PREFATORY EXCHANGE AT PHILADELPHIA

General Washington, taking Dr. Franklin aside, inquired:
“I have now spent considerable time among soldiers, farmers, and statesmen and find that all now speak of a new science they call Artificial Intelligence. Could you please take the time to explain to an old Virginia planter, revolutionary soldier, and reluctant politician what this engine truly is, how it functions, and what is required to make it productive in the practical affairs of men?”

Dr. Franklin replied:

“Your modesty does you an injustice, General. A commander who has surveyed the wild lands of the Ohio, balanced the seasonal husbandry of Mount Vernon, and calculated the ordnance and flour required to sustain a continental army already commands the exact faculties needed to master this subject. Let us set aside the mystical jargon of the computer scientists and examine the apparatus not as a sorcerer’s charm, but as a vast, disciplined extension of human labor and recorded precedent. 

What follows, General, is the plain anatomy of that engine.”

I. The Essence of the Mechanism: Precedent over Intellect

Artificial Intelligence is neither a living mind nor a conscious spirit. It is an automated library endowed with the capacity to cross-reference its contents at the speed of light. It does not think as a human thinks; it compares, synthesizes, and echoes.

Consider the compositor’s stick in a traditional print shop. A typesetter does not invent the alphabet anew each morning; his hand moves across the type-case by familiarity and habit. After a "q", he reaches automatically for a "u"; following the words "Continental" and "Congress", he anticipates a comma or a period.

The artificial engine operates upon this identical principle of precedent, expanded to an unimaginable continental scale. When presented with an inquiry, the machine does not engage in private reflection to discover what is true. Instead, it searches its colossal record of past usage and calculates, token by token, what fragment of a word is statistically most probable to follow the last. It is not an intellect discerning truth; it is an automated compositor assembling the mosaic of what humanity has already recorded

II. The Quarry: Assembling the Ocean of Human Knowledge

 Before a single gear can turn or a single calculation can occur, the builders must gather the raw fuel of the machine: the recorded knowledge of mankind.  I know this may sound audacious, but it is what happens.

To achieve this, the architects scrape and assemble an ocean of digitized text spanning virtually everything set down in print or digital form. While the iron transistors at the bottom of the machine register only the opening and closing of electrical circuits—the simple ones and zeros of the artificers—the engine itself transcribes our language into rich mathematical coordinates. The text is first sliced into discrete syllables and word fragments, denominated tokens. Each token is given a simple number tag, which the machine then plots like latitude and longitude upon an immense mathematical map of human meaning., and the colloquial chatter of the global public square.

Into this vast reservoir flows the summit of human accomplishment—the laws of nations, judicial treatises, scientific transactions, histories, medical discoveries, mathematical treatises, and philosophical discourses. Alongside this legacy of formal thought, the builders pour in the vernacular record of daily human existence: millions of lines of computer code, open gazettes, private correspondences, and the colloquial chatter of the global public square.

This process constitutes nothing less than the digital centralization of human civilization's collective memory. The machine does not inspect physical parchment or paper volumes; it absorbs hundreds of billions of words transcribed into a unified, accessible digital sea. Yet this ocean is not organized by subject catalogs or index cards. It is deposited as an immense, unstructured quarry of raw language from which the machine must extract the underlying order, grammar, and cadence of human reason.

In modern data centers, this vast library is stored across a tiered system rather than on a single drive: the full ocean of raw, cleaned text sits permanently inside massive cloud storage archives; the tokenized words needed for training are staged on ultra-fast flash drives wired directly to the supercomputer network; and the immediate words being read that split-second reside directly on the processor chips inside high-speed memory, feeding thousands of computing cores without delay.

III. The Architecture: Abandoning the Rulebook

For decades, computing systems failed to master human language because engineers attempted to construct rigid catalogs of rules for every conceivable thought. That effort collapsed under its own weight; human speech and circumstance contain far too many exceptions for a fixed manual to anticipate.

The breakthrough succeeded by abandoning manual rules entirely. The builders constructed an immense, flexible web of mechanical connections—hundreds of billions of adjustable numerical dials—and permitted the machine to read.

Do not picture these dials as brass knobs turned by human fingers; they are microscopic mathematical multipliers, what the modern mechanists call "weights". Think of them as the finely threaded screws upon a surveying instrument or the elevation gears of a brass cannon. Each dial controls the strength of the passage between two related ideas. When set to zero, the path is barred, and the machine recognizes no connection whatsoever; when turned slightly, a faint association is permitted to pass; and when opened wide, the current flows with absolute conviction.

At the outset of the work, every one of these hundreds of billions of dials is spun entirely at random, leaving the machine blind and foolish. The builders then feed in unfinished fragments of human text—such as "The Declaration of..."—and prompt the engine to guess what word follows. At first, its randomized connections spit out wild errors. But with every blunder, the machine measures how far its guess missed the mark, tracing the error backward through the network to give the dials along that path a microscopic nudge. By turning these countless tiny screws just a fraction of a hair with each mistake, the entire web gradually tightens, until the once-random maze mirrors the natural cadence, grammar, and logic of human thought.

 It translates human speech into arithmetic, measuring the relational affinities between concepts as physical forces.

IV. The Forge: How Nvidia’s Processors Power the Machine

 To read such an ocean of text and balance billions of connections requires computational power of an unprecedented order. This capability emerged from a fundamental division in computer architecture: the difference between the Solitary Scholar and the Grand Battalion.

• The Solitary Scholar (The Central Processing Unit, or CPU): For half a century, computing rested on the CPU. This processor functions like an extraordinary polymath seated alone at a desk—capable of solving intricate calculations with great speed, but restricted to working sequentially, one problem after another. To require a solitary processor to adjust billions of interconnected dials one by one would demand thousands of years of continuous computation.

• The Grand Battalion (The Nvidia Graphics Processing Unit, or GPU): Nvidia originally designed chips not for statecraft or philosophy, but to render moving pictures on video screens. Painting a complex digital scene requires calculating millions of tiny points of color simultaneously. This task does not need a single deep thinker; it requires thousands of simple calculators executing basic arithmetic at the exact same instant. Nvidia discarded the solitary scholar in favor of a regiment of thousands of simpler processing cores operating in parallel.

Around 2012, researchers realized that the mathematics needed to adjust billions of dials in a neural network are identical to the mathematics used to project millions of pixels on a screen: vast grids of parallel matrix arithmetic. Handing the learning process to Nvidia’s parallel processors compressed what once took centuries into mere weeks.

Nvidia paired this hardware with CUDA, an operating software platform that gave developers a unified system of command to coordinate thousands of parallel cores at once. Today, building a frontier model requires linking tens of thousands of these processors with high-speed cabling into unified computing clusters that function as a single massive installation.

V. The Discipline of the Dials: How the Machine Learns from Error

The internal 'dials' of this engine are numerical values called weights or parameters. They function like valves across a municipal water system, where the incoming signal is human text converted into a stream of numerical tokens.

When an unfinished prompt—such as "The Declaration of..."—enters the pipe network, it flows upward through deep layers of valves. At the outset of training, every valve is set completely at random. Lacking any prior guidance, the current sloshes indiscriminately through the network and splashes out a plainly erroneous continuation, such as "Cabbage".

The system immediately routes this flawed output to a Loss Function, which measures the exact mathematical error between the wild guess and the true target text, "Independence".

Once the magnitude of that mistake is measured, Backpropagation steps in to trace the blame backward through every layer of the architecture, inspecting each valve along the path to determine whether it opened too wide or remained too constricted.

Finally, guided by this backward trace, the engine applies an infinitesimal dial adjustment—governed by the Learning Rate—across the entire network. These corrections must be kept microscopic; large, abrupt adjustments would destroy the system's equilibrium, breaking the delicate connections required elsewhere to produce phrases like "Declaration of War" or "Declaration of Rights".

By repeating this drill across trillions of words drawn from the digital ocean, the errors diminish and the valves settle into permanent alignment with the cadence and structure of human expression.

For a flagship frontier model, this training drill requires three to six months of uninterrupted, around-the-clock computation distributed across an interconnected cluster of 20,000 to over 50,000 advanced processors firing in parallel. Any prolonged hardware breakdown, power loss, or network bottleneck during this phase corrupts checkpoint states and forces the operators to roll back the machinery, adding weeks of delay to the campaign.

VI. Infrastructure and Human Alignment

Operating this machinery requires significant material resources. Modern training facilities demand specialized microchips etched at nanometer precision and industrial installations that draw continuous power on the scale of municipal grids.

Furthermore, raw training alone does not produce a usable tool. Because the base model absorbs the entirety of public writing, it ingests incoherence, hostility, and errors alongside factual knowledge.

 To turn this raw statistical engine into an effective instrument, the system must undergo intensive human alignment and fine-tuning—a process that spans several months of dedicated preparation and extends into perpetual oversight:

• The Core Tutoring Stage (3 to 6 Months): Before a model can be released, thousands of specialized human annotators, lawyers, doctors, and engineers spend months writing verified demonstrations of ideal inquiries and responses. They evaluate competing answers and penalize falsehoods and erratic behavior until the engine adopts the reliable demeanor of a disciplined assistant. Running the algorithms that adjust the model’s weights to match these human preferences takes weeks of compute on secondary clusters, backed by intensive manual verification.

• The "Red-Teaming" and Safety Audit (1 to 2 Months): Once the engine appears disciplined, outside specialists—security researchers, ethicists, and adversarial testers—spend weeks attempting to trick, confuse, or subvert the model to verify that it will not produce illicit, dangerous, or combustible instructions.

• Perpetual Alignment in Operation: Unlike the initial reading phase, alignment never truly terminates. Once deployed, hundreds of millions of daily user interactions generate fresh feedback, requiring permanent teams of human auditors to patch flaws, curb emerging hallucinations, and prepare subsequent generations of the system.

VII. Operational Limits: The Source of Hallucination

A fundamental characteristic of this architecture is that it possesses no direct connection to the physical world. It understands how words associate with one another in sequence, but it has no empirical experience of physical objects, materials, or events.

Consequently, when queried on topics where recorded documentation is thin, conflicting, or missing, the machine will still generate an answer that matches the cadence and structure of authoritative prose, even if the underlying claims are entirely fictitious. This behavior—known as hallucination—is not intentional deception; it is the natural consequence of a system designed exclusively to produce statistically fluent continuations of text.

 Treated as an analytical instrument for transcribing, cross-referencing, and synthesizing recorded human knowledge, the artificial engine provides unprecedented speed and leverage. However, because it reflects only the statistical averages of past text, empirical verification and final judgment remain the strict responsibility of the human operator.

VIII. The Capital Expense and Balance Sheet of a Frontier Model

Bringing a frontier model from raw concept to operational readiness is an industrial undertaking measured in billions of dollars in capital allocation. The financial balance sheet required to sustain this pipeline represents one of the largest concentrations of capital expenditure in modern history:

• Hardware and Cluster Capital: A single cutting-edge processor costs tens of thousands of dollars, meaning a cluster of 30,000 chips represents roughly $1 billion in silicon alone, excluding optical interconnect switches, liquid-cooling distribution manifolds, and custom server chassis.

• Energy and Facilities: Housing and energizing these clusters requires specialized data centers that demand between 50 and 150 megawatts of continuous electrical load, yielding electricity and utility cooling bills that run into tens of millions of dollars for a single multi-month training run.

• Human Tuning and Data Operations: The alignment phase—employing global data operations firms alongside thousands of credentialed specialists, attorneys, and engineers to curate demonstration datasets, audit model outputs, and execute preference scoring—adds tens to hundreds of millions of dollars in recurring labor overhead.

When aggregate depreciation, elite research engineering payroll, repeated failed runs, and post-training refinement are tallied, a hyperscaler commits anywhere from $500 million to well over $1 billion to train a single generation of a flagship frontier model. Over multi-year deployment cycles across multiple generations, major technology firms invest tens of billions of dollars annually in capital expenditures simply to construct, power, and upgrade this colossal physical apparatus.

 The General then asked who the players are in this race, and Mr. Franklin replied.

The primary global hyperscalers committing tens of billions of dollars annually to this physical apparatus and computational infrastructure are:

·      Microsoft (Azure): Anchors massive AI training clusters, serving as the core infrastructure and compute backbone for OpenAI while deploying hundreds of thousands of Nvidia GPUs alongside its own custom Maia AI silicon.

·      Alphabet (Google Cloud / Google Core): Operates global-scale data center infrastructure powered by dense clusters of Nvidia GPUs as well as its internally designed Tensor Processing Units (TPUs) to train and serve its flagship Gemini models.

·      Amazon (AWS): Deploys large-scale GPU infrastructure to support external enterprises and partners like Anthropic, while augmenting clusters with its proprietary Trainium and Inferentia acceleration chips.

·      Meta (formerly Facebook): While primarily a consumer platform rather than a public cloud vendor, Meta operates hyperscaler-scale infrastructure, assembling hundreds of thousands of frontier processors into centralized training clusters to build and open-source its Llama model families.

·      Oracle (Oracle Cloud Infrastructure / OCI): Has emerged as a major provider of bare-metal GPU clusters and optical supercomputing fabrics, securing multi-billion-dollar deals to train models for frontier AI research labs.

Outside the West, the dominant hyperscalers making comparable capital expenditures in infrastructure and silicon are Chinese cloud giants Alibaba Cloud, Tencent Cloud, and Baidu.

If the hyperscalers—Microsoft, Google, Amazon, and Meta—are the railroad barons who own the steel rails, locomotives, and rights-of-way, OpenAI and Anthropic are the master locomotive builders and elite patent holders.

They occupy a distinct, highly leveraged, yet vulnerable position in this ecosystem:

1. The Pure-Play Frontier Labs

Unlike Google, Amazon, or Meta, OpenAI and Anthropic do not sell cloud hosting, office productivity software, e-commerce, or consumer social feeds. Their sole business is pushing the frontier of artificial intelligence itself:

• The Pioneer Engine: They set the pace of invention. OpenAI’s GPT series and Anthropic’s Claude models have repeatedly defined the bleeding edge in reasoning, multi-step problem solving, and human alignment.

• The Talent Magnet: They operate with the agility of focused research institutions, attracting many of the world’s foremost machine learning minds who prefer dedicated research cultures over sprawling Big Tech conglomerates.

2. The Computational Trap: Dependent on Hyperscaler Rents

Because OpenAI and Anthropic do not own power plants, physical land, optical fiber grids, or massive server warehouses, they cannot train or serve their models without renting massive clusters from the hyperscalers:

• OpenAI has been historically tied to Microsoft Azure for its core compute and commercial enterprise distribution.

• Anthropic has raised billions in strategic capital and compute commitments from both Amazon (AWS) and Google Cloud.

This arrangement creates a complex dynamic of "co-opetition":

• The hyperscalers invest billions of dollars in cash and cloud compute credits into these labs.

• The labs hand that capital straight back to the hyperscaler to pay for GPU runtime and data center electricity.

• Meanwhile, the hyperscalers use that relationship to bolster their own cloud revenues while concurrently developing their own competing internal models (such as Google’s Gemini or Microsoft's smaller internal AI systems).

3. What Happens to Them When the Dust Settles?

As the foundational models mature and the major hyperscalers achieve comparable frontier capabilities, OpenAI and Anthropic face three distinct potential futures:

• The Super-Application Layer: They can evolve into enduring consumer and enterprise platform giants in their own right (akin to how Apple or Salesforce built massive software empires on top of commodity infrastructure), leveraging brand recognition, developer loyalty, and advanced reasoning tools.

• The "Intel Inside" Supply Layer: They become the outsourced elite R&D engines for global industry, licensing their proprietary weights, specialized agent architectures, and enterprise security tools through multi-cloud distribution.

• Capital Squeeze or Acquisition: Training frontier models requires ever-expanding capital outlays—eventually demanding tens of billions per generation. If public cloud commoditizes basic intelligence, pure-play research labs without diversified revenue streams (like advertising, enterprise software suites, or consumer hardware) could face brutal margin compression, leaving them vulnerable to closer integration or absorption by the very tech giants funding their compute.

The race will not yield a single imperial victor, General, because the staggering capital and computational resources required are being matched simultaneously by several rival giants. What emerges will not be a lone monopoly, but a consolidated oligopoly—an infrastructure cartel akin to the railroad barons, electrical utilities, or oil trusts of past centuries. Once Microsoft, Google, Amazon, and Meta all achieve broadly comparable, world-class frontier intelligence, raw model capability will cease to be an exclusive prize; it will become a shared baseline commodity. The genuine contest then shifts entirely to distribution and lock-in: who controls the proprietary software pipelines, the exclusive corporate workflows, the national security contracts, and the daily devices through which the rest of human society must interact with this intelligence.

When that collective victory is finalized, AI will function as the universal operating system of global commerce, medicine, and statecraft. For the hyperscalers, this shared summit guarantees an immense, perpetual private toll on the daily decisions and transactions of the entire world, accompanied by colossal barriers to entry that prevent any outsider from ever catching them. For the broader public, the economic consequences will be profound and destabilizing: sudden, staggering leverage over professional labor, unprecedented velocity in scientific discovery, and a severe concentration of political and economic influence held by a handful of corporate boardrooms. The ultimate question will no longer be which firm built the fastest engine, but whether a free, republican society can maintain its independent judgment and democratic sovereignty when its essential information conduits are owned and mediated by an oligarchy of machines.”

The General then thanked Mr. Franklin and replied calmly and honestly, “You have given me much to ponder and wonder at.”