CALLING COLLABORATORS: I don't have the expertise to just dash the contents of something like this off from memory, and I don't have the time to do a satisfactory job of researching it by myself. Where I do have expertise is in representing the result in both machine-actionable and graphical formats. This is very preliminary, indeed just a sketch of a README for now, but I hope to put it together with anybody who is interested over the next few months. If you wish to participate, leave a comment on issue #1.

Since the term was first coined in 1955 by John McCarthy in his pitch for the Dartmouth Summer Research Project the following year, "Artificial Intelligence" has meant a great many things, often vague, sometimes mutually incompatible. The purpose of this project is to create a representation of the totality of "AI" — as it has been actually implemented, rather than fictionalized — to educate the public, and so that people can be specific about what they're referring to.

The Story #

I see a fundamental schism in what has at one point or another been called "artificial intelligence", which has to do with the essential implementation strategy of the system in question: is it primarily deterministic, or primarily statistical?

Early attempts at artificial intelligence were strictly deterministic, and the Perceptron, which is the direct ancestor of all — eminently, characteristically, conspicuously statistical — contemporary chatbot products, was excluded from the AI pantheon. It was only with the invention of backpropagation that neural networks redeemed themselves. I view this as significant.

Also significant is the fact that deterministic and statistical systems are constructed, and thus behave — and fail — in fundamentally different ways. Deterministic systems are based on rules, while statistical systems are based on examples. Thus, if a deterministic system fails, it's because there wasn't a rule to handle the situation (or the rule wasn't properly followed). Indeed, statistical systems exist because some rulesets are so complex that it's just easier to derive "rules" from a number of examples. However, all statistical systems invariably exhibit both false positive and false negative errors, which lead to misinformation.

So what we currently call AI is fundamentally behaviourally different from what we used to call AI, in this way and a number of others.

Memory-holing of deterministic strategies #

Decision trees, A*, expert systems, theorem provers, knowledge graphs, all these still exist and still work (quite well, actually), they're just not called "AI" anymore.

Distinctions within statistical strategies #

There are a bunch of statistical "AI" techniques besides neural nets (Bayesian classifiers, support vector machines, Markov chains…) and they're still useful, in fact a lot of them work better for certain tasks than neural nets do (in general, to say nothing of specific LLM products, and a lot cheaper to create and operate too)

Opacity of neural network state #

You don't know why a neural net did a thing because the state space is so big; it also doesn't help for consistency of outputs that they throw a PRNG in there…

Discriminative vs. generative neural networks #

Discriminative neural nets have been around a lot longer than generative ones, are a lot more mature technology, and are a lot cheaper both to create and operate. In fact, a generative neural network is just two discriminative ones in a trenchcoat.

Discriminative neural nets are the things that do all the spooky stuff like facial recognition, developing bioweapons, adjudication on paroling inmates, and military target procurement. These are also almost all purpose-made, task-specific systems. Despite the bloviations of various pundits, ChatGPT (or Claude, or Gemini, or whatever) can't do any of that stuff; it's not even the same species of thing.

The transformer and the AI boom #

the transformer shifted the bottleneck from time to money

Unsupervised learning, RLHF outsourced to the global south, ingesting the entire internet, massive data centre build-out, all these are choices; they're business decisions

Technologies vs. products #

Again, separate specific business decisions made by specific people from raw capability repertoires. There is nothing inherently evil, for instance, about creating a transformer or diffusion model; the ethical issues are in how you do that — specifically the training data, its quantity, its origin, and the means to operate over it — and then how you deploy the result of that process.

There is also the matter that "AI" products, even when narrowed to mean those based on generative pre-trained transformers, now exhibit quite a bit of packaging that is plain ordinary run-of-the-mill software (e.g. the chatbot interface, and "agents" in particular).

Provisional/eventual fusion of strategies #

RAG (current) and neurosymbolic (future)

The Project #

The goal of this project is to produce an artifact to help people ask better questions and make better arguments about the use of so-called "AI" in society. We want to be able to ask questions like:

"Why do we need a statistical solution here? Will a deterministic one not do?"
"Why do we need a neural net here? Will a (e.g.) support vector machine not do?"
"Why do we need a generative neural net here? Will a discriminative one not do?"
"Why do we need a large model? Will a small one not do?"
"Why do we need to involve this vendor? Can we not partner with a different one, or fabricate the component ourselves?"

Disambiguating the anatomy of "AI" gives us hard targets to point at. Grounding the argument in parsimony of engineering complexity, economic outlay, and contractual relationships, enables us to interrogate, from the perspective of a policymaker or prospective adopter, the aggressive marketing of particular commercial products, at the distant end of a particular branch, of a specific limb, of an enormous tree of capabilities.

What I am envisioning, ultimately, is some kind of diagram — or set of diagrams, tailored to specific audiences — generated from an underpinning knowledge graph structure. I imagine the topology as described: an initial bifurcation of overarching technical strategy, and then different inventions and milestones on a roughly temporal basis within those. I imagine the data structure to encapsulate the full set of all entities and relationships, as well as prospective audience comprehensive levels, so it can be used in other contexts and applications. The final artifact will probably be an RDF graph (e.g. Turtle or JSON-LD) file, plus some scripts to generate a visualization, if not a fully interactive piece. I don't know yet.

The story sketch above is undoubtedly incomplete and likely inaccurate. I suppose I could just "ask Claude", but I think it would be more fun and rewarding to involve a community of contributors.

Targets #

I am ultimately imagining this repository will eventually contain a serialized knowledge graph, featuring:

Concepts: my tendency is to make every entity a SKOS concept if it can't be classified as a more specific kind of thing.
Inventions: by which I mean well-defined techniques or ways of doing "AI", or otherwise supporting it (e.g. hardware), their inputs, their general properties, and their applicability range. I see these being a subclass of concept and being also roughly hierarchical in structure (although they may fuse) and of course temporally oriented.
Audiences archetypes that represent groups of people who will ultimately be reading the diagram(s) or other finished artifacts, organized by their respective agendas and what concepts they understand.
Events: both one-off (e.g. when this or that groundbreaking paper was presented at this or that conference, or when this or that product was released) and recurring, like the actual conferences themselves.
People: who wrote the papers, invented the technology, created the products, research at the universities, run the companies, etc.
Institutions/Companies: where said people work; where research is done; where products are made; consortia, governments, etc.
Products: that operationalize the technologies and use other products (E.g. NVidia chips) as inputs.
Works: Scholarly works — papers, books, etc., may as well throw a citation network in there while we're at it; scicomm works that are particularly good at engaging the public; also things like news articles involving entities and/or products, policy documents (although these may be scope-creep-y).
Epiphenomena: suicide epidemic, disruption to education system, GDP-bending data centre build-out, circular investment frenzy, environmental impacts, the various labour impacts, &c.
Annotations: plain-text notes that are none of the above.

A lot of these types already have ontologies available to represent them but I wouldn't rule out needing to patch in some classes and properties. We'll probably determine a mechanism for laddering up e.g. spreadsheets or some other more manageable format (YAML?), since RDF graph data is a bit messy to work with sans special tooling.

As with all knowledge graphs, I don't expect it to ever be "done", but rather can accumulate piecemeal on demand. When there is enough material, we can determine what pictures we can draw from it.

Copyright & License #

This work is licensed under the Creative Commons Attribution 4.0 International License (CC-BY-4.0).

Clone this repository

AI Genealogy Graph #