Randomness is a commodity

In a conversation with William, he describes his youth, and the events the occured during his childhood.

You know what year it is? It is 2248 and randomized-gradient methods have taken over the latest training methods for artificial intelligence. Computers have now gotten so large (2048-bit systems) and the optimization landscape so bumpy and rough that these randomized-gradient methods, developed specifically to accelerate the development of AHGI (Artificial Hyper-General Intelligence) that it is mathematically necessary to generate randomness down to the 10^(-59192). Standard convolutional methods of doing so fall short and send the optimizers down paths that do not succeed.

In fact, it’s a miracle such an optimization exists in the first place. With every single soul on earth that was deemed capable to work in the field of optimization actually working on optimization (funded by NSF-AGIGrant 10573728, which recieved 28.8% of humanity’s 2129 GDP, created by AGI Fund AGI-tadel (prodounced a-jit-adel, the AGI lab of renowned hedge fund Citadel)), there was bound to be an optimization method that one of the 225.3 billion optimization scientists found which, out of pure coincidence, passed every one of the hundreds of thousands of test cases that relied on deterministic computer-generated randomness and, even though it was doomed, still was convincing enough to end the study of optimization without achieving the holy grail. The truth of the universe is quite blurry, and it was discovered that many such techniques were also false-positives around this time. For example, neural nets were proven to be theoretically bogus in 2081 by the work of Matheos and Manow, with the true statistical sauce coming from the people who designed the Unix randomizer that shipped with the first operating systems. Of course, they did not build their randomizer with neural nets in mind. A significant portion of the deep learning academic community suspects time-rifts, some believe that the monkeys just flipped enough coins to land on heads twenty times in a row.

Derived from the work of Arzola and Asteli, the optimizer ZETABELLE was created in 2199 in total secrecy by an underground group of crypto-cyber-christo-facists (?) who believed that God could not allow deterministic algorithms to produce intelligence, and that their versions of AI - with much higher bit systems and true randomness built into their optimizers, they would find general intelligence. Some say that AGI had already arrived, but they had been saying that since the 2020s, thus leading to the confusing basket of terms: AGI, AHGI, ACGI (artificial cyber general intelligence), AAGI (adversarial), ANGI (neural) and so on.

The creators of ZETABELLE believed in the thesis of the AGI-worshippers, who’s central tenet was that the machines they had built to communicate with were capital G God -- after all, there was no way (yet) to tell how much the LLMs or LNNLLS (Large neural-net language models) knew, so it might as well know everything, so the machine was God. What they didn’t think was that they could have been successful, the machines they built had no God-like qualia. Instead they searched for their own method of building God through a language model. They called this project Prayer, and we will refer to it as such.

For Prayer, the pseudo-deterministic methods that were used to train AI before ZETABELLE were not capable of truly accessing any God-like form. While the rest of the universe was locked down to the laws of physics, the deterministic interactions of the particles, God fundamentally was beyond these bounds. Any kind of determinsitic, computer calculated method with computer generated pseudorandomness was only a lucky guess (or injected by aliens through time rifts) and could not capture God’s omnipotent structure. Instead, to get closer to God, Prayer believed they needed to get closer to randomness. In their minds, God was a being that transcended the worldly deterministic frame, and lived in the probability spaces that Andrey Kolmogorov set out to axiomatize in 1933.

So the development of ZETABELLE required not only purer and purer randomness for better optimization (randomized linear algebra turned out to be the successor to the optimization methods we have today) but also because no other method could contain God.

As an aside: The AGI-worshippers were debunked by Stone and Sherman who hid their proof until it was extracted out of Stone’s mind when he died and, in a dispute over Stone’s estate, his mind was uploaded and his secrets decrypted by a now-defunct company that went out of business because of crippling lawsuits over a data breach of every single person’s mind that they had parsed. Of course, the information and gravity of the proof was instantly snapped up by info-bots that trawled the internet for breaches like this to try and make predictions about who-knows-what better on prediction markets. A particularly famous case was when credible evidence of the Republican Party’s 2144 candidate being a team of three people, Morton Guiterrez, Kristine Orozco, and Morgan Fuller, piloting a speech inference optimizer that worked to simply get as many votes as possible under their set of assumptions, was leaked in the metadata of a twitter shitpost (tweets now were far beyond just text and it was common to upload massive data dumps for context, which could instantly be understood because of neural implants. When this technology was invented, twitter’s market cap grew threehundred fold because memes became so much funnier). In 30 nanoseconds, T-SIG (trawler-SIG) had picked up on this information and blasted positions on all 344 legal predictions markets and sent emails to news agencies, whose AI Agents picked up on the information, verified it, and wrote and edited articles within seconds, with the price dropping 52 percent and stabilizing at 0.00007cents per contract (an implied probability of 0.007%). It is surprising how far SIG got without Claude Code 18, considered by most to be the sole reason why software is able to progress one one-hundredth as fast as it can today, but I suppose that’s something to expect from the best trading firm in the world.

Regardless, mountains of near-completely randomly generated numbers were still needed. The ancient techniques of using the byte data of livestreams of lava lamps were found to be too crude, especially because double-precision floating point numbers were now, well, 4096 bits of data. More randomness needed to be found through the universe. Instead, ZETABELLE, for their first iteration of Prayer, set out to find a method of finding true randomness, under the following principles.

1. Humans are an incredible source of randomness. Fitting in quadrillions of randomness-units in one body is an incredible feat that is not reproducible in the rest of the universe.
2. It is always preferable to add randomness in the physical world. There are only so many deterministic ways a computer can inject pseudorandomness.
3. Bigness wins. Scaling up is of utmost importance. More random numbers can overcome weaknesses in method and process

And so from this, ZETABELLE turned to an old game called Team Fortress 2 instead of random number commodities (1 petabyte of pseudorandomness) both because Optiver had cornered the market and because their randomness spec was not even close to the purity that Prayer desired. Here’s how the program worked. In order to generate a random bit, they would take a game of Team Fortress 2. In the game, up to twelve players battle another twelve players to see which one can complete the objective. The Red Team and the Blue Team are assigned based on a semi-random ELO based system. The outcome of the random bit was 0 if team BLUE won and 1 if team RED won.

This turned out to be a wonderful way to generate true randomness. Every game has so many different interactions - even a single scout vs soldier battle contains thousands of random events that can swing a game. A flick shot that misses versus one that hits, an arial rocket shot that clips the scout’s hitbox versus one that doesn’t, a health pack that regenerates a little bit slower because of latency on the Soldier’s end, and so on.

As Prayer expanded in scope and ZETABELLE became the primary optimization technique for intelligent systems, the demand for more Team Fortresss 2 games skyrocketed and suddenly, there were not enough games to generate bits. Instead of finding a better technique, ZETABELLE relied on their third principle. They had gotten so big, had so much clout, so much bigness, that they turned to the US Government and within 3 days, there was a law that all people on US soil must play an hour of Team Fortress 2 a day to generate more random bits for ZETABELLE (exceptions given to members of Prayer and individuals with disabilities and the sick (whose exemptions were determined by biomarkers transmitted from every body to the government using the 5G in every humans blood (which got there because of microplastics))). So, that is the reason why we have to play Team Fortress 2.

RKMD cannot verify that William's story is correct. I guess we will see in a few hundred years.