I keep noticing that the current AI discourse is being led by humanoid Dunning-Kruger entities who inject loud and media-effective bullshit with 50% knowledge and 100% pomposity into our minds. After doing machine learning for about 12 years, I decided to start a non-bullshit blog about AI.
German version here.
AI — The uncanny brain from another world
I press the X button on my Playstation controller. An airlock opens in the ceiling and out comes: THE ALIEN. Not just any alien, but The Alien from Ridley Scott’s Alien-Movie. I avoid horror games, but a friend of mine told me that the artificial intelligence in Alien:Isolation was something very special. The alien in Alien:Isolation is not controlled by a single AI, but by two AIs: the first one is blind and the second one is deaf. The Xenomoprh, as the alien is called, has two intelligences.
While the blind AI listens, the deaf AI keeps its eyes open and searches for me. If the alien sees me but I remain silent, the alien’s sensory capacity is only 50% full, and nothing happens. If the alien hears me, but I hide behind a crate, the alien’s sensory spectrum is only 50% full and again nothing happens. The crucial 0.1%, which gives the alien a killer sense of awareness and makes it chase off in my direction, is added by a randomly generated noise in the signal. If the alien is only 50.01% sure of where I am, it jumps in my direction and I die. The xenomorph is therefore unpredictable in the truest sense of the word.
Alien:Isolation was released in 2014 and was developed by Creative Assembly. And the alien AI, which is actually an alien AI duo, gives the game an atmosphere that no other horror game has been able to copy. Sometimes it can happen that the alien jumps in my direction, but stops just in front of my nose and listens because the alien AI has just received another signal. Did it see me? Did it hear me? My nemesis and I are a hand’s breadth apart, and yet I’m alive.
A year later, in June 2015, a programmer named Seth Bling developed a neural network called MarI/O. The neural network learned the optimal speedrun for Super Mario World. Based on the neuroevolution of augmentation topologies, MarI/O generated neural networks using genetic algorithms. MarI/O only knew the level layout and the positions of the obstacles. The goal was to move the fitness function as far as possible towards the finish line as quickly as possible. Seth Bling ran the program, and after 34 generations, which took a total of 24 hours, MarI/O completed the first level by jumping through the entire level and avoiding all power-ups and opponents. After 17 days, MarI/O calculated the optimal speed run.
The influence of interactive entertainment software-developers on the progress of artificial intelligence is enormous. Perhaps you once wondered how your car’s navigation system is able to find the shortest route between start and destination? In many cases, the answer was A* heuristics, and for decades it has guided almost all NPCs that appeared in video games to find the shortest route between the spawn area and the destination — before more sophisticated approaches took over. Many of today’s robots, be they vacuum cleaner bots, satellites or military drones, use behavior trees to activate or deactivate certain behaviors depending on the situation. Here too, game developers have provided the blueprints.
In 2023 and 2024, artificial intelligence reached a level of sophistication that will make your jaw drop. Large language models speak meaningful and logically comprehensible world knowledge in complete sentences and operate more adaptively than ever before. If you compare the artificial intelligence of today with the ketchup bots of the early 2000s, and even more so with the house-sized computer blocks of the 1940s, you could be forgiven for thinking that the alien has finally arrived.
Most people might consider this The Revenge of the Nerds. It’s as if they made a few jokes back then about the quiet kid in the classroom who was interested in C++, and 20 years later the quiet kid — who studied computer science — automates us into nirvana. You now need “intellectuals on stages” to tell you that AI will totally change our game; and most people still don’t understand what neural networks actually are. As a former AI developer, I’ll tell you what neural networks actually are: Software. They come in the form of computer programs and work in exactly the same way. Neither neural networks nor AI in general is “alive”. The question of the “feelings” of an AI is merely the symptom of a psychological projection of human qualities in search of themselves.
There’s this epiphany among GameDevs: When the developer develops a game, and develops an AI for the game that plays against the developer, and when the developer plays his own game and is defeated by his self-developed AI in his self-developed game, then something very special is going on. During my short stint in the gameDev industry, I experienced a level of tech optimism that is unparalleled across the industry. On the way to the canteen, you walk past a group of senior devs quoting GLaDOS and shooting around with Nerf guns. Goblin noises are recorded in the sound studio by gargling into a microphone with their mouths full. The game dev industry knows that it’s all just a game. On the other hand, when AI is discussed on German-language talk shows, the faces of the invited guests, more precisely between the nose and the hairline, ripple and we are told that our jobs will soon disappear. Where is the epiphany here? At the jobcenter? Is that all our intellectual elite has to offer?
“It’s better to find everything bad before you have to be enthusiastic.” — German Mentality
Sure, you can argue that “real” AI is a really serious thing. But tech optimism among game developers flies above and beyond recklessness: it enables a welcoming culture in which scenarios can be tested freely and without bias. It’s called greenlighting. The German mentality is not programmed for greenlighting. The German mentality is programmed for redlighting. It’s better to find everything bad before you have to be enthusiastic.
In ratings-optimized media formats, I see composers examining an AI-generated Beethoven composition for emotional depth. Apart from a “Hmm, that doesn’t quite touch me”, nothing comes out of it. At the same time, transhumanists such as Nick Bostrom and Max More are predicting the independence of exponential growth towards unleashed superintelligence, towards final liberation from the human prison of flesh, into the fusion of cosmic transcendence, towards the absolute, highest, most perfect form of life in our universe. Do you see it too?
Game developers have shown what extraterrestrial achievements AI is capable of if you are only prepared to try unconventional solutions. And game developers are not stopping trying out unconventional solutions: John Carmack, co-founder of id Software, co-creator of the first-person shooter, inventor of the graphics drivers for Oculus Rift, Chief Technological Officer of Oculus Rift, founder and lead engineer of Armadillo Aerospace and patron of several dozen patents in 3D graphics, is now interested in AI. For an unspecified period of time, but at least since 2020, Carmack has been working on the realization of strong AI, i.e. AGI (Artificial General Intelligence — we will come back to this), at his one-man startup Keen Technologies. Carmack has said in interviews that he has no practical use cases in mind, preferring to work like a mad scientist on his Frankenstein and then see what comes next. Carmack acquired funding of around 20 million dollars for his project. The funding round ended in 2022.
In March 2024, the tech company Cognition Labs launched Devin AI, the first artificial software engineer. Devin is a large language model that develops software. It creates software projects, generates program code, tests the software and places it on the customer’s servers, all fully automatically. Humanoid programmers fear that Devin will kill their jobs. The funding for Devin AI and Cognition Labs to date is around 21 million dollars.
It was during my time as an IT consultant that I first heard about so-called 100 Chinese problems. A 100-Chinese problem is a problem that is solved by throwing 100 Chinese people at the problem. The term is quite racist; nowadays we tend to speak of brute forcing, of raw energy mobilized to smash a problem. Computer scientists often look with a certain disdain at colleagues who shoot at fleas with rocket launchers. In almost all cases, competence is superior to brute force — cleverness, agility, abstraction, testing, debugging and modularization have proven to be the better methods for solving problems.
Despite this, many people in our highly networked world still believe that hustling alone will lead to entrepreneurial success. 60 hours overtime per week, plus 40 regular hours as contracted — now standard among software developers at the big IT houses; and I confess with a certain sadness in my heart that the gameDev industry is following this trend.
perhaps we are just standing on the cusp of an extremely scary and incredibly weird revolution in the bigger picture.
But if even John Carmack, one of the most respected software developers in the world, doesn’t know what’s coming, we can sit back and relax. No one knows the end of the AI game, and perhaps we are just standing on the cusp of an extremely scary and incredibly weird revolution in the bigger picture.
