In 2024, Alex Karp, CEO of software company Palantir, stated in his letter to the investors that the main value that his company was creating was going to be based on ontologies. Since Palantir is in the business of making software models for private and public organizations, I wondered why the CEO would specifically mention ontologies as a key product.

Ontologies have a few defintions, only overlapping to some extent, ranging from philosophy and information sciences to software engineering. So which one are these? What are those ontologies that he talks about? How do they perform, and how do they affect the future?

In order to clarify these questions, I will start by providing a general understanding of software ontologies, then replacin them in the context of modeling through a comparative analysis of the ontologies in two models, the World3 model and the Latin American World Model.

From there, we'll circle back to ontologies as they are presented and implemented within Palantir's sofware offerings, to sketch out the kind of representations latent in their use of ontologies.

We'll see how the predictive power of these world models can actually be understood as a form of fiction.

Let's start with what ontologies are from a computer perspective.

It first looks at the knowledge engineering sources of the term (AI, information sciences), then broadens it up to include any kind of data structures, before comparing it to a philosophical definition to highlight the role of the linguistic system in which these ontologies exist.

This is a quote from a 1993 article by Tom Gruber, which offers a very practical approach to ontologies: how do we design categories of things, such that these categories and their relations enable humans and machines to do productive work?

Gruber takes care to distinguish his conception of ontologies from a philosophical one, particularly along the lines of the rigorousness of definition. While philosophers (and particularly metaphysicians) gratefully indulge in minute definitional work, the point here is not to be exact, but rather to be useful.

This approach also stems from the field of artificial intelligence. At that time, AI was supposed to be reached by being able to formalize all things in the world through catalogs of symbolic concepts ("chair", "color", "obstacle", "friend"). The idea is to be able to reason about the consequences of what was built and stated.

This categorizing of the world from AI would be taken up in information sciences, and is inscribed in markup languages.

The ontology linguistic system presented here is OWL (Ontology Web Language). Specificially, it is a representation of the International Classification of Diseases, one of the main objects of Starr and Bowker's Sorting things out. For things to exist and act in the world, they must be pinned down, sorted out and formalized.

In the case of the ICD (international classification of diseases), Bowker and Starr show how this informational system introduces interesting interplays with the social and labor agencies of constituting and using this ontology in the first place.

Computers being ordering machines, they are particularly apt at this kind of precise, unambiguous (but open!) taxonomy work. You can inscribe generic ontologies about any application domain via proramming languages rather than any non-linguistic systems (such as, e.g., cooking). Particularly if your linguistic system is explicit and formal (logic vs. slang).

All of this labor and controversy is then reified in a linguistic system (that is supposed to work without problems).

In fact, computer programs already provides us with a certain kind of ontology.

The important distinction to make here, and the one that was overlooked in information sciences, is that ontologies can be approached at either the syntactical or semantic level. You can represent a lot of things with OWL but it has to be done with tags.

It is something programming language Niklaus Wirth (and the majority of programmers after him) refer to as "data structures" or "primitives". Wirth's seminal work in applied computer science starts from the observation that much time is spent thinking of the algorithm and the raw data oxymoron, and much less time is spent thinking how exactly is that data modeled (what kinds of entities and relations it has)?

So he proposes to think about the the interaction between technical syntax (int, record, array) and human semantics (which appears upon execution and interpretation).

A system such as OWL focuses on semantic ontologies, on the "what" of ontologies (in our previous examples, it focuses on diseases). A system such as Wirth's data structures focuses on syntactic ontologies, on the "how" of ontologies. In his view, data structures abide by a sort of gestalt effect, by which the whole is bigger than the sum of its parts: complex, dynamic structures can be built from simpler primitive structures.

A similar argument is made by Guy Steele, a prominent programming language designer, whereby any language can be drawn up from basic elements.

What you see here is an instance of a record, also called a class, or an object. This would be an ontology upon which a HR system would model a company.

And it turns out that one of Wirth's fundamental data structures, the record, is particularly good at modeling things from our world.

After emerging from languages specifically designed to simulate the real world, the object (which is basically a record which can also dothings) created a "paradigm shift" among programmers in the 1990s. Everything in our world could be an object, just like everything in OWL is a tag.

Here is another example, from real-life this time. This is the ontology used by the teaching system at the University of Basel. On a day of a software malfunction, the server-side code transpired to the client-side, and showed me how I, as a user attempting to login, was modeled by a combination of the university and the software vendor (who exactly gets to define what in this case would be a fruitful avenue for further research).

Ok, so much from engineering ontologies, let's say a word about philosophical ontologies.

So let's look at concrete cases of the influences and interplay of linguistic systems when it comes to predicting the future of the world.

We'll look at the data structure ontologies, meaning what are the variables used, what are their types, and what is their definition of evolution.

World models are computer models particularly popular in the 1970s, which intend to compute particular scenarios (unfolding of states through evolutions).

A world model has specific characteristics.

It represents and enacts cross-domain states at a macro level whose evolutions are interlinked.

It ambitions to provide insight in the long term (from 50 to 100 years in the future).

And it is made to be enacted by a computer (i.e. it is programmed).

Additionally, it needs to be intelligible to the student of the system while remaining compatible with the limitations of the computer.

So world models are fundamentally about predicting the future. But how exactly does that unfold? Especially given the fact that no model is every truly accurate (even if they are useful, as the saying goes).

The basic concepts of simulation languages are those of event scheduling [simscript], activity scanning, and process interaction [simula])

Event scheduling provides locality of time: each event routine describes related actions that may all occur in a single instant

Activity scanning provides locality of state: each activity routine describes all actions that must occur because a particular model state is reached.

Process interaction provides locality of object: each process routine describes the entire action sequence of a particular model object.

All of these are characterized as worldviews by Michael Lackner in the early 1960s, clearly acknowledging that how you encode the world makes you see things differently. Choosing different technical scripts is choosing different representations.

So what are the fundamental ontologies of World3?

These are positive ontologies, which you see here. The central values from which everything else derives.

From there, you can combine those ontologies and propose the existence of, e.g.,

• food from pollution multiplier (food production decreases with pollution increase, which estimates food production to fall by the assigned values on the first line)
• natural resource from material multiplier (use of resources rises with industrial output divided by population—material standard of living), which interestingly

or human welfare and footprint:

• human ecological footprint
• human welfare index

200+ equations, 80k+ numbers -> needs to be abstracted

Here is how all of them interplay, taken from Jay Forrester's documentation.

You can see in particular on the left side:

there is a self-limiting constraint on population and no self-limiting constraint on capital.

In other words, human growth is not exponential, but human greed is.

These are variables with particular status: they almost negatively affect the other values that depend on them. Less limits does not necessarily imply more production, but more limits always does (significantly, Forrester and his team do not put any limit on capital investment, such as depreciation). These are deliberately thought through as a centerpiece of the system.

And finally, what is usually the hardest to spot, is what it does not show. As a computer model, World3 has to abstract, and abstracting is always excluding. Here are some things that it excludes.

Growth is a primary goal because capital investment is only dependent on environmental factors and not on social factors (which is not even charactized as a relevant entity).

The model traces broad trajectories, and gives results as regular, arbitrary moment. The role of the user of the model is thus to interpret the different outputs of the model and draw consequences freely (such as in the case where there is a rapid collapse of most variables around 2040).

Finally, the abstraction level is quite high: the outlook is global, and the ease of interpreation of the model is due to the broadness of the entities at play.

The success of World3 was not exactly technical per se, even though it did set standards of design strategies and implementations for world models in the 1970s.

What it did rhetorically, is that it introduced the notion of limits as something to be reckoned with in public, economic and governmental discourse.

By entering public discourse, it also prompted reactions.

To answer World3's prediction of a collapsing world under unchanged conditions, and the perceived implication that the south would have to pay for the mistakes of the north, the Latin American World Model was released in 1975.

It is considered to be "bottom-up" in two senses: it is more based on empirical data than World3, and it is done (mostly) by people who come from a southern country (argentina)

This quote is from Graciela Chichilnisky, argentinian economics professor at Columbia University, who designed the economic subsystem of the Latin American World Model.

It deliberately arguments against World3 at the level of ontologies. This example shows the contestation of ontologies, some that might not often be open/visible to contestation. The first implicit question in this statement is rhetorical: is World3 a "world" model or "north-american world" model?

There are four basic needs, which are taken as both starting points and as goals, as opposed to the gross 'growth' of the world3, also starting from the observation that World3 (and thus northern developed society from where it comes) makes no difference between needs and wants.

The method used for measuring output in the dimensional space where the per capita values of economic model is not the same is in World3. It is not a single number of goods produced at market prices. Rather, it studies output as a vector of real goods. There is a multiplicity in the conception of what economic output is that is absent from World3, and visible in the choice of choosing the representation as a vector, rather than an integer.

The intent was not so much to model the future unconditionally (i.e extrapolative, or projective), but rather optimize for some objective function (this is why it is called a normative model (the norm is pre-existing))

Note, this section is very similar to this presentation of Giovanni Fava, who died yesterday.

The conception of the world is to "make it" to reach all basic needs for a given population.

What LAWM models are the development of "feasible paths", aka evolutions towards goals: the point is to reach a given minimum for these three at the same time.

Those goals give a different flavor to the predictive modelling: it's not what happens, as with world3, but how it happens/whether it can happen.

The conclusion of the report is that: "in standard runs, africa and asia don't make it, but in a run where developed countries dedicate 2% of their GDP to aid (instead of warmongering), then african and asia make it". By modeling cross-region captial flows, it implicitly argues for developmental aid from north to south.

The assumptions baked in the model are called policies; in this sense it's also a predictive model of policies.

So what is at stake here? It is about exploring alternatives conceptions of the world.

LAWM sparked a discussion (particularly unesco and ilo), and was attempted to be used as actual planning (see Cole) but it was taken too "seriously", and thus there was a backlash in the 1980s (see Steenbergenen 1994s)

There is a distinction between the conceptual model of an 'ideal' society (hidden) and the mathematical computer model used in the supporting analysis (formal).

World3 and LAWM are not just operating formally, through programs, but as parts of discourses as well, and such positions can only be partially reflected in a mathematical model.

We should note that the LAWM faced more criticisms than World3, given the lack of measurement to conjure up these entities (Life expectancy as a gold standard for policy is more contested than choosing GDP).

Ultimately, both of these models were very limited technically, and criticized for it, but their structure hints at a different aim: influencing the imagination of the users and the public.

And this brings us back to Alex Carp and his letter to the shareholders.

So if ontologies are a foundational aspect of the work of models, what are their role in modeling software 50 years later? How much of it is technical rhetoric or organizational rhetoric?

On the organizational part of the company, it is often referred to as "the ontology".

A survey of the corporate literature coming from the company (reports, marketing documents, tweets) also mentions the term in the singular, with the first letter uppercase. This suggests a sense of importance, abstractness, authority and even primordiality, in the sense that it pre-dates, and thus frames, any further discourse.

As such, the Ontology looms over the public of Palantir's discourse, drawing strength from its un-definedness. So given that Palantir is a software company, and thus in the business of selling software, what is the reality of ontologies on the programming level?

Foundry is one of the products of Palantir, which promises to enable its clients to model their whole business. What they provide is a way to construct objects, in the very old and banal sense of object-oriented programming, which saw an example of earlier.

You can see this on the Titanium product page (which is a fancy name for a Linux distribution). There is a JavaScript console that imports an object from a SDK called Ontology. Check out the actual practical details of providing an SDK and a documentation

In other companies, they would usually call this a model, or a schema, but functionally it appears to be the same thing, but Palantir elevates the ontology to the whole problem/business domain that their sofware models.

Palantir must also contend with the fact that, today, models often mean machine learning models. This use of ontology suggests a layer a absolute categorization/definition over the incertainty of stochastic systems (the same way the softmax function operates). This is interesting but unfortunately there is no time to get into it here.

So what Palantir proposes is a way to model the current state of the problem domain, in order to assess short-term-ahead dynamics and trajectories to give an apparently rigourously defined and reliable big picture.

Interactive, OOP modeling enables the constitution of epistemic artefacts whose manipulation allows users to answer "what if" scenarios (what if a supplied raises their prices, what if market access gets cut off, or what if this target is eliminated, or this police force deployed in this location at this moment?). We're no longer thinking 100 years ahead, and the world is only the company.

This narrow relationship to the future, especially compared to the scope and breadth of the world models of the 1970s, reminds us of the work done by predictive policing software. As Bilel Bendouzid has shown, predictive policiing software is less about predicting where crime will happen, but rather for the police to efficiently manage the deployments of officers on the ground.

But the technical feature of the model is only part of the story. Carp's mention of ontologies is only very loosely correlated to what his software enables.

What Palantir is selling is the dream of dealing with perfect data (neat data structures) than with the messiness of everyday data management (incomplete excel spreadsheets—in fact, one of Foundry's most prominent feature is the cleanup and reconciliation of patchy spreadsheets).

The way Palantir presents ontologies echoes how other organizations have presented digital twins: abstract, vague entities that denote holistic technicity and the promise of productivity and efficiency to decision-makers.

This is more true for world models than, say, engineering models, but according to World3's own authors, world models are "mismatches of methods with purposes, sloppy documentation, absurd assumptions buried in overcomplex structures, conclusions that do not even follow from model output, and project management strategies that destroy the possibility of influencing actual policy".

Computer models never work perfectly, and require large investments of time, capital and effort to develop and maintain at a level of accuracy that would render them indispensable to decision-making.

Since models involve some sort of abstraction, or idealization, they can only be grasped in the hypothetical, the imaginary, while also retaining a strong sense of hypothetical realism. They lay out some explicit rules, and they let us imagine the more implicit consequences (in this sense, they are also not so different from video games).

Models, just like fiction, are a form of communication which involves interpretation.

Meadow, computer simulations as oracles:

In a fiction, characters have an ontological status (does Anna Karenina exist?), and morals of story is the output of dynamic functions (was her suicide unavoidable?).

In the case of World3, the implied statement is that there is that human greed is a fundamental factor to reckon with. In the case of LAWM, the implied statement is that sociopolitical change, not technological change, is required to live within limits.

In both cases, these models provide learning by enabling us to see how a system operates under certain conditions.