Artificial Intelligence has rocked the world lately, but that's a rather loose term. Some would say it is even an inaccurate term.
What has rocked the world lately is Generative AI - models that generate content on a level that can seem like genuine human-generated art to an untrained or gullible eye.
Stages Of Artificial Intelligence
Let's briefly go through the rough phases of AI. Be warned, the terms are not very precise, and different people involved in the field carry different definitions for each term.
For example, what I refer to as artificial intelligence, some would call machine learning.
Machine Learning
What I call machine learning, some might call "best-fit models".
These are simple mathematical models where an algorithm is defined, but some coefficients are left unspecified. The training process then calculates (not "estimates", "calculates") the coefficients that fit the training data best. The inner workings of these models tend to be fairly simple for humans to understand.
Deep Learning
These are neural networks, which is another way of saying "thousands of simple mathematical models working to create one big model". From this point onwards in the complexity scale, the training process estimates the best coefficients, because calculating the exact numbers ranges from impractical to impossible. This also steps into a region of complexity where humans cannot understand the inner workings of the models.
Generative AI
A catch-all term for large language models (chatbots), latent diffusion models (text-to-image), and video diffusion models (text-to-video), which can create content that is remarkably complex. This is the current state-of-the-art in AI.
Artificial General Intelligence
Some would say AGI is the only true AI, and it is indeed the ultimate goal we all picture when we think of AI. A computer algorithm that can do anything, with intelligence coming not merely from computing power but an ability to understand the world.
Does ChatGPT Understand The World?
There was a large language model that could provide really accurate navigation within New York City. However, its performance plummeted when some streets were closed, because it simply hallucinated a lot of non-existent streets. Navigation apps are built with a model of reality specific to their one goal, and so is their training data. The same is not done with generative AI.
GothamChess did a similar evaluation in a video pitting ChatGPT against Stockfish (the best chess playing and evaluating algorithm), where ChatGPT was unable to keep track of the state of the chessboard, bringing in pieces out of thin air.
Large language models in particular are only good at determining what word comes next in a stream of text. That doesn't mean that they actually understand what they write, reminiscent of the urban legend about Max Planck and his chauffeur.
What Does An Open Source AI Look Like?
At first glance, one would say that any model whose underlying mathematical model and its coefficients can be verified can be called an open source model.
However, that definition rings hollow when you consider the collaborative and transparency aspects of open source. We do not start from a god-given definition of open source which we worship - we start with a set of principles and behaviour we value, explore how they apply in different fields, and then create field-specific definitions. Hell, "open source" is just "public domain intellectual property" applied to software.
The Open Source Initiative, an organisation that aims to be, and is widely regarded as, the steward of all things open source, has defined open source AI as AI which provides enough information to substantially rebuild it. This includes a training dataset, and how it was obtained and processed.
This definition is intended to evolve over time as the field expands.