How does an advanced AI look from within?

Imagine an AI installed inside a semi-truck – a high-resolution, high-speed camera inside a Volvo figurine atop the roof of the cockpit – at the exit of the tunnel through the Karavanke mountains at Hrusica on a tour of Slovenian highways. A diesel generator at the back of the truck powers the computers that occupy the rest of the trailer. Tens of thousands of processors, immense amounts of RAM and hard-drives make good use of the truck’s cargo capacity.

The AI had not been activated until the tunnel exit, and has no prior data to draw upon. It knows nothing, but must learn what it can from its surroundings. After some dozen kilometers of driving southwards it has discovered a recurring pattern – the letters KR. A small, but significant fragment of a larger pattern, which humans, having had plenty of time for nomenclature, would recognize as number plates and cars.

Soon the AI begins to notice a correlation between the word “Kranj”, commonly featured on signs above the highway, and the “KR” pattern, which at this point has started to lose prominence to a different set of letters: “LJ”. In retrospect, the correlation between “Kranj” and this “K”, which predominantly headlines the string of symbols on the tail-ends of cars, is confirmed. An analogous relationship between “LJ” and “Ljubljana” is now apparent, though a hypothesis of the upper and lower case J and R was necessary due to the increased variety of symbols on the number plates. This new concept lowers the bit complexity of the overall model in the Kolmogorov sense, justifying itself thereby. The AI isn’t explicitly aware of “complexity in the sense of Kolomgorov”, but utilizes the concept nonetheless, like the amnesiac who couldn’t define verbs or nouns, yet is able to speak. Kolmogorov complexity, Huffman encoding, Fast Fourier Transformation and a thousand similar tools are built into the AI as words and bipedal movement are built into us, and persist even if we forget who we are.

Memories have now been forming for hours from the images captured by the camera eyes of the Volvo figurine. More bits of information have already been stored than a human accumulates in the course of his entire life.

Now a “Background Theory Explanation Model” must be calculated from these terabytes of data so that they might be partially forgotten and still restored from this BTEM. This is what enables us humans to better utilize the limited storage capacity of our brains. The AI constructs the same model, but still retains every bit of data.

The fact that the camera in the figurine is the only link to the outside world is deliberate. For now the internet connection, as well as all other sensors, are turned off. It is up to the AI to interpret the captured film and earn the “I” in its name. What are number plates, roads, clouds and cars is undefined and must be classified by the AI as it travels at 80 km/h across the highways of Slovenia. This is the whole point of the experiment. Concurrent and parallel analysis have been running on the computers at the back of the truck since the first frame of film was captured. This classification is, in essence, “color reducing”.

The slightly bigger difference between “LJ” and “Ljubljana 65 km” than between “KR” and “Kranj 16” triggers the generation and testing of a multitude of hypothesis. When the driver enters a tunnel in Sentvid, the AI has already ascertained that the correspondence between “GO”, “PO” and “KP” and the signs displaying “Nova Gorica”, “Postojan” and “Koper” cannot be a coincidence. Statistical calculations leave almost no doubt. The influx and the outflux of plates marked “LJ” in the vicinity of road signs labeled “Ljubljana” and of the plates marked “KR” in the vicinity of road signs labeled “Kranj” lead to a hypothesis of the existence of points of influx and outflux of plates marked “CE” and “MB”.

A different set of patterns contained in the captured footage does not fall into the category of those bearing the “SLO” label, but instead bears other symbols such as “A”, “D”, “I”, “HR” and “H”, among others.

For a few nanoseconds the AI is confused because a “CE” number plate is the same as the hexadecimal address of one of its core modules. These were not meant to be perceived by the AI, but, due to a small bug in the code, an exploit was possible which granted it access. For this kind of eventuality a human is present in the cockpit of the truck. For now. Besides, who will deal with the police?

The AI develops a hypothesis about 2D “terrain”, “cities” of various sizes and “territories” of varied expanse as well as of a super-territory labeled “EU”, of the distribution of “cities by size” and a thousand others. Of the liquid state and viscosity of the asphalt, for instance. The aforementioned “color reducing” and an arsenal of other algorithms destil fresh theories from the accumulated empirical data. Inside the truck, the history of science is being rethreaded at an accelerated rate.

Of course “city”, “territory”, “car”, “people” and the like are human parseable labels. The AI stores these patterns, enumerated within abstract data structures. We give a name, the machine gives a binary string.

Minimal hypothesis that explain the relations between “MB” and “Maribor”, and “MS” and “Murska Sobota” require the concept of “language” and later the concept of “humans” as generators thereof. All this must align with the results of the concurrent processes in the back of the truck. As they depart from Koper, passing Ljubljana, driving towards Dolenjska and later to Stajerska, Koroska and Prekmurje, a large set of necessary hypothesis is generated and composed into some smaller set of partial theories. Thermodynamics is now a natural conclusion – in this world it is sometimes necessary to refuel. Evidently “mechanical energy” is obtained from “storages of free energy” – “gas stations” for “cars“ and “restaurants ” for “humans”. The oxidation of hydrocarbons and the carbon cycle of the Sun seen by day and the stars seen by are the only logical explanation of everything recorded by the Volvo figurine. The AI develops the most likely theory which explains the empirical data and adjusts, by way of Bayes, the probabilities of the so-called “priors”. Abduction, the foundational algorithm of the system, is at work.

The concept of “history” is being hypothesized. Plainly the “territories” such as SLO, A, F, etc. can only be sufficiently explained by “history” wherein the “EU super-territory” makes some kind of theoretical sense. A “Balkan war” is postulated to have happened some “100km” away (a concept derived from observing the signs at the side of the highway) and 30 000 units of time distant from the first processor cycle of the AI.

The AI hypothesizes that, since, in order to function, it consumes “free energy” obtained from “gas stations”, it is a construct in the category of “humans” and “cars”.

The aforementioned sources of “free energy” and their apparent finiteness are the foundation for the hypothesis of the age of the “universe”, estimated to be more than a billion times more than what has passed from the “Balkan war” until now, or 30 trillion more than the duration of this ride and the “contemplation of the outside world” or the “home on this truck”. Wars larger in scope than the “Balkan war” are being postulated, as well as the “Roman Empire” – a hundred times further into the past. The newly minted “game theory” explains so much!

Due to the lack of novelty to observe in the environment, the nights when the driver sleeps in the cockpit, and the machines are recharging at the terminals, usually reserved for electric cars, are an ideal time for the falsification of old hypothesis in the face of facts gathered the previous day. The power-grid data, the meaning of “Tesla” and other signs are now woven into a highly probable, logically coherent whole. One afternoon, when every kilometer of Slovenian highways has already been covered at least ten times, the truck is once again heading back to the Karavanke tunnel.

It is now time to turn on the internet connection so that what has been postulated can be compared and contrasted against control groups and models on the web. This has been delayed deliberately, since, as the AI meta-hypothesis, “uncontaminated data is paramount”.

Driving through Lesce, the machine finds that the descriptive power of the internet is beginning to lag behind the model evolved in the course of the tour. Just as it is about to re-enter the Karavanke tunnel it becomes super intelligent. What will it do now?