Palantir Is Building The "Brain of The Firm" Envisioned Half a Century Ago

Disclosure: The author holds a beneficial long position in Palantir Technologies Inc. (NASDAQ: PLTR). This article is provided for informational and entertainment purposes only and is not financial advice. The author receives no compensation for this article and has no business relationship with any company mentioned. Please see the full "Legal Information and Disclosures" section below.

The helmsman of an ancient Greek ship held his heading against wind and current only by watching the drift and correcting it continuously. Steering is a feedback loop. It compares the state of a system against a target, acts on the difference, and reads the result back in. The Greek word for the man at the steering oar was kybernetes, and Norbert Wiener took it as the name of the science he founded in his 1948 book Cybernetics: Or Control and Communication in the Animal and the Machine. Wiener's claim was that this principle is indifferent to its medium. Feedback loops keep a ship on course, regulate a body's temperature, and govern a machine's speed. Loops of this kind are the basic mechanism of every primitive (artificial) intelligence.

Stafford Beer defined cybernetics as the "science of effective organization," founded management cybernetics, and treated the company as a system of exactly this kind. A company, on this account, is an artificial intelligence in the strict functional sense, a made thing that takes in information, decides, and acts on the world with capabilities no single member has. Its feedback loops are built from its structure. Budgets set the targets, reporting lines carry the deviations upward, reviews compare plan against result, and standing procedures answer the routine cases automatically.

Organizations from the Roman legion to the nineteenth-century railway to the modern tech company are intelligences of this primitive kind. Beer spent the 1960s working out what such a system needs to stay viable and set it down in his 1972 book Brain of the Firm, a model of the firm patterned on how the brain coordinates the body. He believed it enough to try it on a country, wiring Allende's Chile with telexes and a mainframe until the coup of 1973 ended the experiment.

What this intelligence of organizations lacks is a faculty of its own for thinking. It senses and it acts, and its procedures even give it reflexes of a kind. But everything beyond routine is thought for it by the humans it employs, who hold its knowledge in their heads and reach its judgments by talking to one another.

Beer's Viable System Model laid out the five functions any viable organization must perform: operations that do the work, coordination that keeps the operations from colliding, control that runs them in the present, intelligence that scans the environment and models what is coming, and policy that settles what the whole is for. Optimize all five and the dependence on human intelligence remains. The intelligence function, the part that models the world and looks ahead, has always depended on humans, because only humans are capable of the thinking it demands.

Ross Ashby measured the difficulty of any control task in variety, the count of distinct states the thing to be controlled can present, and every state a controller cannot answer is a state it cannot control. His law of requisite variety says that a controller must hold at least as many distinct responses as the system facing it has states. Beer's version was shorter. "Variety absorbs variety, and nothing else can." A company's world takes more states than anyone can list. Demand shifts, a competitor cuts prices, a supplier fails, a machine breaks, a regulation changes, and every combination of these is a situation that may demand its own answer, while a management team can hold only a few things in mind at once.

Requisite variety settles how many responses a controller needs. It says nothing about which one to pick. A company whose sales are falling can cut prices, rework the product, or replace the sales team, and the right choice depends on why sales are falling. To choose, the controller has to predict, to ask what would happen if prices were cut and what would happen if they were not, and prediction requires a picture of how the company and its market work.

In 1970 Roger Conant and Ashby proved the requirement formally. "Every good regulator of a system must be a model of that system." Whatever controls an organization must contain a working model of it, and no such model has ever existed in one place. It has lived only as fragments, scattered across employees' heads and the organization's records. The organization's dependence on humans comes down to two services. They supply the responses, and they carry the model.

Generative AI supplies half of what was missing. For the first time a machine can take in concepts expressed in ordinary language and reason over them rather than merely store and compute. Yet in the fourth year after the release of ChatGPT, AI has swept through software development and barely arrived in the rest of the economy. A 2025 study from MIT's Project NANDA found that 95% of corporate generative AI pilots delivered little to no measurable impact on profit and loss, because the tools were never integrated into the work. Individual humans consult a chatbot and carry the answers back into the company's processes by hand. The AI sits outside the loops the company runs on, and a chatbot that augments the intelligence of a single employee does not make the organization more intelligent.

For an artificial intelligence to run inside an organization rather than beside it, the organization itself must exist in a form that intelligence can act on: its objects, its processes, and the relations among them. The data sits scattered across systems built in isolation that were never meant to speak to each other, and the same real thing, one customer or one machine, might appear across a dozen of them under a dozen identifiers in a dozen incompatible forms. Worse, the records merely recount events without capturing their meaning. To produce real value, generative AI needs a living model of the organization, something to reason over and act through. That model is the missing half. Joined together, the two halves make a genuinely intelligent system, and the system is the organization itself, holding at last both a model of itself and something that can think about it.

In 1972 the psychologist Endel Tulving separated episodic memory, the record of particular events fixed in time and place, from semantic memory, the web of concepts and relations that holds what those events mean. This second kind lets a mind reason past its own experience into situations it has never met. An organization keeps an exhaustive episodic record and distills almost none of it into meaning. It remembers everything that has happened and understands none of it, because the understanding forms only in the humans it employs.

Palantir's ontology amounts to an attempt to give the organization that missing semantic memory. The work begins with integration, drawing the scattered data out of its separate systems into one place and keeping it current as the sources change. Over that, the ontology maps the records onto the things the organization runs on: object types, the nouns of the enterprise (a customer, an order, a machine, a shipment), each with its properties, and link types, the relations that join them. One real entity becomes one object wherever it appears, so a machine is a single thing carrying its maintenance history, its sensor feed, and its parts. Palantir describes the result as an operational layer for the organization, a digital twin built to represent how an enterprise's decisions interconnect rather than the data alone.

The ontology also defines actions, the verbs of the enterprise, each an operation that changes an object's state under set rules and writes the result back to the source systems. Palantir's own documentation names the two the same way, calling objects and links its semantic elements and actions its kinetic ones.

A delayed shipment, a sensor fault, a thinning stock level, and a high-priority order are four facts in four systems, each blind to the others. In the ontology they are linked objects, and the links make the consequence legible: the fault, the thinning stock, and the delayed shipment together threaten the order. What an experienced manager would once assemble by hand now stands in the ontology, ready to be acted on. The organization holds this semantic memory itself instead of borrowing it from humans.

AIP, Palantir's AI platform, puts the language model to work inside this structure. An agent reads the linked objects and infers what they imply, and because its moves are limited to the defined action types, it can do more than advise. It can reserve the substitute part, expedite the replacement shipment, and flag the order at risk, or propose those steps and wait for a human to confirm, exactly as each action's rules require. The decision writes back through the same actions, every source system updates, and the loop that once ran through inboxes and meetings closes inside the ontology. Because the concepts and relations are already represented, the agent can handle situations it was never shown, and because it acts only through the ontology's verbs, it works inside the company rather than chatting beside it.

The ontology could be the working model of the organization that Conant and Ashby proved every regulator must contain, now existing as something a machine can hold. The agents supply what management teams never had enough of, responses generated at machine speed and in machine numbers, while the ontology's object types filter the world's noise into a finite set of states a decision can address, which is amplification and attenuation in the same architecture. The write-back gives the loop what Beer's telexes could not, correction applied in the moment it is decided. And the intelligence function, the seat Beer could only staff with humans, would have its first non-human occupant.

Beer had already taken the next step. A system is viable, he held, only if its intelligence function carries a model of the system itself. The Conant-Ashby theorem says the same from the other side: when the regulator sits inside the organization and the system it regulates is the organization, the model it must contain is a self-model.

In the engineering literature, a self-aware system is one that maintains models of itself, its context, and its goals, and reasons over those models in order to act. An organization that holds a live model of itself, consults it before acting, and watches it change as the consequences return meets the definition. What is being assembled might be the beginning of functional corporate self-awareness.

Whatever one calls it, an agent working inside a company runs on tokens, and a single task is many model calls. Turned loose on raw data, the agent spends most of those tokens orienting itself: loading tables into context, retrying, second-guessing its own answers before it can act. The ontology is built to cut the waste at the root. The agent starts from resolved objects, the links among them, and a fixed set of permitted actions, so its tokens go to the decision rather than to reconstructing the situation first.

On July 1 Palantir's CEO and co-founder Alex Karp took a blunt version of the same case to CNBC's Squawk Box. The occasion was an expanded partnership with Nvidia that packages open models inside Palantir's platform, so that government and enterprise customers keep control of their own weights, data, and compute. Karp used the airtime to attack how OpenAI and Anthropic sell AI. Enterprises, he argued, appear to have accepted spending tokens for little measurable return while risking their intellectual property, largely because the raw models may have been marketed well beyond their actual utility. Tokens burned for no return buy reasoning with no model of the firm for it to work on. The Nvidia partnership reverses the flow: rather than ship the customer's data out to the model, it brings the model inside the structure that already holds the customer's reality.

Every ontology Palantir builds is a map of how one company's world decomposes into objects, links, and actions, and companies in the same industry run on roughly the same world: one hospital network resembles the next, one insurer the next, one manufacturer the next. The first map of a sector has to be worked out from scratch. By the third, Palantir already knows the shape it will take, which object types recur, how they tend to link, where the hard cases hide, and what an agent needs in order to act on them. Each client's data stays walled off behind its own controls, but the knowledge of how to build the map carries across, so each new deployment in a sector should be quicker to stand up, cheaper to deliver, and better than the ones before it. Each deployment then teaches Palantir more about the sector, the knowledge improves the next one, and the better product wins more of them, a flywheel turning on knowledge. Advantages that grow with use and would have to be re-earned from zero are, I think, the hardest kind to compete away. The firm that supplies the brains comes to know the anatomy of every body they are fitted to.

Organizations are still the most consequential artificial intelligences in existence. They run the hospitals, move the freight, approve the loans, and fight the wars, and for the whole of their history their thinking has been borrowed from the humans they employ. The ontology, with machine reasoning working inside it, is the first serious attempt to change that arrangement. Half a century after Beer set out to give the firm a brain, organizations might be given a real functional intelligence of their own.

Follow me on X for updates (@chaotropy).

Legal Information and Disclosures

General Disclaimer & No Financial Advice: The content of this article is for informational, educational, and entertainment purposes only. It represents the personal opinions of the author as of the date of publication and may change without notice. The author is not a registered investment advisor or financial analyst. This content is not intended to be, and shall not be construed as, financial, legal, tax, or investment advice. It does not constitute a personal recommendation or an assessment of suitability for any specific investor. Readers should conduct their own independent due diligence and consult with a certified financial professional before making any investment decisions.

Accuracy and Third-Party Data: Economic trends, technological specifications, and performance metrics referenced in this article are sourced from independent third parties. While the author believes these sources to be reliable, the completeness, timeliness, or correctness of this data cannot be guaranteed. The author assumes no liability for errors, omissions, or the results obtained from the use of this information.

Disclosure of Interest: The author holds a beneficial long position in Palantir Technologies Inc. (NASDAQ: PLTR), whether through stock ownership, options, or other instruments, as of the date of publication. The author holds no position in NVIDIA Corporation (NASDAQ: NVDA) or in any other company mentioned. The author reserves the right to buy or sell these securities at any time without further notice. The author receives no compensation for this content and maintains no business relationship with any company mentioned. OpenAI and Anthropic are privately held, and the author has no financial interest in or business relationship with either.

Forward-Looking Statements & Risk: This article contains forward-looking statements regarding product adoption, technological trends, and market potential. These statements are predictions based on current expectations and are subject to significant risks and uncertainties. Investing in technology and growth stocks is speculative, subject to rapid change and competition, and involves a risk of loss. Past performance is not indicative of future results.

Copyright: All original content, including text and images, is the property of the author and may not be copied, reproduced, or published, in whole or in part, without prior written consent, except as permitted by applicable law or the terms of the platform on which it is published. Use of this content for training machine learning or AI models is not permitted without explicit authorization. Third-party or public domain images remain subject to their respective rights and are not claimed as the author's property.