AI, My Current Thinking
Why I believe we need to slow down, regulate AI, and buy democracy time to catch up
Last summer, I went to a destination wedding. The bride and groom are both Ivy League graduates and invited about 100 of their closest friends. One night, they hosted a “Promethean” dinner. The rules are simple: no small talk.
We sat at a table of 10 people. Across from me, a woman in her thirties is starting a business using micro-AIs. Next to me, a man the same age who blogged about AI. The bride put me at this table to make sure I was educated. We talked about AI and the singularity for about two hours. It was eye-opening.
I am retired. When I worked, I programmed and led teams of programmers. I can code in over 25 programming languages, from LISP to assembly. Technology, per se, does not scare me. Though I am 70, I still think I can out-technology almost anyone.
When I got home, I put away my 3D printers and ordered a “gamer” desktop with lots of memory and massive GPUs. I started downloading open-source LLMs (often a generation behind commercial ones) and built several Python systems. I wanted to see, at a technical level, what LLMs were and why these 30-year-olds were convinced the end was near.
I am on the fence about their fears. But the risks are real.
In October, I started writing A Cold Civil War because arguments don’t change minds very well. Stories do. Before the Civil War, the Lincoln-Douglas debates were brilliant but changed almost no one’s thinking. Uncle Tom’s Cabin changed a nation. If we are going to address the risk, we need to address corruption, authoritarian power, and institutional collapse first.
Strongmen, tech billionaires, dictators, and authoritarians are not going to help us. Churchill said that democracy is the worst form of government, except for all the others. Elon Musk, Sam Altman, Bill Gates, Trump, Xi, and Putin are not going to save us from a potential singularity. As Lord Acton wrote in 1887, “Power tends to corrupt, and absolute power corrupts absolutely.” It is a universal rule in the same strict sense that gravity is a rule.
I think this is where the Cheney Doctrine applies: if there is even a 1% chance of a catastrophic threat, we need to treat it as a certainty in terms of our response. I don’t know the percent risk, but it’s clearly more than 1%.
My position is that we need to prohibit all new LLM development at its current stage, just as we have prohibited research into bioweapons.
I use that analogy knowing that there are research facilities where smallpox is still studied, but it’s done in BSL-4 containment, containment that is about as good as we can make it. When I say “freeze,” I don’t mean an absolute ban. I mean, we must limit research to very tightly controlled environments, as we do with smallpox.
This isn’t a new idea. We’ve done it before. In 1972, 183 nations signed the Biological Weapons Convention, banning the development and stockpiling of bioweapons. Inside the US, the CDC and USDA regulate who can even possess dangerous pathogens. The most dangerous organisms can only be studied in BSL-4 labs: sealed rooms, decontamination chambers, and pressurized suits. Research that could make a pathogen more dangerous requires federal approval before it begins. We built a legal and physical framework to make sure curiosity doesn’t become a catastrophe. We need the same framework for AI.
In the same way, we need to limit future research into LLMs and AI. What we need is an AI Research Convention, modeled on the Biological Weapons Convention. An international agreement that draws a line between safe research and dangerous research. It would establish containment tiers, as biosafety levels do for pathogens. Training a chatbot for customer service is not the same as training a system that can recursively improve itself. The risk and oversight are not the same. We did this with bioweapons, with nuclear testing, with chemical weapons. Every time, the argument against regulation was the same: we’ll fall behind, the other side won’t comply. Every time, we regulated anyway, because the alternative was worse.
There is a second question: people. AIs at their current level hold great promise but also pose risks of mass unemployment, power issues, cooling issues, and more. Mass unemployment, power consumption, and resource strain are political problems that require a functioning democracy and time to solve. The tech billionaires, dictators, and authoritarians are not going to address the risks. For them, AI only makes them richer and more powerful.
My position is to freeze new development, not ban current models. I use current LLMs heavily. This piece has been passed through Claude a lot. It’s made many serious suggestions for improvement, most of which I accepted.
I have been writing about my AI writing processes. I will continue to develop the AI writing section on Substack so everyone can understand how I use them, how they help, and how much of a crutch they are.
I am neurodivergent. That may explain my facility with programming languages. I have used computers to fill gaps since about 1979. I have felt the sting of rejection because I misspelled a word, used the incorrect tense, or made grammatical errors. Now with LLMs, I can guard against incorrect facts or convoluted prose.
I have heard “the computer is doing the writing” many times. Please forgive me, but to me this sounds like “he did not use a quill and parchment, thus the writing is slop.” I heard it when I wrote on a word processor and built a dictionary. I heard it when I programmed an electronic typewriter to print my output. Every time. I understand why people reject AI or call it ‘slop’. That’s one reason I advocate slowing down. We need time to digest this technology. How should we use it? When should we use it? How should we value it?
The goal of my novel is to change how people see the political world we live in. I feel a great urgency. It is impossible to write this novel in the time I have using traditional methods. My writing style, enforced by LLMs, is sparse. Here is a link to the instructions I use: Prompt-1 Style Check:
Active voice where possible
Short, direct, concrete sentences
Emphatic word at the end of the sentence
No qualifiers (very, quite, somewhat, rather) unless character voice requires them
LLMs produce smooth, polished prose that sounds authoritative. They have their own default style. It’s not mine, and that is not my goal.
If someone comes away from a scene feeling, “I did not realize it would hit someone that way,” then I have succeeded.
LLMs and AIs are a means to an end. I would be over the moon if my novel changed how people see this world and was never heard of again. It’s not about me or the sound of the words. It’s about how we see the world.
If AI-assisted writing offends you, don’t read my novel.