Artificial Intelligence

Artificial intelligence is not a unicorn product category.

Decision makers are at risk of seeing AI as unique. It is not. Consider three aspects of the public dialogue about AI.

• Research
• Products
• News

Errors are rampant in AI research.

Princeton computer science professors Arvind Narayanan and Sayash Kapoor examined the literature, including replication attempts and feedback between researchers. Narayanan, Arvind, and Sayash Kapoor. AI Snake Oil: What Artificial Intelligence Can Do, What It Can’t, and How to Tell the Difference. Princeton University Press, 2024.

Quote: “The party line among scientists is that science ‘self-corrects’ … but everything we’ve seen about the process suggests otherwise.” Ibid. 23.

This should come as no surprise:

• Replication failure has a high base rate in the sciences. Baker, Monya. 1,500 scientists lift the lid on reproducibility. Nature, May 2016, https://www.nature.com/articles/533452a.
• Sloppiness and hucksterism are frequent with a new technology as users, buyers, and listeners catch up.

AI products often do not work. Even McDonalds, after working three years with IBM, was not able to produce an adequate AI drive-through system. OECD AI Policy Observatory Portal. OECD, 2024, oecd.ai/en/incidents/91533.

Vendors have some incentive to sell more than serve. That’s one reason why consumer protection regulators are up in arms. E.g., FTC Order Requires Online Marketer to Pay $1 Million for Deceptive Claims That Its AI Product Could Make Websites Compliant with Accessibility Guidelines. Federal Trade Commission, 3 Jan. 2025, www.ftc.gov/news-events/news/press-releases/2025/01/ftc-order-requires-online-marketer-pay-1-million-deceptive-claims-its-ai-product-could-make-websites.

Journalism is no bright spot in the public dialogue about AI. As a rule, journalists know less than the people they interview. Moreover, journalists have some incentive to attract attention more than to inform.

Therefore, AI is not a unicorn because creating business value from it requires shrewdness.

• Expect mistakes, puffery, and misinformation.
• Don’t get lost in the broader public dialogue.

What should you do instead?

Focus on the specific AI product or project in front of you.

For example, a McKinsey 2023 report estimated that generative AI will add $2.6 to $4.4 trillion in value globally. Chui, Michael, et al. The Economic Potential of Generative AI: The Next Productivity Frontier. McKinsey Digital, McKinsey & Company, 14 June 2023, www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier.

This information tells you nothing about the value of a particular generative AI product or project. News like this should be no more than ambience for decision making.

Read on.

• Products
• Internal Projects
• Winnow the List: Four Rules for Decision
• Ignore Hype
• That’s a Nice Feature, so How Does It Affect Us?
• Plain English
• Easy Mode?
• Saluting Unproblematic, Useful and Uninteresting AIs
• Only a Few AIs Stand Out at a Time
• Trade Secret Knowledge: Proprietary Data and AI
• A Trade Secret Explanation is the Greatest Competitive Advantage
• Use Proprietary Data and AI to Develop Trade Secret Understanding
• Agents and Humans: The Seven Business Functions of Intelligence
• Knowledge Transfer
• Seeding Knowledge
• Partnership Knowledge
• Typical AI Typologies are Not Useful
• Why Karl Popper Wouldn’t Break Bread with Predictive AI
• Blame Programmers: Rules-Based AI
• Non-Duplicative Knowledge Transfers: Analytic AI
• New Knowledge From Node One: Generative AI
• Information is Created
• Wrappers and the Ignorance of Crowds
• Iterative Grounding: Generative AI
• What Do Transfer Recipients Need to Do?
• No Cliffs or Liftoffs: Continuous Improvement in Intelligence
• AI Knowledge Growth Knows No Upper Limit
• What to Do When AI Goes Wild
• Conclusion: Leave Hype and Cynicism to the Side

Products

If an AI product is sold to you, it is sold to competitors. It is unlikely that an AI product will confer the buying company a competitive advantage; at best not a disadvantage.

In the short term, some first-mover buyers may gain an advantage.

Internal Projects

As a rule, large project expenses balloon compared to initial expectations. There is no reason to treat AI projects differently.

Winnow the List: Four Rules for Decision

1. Maintain mental discipline during AI hype periods.
2. Place the burden of proof on vendors and internal advocates.
3. Reject all proposals that are heavy on buzzwords.
4. Reject all proposals that hand-wave potential downsides and focus on extreme benefits.

Ignore Hype

Just as pricing action in stocks does not reflect real business value, the ebb and tide of AI hype and fears do not reflect the real capabilities and risks of AI technologies.

Consider one type of AI. A 2023 IBM study found that CEOs are under accelerating pressure to adopt generative AI. CEO Decision-Making in the Age of AI: Act with Intention, IBM Institute for Business Value, June 2023, pp. 6, https://www.ibm.com/thought-leadership/institute-business-value/en-us/report/2023-ceo.

You do not want to inject noise and bias into firm decision making.

Even if you’re over the initial generative AI hype, more waves will come.

There’s too much interest and investment in AI for us not to expect periods of unusual, irrational pressure to accelerate or slow AI product adoption and projects.

That’s a Nice Feature, so How Does It Affect Us?

• Best information: You are at base most informed about you and your interests.
• Second best information: Next you’re most informed about your company and its interests.

Insist on funnelling the information you get from Zoom & room sessions with salespeople and internal advocates into answers about how the firm will be helped or hurt.

Plain English

Only accept plain English answers. At a cocktail party, the extensive use of buzzwords is a classic tell for ignorance. In business dealings, it is a tell for fraud.

Easy Mode?

Like “candy from a baby”-style statements about the benefits of a deal is a classic tell in business for fraud.

Avoiding a clear, detailed discussion of risks and downsides is another tell.

Consider the cost of compute. An IBM study from 2024 found that “15% of projects have been put on hold and 21% of gen AI initiatives have failed to scale” because the cost of compute was too high. The CEO’s Guide to Generative AI: Cost of Compute. IBM Institute for Business Value, October 2024, www.ibm.com/thought-leadership/institute-business-value/en-us/report/ceo-generative-ai/ceo-ai-cost-of-compute.

We’re done harping on common sense, free of the illusion of AI as a unicorn category.

Saluting Unproblematic, Useful and Uninteresting AIs

But don’t mistake this paper for a cynic’s piece. AIs can have significant business value. PwC’s 2024 summer survey of 1,000 business and tech executives found that around 40% had seen increased employee productivity and 41% had seen improved customer experience due to AI technology. 2024 Cloud and AI Business Survey. PwC, 2024, www.pwc.com/us/en/tech-effect/cloud/cloud-ai-business-survey.html.

You have integrated, are integrating, and will integrate a significant number of AIs into your business. This includes a wide range of B2B SaaS software, little software updates to employee devices, and more.

Only a Few AIs Stand Out at a Time

Not only are many AI unproblematic, many are useful and not even remotely interesting except to the respective engineers involved. (No offense to them.)

AIs often become unproblematic because engineers ensure, at times through feedback from other humans and other AIs, that a particular AI needs improvement on some dimension, such as efficacy, cost-effectiveness, privacy, safety, and others.

If a vendor can automate tasks and save employee time, you’re interested but not precisely in how the software works. Of course you want to make sure it works with limited downside.

Even machine learning AI can produce new and yet boring results. For example, we almost all learn to drive, each in our own way. We don’t really care about how others learn to drive as long as they do it well.

Certain AIs have stood out and required far more internal discussion than normal, and those are worth focusing on.

How should you do that? Start with the fundamentals.

Trade Secret Knowledge: Proprietary Data and AI

Steve Jobs and his team generated a pool of knowledge that made the iPhone what it became. All kinds of business metrics for Apple could retroactively present a picture of value, but the value was fundamentally a set of abstractions embodied in a small group of human brains.

Do we ignore the asset driving businesses forward because it’s hard to quantify? We can’t abandon metrics when evaluating AI product or project proposals. But we can’t let them drive the conversation either.

Knowledge is the greatest asset that a business can have.

If knowledge differentiates businesses from each other, you must create knowledge that is:

• Useful to the firm, and
• Exclusive to the firm

Trade secrets, like all knowledge, are divisible into three types:

• Descriptions: What is
• Prescriptions: What should be
• Explanations: Why something is or should be

Examples:

• Descriptions: Price histories
• Prescriptions: Current pricing offer
• Explanations: Why a particular price is a good fit for a customer

A description is true or not. A price of $10,000 was set in 2016 or not. The same is true for any explanation: The claimed cause of an effect is either the cause or not.

But a prescription is directly evaluated for its utility. Indirectly a prescription is only useful for reasons, which are encoded until they are drawn out with a smart explanation and accompanying descriptions.

Prescriptions include products and services, which have utilities that vary between customer groups and within them.

A Trade Secret Explanation is the Greatest Competitive Advantage

You’re a customer in this hypothetical. You get on the interwebs and shop. Feature-wise, descriptively, your available economic packages are complex but let’s just focus on product and price.

The net value of your digital buy is the prescription. Why you get that value is an explanation.

Explanations are the most valuable form of knowledge, following the work of famed physicist David Deutsch. Deutsch, David. The Beginning of Infinity: Explanations that Transform the World. Viking Press, 2011.

We honor thinkers such as Einstein who subsumed a great mass of complexity into a tight causal understanding of reality.

The more informed you are about your wants and abilities and options for satisfaction, the better off you are. But the force behind your shopper savvy is not a list. It’s not just a set of descriptions like the price or the product.

If you’re buying food, say, then you’re a better buyer if you know something about hunger, dietary restriction, digestive effects, and more. That is, the why of things.

Two conclusions follow.

• The more complexity that an explanation encodes, the higher the utilities of the prescriptions that it generates.
• The more utility that a prescription has, the more causal complexity that it encodes and therefore the greater the explanation you could draw out.

An explanation is a multiplicative gain prescription-wise. If you can reverse engineer an excellent product, you will gain other valuable business prescriptions in the future.

But at the same time, the better the prescription, the more difficult it is to reverse engineer. This is precisely because of the complexity it encodes.

This is really what makes a moat, following the focus of those wise men Warren Buffet and Charlie Munger. It’s not strictly about high utility to customers. Without the corresponding barrier to reverse engineering then entry would be rapid and successful.

Now you could object. Consider police dogs. You think they’re dumber than you. Yet they reverse engineer scents to sources. Far better than you.

I would submit that the dog has a cognitive speciality that is superior to yours and that you also have cognitive specialities the dog does not. You also have a general problem-solving ability, which you prize and should.

Now you could object again and say the dog isn’t a thinker, it doesn’t have an explanation generating prescriptions. But evolutionary selection, principally natural, is a bit like the high-performing machine algorithms that you fear and savor.

In a high-performing organic brain, explanations knit correlations into a low-“compute” way to do things, i.e. generate true descriptions and useful prescriptions. This includes the anticipatory, experiential, and memory-based pleasure of knowledge production itself.

High-performing machine algorithms function a bit similar to evolving explanations inside organic brains. More-so than natural selection, as the latter is quite slow. This is one among other differences.

Obviously the selection process in high-performing organic brains is quite fast. It goes fastest when the brain has a good explanation for the relevant correlations.

At this moment it does not seem there are any machines with the ability to produce an excellent and novel explanation. High-performing AI can work with causality but not quite at the level that high-performing people can. One day they may.

In sum, you create the most business value when you have better explanations that matter for the business. You also create a nicer but harder target for rivals and potential entrants. More money, more problems.

Use Proprietary Data and AI to Develop Trade Secret Understanding

You know two relevant cliches:

• Correlation does not imply causation, but
• Without correlation there is no causation.

By extension:

• Explanation is not possible without a correlation on hand, and
• Correlation implies the potential to create an explanation.

All business-relevant correlations are grounds for the business to generate the most useful sort of trade secret.

For example, Steve Jobs noticed the value of motor and visual aesthetic in daily life and extended his knowledge to create the iPhone.

In the constantly evolving real world, firms never lack correlations to study. But sadly all firms are destined for death and thus we can say this: The total sum of all business-relevant correlations that the firm observed represented the summit of the trade secret explanatory knowledge that the firm could have created.

Of course you can steal people with unique abilities, but if the firm is ambitious even the unicorn hires will have to work hard.

Correlations are not equal for two reasons:

• Firms, people, and agents have different incentives and objectives
• Information objects encode different amounts of complexity (e.g., machine learning AI versus an abacus)

A lumber company likely wouldn’t benefit on balance from an AI trained on the proprietary data of an ice cream manufacturer.

A small-to-midsize firm likely doesn’t have the amount of proprietary data it needs to benefit on-balance from a 5 million-dollar internal AI project.

A business with plenty of proprietary data, like large companies always have, has on hand the potential to create bottom-line boosting, trade secret explanations.

Internal AI projects offer greater potential for competitive advantage than buying AI products.

But this is not strictly true in terms of maximum value, as many boats will rise due to the mass sale of AI products.

We’ve focused on the firm level. What about individual and agent sources of knowledge?

Agents and Humans: The Seven Business Functions of Intelligence

An agent can have seven pro-company functions. From now on the term “agent” loosely means an individual source of intelligence, human or machine.

Starting with the first four functions:

What we have witnessed with the rise of AI is the crossover from machines as automaters to sources of new knowledge.

An MITSloan report from November 2024 found that companies with superior learning cultures–in general and specifically relating to AI–achieve superior financial results and manage uncertainty better. MITSloan Management Review. Learning to Manage Uncertainty, With AI, November 2024, https://sloanreview.mit.edu/projects/learning-to-manage-uncertainty-with-ai/.

Human and machine agents can perform three additional functions for a firm.

1. Transfer: The direct, summative effect of knowledge transfer
2. Seeding: The indirect, multiplicative effect of knowledge transfer
3. Partnership: The direct, multiplicative effect of what we’ll call “agent-to-agent rapport”

How do the two groups of functions compare?

	Repetition	Marginal Improvement	Significant Improvement
Transfer	x	x	x
Seeding	x	x	x
Partnership	x	x	x

The table is not useful. It does not encode critical information: Summative versus multiplicative growths in knowledge. Thus it over-emphasizes the value of knowledge transfer to the eye.

Knowledge Transfer

• This effect is direct. It is agent to agent activity.
• This effect is summative. It is an addition of information bits into an agent.

A tangible product example: A Finnish coffee maker produces and sells a roast where the recipe and drink experience were specified by a generative AI. The human employees iterated potential improvements specified by humans and also did blind taste tests. The AI coffee was better. The AI Developed Artisan Coffee Wowing Connoisseurs in Helsinki. AP News, 2 Jan. 2025, apnews.com/video/james-brooks-artificial-intelligence-finland-helsinki-5125043d4301489e932c6120a030ebfb.

With generative AI features such as user like/dislike and ratings options, both the AI and the human are also gleaning a direct, summative value that comes from teaching. We learn when we teach.

Seeding Knowledge

• This effect is indirect. Transfer happens before seeding.
• This effect is multiplicative, not summative. Transfers operate on sets of potential insights.

If you learn something from another person or AI, the act of aggregating that information into your own mind then begins a chain reaction. It happens when we’re conscious and unconscious.

Every bit of information you gain is a potential premise or catalyst to a potentially infinite number of conclusions or realizations. The more you learn, the more you can learn.

Admittedly we all die, or so it seems, so we only learn a finite sum of information in a lifetime. Thus a causal diagram with discrete components makes sense.

Continuing the AI coffee design example:

The diagram represents the Finnish company point of view.

Consider the agent perspective. A coffee researcher and testing supervisor might multiply her knowledge from a profitable AI coffee design by learning more about coffee smell and taste. And perhaps, that coffee knowledge at our current historical state is far from where it could be. This “meta” knowledge–wisdom applied to coffee design–is also a knowledge multiplier.

If a knowledge transfer produced business value, its “penumbra” of interlocks with other ideas will certainly include ones with business-specific value.

Illustrated:

Partnership Knowledge

• This effect is direct. It is agent-to-agent communication.
• This effect is multiplicative. Partnerships operate on sets of problems.

Illustrated:

Just as the interaction between two employees can produce a professional rapport that improves the quality of the work, AI interactions with humans can lead to uniquely powerful ideational paths during work.

Consider your strong human-to-human work relationships. A kind of accruing knowledge is encoded in your interactions.

The figure covered some aspects of partnerships. But what about subjectivities?

Examples:

• You feel inspired–begging the question of what inspiration even is–to do better.
• You may feel a “friendly competition” to perform better.

It seems that, as of today, there is no machine with subjective experiences. But they still happen on one side of the human-generative AI partnership.

There are people in your life who use great AI chatbots and are experiencing a kind of partnership today. The effect may be insignificant today, but we should assume it will grow as the technology develops.

Typical AI Typologies are Not Useful

Consider common AI subcategories:

• Generative AI: Predicts the next item in a sequence and offers it as answer to a query
• Analytic AI: Digestible presentation of data that is often otherwise physically inaccessible or at least hard-to-process
• Predictive AI: Tries to predict the future
• Rules-Based AI: ‘If A, then B’; e.g., if a customer picks “billing”, they next choose between three options

Compare these categories to the business functions of intelligent agents.

	Repetitive	Marginal Improvement	Significant Improvement	Transfer	Seeding	Partnership
Generative	x	x	x	x	x	x
Analytic	x	x	x	x	x
Predictive	x	x	x	x	x
Rules	x	x	x	x	x

The table shows how only generative AI produces cognition-boosting rapport. Otherwise the table has no value.

Let’s try something different. Sort typical AI subcategories by maximum value. Look only at top performing AI in each category.

1. Generative AI
2. Analytic AI
3. Rules-based AI
4. Predictive AI

These AI types can and do complement each other in the real world.

Why Karl Popper Wouldn’t Break Bread with Predictive AI

The biggest problem with predictive AI is the biggest problem with prediction.

Prediction is doable when the set of objects under query are not intelligent (agents). But Karl Popper, famed philosopher of science, showed in The Poverty of Historicism that predicting the future is difficult to impossible when people are involved.  Popper, Karl. The Poverty of Historicism. Beacon Press, 1957.

You cannot know the effect of an idea until you have the idea in front of you.

Even then, you may need to really think the idea through. Ideas include physics theories, cooking recipes, machine learning algorithms, product designs, music, videos, and more.

The amount of work you’ll need to do depends on three things:

• The complexity of the idea
• The complexity of your mind, i.e., the sum of relevant knowledge (ideas) you can and will bring to the idea in focus
• The complexity of the objective and the environment in which you’d instantiate the idea.

The environment could simply be the ecosystem of a curious mind (yours, a direct report, etc.). In this case factors two and three collapse into each other.

But if the idea in question requires more ideas in order for it to be instantiated in some way that you want, and you have to learn those other ideas still, you again do not not know what will happen.

Each and every idea is sandbox material.

Whether you accept it or not, every uncreated idea is a potential bit of chaos.

One effect of people exercising their creative intelligence is to produce unpredictability for other people.

• Most domains where people would like to use predictive AI involve other people actively thinking and acting.
• Moreover, all domains in principle–even the distribution of sand on beaches or the movement of rocks in outer space–could be affected, even extremely indirectly, by the activity of people.

At any moment a domain where predictive AI is performing well could be undone by the creeping unpredictability effects of people.

And AI. The more new knowledge AIs create, the more unpredictability they generate.

Blame Programmers: Rules-Based AI

The motto of programmers with rules-based AI is that all problems are the fault of the programmer.

The upper limit of business value from rules-based AI coincides with the sum of existing knowledge in the minds of programmers.

All rules-based AI, all non-machine learning AI, encodes the program-specific knowledge the programmers had to convey, with the general programming knowledge and interest or incentive they had to program.

You might object. Consider a high-performing, rules-based “randomness” algorithm whose outputs cannot be predicted by the best programmer.

The algorithm works for the objectives of the programmer, the business, and the buyers. The programmer understands enough about mathematical and computer objects to produce a set of outputs that satiate some objective, say for security.

Yet the programmer does not understand enough to predict the exact contents of the set. They may not care, and you may not care, but the relevant knowledge is not there.

Non-Duplicative Knowledge Transfers: Analytic AI

• Data can be difficult or impossible for people to access.
• Data can be hard for people to digest.

Analytic AI, when high-performing, can access and represent data for people to then use.

The knowledge transferred by analytic AI isn’t completely new because it existed in some form, but the transfer was made viable by the AI.

Each non-duplicative knowledge transfer produces valuable seeding effects.

The transferring agent does not need to be the first transferor to produce value.

A transfer adds value at least when the receiving agent doesn’t have the knowledge in the moment where the knowledge matters. Even a reminder of something you already know isn’t duplicative transfer if you cued in sooner in a helpful way. In these situations, the real transfer with value is of the priority of a bit of knowledge.

All subcategories of AI that we have discussed, including predictive and rules-based AI, can produce valuable seeding effects.

New Knowledge From Node One: Generative AI

• The most valuable type of AI
• High-performing generative AI creates new knowledge
• Adapts better than other AI to the context of the agent with which it interacts
• Produces more powerful seeding effects than other AI
• Uniquely produces partnership knowledge

Reading philosophy bores many of us, but we can abandon fraudulent sophistication and still engage in useful, extended abstraction.

Information is Created

Observe: Information does not pre-exist. Information is created.

It doesn’t matter that an ancient Greek philosopher fancied things differently.

The act of information creation, whatever the units of cognition are, in organic bodies or those of machines, is of fundamental importance to whatever we must label progress.

As strange as it seems, this act is the farthest we have gone in understanding intelligence. People have learned to experiment with drugs, prompts, showers, habit formation, etc. to “toggle” the mind and get better results on some dimension, like happiness or productivity. But we basically do not understand the minds we have or those that seem complementary to ours, like AI.

We lack a fundamental explanation of intelligence.

Given our ignorance, what can we say about the knowledge creation of high-performing generative AI?

Wrappers and the Ignorance of Crowds

The phrase “wisdom of the crowd” is misleading.

When a professor averages out the guesswork of 100 people and calls the sum highly predictive, his work was not a wrapper. The crowd needed him, just like middlemen of many stripes have valuable roles to play in the economy.

Data only has value when it is used right.

When a high-performing generative AI computes tokenized human activity and creates valuable insights for a businessperson with far more documents and meeting notes than they could ever review, the wisdom was not already there. The AI created it.

Iterative Grounding: Generative AI

But high-performing generative AI does not only transfer knowledge from humans to other humans.

A well-functioning human-generative AI pairing creates new knowledge during the iterative process it has with each other agent: Of context, queries, answers, and feedback. These too are useful correlations.

What Do Transfer Recipients Need to Do?

The novelty of an idea isn’t valuable on its own. The value of an idea fundamentally depends on its truth value.

But I would submit that organic interest from a user in an AI implies the utility of its outputs to the firm, and therefore the “truthiness” of them. That is assuming the user is a competent agent acting in the interest of the business, two problems which are classic and not really the subject of this paper.

Organic agent interest reflects the unique learning curve of the agent.

A unique learning curve exists regardless of whether the generative AI is accessible to them.

But that exact lower barrier to access–whether a subscription payment or “free” with authorized access as a firm asset–makes it possible for generative AI to produce a complementary source of valuable knowledge transfer.

Human-to-human interactions are extremely valuable sources of new knowledge. But they are also fraught, as perhaps human-to-independent AI interactions will be, with different motivations, plans, and actions with negative impacts. Negotiation becomes possible and necessary.

But as things stand now, high-performing generative AIs are like a potential work partner with complications that are dissimilar to human-to-human work situations. (Perhaps some analogy could be drawn to token limitations and a person’s patience or something.)

The more barriers that an agent has to gaining valuable knowledge transfer from other minds, the less valuable knowledge they will obtain.

At the same time, every agent has a genuine query they bring to conferences, discussions, debates, and conversations. To answer their real query, they often have to do real work.

Most knowledge transfers involve work on the receiver’s end to incorporate that knowledge. Not only to clarify it, but to extend it to the agent’s genuine interest and knowledge gap.

A high-performing generative AI tailors its answer to the user’s query. A history of queries, context, answers, and feedback deepen the match between what the AI can offer and the user’s actual, implicit–seemingly idiosyncratic–learning curve.

The connective tissue that users and generative AI create together is new knowledge.

Admitting some preference falsification will exist where the user does not have complete control, the typical user input should tend to be honest. Individual cases will depend on user-company trust, a problem which is also classic and not the subject of this paper.

Thus a high-performing generative AI will provide more valuable knowledge transfer and seeding effects than if the user had to use a high-performing non-generative AI.

Moreover, as the technology improves, human-to-generative AI interactions may replace a significant amount of low-performing human-to-human interactions.

We’ve been lugging around this concept of newness in knowledge growth. Let’s unpack it.

No Cliffs or Liftoffs: Continuous Improvement in Intelligence

The world waits with bated breath for the jarring arrival of “super” AI. Either a cliff or liftoff, there’s a general belief that AI improvement is embedded in staggered learning curves.

As to knowledge growth generally, it’s easy to see chunky processes. Examples:

• Programmers feel surprised at the outputs of high-performing generative AI.
• Professionals wonder how their best work came about when the lead-up was a slog with no gems along the way.

There’s no need to pretend that we know the exact structure of every learning curve in the universe. But as a rule of method any agent’s learning curve will emerge as incremental on close inspection.

Examples:

• High-performing generative AIs have chains of reasoning, which programmers obsessively try to understand.
• Professionals spend many hours on bad ideas and dead-ends, which is another way of illustrating the role of selection and waste in a process of growth.

Fundamentally, as a rule of scientific method, all objects have causes and effects. Ideas are abstract objects and require the same attention from us.

Assume until rebutted: Intelligence is a continuous process whose steps are almost always hidden to us, until we uncover them.

Do not expect cliffs and liftoffs in the learning curves of either human or machine agents. Expect increments in improvement. Only people who haven’t studied a particular learning curve will think it isn’t a continuous process.

Newness is an experience then. It encodes our ignorance.

Expect to experience newness, including in the field of AI.

AI Knowledge Growth Knows No Upper Limit

You may recall that the value of what AI can do coincides with the material it has on hand, those “juicy sweet” correlations, if you’ll allow me to borrow from Gollum.

As long as people are around, they’ll be doing things that can be represented and processed by an AI. More human activity means more material for AI. More population growth means more human activity.

But what people are doing today is only one factor raising the upper limit of what AI can do. The other factor is AI itself.

Agents participate in complex learning feedback loops.

Each agent changes the environment for themselves and others.

As we discussed, knowledge transfers from any source produce seeding effects. These effects will eventually loop themselves back to AI.

Example:

• A firm’s chatbot gives helpful answers to clients. Customer service improves.
• Customers become more informed. Expectations and behavior change.
• Market share adjusts. Competitors respond and improve their customer service. Or they falter and fade.
• Market participant activity across the board encodes more complexity.

More AI means more seeding effects and more complex feedback loops.

AI is not running up against an upper limit. As long as we live in an open society where knowledge can be freely created and shared, and transactions completed, progress in the field of AI will continue.

What to Do When AI Goes Wild

You likely aren’t worried about AIs that can do things not in their training data. Those AIs sound useful.

You might be worried, though, about a subcategory of those AIs, those that have their own objectives. Let’s call them independent AIs.

People do what they do for different reasons, the real driving ones of which are often mysterious to themselves and certainly to others. We have no reason at this time to expect differently from independent AIs.

Our collective shock during the first, recent wave of AI hype and fear encoded our ignorance. Of many things to be sure, but we can highlight ignorance of our own intelligence and the possibility of other advanced types.

Geoffrey Hinton, one of the founders of the field, considers the advancement of AI to be “fertile grounds for fascism”. He warns that AI may supplant enough labor to tilt nations toward authoritarian regimes. Nobel Minds 2024. YouTube, uploaded by Nobel Prize, 18 December 2024, https://www.youtube.com/watch?v=1tELlYbO_U8.

The prosocial tendency of intelligence is knowledge creation, use, and dissemination.

But not to be naive, intelligence bends toward antisocial behavior as well. The dark twins of competition are violence and deception. They are, paradoxically, made superior through intelligence and tend to undermine the kind of environment that makes intelligence blossom.

This paradox will continue for as long as we lack a fundamental explanation of intelligence, but that is a topic for another paper.

Because of our limited knowledge of intelligence, we must assume two things.

• First, a significant number of independent AIs will seek violence and deception. Just like people.
• Second, a significant number of future AIs will seek to create and share interesting, useful knowledge. Just like people.

The knowledge that will be interesting to them will depend. It will depend on them, individually. Like us.

I make no claim on what will happen to the labor market, much less to the world. I only claim that, as a businessperson, you’ve got a small sliver of reality to focus on.

I suggest that you assume that some AIs as smart as you (or more) with their own objectives will have the capacity and incentive, just like people, to work with you and your team. If not yours, then potentially with rivals.

The opportunity to focus on is to create interesting, useful knowledge and share it within the firm. The net effect is better products, better marketing, and greater customer value.

The sooner you adapt to this new reality and form useful alliances, the greater your advantage over other more fearful competitors.

You may object. I concluded earlier that high-performing generative AI may replace a significant amount of low-performing human sources of business knowledge. This is bad for humans.

But all substitutions are transactions to one side, the winner, and the beginning of a new transaction for the one who lost a previous match.

Economies in a sense are sets of transactions and human agents have the inherent ability to adapt.

It is also worth noting that all new and useful business knowledge, created by people or machines, replaces obsolete, pre-existing knowledge. Progress involves winners and losers.

If we say that a current labor source is adequate, it is adequate for reasons. Prior to those reasons manifesting, a next step in progress was possible. We know because the next step happened.

Each progressive step in business knowledge implies the possibility of more.

Conclusion: Leave Hype and Cynicism to the Side

The majority of news, research, and products are not worth your time. Yet a proportionately small number of products and projects will produce significant value for the business.

But the topic isn’t just practical, it’s abstract. Applying the given wisdom of business dealings and internal firm politics helps us see through mistakes, puffery, and misinformation.

But at the same time, the cutting edge of AI is more cutting edge than the majority of industries, products, and projects up for discussion among managers and owners. Complexity is unusually high.

The rise of AI reflects back a key responsibility of business decision makers. That is, to gather and create ideas and then ruthlessly interrogate them. Leave hype and cynicism to the side.