthetically
• be open-loop — it will ask for help when it gets stuck
• experiment — test in the real world

Sound magical? We believe it is just about the closest thing to magic that exists and it is only made possible due to several cross-disciplinary insights and how we have framed the learning problem.

Our approach is to do as Turing did when he invented the computer, he re-framed the problem he was working on so that it could be solvable. The Universal Learning definitions below can be represented mathematically and thus are solvable.

A Universal Learning Machine comes in a few flavors, this is the ladder:

• Bounded Universal Learning — If a machine can generate it’s own curriculum (a theory) and then test it (experiment) then it can learn recursively and learn everything coupled to the dimensions it understands.
• Weak Universal Learning — The ability to expand the dimensions of understanding by plugging in new sensors and new actuators that are given to it ad hoc. Think something like a universal transformer.
• Strong Universal Learning — Invent new sensors and actuators. At this point, the machine truly eclipses the capabilities of humanity.

We will build the smallest possible example of bounded universal learning system in the fastest and highest signal environment — the real world. Then expand up the ladder of the Universal Learning Machine to ever more abstract environments ending with Strong Universal Learning and Droid Capitalism.

Think of it like this, we will produce “baby” Universal Learning Machine and then grow it up while the rest of the world is trying to squeeze an adult through the birth canal.

Throw in a complex viral loop ala The Droid Flywheel Coldstart to drive federated tutoring (rather than in-house limited tutoring) and we have something truly special.

Closing

In a decade there will be tens of millions of droids operating under with an early form of Droid capitalism and Weak Universal Learning. In 50 years there will be at least 3-to-1 per human and arguable Strong Universal Learning.

While the rest of the world is caught up with the slow world of abstract thinking we will blitz krieg the fast world of physical evolution in silicon.

Welcome to the Droid revolution.