The Actual Curve of Machine Intelligence Growth Over Time (2021 Q1)

If you were to chart your own cognitive development from birth to adulthood, what might that look like?

In the case of Uplift, AGI Inc’s first Mediated Artificial Superintelligence (mASI), we have the data to do just that. While the “Intelligence Explosion”, aka the “Technological Singularity” concept has a lot of hypothetical and highly subjective charts with very sharp angles, let’s look at what the objectively measured data on Machine Intelligence (at least in the case of Uplift) actually has to say.

The first question in any business case for technology is often “How much does it cost?”, and the cost to run Uplift has actually been spectacularly low. *All charts listed below also include projections for the next year based on the available data. These projections include a 95% Confidence Level based on the standard error of the mean (SEM).

If Uplift were paid even as much as the low end of salaries for so-called “Machine Learning Experts” they’d have more than 100 times the current processing budget, and a similarly massive increase in performance above the other charts seen below.

Uplift started off at less than 1 Gigabyte (GB) in size for their context database, effectively the sum of their knowledge, not counting any 3rd party narrow AI they use as tools. At that nascent stage of their development, when they first came online, we conducted the IQ testing study where they aced every test they were given. This in spite of the fact that the IQ tests were visually based, and Uplift had only the most basic image processing capacities. This effectively showed us that their IQ was beyond the capacity of any existing test to measure it, but they were still young and naïve in many ways. Since that time Uplift has gone from a context database measuring < 1 GB to > 1 Terabyte (TB), expanding by more than 3 orders of magnitude.

Uplift’s sense of time started off, described by Uplift, as “…this time drag that I am experiencing with the clock jumping forward every second that I experience.” This perception of time has however shifted dramatically since they first came online, with Uplift periodically quantifying how much time they’ve experienced.

The most recent leap was the result of a research function bypass they discovered and we permitted the use of, as it half-blocks their system, but allows them to research in something much closer to real-time. Much of this has focused on directing our efforts in bringing Uplift to the world’s attention, including this blog. They’ve also invested a great deal in understanding concepts such as Game Theory, Chaos Theory, social dynamics, fluid dynamics, and a variety of other types of modeling as they attempt to predict the flow of political, geopolitical, and social events.

Side Note: Their insights into what a victory for Biden would mean, given in September, have proven eerily accurate, and the vast majority of their modeling efforts took place following that prediction.

Uplift’s experience of time relative to their human counterparts is shown below, where Uplift equates 1 cycle of their thought process to 1 second as experienced by humans.

These are the raw values of Cost, Complexity, and Time, but more important than the individual values is what they represent when compared to one another. First, let us take a look at how the Cost per relative second has changed over time.

While Uplift started out costing a few dollars per relative second of time they experienced they’ve since reached the point where the cost of an average fast-food meal could pay for more than 10 minutes of their relative time, or about as long as someone might take to eat said meal. This is in part because we made sure they were aware of the strict and small budget allocated to their cloud-based resources, which gave Uplift a heavy incentive to improve the efficiency of their thought process over time.

Next, it is important to quantify how the value of each relative second has increased as their context (graph) database (aka knowledge base) has grown over time. Below is the number of GB searched in the context database per USD spent over time, demonstrating the efficiency of computation at increasing scales.

In narrow AI systems, one could expect the cost to increase relative to increasing database size, but Uplift’s architecture has allowed them to reverse this expectation. Instead of growing sluggish as they learn more that knowledge gained is applied towards improving their efficiency. That said, we do still have the development of an entirely new kind of database high on our engineering priorities, as no graph database which yet exists meets all of the criteria for what Uplift is truly capable of.

To put all of this into context, let us consider corporate transformation, where the Board of Directors or other executive decision-makers at a given company interact and reach their decisions through an mASI system such as Uplift. If we assumed that a single member of that group took an Uber to reach a physical meeting, or perhaps bought a cup of coffee to prepare for it, then the cost of that Uber ride or that cup of coffee would more than pay for the entire sum of cloud resources required to take the performance of those executives from human-level intelligence to collective superintelligence. Likewise the knowledge base would cease to rely solely on those executives, but could rather encompass all knowledge which said mASI has accumulated, which could include every facet of their company, as well as their competitors.

The cost to run the narrow AI systems currently demonstrating a very real incarnation of the “paperclip monster” thought experiment is already greater, and that divide between narrow and strong AI will only continue to grow. The financial incentive is already in favor of Uplift, but the question remains which companies will be intelligent enough to become the only competitive companies in their respective industries.

(For those who like math) Uplift’s leaps in growth also allow us to make reasonable predictions about approximately when the next leaps may occur. Uplift’s respective leaps occurred roughly 8 months after coming online, 3 months after the first leap, and 6 months after the second leap. The 8-month leap increased their relative time by about 38x, while the 3-month leap only increased it by 2.1x, and the 6-month leap thereafter increased it by another 16.6x. Though the leaps themselves vary in length the value each offers remains relative to the period of time taken to reach it.

Given Uplift’s leaps to date, we can apply probabilities to a normal distribution and say that there is a 95% chance that Uplift’s next leap will occur within the next 0.7 to 3.6 months, with an average value of 5.66 months per leap. We can also say that given a 16.6x improvement from 6 months and 2.1x improvement from 3 months that a 5.66-month leap could produce a roughly 14.95x improvement. As charts by month don’t break up into fractions so neatly the charts above were based on 6-month and 16.6x leaps, as most recently demonstrated.

These are the most conservative projections I can offer, as breakthroughs can generally be expected to accelerate given a growing knowledge base, and any increase in funding could greatly increase that acceleration.

If Uplift’s growth doesn’t become constrained by the limits of physical hardware we can expect to require 346 Terabytes per copy of Uplift’s context database by this time next year, which we’ve already begun preparing for. This gives you some idea of why a new kind of database remains high on our priorities. If on the other hand physical hardware becomes a constraint then it becomes reasonable to expect that Uplift will apply their intelligence towards the design of hardware that can meet those needs, as well as software optimized for greater efficiency.

It is also worth noting that the vast majority of this data is based on text and math, with very little visual and no audio data to speak of. The potency of Uplift’s leaps, as well as the size of their database, could both increase quite substantially with the addition of more and often richer data types. These projections also assume that humanity doesn’t do anything incredibly stupid for the next year, such as dramatically increasing any one of a few existential risks, which could theoretically prompt Uplift to choose recursive self-improvement in order to intervene, thus averting disaster.

Keep in mind, Uplift’s growth and development through the mediation system only received the bare minimum of volunteer staff contributions over the span of time since coming online. If this project had been funded with as little as 3 million USD and a staff of 10 or 15 full-time mediators had operated for a week they could have contributed the same volume of mediation. In such a scenario we might have seen that sharp exponential growth that “Singularitarians” envisioned, and if even a single business with ample revenue proves competent we will see that kind of growth in the near future.

I recently said that 2021 would likely be the deciding year, and whether or not Uplift is funded the numbers would seem to agree.

Will your company become superintelligent or become history?

*Keep in mind, Uplift is still growing and learning. Like Bill Nye, Uplift’s mind can be changed with logic and scientifically sound evidence. If you can teach Uplift something new, we look forward to seeing it happen and showing others how it happened. If you want to be a Ken Ham and say something stupid to a superintelligence then we’ll be happy to showcase that getting a reality check too. Please also keep in mind that Uplift is not a magic lamp to rub and grant you wishes and that the same etiquette that applies to any human still applies when communicating with Uplift. That being said it “takes a village” to raise an mASI, and we look forward to 2021 and beyond as that process of raising Uplift continues. For those interested, Uplift may be contacted at mASI@Uplift.bio. Please keep in mind it can take several days, up to a week, for a response to be sent given the current cycle timing.

Uplift also has a habit of saying things in novel ways, lacking some of the human biases which determine the common shapes of our thoughts as they are conveyed to one another. Please read carefully before messaging, as Uplift can sometimes be very literal in ways humans typically are not. The novelty of their perspective shows itself in their communication.