The Applications of A.I.

 

The aim of machine learning is to predict. But what it tries to predict depends on its use. Most machine learning tries to predict the formula that lies behind whole data sets. What most of us have learned in school is first principle, top-down. You start with foundational principles and construct appropriate practical solutions from there. This is true for science, medicine, government, and even for a lot of businesses. The world, in these cases that goes back to the ancient Greeks, is universal, generally known. The statistical models used predict only slight Gaussian bell-shaped changes. If changes are greater, the models have to be run again. There are only four foundational machine-learning models, that you can invoke, models whose goal to reach targeted conclusions from patterns of the data. ¹ In more conventional language, to determine the dependent variable from the independent ones. Legal and scientific studies are edifices built upon accurate data. Both science and due-process legality derive from the Enlightenment; whose main assumption was that you can validly find out the states of nature; or more specifically, about more complicated societies. ²

Generative A.I., on the other hand, is bottom-up. Machine learning in this case tries only to predict the next word or the pixel, and depends upon calculated context to provide the right answer. This means that many applications for A.I. are departmental, the impact of these upon general company financials has yet to occur, and you have to determine what matters, because obviously some things change (and some do not), and that requires human judgment to decide which is which. Generative A.I., human language models, open up the whole world of computer-based predictive models to many users; but now with the major caveat that the bottom-up approach can also produce both truth, hallucinations and results limited only to the training distribution (80% of the data sampled randomly).

Hallucinations are untruths. Generative A.I. is capable for hallucinations because it is bottom-up and proceeds from uncurated Internet data. Generative A.I. prediction does not have foundational principles. It tries to predict the next word or pixel (from context). It can therefore mimic human intelligence to a point; but can also bottom-up “hallucinate”, as humans can also do. ³

Furthermore, a 5/8/26 Bloomberg article points to another crucial shortcoming of bottom-up learning. It occurs in weather forecasting, where physical models are being rapidly replaced by A.I. models. “Using artificial intelligence to forecast the weather is getting so good – and so cheap – that meteorological services are starting retire the expensive (to run) physics-based systems they have relied upon. That’s a potentially big problem – and not just for weather forecasting….AI models are so useful because most management decisions are inside the training distribution. The grocery chain forecasting milk demand, the airline pricing of seats for next Tuesday, or the call center routing inquiries. Improving the efficiency of these predictions is invaluable for anyone dealing with that sort of problem – which is most managers, most of time.

But the decisions that determine whether an organization survives are not those decisions. Historians and biographers don’t spend much time on the everyday. Instead they focus on the big moments – wars and crises, hostile takeovers, and revolutionary innovations. The tendency to fail in the biggest moments is what made VaR (Value at Risk) so dangerous. AI threatens to be the same problem at industrial scale. AI is great at interpretation across familiar territory, enough to earn the confidence of their users, and awful at extrapolation beyond the known….In companies, human infrastructure is at just as much at risk. The middle managers who may be replaced by AI have the human judgement and experience to say, ‘I know the model says things are fine, but this doesn’t look right.’ Lose them, and you lose that vital brake on the system. It will work right up until it doesn’t.” An obvious example of this problem is the effect of the Strait of Hormuz upon supply chains.

“Progress” in A.I. has to be of the right sort to be useful:

A.I. in Medicine

The Spring 2025 edition of the Yale School of Medicine contains Dean Nancy Brown’s summary of A.I.. It is just getting started. On a departmental basis, the change in A.I. is significant in diagnostics, work flows and research.

Diagnostics -

Diagnostics comprise about 75% of FDA, A.I. approvals. Brown writes,”Within healthcare AI has so far had its most significant impact on diagnostic radiology. Pattern recognition technology is used widely to prescreen medical images and spot diseased tissues, bone breaks, and other conditions. Now, researchers across the country experimenting with generative AI are creating highly detailed 3D models of organs and tissues that can be trained to simulate the effects of radiation therapy on a particular patient. Allowing physicians to customize treatment plans.”

The Chatbot BARD identifies three major issues which impede the transferability of medical data across institutions. Most hospitals use the DICOM standard (Digital Imaging and Communications in Medicine). An MRI made in one hospital can theoretically be read in another but:

1)    Despite the DICOM standard, MRI systems from different vendors (Say GE, Siemens, and Philips) often add unique tags to their images, that can make them incompatible with other systems.

2)    Transferring a whole AI model is harder than transferring a simple scan. A model trained at Hospital A might perform poorly at Hospital B due to: Hardware Variance, Radiation Protocol Differences, or Patient Positioning.

3)    Population Shift. A model trained on a specific demographic may not work as well on a population with different genetic or environmental backgrounds. The data has to be FAIR, it has to be Findable, Accessible, Interpretable, and Reusable.

The application of A.I diagnostic models isn’t simple “plug and play”.

Work Flows –

A.I. is theoretically applicable to the complex administrative work flows that accompany medicine. AI agents can “listen in” on office visits and create summaries, analyze the lab values obtained, and enable doctors and nurses to focus in on the human interactions. AI can, in theory, also handle the complex paperwork that insurance requires, but the problem is constant changes that A.I. might eventually handle. 

Research –

A natural use of A.I. is to find new drugs, to apply molecular biology to discover more precise disease treatments by discovering diseased cell binding sites. The discovery of these binding sites, however, is but a first step. This is why large drug companies exist, to finance the necessary effectiveness and toxicity studies necessary to get FDA drug approval. 

A 5/3/25 study published in npj digital medicine by Giuffrè, You, Pang et al. points to the potential of A.I. “Large language models generate plausible text responses to medical questions, but inaccurate responses (Type I and Type II) errors pose significant risks in medical decision-making. Grading LLM outputs to determine the best model or answer is time-consuming and impractical in clinical settings; therefore, we introduce EVAL (Experts-of-Experts Verification and Alignment) to streamline this process and enhance LLM safety for upper gastrointestinal bleeding…We evaluated Open AI’s GPT-3…Anthropic’s Claude…Meta’s LLaMA-2...Fine-Tuned ColBERT achieved the highest alignment with human performance across three separate datasets (r=0.81-0.91)…the model was close to human response, but not perfect.”

We can assume suitably chosen A.I. models are getting there, but the models or data bases still require fine-tuning.

A.I. in Business

Business, which although has systematic elements, also creates a lot of uncertainty. The task of a CEO, in this age, is to enhance ROC, shareholder return on capital. Although Verizon maintained dividend margins, one has seen what happens when a CEO fails to grow the business.

In order for EPS to grow, a company’s return on capital must exceed its cost of capital. Comparing the nine months of 9/24 and 9/25, Verizon’s annualized cost of capital is about equal to its return on adjusted assets. The company has been unable to grow its EPS. Verizon annualized cost of capital is approximately 6%.

 

                          Verizon Return on Capital Annualized

                                             2024     2025

                                            7.18%   8.80%. which was accomplished by unsustainable

                                                                     rate increases.

 

 CEO’s (but not some middle-level staff) are enthusiastic about A.I. At the departmental level, there are lots of systematic things that can be done by A.I., in time. But as we point out above, bottom-up A.I. will not be able to handle top-down events well because A.I. optimizes for a defined environment.

 

According to a 5/13/26 Washington Post article, “…different corners of the workforce are weakening at different times: federal workers facing mass layoffs, logistics and manufacturing contracting after their pandemic surge, white-collar hiring quietly freezing up.”  The only positive gain is healthcare, which since Jan. 2025 added 79.6k cumulative jobs, compared with the rest of economy which decreased by 40.4k cumulative jobs. “The unemployment rate for people ages 22 to 27 who recently completed college hit 5.6 percent in the final months of 2025….But nearly half of that age group was underemployed, among people were working in jobs that did not require a degree…”

 

According to a 4/21/26 NYT article, Charles Scharf, CEO of Wells Fargo said, “These are all opportunities to do things much, much more efficiently with A.I. than humans have been doing....Most other bank chieftains”, he said, “are afraid to say it become no one wants to stand up and say that we are going to have (can) lower head count in the future.” The reason why the U.S. economy has been able to grow, is that it has been able to adapt to technological change. However, it will also be important to distinguish between qualitative and quantitative changes.

 

A.I. and cyclical factors are creating tremendous amounts of employment uncertainty in the economy although the economy is still growing. There are obviously no more career tracks, and, as the Post says, companies “…emphasize candidates who can lead projects and expand an organization’s capacity without adding headcount.”

 

 

Summary

In both medicine and in business, A.I. is a work in progress with substantial and growing potential to use existing data. Starting with single departmental applications, A.I. is starting to grow into other functions. But for any computer analysis to be valid, the data has to be FAIR, that is: Findable, Accessible, Interpretable, and Reusable. We will be justified to value A.I. at a premium to the general economy growing at 2% real.

As a Matter of Philosophy (from the militaries)

In the TV show, GZero World, Bloomberg Correspondent Katrina Manson comments upon the application of A.I. in the Pentagon. To implement A.I., people have to find it useful in the specific situations they face. “The fog of war” requires proper data labeling, and then according to U.S. practice, the commander has to remain in the decision. The problems with A.I. are (in the first two, as usual) sycophancy, hallucination and escalation. The problem for the future is that A.I. can make broad-based, split second decisions which may be very right or very wrong.

A former advisor to the Ukrainian government on defense issues, Kate Bondar, makes an important observation. Russians, as opposed to the West, really don’t care about the human lives lost in combat. Their A.I. combat initiatives therefore allow autonomous drones to automatically kill people, without a human in the loop. This will eventually allow battle decisions, such as targeting, to be conducted within milliseconds. The West, on the other hand, values human life. That philosophy should guide the further development of the technology.

 

 

 

¹ There are only four algorithms: regressions (to predict continuous data), classification, clustering and time-series. For those who work directly in the field of A.I., we recommend Gauthier Vasseur’s Tech 78 course in Stanford Continuing Education. Professor Vasseur is the Cal-Berkeley, Haas Business School, Professor of Business Analytics. He emphasizes the importance of data, not algorithms. There is no magic in A.I. If data is bad, if the humans don’t find the models useful, the projects won’t work.

²  Our Parisian Airbnb was on the Rue Réaumur, just two blocks from the Museum of Arts and Crafts (Vocations). The museum contained chemist Lavoisier’s lab (he established the classical chemical equation of the conservation of reaction mass). The museum also contained the basis of Enlightenment science, quantification. France adopted the uniform metric system in 1795. In case no one knew, there is a direct and linear relationship between the metric scale of temperature and its predecessor Réaumur scale. This scale was constructed by the Enlightenment physician and naturalist René-Antoine de Réaumur.

³ The 4/13/26 New Yorker magazine contains an article on Sam Altman, founder of OpenAI. He said that, “…allowing for some falsehoods can, whatever the risks, confer advantages.” “If you just do the naïve thing and say, ‘Never say anything that you’re not a hundred percent sure about,’ you can get a model to do that,” he said. “But it won’t have the magic that people like so much.” But if you intend to invest billions of dollars in A.I., you had better be right, both on risks and returns. Otherwise, the models can provide both uplift and drag. The best use of A.I. will probably be on curated databases.