Brave New AI World: Theoretical AI, Applied AI, and the Case for Transparent Models

09/22/2026
By Ed Timperlake

In several published articles I have argued that all things AI are best understood through an algorithmic lens, broken down for clarity into Theoretical AI and Applied AI. There can be intellectual bleed-over between the two clusters of research, but that is mostly seen as it happens.

In this white paper I offer a real-world example of my argument for a research approach which focuses on the difference between the heoretical and Applied, and point toward a way ahead for AI oversight.

My field for introducing AI engagement is mathematical modeling: the process of turning a real-world system into mathematics so one can analyze it, predict behavior, optimize decisions, or have a computer interact with the system.

For Theoretical AI, brilliant thinkers can look at AI enhancing all mathematical modeling techniques. The possibilities are limited only by imagination and by an honest willingness to conclude that AI may not actually be that helpful.

The fields include:

  • Applied Mathematics: differential equations, optimization, numerical methods, and mathematical analysis.
  • Statistics & Data Science: probability, statistical modeling, inference, machine learning, and predictive analytics.
  • Operations Research: optimization, decision theory, scheduling, logistics, and resource allocation.
  • Computational Mathematics: numerical simulation, scientific computing, algorithms, and high-performance computing.
  • Physics & Engineering: modeling physical systems, mechanics, fluid dynamics, electromagnetism, and control systems.
  • Biology & Bioinformatics: population dynamics, epidemiology, genetics, ecology, and biological systems.
  • Chemistry & Materials Science: molecular modeling, reaction kinetics, thermodynamics, and materials behavior.
  • Economics & Finance: econometrics, financial mathematics, economic forecasting, risk modeling, and game theory.
  • Computer Science & AI: mathematical foundations of algorithms, optimization, neural networks, and computational models.
  • Environmental & Climate Science: climate models, environmental systems, pollution transport, and ecological forecasting.
  • Medicine & Public Health: disease transmission models, medical statistics, pharmacokinetics, and health-system modeling.
  • Social Sciences: models of population behavior, networks, voting, migration, and social dynamics.
  • Industrial & Systems Engineering: manufacturing, supply chains, reliability, process optimization, and systems modeling.
  • Geosciences: models of earthquakes, groundwater, geological processes, and natural hazards.

Across these fields, the most important modeling approaches include differential equations, probability and stochastic processes, statistics, optimization, numerical analysis, linear algebra, dynamical systems, discrete mathematics, simulation, and machine learning.

My Applied AI test case is very personal.

In the late seventies and early eighties, I led a group of brilliant thinkers who invented a multi-attribute math model called TASCFORM, the Technique for Measuring Comparative Force Modernization.

Our challenge was to build a very transparent way of comparing fielded U.S. and Allied combat aircraft, and those projected only five years out, with the Soviet/Warsaw Pact combat aircraft inventory.

The model was built to give strategic insights into both quality and quantity at the height of the hot/cold war.

It was created for the Office of Net Assessment, validated by the CIA, and ultimately also used by Congress.

For more detaills I refer the reader to my article: The US vs. USSR in TacAir: Lessons Learned from a Hot Cold War.

This is taken from the historical record as indicated in published sources and assembled by an AI syssten:

On Edward (Ed) Timperlake and TASCFORM, there is a fairly direct connection between mathematical modeling and military force assessment.

Edward Timperlake developed the TASCFORM analytical methodology while working at The Analytic Sciences Corporation (TASC). It was used for assessing the modernization and comparative capability of military forces, particularly tactical aircraft, for the Office of Net Assessment and the CIA.

What TASCFORM did

TASCFORM can be understood as a multi-attribute quantitative decision/modeling framework.

Rather than judging a weapon system from one characteristic, such as speed, it combined multiple factors into measures of military effectiveness.

For tactical aviation, documented factors included things such as:

  • Payload
  • Range and standoff capability
  • Maneuverability
  • Speed/mobility
  • Target acquisition and fire control
  • Survivability
  • Navigation
  • Countermeasures
  • Weapon enhancements
  • Obsolescence and productivity

These factors could be combined into increasingly comprehensive measures such as Weapon Potential (WP), Weapon System Potential (WSP), and finally Adjusted Force Potential.

Applied mathematics → operations research → decision analysis → systems engineering → defense modeling

Conceptually, it was a multi-criteria utility model: different attributes were normalized, weighted, and combined to produce comparative measures.

Taken from the historical record:

Among the major mathematical modeling fields, and specifically through the lens of Ed Timperlake/TASCFORM, the most relevant academic subjects would be operations research, multi-criteria decision analysis, utility theory, systems engineering, statistics, optimization, simulation, and defense operations analysis.

The transparent, multi-attribute weighting factors were balanced in several “Delphi” sessions with very experienced combat pilots. Trading off range against payload is a good example, since an aircraft can carry only so much weight in fuel and weapons. The lessons of the Vietnam War were extremely important for getting the attributes right for future aircraft design.

The Delphi Technique

The Delphi technique is a structured method for obtaining and refining expert judgment, especially when reliable quantitative data are limited or the subject involves uncertainty about the future.

It is widely used in forecasting, technology assessment, policy analysis, operations research, strategic planning, and mathematical modeling.

A typical Delphi study proceeds through several rounds:

  • Select experts — Assemble a panel with relevant knowledge.
  • Round 1 — Experts independently answer questions or estimate future developments.
  • Aggregate responses — Researchers summarize the responses, often using statistics such as the median, range, or distribution.
  • Feedback — Experts receive an anonymized summary of the group’s responses.
  • Round 2+ — Experts reconsider their estimates in light of the group’s aggregated judgment.
  • Convergence — Several rounds may be conducted until judgments stabilize or a predetermined stopping criterion is reached.

The important feature is that experts generally do not need to meet face-to-face, and individual responses can remain anonymous.

This reduces some effects of hierarchy, reputation, and dominant personalities.

Delphi vs. Mathematical Modeling

Delphi is usually not itself a mathematical model. Instead, it can provide inputs to a model when empirical measurements are unavailable.

The era of studying many combat aircraft is over because there are far fewer Type/Model/Series (T/M/S) in combat inventories. But for the unique design wars of the fifties, sixties, and seventies, TASCFORM was very useful for creating strategic insights into the quality/quantity aviation technology race.

The good news is we won!

Now add the power of AI, using the TASCFORM example, to the next use of multi-attribute utility functions, where it might be very useful in making comparative measures of current and projected insights across many of the research fields listed above.

It is simple: ask AI to critique AI.

AI could modernize TASCFORM by augmenting its traditional force-assessment framework with machine learning, Bayesian inference, natural-language processing, and AI-assisted Delphi techniques. AI could extract and organize technical information, estimate uncertain parameters, identify relationships among performance attributes, aggregate expert judgments, and conduct large-scale sensitivity and scenario analyses, while TASCFORM would retain the transparent mathematical structure for combining attributes into comparative measures. Rather than replacing human judgment, the approach would use AI to quantify uncertainty, expose disagreements among experts, test assumptions, and improve the speed and consistency of analysis, producing probabilistic assessments rather than relying solely on fixed point estimates.

Now consider a case far from measuring fielded aircraft performance. Applied AI might bring significant clarity to the recent, horrific debate over which medications to use in treating COVID. That “TV Delphi” process was flawed by politics and fought out in public, but it could have been modeled mathematically using fact-based reporting of deaths and treatment successes.

It is now time to ask AI for clarity: a multi-attribute utility causation/treatment model, using insurance law-of-large-numbers analysis of COVID deaths, vice “Delphi Docs” fighting it out politically on national news.

AI can be applied to the insurance law of large numbers during COVID-19 by combining large claims, mortality, treatment, demographic, and epidemiological datasets to estimate how well observed losses converge toward expected losses as the insured population grows. An AI-enhanced actuarial model could segment policyholders by age, comorbidities, vaccination status, geography, treatment exposure, and other relevant variables, while statistical and machine-learning methods estimate the probability and severity of COVID-related death, hospitalization, and treatment costs. The model could then compare actual aggregate claims with expected aggregate claims across increasingly large populations, measuring the degree of convergence and identifying where heterogeneity or correlated events, such as a pandemic, cause the traditional independence assumptions behind the law of large numbers to break down.

In this framework, AI is primarily a tool for risk classification, forecasting, anomaly detection, and uncertainty quantification, while actuarial methods, similar to TASCFORM math modeling, could provide the formal framework for pricing, reserves, and solvency analysis.

My research point is very simple: Theoretical AI exists, and Applied AI, for good or ill or simply not helpful, is now in play.

Trying to legislate into such an evolving new world of research is a very, very bad idea.

Editor’s note: This white paper extends an argument Ed Timperlake has been developing on our sites. In Organizing for War-Winning AI: Combine AI with Cyber Command, he divided AI into Theoretical and Applied research, and in How AI-Empowered Kill Webs Enhance Carrier Operations he showed what applied AI can do inside a kill web. Here he takes the same distinction beyond the military and tests it against his own experience building TASCFORM, a multi-attribute utility model and the same family of thinking behind the “payload utility function” language in his kill web writing.

The central point echoes much of our recent AI coverage, including Murielle Delaporte’s series on AI and military command, which we discussed in Who Really Commands?, and our August piece on the trust AI depends on: AI serves best when it strengthens human judgment inside a transparent structure rather than replacing it.