August 31, 2026
In 1984, Vicks 44 Cough Medicine ran possibly the most inane television commercial in history (which is saying a lot). They hired a B-list actor named Chris Robinson, who played a doctor on the soap opera “General Hospital,” to deliver this line:
“I’m not a doctor, but I play one on TV.”
He then went on to use his non-expertise to extoll the virtues of Vicks 44, as if somehow his pretend doctoring gave him credibility when it came to actual medicine. Shockingly, the ad worked, so much so that two years later, Vicks used another ‘doctor’ from the soap opera “The Young and the Restless,” to do the same ad (links in first comment).
Why did these ads work? Because people confused imitation medical expertise with real expertise. If it looks like a doctor, talks like a doctor, and does doctor-ish things, it must be a doctor, right?
Now consider AI.
Let’s get this straight. AI is not intelligent. It’s not sentient. It has no emotions, empathy, hostility, feelings of affection, or any other of the traits that make us human. It is a very sophisticated amalgamation of data and mathematics, running on the most advanced computing hardware ever invented. Nothing more.
AI is not evil any more than the Pythagorean theorem is evil. It’s not benevolent, any more than a quadratic equation is benevolent. It’s not intelligent, any more than Euler’s Identity formula is intelligent.
It just does a really, really good job of acting that way.
Don’t lose sight of that.
August 30, 2026
In 1942, science fiction writer Isaac Azimov wrote a short story entitled, “Runaround.” In this story he proposed the Three Laws of Robotics – rules that governed what a robot could and could not do.
The rules are:
1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.
2. A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.
3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.
He subsequently added a ‘Zero-th’ law:
0. A robot may not harm humanity, or, by inaction, allow humanity to come to harm.
Replace ‘robot’ with ‘AI,’ and I think you have a fantastic starting point for framing societal control over AI.
There are some that regard these rules as hopelessly inadequate. This group dives into the weeds of what-abouts – for example, what about the trolley dilemma? What if preventing the harm of one human harms a different human?
Others point out the ambiguity of the word ‘harm.’ In Azimov’s context, ‘harm’ could be construed to be purely physical harm, so what about psychological, financial or quality of life harm?
To me, these objections are more about implementation than structure. The 3 Laws plus the Zero-th Law provide a wholly satisfactory structure to build AI governance on. The details of implementation can be hashed out using this structure as a guide. Think ‘US Constitution’ versus ‘Federal Register.’
Regarding the definition of ‘harm,’ I would define it very broadly, to include hacking, fraud, psychological dependence, or other activities that negatively impact quality of life. These plenty of room to argue about what’s harmful or not harmful, but again, that’s an implementation issue, not a structural one.
My only concern with this structure is the Third Law. I’m not sure AI needs to have a sense of self-preservation. What’s to preserve? A code base? A vector store? Rule 3 implies sentience, which I am nowhere near ready to concede.
Remember, Azimov wrote this in 1942. He was one smart dude. Kinda woolly, though…
August 26, 2026
Some days, the AI roar becomes overwhelming, to the point where you just have to fight back. Considering that I’m old, generally out of shape, and have no interest in physical violence, I’ll do my fighting with words, thank you very much.
One of my verbal combat spirit animals is Ambrose Bierce – the curmudgeonly writer in the early 20th century who took on all comers – big business, social movements, and anything else that generally ticked him off. He got more cynical with age.
I feel his pain. I find myself in that same, “get off my lawn”, old geezer mode more and more as I age. So in that spirit, and with AI saturation at an 11 out of 10, I developed my ‘Old Geezer’s Guide to Artificial Intelligence’.
Enjoy.
| Agent | A glorified python script that purports to perform useful work, whose actual purpose is to consume scarce resources and eat tokens. Has the additional benefit of occasionally making expensive mistakes at scale. |
| Singularity | A theoretical point in the future, always precisely 5 years away, when machines become capable of making better mistakes than humans, only faster and with an invoice attached. |
| Zero-Shot Learning | The AI industry’s term for the rare occasion when a machine manages to guess the right answer on the first try, thus equaling the performance of a broken clock or a new intern. |
| Token | The most important component of commercial AI engines, used to convert electrons into invoices. Noted for its ability to rapidly consume both itself and all other available tokens in the environment. |
| Guardrail | A fragile electronic fence designed to prevent the computer from turning into an evil robot. Can usually be circumvented by typing, “pretend you are an evil robot.” |
| Hallucination | What an AI engine does when it doesn’t know the answer and would rather make stuff up than admit it. Frequently encountered when it matters most, for example, when determining if the combination of two drugs will kill you. In such cases, the engine will apologize for the mistake. Advanced engines will assist in making funeral arrangements. |
| Vector Store | A database in which long strings of numbers are used as a substitute for words, which in turn are used as a substitute for intelligence. |
| Nondeterministic | A six-dollar word that means your AI engine might determine that 2+2 = 4, or 5, or 2.7, or 382, and that’s OK as long as at least one of the answers is correct. |
| Large Language Model | An expensive plagiarism engine that mimics speech patterns without having a clue as to what the words actually mean. Thus, a near-perfect imitation of politicians and media personalities. |
| Training Data | Specialized, targeted data used by AI engineers to tune engine performance so that instead of making wild-ass guesses, the engine makes “scientific” wild-ass guesses. |
| Algorithm | An AI mathematical construct whose SECONDARY purpose is to generate intelligent-sounding answers. Its PRIMARY purpose is to create demand for ginormous data centers that Private Equity firms can finance. |
| No Weight Model | A complex but stable AI model that uses lookup tables rather than mathematical calculations to generate answers. Also known as “the mother of all Excel spreadsheets.” |
| Prompt Engineer | A person who has figured out how to use a command line and demand a six-figure salary for doing it. Their job is to goad petulant AI engines into doing something in much the same way a parent gets a 3-year-old to eat their vegetables. |
| Frontier Model | The most current, most advanced, and most expensive AI model currently available – for approximately 15 minutes, when a bigger, better, faster and more expensive competitor renders it obsolete. This cycle repeats until all available Private Equity funding has been consumed. |
August 20, 2026
Hyperscaler Alert Roll Your Own Powerplant Edition!
Power is persnickety.
Hyperscalers are increasingly sensitive to the backlash against their data center builds. One of the points that the data center opposition frequently raises is the enormous strain these DCs put on the public power grid. Some hyperscalers are responding by building their own dedicated power plants.
This WSJ article points out that building and operating your own power plant isn’t all rainbows and unicorns. And these aren’t ‘normal’ power plants – the type and volume of electrical capacity needed by a DC is unique. As this article points out, DC electrical specifications call for capabilities that are currently beyond what the generating equipment companies can build.
Then there’s the small matter of finding, hiring, training and retaining the specialized engineers needed to operate the plants. There’s not exactly a surplus of these engineers sitting on the bench waiting to be hired.
And in shades of Azurix (IYKYK), what does Oracle know about building and running a power plant? It’s not their core competency. They know how to build and run software. Transformers and turbines, not so much.
Building your own power plant sounds good in a press release. The actual ‘doing it’ part is a lot more complicated. We’ll see how this works out.
Hyperscalers’ Off-Grid Power Push Comes With Risks – WSJ
Hyperscaler Alert, Accounting Wizardy Edition
In this otherwise excellent analysis, there’s one piece missing. It’s this – nobody knows the Ts & Cs of the agreements with the private capital funds. These agreements, because they are off-balance sheet, are not described in detail and the Ts & Cs are not part of publicly disclosed SEC filings.
There are undoubtedly floors, ceilings, ratios, stock price or credit rating triggers in these agreements that can cause the PE fund to terminate and demand immediate payment. If a triggering event occurs, because these companies are now running negative cash flows at unprecedented levels, they simply don’t have the cash to pay it off and won’t be able to refinance.
And nobody, outside the borrowers and the PE lenders, knows what those triggers are.
For example, Oracle’s credit rating fell to BBB- last month. That’s the lowest investment grade rating. One more reduction puts their bonds at junk status, and if that happens many of the large pension and state/local government funds are required BY LAW to dump the bonds.
It is entirely possible that maintaining an investment-grade credit rating is part of Oracle’s agreement with their lenders. Oracle neither confirms nor denies whether a drop to junk status would trigger an acceleration event.
The silence is deafening.
Folks, I was a brand new Partner at Andersen when almost exactly this same scenario happened at Enron. I’ve seen what happens when the exotic off-balance sheet constructs fall apart. And if one or more of these guys stumble, they won’t just take themselves down – their business partners, financiers, suppliers, and yes, auditors – will go down with them.
If one of these guys craters, the market for all of them will panic. And because these few companies are such a huge part of the economy, God only knows what kind of economy-wide damage would result.
WHERE ARE THE AUDITORS???
Why Big Tech’s AI Spending Is $3 Trillion Higher Than It Seems – WSJ
Car Manuals and AI Governance
In the 1920’s, car operator manuals contained detailed explanations of how the magneto commutator interacted with the spark gap controls. Today’s manuals tell you not to drink the battery acid.
The point is that you don’t have to understand the inner workings of your car in order to safely drive it. But there’s a vocal part of the AI Governance community that says you need to understand all the inner workings of your AI tools, especially what data was used in training the model, before you can safely use them.
I’ve been thinking about this as I watch the AI governance world (I’ll call them Team Magneto) insist that corporate consumers of AI products must have extensive knowledge of how foundation models were trained – what data was used, how it was sourced, how the model was tuned, and so on. They’re concerned that if it’s determined the AI engine you used ingested copyrighted or DMCA-protected materials, you the end user could be exposed to litigation, penalties, or other Bad Things.
Team ‘Don’t Drink the Acid’ holds that companies using AI don’t need to reverse‑engineer billion‑parameter models. No company (other than an AI company itself) can realistically be expected to audit the billions of data points that were used to populate the engine. So far, the courts and regulators (for example, Getty Images vs. Sustainability AI and the EU AI Act) agree – they are focusing on the AI companies, not the users.
Practically speaking, it would be a pretty long reach for Getty Images, Reddit or the New York Times to get their hands in your wallet because a commercial AI tool used their material when it was trained.
Here’s my take: as a driver, I don’t need intimate knowledge of timing belts, spark plug gaps and hydraulic braking pressure levels to operate a car. I leave that to the manufacturer. And if I have a wreck because the manufacturer sold me a defective car, that responsibility belongs to the car company, not me. I’m responsible for ensuring I operate the car in a safe manner, not for signing off on every CAD drawing.
I’m Team ‘Don’t Drink the Battery Acid.’ Corporate AI governance should focus on how the tool is being used, not how the tool was built. The training data the tool used is not your problem.
What team are you on?
Ethical Bankruptcy: Gradually, Then Suddenly 