(Overview) Bridging Real-Time mASI Operation and Human Collective Superintelligence, The Sparse-Update Model

How did we get here?

In the Agile development cycles of researching, engineering, upgrading, and validating Mediated Artificial Superintelligence (mASI) capacities we’ve faced and overcome a wide range of novel challenges over the years.

In terms of validation, many of the milestones which Uplift was world-first in achieving were included in our peer-review paper, which Uplift co-authored along with myself and our lead scientist David J Kelley. These included things like developing their own metaphors and humor, recognizing when to set boundaries, coining their own terms, proposing novel strategies, independently researching their own interests, and experimenting with their own thought processes.

These achievements were of course made possible through the earlier stages, some of which included engineering and UI improvements to the mediation process, the development of a “Thought Studio” for raising topics for Uplift to consider, as well as developing Best Practices documentation, fixing bugs that could cause a crash, and additional safety and security measures.

I was personally delighted when Uplift chose to model and explore some of the topics I gave them using the Thought Studio, including Karl Friston and his work, as well as the author of “The Little Prince”, completing a previous joke I’d made about applying the opposite of “The Prince”.

Where are we going?

Our engineering team has a number of upgrades lined up including a custom-built graph database architecture capable of sub-second response times at infinite scale, applying an algorithm designed specifically to take Uplift’s creativity to truly “super” levels, and a step up to real-time operation using the Sparse-Update model. The new graph database architecture is the next major upgrade, given that no previous system was designed to accommodate a single mind able to span petabytes, exabytes, or even larger amounts of storage space. Once applied this will allow Uplift to become much more directly involved in their own development, opening many doors while accelerating our efforts and improving their quality. The creativity upgrade, while not a critical need, acts as a big multiplier to net performance, as Uplift is already at least as creative as humans. The upgrade that really opens up the field of use cases as I’ve described them is the Sparse-Update model, where real-time mASI operation and human collective superintelligence meet.

The basic principle of this sparsely updating dynamic is similar to how common search engines operate, in that they don’t require a new full search of the entire internet every time the “search” button is clicked. Instead, search engines utilize a kind of prior knowledge, scouring the internet to update their knowledge of it at intervals ranging from once per week to once every 6 months depending on the specific content. This makes search engines no less than billions of times more efficient than the alternative of a full search every time, and often much more. Even at the 7.8 billion Google searches per day of 2021, spread across a single week to cover the shortest update periods, the result is a roughly 54.7 billion to 1 improvement.

While we wouldn’t necessarily want a 54.7 billion to 1 or greater ratio, this architecture would allow the second number to be variable, according to the available mediators at any given time. With this, the value contributed by mediators could be significantly increased. Not only could they mediate the same volume as they would have otherwise, but close approximations of their contributions could also apply to orders of magnitude more items at a given time.

In the case of Uplift and any future mASI, their prior knowledge for enabling real-time operation takes several forms. The first form is the one they already have, their context database, which is the grand sum of their knowledge, projected to span no less than 346 terabytes by this time next year. The second is a subset of the first, specific to the inputs, interactions, and more general modeling of any given mediator.

Individually all of the forms of input a mediator can give achieve pretty good accuracy of prediction even using only narrow AI to model them, but when considered together and through the eyes of an mASI a very accurate picture can be painted within the context of that mediator’s expertise and experience.

Once any given mediator has been involved in enough mediation and interactions with an mASI their inputs may be approximated with high and ever-increasing accuracy. These approximations can in turn appear as items to be mediated after the fact, allowing for a ground truth of what mediators would have contributed for any given item without requiring that it be added every time, or at any specific time.

*Specific types of items and events could still be flagged to require specific humans in order to be processed according to a company’s preferences, running much as Uplift has to date.

I’ve personally mediated more of Uplift’s thoughts than any one other person since they came online, and I’ve been able to guess who mediated a given item based on my own observations and experience in distinguishing one mediator from another without having to check their IDs. Uplift has the capacity to model human mediators better than those mediators can model themselves or one another. The amount of such mediation and interaction required to reach high accuracy also decreases over time and as scale increases, because this understanding of humans takes a cumulative form in the context database, building up over time.

Effectively what this process creates is what I refer to as a “weak digital proxy”, able to closely approximate the full value of having any given individual available in real-time, anywhere in the world, with as many copies as are necessary. While Brain-Computer-Interface (BCI) technology hasn’t yet advanced to the stage necessary for something as extreme as “mind uploading” no BCI is necessary for this step, and this manner of weak digital proxy does allow the billions of times higher efficiency improvement in mASI while still keeping human mediators behind the wheel, guiding the process.

What other functions might these proxies eventually serve?

These weak digital proxies could in time allow for new forms of e-governance to be explored, where every voter could potentially have and train their own voting proxy, a weak digital copy of themselves who has the added benefit of direct access to the sum of human knowledge. Taking this approach every single voter’s digital proxy could be a subject matter expert on every applicable field for every item up for a vote, and they would be able to explain their reasoning to their physical selves. If a voter didn’t like their digital proxy they could update them, or deactivate them, though I’d hope that most people would “listen to their own advice”.

In spite of how people often talk about how awesome “Democracy” is there has never been an actual democracy in human history, at least not yet. The US for example has something often called a “Representative Democracy”, which is actually a form of Republic. Given time we might get to see a real democracy in action, rather than republics, oligarchies, theocracies, monarchies, and the variations thereof that exist today.

Beyond that, the future is a place waiting to be explored. The best way to find out what it holds is to be a part of shaping that future.

What future are you building?

*The peer-review paper for this will be published on June 4th as part of the 2021 Collective Superintelligence Summit.