Akinator has been floating around on the internet in his ephemeral genie form since 2007, and continues to wow people even today. Think of any character, object, or place, and just answer Akinator’s volley of questions. And, hey presto, the genie has already read your mind!
Alright, it’s not magic. It’s just a simple tree search algorithm. Akinator’s guesses resonate with Bing Chat or ChatGPT, or when trying to create an image with Stable Diffusion or Midjourney. Modern AI algorithms often invoke a sense of incredulity and curiosity, much like watching a magician perform without revealing his tricks. This leads to the question, at what point does a program cross the threshold from being a simple algorithm to an AI agent?
What Makes Akinator Tick?
Akinator was first developed in 2007 by a company called Elokence, which specialises in ‘man-machine dialogue solutions’. According to their website, the genie accesses a database of over 100,000 characters updated daily and combines it with an ‘algorithm that manages the choice of questions’. While the company remains tight-lipped on the nature of the algorithm, a dig through Stack Overflow’s forums spills the beans.
In this thread from 10 years ago, there is a split consensus on how Akinator’s algorithms seem to work. While many responses have stated that the genie uses binary trees, some others concluded that it was a fuzzy logic expert system. Even as the true algorithm behind the service remains Elokence’s ‘little secret’, it does not seem to be what we call AI today.
Akinator seems to use some form of decision tree algorithms, mixed with Elokence’s secret sauce and tries to guess a character based on its database. Currently, the only algorithms that can use AI to do Akinator’s job are neural networks, but these were not widely used in 2007, certainly not for a browser game.
Even though the service uses a fairly basic algorithm, it punches above its weight class in terms of evoking curiosity about its inner workings. The fact that the character is a genie is, of course, part of the charm, but every user playing the game is left stumped at how Akinator guessed their character. For someone using ChatGPT for the first time, watching lines of coherent, well-written text appear in front of their eyes is nothing short of a magical experience.
Even though ChatGPT was built after years of research and development by OpenAI, the end user only sees its text generation capabilities. It is also difficult to explain the concept behind LLMs without either oversimplifying or providing a condensed machine learning seminar. This technological marvel takes us back to Arthur C Clarke’s third law: “Any sufficiently advanced technology is indistinguishable from magic”.
This explains why Akinator and ChatGPT might seem like magic to a layman, but comparing the algorithms behind them is akin to comparing a stone wheel to a supercar.
What makes modern AI
Assuming that Akinator uses decision trees, it is built on one of the most fundamental concepts of AI algorithms like recommendation engines, prediction systems, and fraud detection mechanisms. In fact, decision trees are just one of the components that make up modern AI algorithms.
While decision trees and other statistics-based computing methods have been used to build the foundation of AI, modern algorithms have progressed to more complex concepts like diffusion and transformers. Basic concepts like decision trees have been relegated to the wayside, consigned to the arena of being beginner projects for budding machine learning engineers. State-of-the-art models, such as how Akinator would have been at one point in time, now combine multiple complex concepts to reach a closer approximation of human intelligence.
Let’s take Stable Diffusion for example. The core of the model is already a complex piece of technology that employs reverse diffusion to create an image from noise. To make this expensive algorithm run effectively, the agent makes use of another neural network called a variational autoencoder that cuts down on compute costs. To ‘understand’ the users’ input, Stable Diffusion also makes use of an NLP algorithm, which comprises a tokenizer and text transformer.
As we can see, using a simple decision tree with a few extra steps is a far cry from Stable Diffusion’s image generation capabilities, but Akinator still wows people the same way that SD does. On the other hand, we have algorithms like NVIDIA’s Megatron, which is extremely capable as an LLM, but no forward-facing execution to expose its capabilities to the layman.
Even as we make strides in creating more powerful algorithms, AI researchers must keep in mind that they are developing these agents for human use. Something like an Akinator still gives the flavour of intelligence even when based on simple algorithms, simply due to the fact that it was developed with the intention to do so. ChatGPT gives human-like responses which wow its users due to its ingrained ‘humanity’ given to it through reinforcement learning with human feedback.
Scientists today are pushing the boundaries of what is possible with AI algorithms, but an integral part of creating these agents is embedding them with human-like traits. Leaving aside the Turing test, what people perceive to be an actual artificial intelligence can be found at the juncture of cutting-edge technology and the measured implementation of human-like traits. One only needs to look at Microsoft’s Bing Chat to see the drawbacks of adding too much ‘humanity’ to an AI chatbot, and countless examples of cutting-edge technology lie in the attention graveyard, waiting to be made into a product which can wow the next generation of AI users.
