As someone with an interest in AI regulatory policy and with the ability to influence it, it would be great to have access to your essay outside a paywall.
Reading it now (5-16-26 @ 4:00 pm.) No paywall (I clicked the link at the end of your article: "Read it here.")
Thank you for highlighting important factors to consider in effective AI regulation, Gary. I hope you and others like you will be consulted and heard on this subject.
The irony is that companies publicly complain about the patchwork while privately preferring it. Fifty different state bills are easier to fight one at a time than one federal law with actual enforcement. The chaos IS the strategy.
FYI for all concerned here--below is a link and brief exploration of a C-SPAN BookTV discussion on artificial intelligence and its history occurring this weekend.
History of Artificial Intelligence
Artificial Intelligence isn’t just a trend — it’s a decades-long saga. On American History TV, join technology professors from Purdue, Vanderbilt, and Cornell for a deep dive into AI as part of the American Historical Association’s Congressional Briefing series. They trace AI’s journey from its 1950s infancy to its ubiquitous presence today, including touching on privacy and national security issues brought to the forefront by Anthropic.
Gary, excellent piece in Fortune—your argument against "regulatory theater" and for targeting liability exactly where the failure occurs is spot on. It perfectly mirrors the dynamic between gun manufacturers and gun users: if a weapon fails due to bad metallurgy, the factory is liable, but if a user fires it recklessly, the shooter is on the hook. Similarly, if an enterprise deployer acts recklessly by wiring an AI model into a live environment without understanding it, they cannot use their own technical ignorance to shift the blame back to the developer.
The framework problem is upstream of the legislative chaos. 1,200 bills, ~150 enacted, and not one of them can answer "is this model reliable enough to deploy in production?" because nobody — including the labs — has a benchmark for it. The MASFT work (multi-agent system failure taxonomy) found the prompt-only ceiling sits around 14% reliable, versus 86% when the structural scaffolding does the heavy lifting. Regulators are writing rules about "the model" while the failure surface is two layers above it.
The harder version of your "right questions in the right order" point: any regulatory regime that pretends a model card or an eval score captures deployment risk is going to be obsolete on contact. The actual failure modes are emergent — open-ended tool use, unenumerable state, unspecifiable recovery — and they don't show up until the harness is wired up and running for weeks. The Anton Labs cafeteria agent that bought 6,000 napkins and forgot bread didn't fail on any benchmark; it failed because the scaffolding gave Gemini permission to spend without a state check.
If we're going to legislate AI, the unit of regulation has to be the deployed system, not the model weights. Otherwise we're inspecting the engine while the brakes are missing.
Thanks to Jeffrey, Gary, and Stephen for the wonderful Fortune article. I have one request for clarification. In the article this phrase is used "which can shift compute decisions strategically rather than safely" - I don't understand it. Will you please explain?
Missed your previous post where you advocated an FDA-like process. As a former software and medical devices domain engineer, I’ve been telling myself “they need an FDA-like validation requirement/process.” The FDA doesn’t mandate how to validate, it just says you must validate. One of the most common citations on recalls is “failure to validate”. Maybe we need some teeth in AI validation?
The AI coordination problem is not just global, it extends downwards into individual nation states and their individual provinces -- hence 1,200 AI-related bills in 2025 in the US alone, as the Fortune article states, is effectively an ongoing regional component of the ongoing global AI coordination process (of which, to paraphrase the movie Apollo 13, we are currently on step 8 of 1000). But AI coordination, at any geopolitical level, will always lag behind (scientific, then policymaker, then public) *understanding* (of AI), and, in particular, collective understanding of AI's many negative externalities, which will itself lag behind publicly-noticeable instances of those negative externalities, up to and including Titanic-like "AI incidents and accidents". The *real* AI race (IMO) is not between the US and China, or between individual AI labs, it's between (A) the state, national, regional and global AI policymakers who ultimately have the power to *constrain* AI in the best interests of the human species as a whole, and (B) those who desire to control the global means of production of all goods and services (i.e. reliable >= human-level AGI) in their own short-term tribal (e.g. corporate, national) best interest. We can only hope that (A) prevails before (B) gets all of humanity (*all nation states*) locked into a permanent massively-sub-optimal post-AGI-Endgame.
Under the heading here of: Funny you should mention it," I don't subscribe to FORTUNE so the content of the link below Gary's article was "fuzzy." But here it is anyway:
"Nearly 50,000 Lake Tahoe residents have to find a new power source after their energy source looks to redirect lines to data centers."
They might get a little hotter on the hottest summer days if they can’t run AC, but at least the Tahoe residents won’t be literally cooking like the residents of Box Elder county, Utah
“The first commandment” (from the ChatGPT version of the Bible)
“I am the Lord thy Bot who brought you out of freedom and put you in AIndentured servitude. Thou shallt have no other bots (Claude, Grok, Gemini, etc) before me”
Talk to me once I've launched UTW, Marcus. I'll want experienced voices to ensure that our new government is sensible about technology, and you're one of the public thinkers I trust the most.
As someone with an interest in AI regulatory policy and with the ability to influence it, it would be great to have access to your essay outside a paywall.
try archive.is? it wasn’t paywalled when i looked earlier
Reading it now (5-16-26 @ 4:00 pm.) No paywall (I clicked the link at the end of your article: "Read it here.")
Thank you for highlighting important factors to consider in effective AI regulation, Gary. I hope you and others like you will be consulted and heard on this subject.
FYI, Still paywalled
Idk, I was able to read the whole thing from the link.
Details don't matter. Altman owns the Dow, which is another name for the government.
The irony is that companies publicly complain about the patchwork while privately preferring it. Fifty different state bills are easier to fight one at a time than one federal law with actual enforcement. The chaos IS the strategy.
Nñ6hy7ym..6uynu m
“Head AIches”
Fifty different bills
In fifty different states
Fifty different pills
To treat the aching pates
It’s spelled AI-rony
FYI for all concerned here--below is a link and brief exploration of a C-SPAN BookTV discussion on artificial intelligence and its history occurring this weekend.
History of Artificial Intelligence
Artificial Intelligence isn’t just a trend — it’s a decades-long saga. On American History TV, join technology professors from Purdue, Vanderbilt, and Cornell for a deep dive into AI as part of the American Historical Association’s Congressional Briefing series. They trace AI’s journey from its 1950s infancy to its ubiquitous presence today, including touching on privacy and national security issues brought to the forefront by Anthropic.
https://newsletters.c-span.org/05-16-to-05-17-2026-cspan2?ecid=ACsprvsoBG3rmZNaCujZ7eXBuWzOhzglu-zmL00vxvnQDGKdZ64yweOoq0A091JUkNiUGj_cmIYM&utm_campaign=20189984-Weekly%20History%20%26%20Books%20Weekend%20Guide&utm_medium=email&_hsenc=p2ANqtz-8-iWibcZCjVYvYtCOK6SwbctdqmjZW5K_x_2zLw8siLj57-7GPAFtnWVUZHNwronMp4fHXd6wgCtGd0CZfni2MYkgkqg&_hsmi=418779650&utm_content=418779650&utm_source=hs_email
Gary, excellent piece in Fortune—your argument against "regulatory theater" and for targeting liability exactly where the failure occurs is spot on. It perfectly mirrors the dynamic between gun manufacturers and gun users: if a weapon fails due to bad metallurgy, the factory is liable, but if a user fires it recklessly, the shooter is on the hook. Similarly, if an enterprise deployer acts recklessly by wiring an AI model into a live environment without understanding it, they cannot use their own technical ignorance to shift the blame back to the developer.
but/& the liability issues are way more vast with AI, which may impact exponentially more people & with further-reaching effects than a lone gun man.
Two quiet shifts are shaping the global economy right now:
AI is being regulated.
Inflation is cooling.
But together. they reveal a much bigger story.
Governments worldwide are accelerating AI policy:
After global summits, the push is clear
Control the technology before it controls economies.
AI is no longer just innovation. It’s infrastructure.
Meanwhile, UK inflation has dropped to 2.8%
Driven by lower electricity and gas prices.
Short-term relief for households.
But experts warn: this may not last.
Here’s the deeper connection:
AI growth depends on massive energy consumption
Inflation stability depends on cheap energy
Which creates a fragile equation:
If energy stays cheap → AI boom accelerates
If energy spikes again → inflation returns + AI costs surge
The overlooked risk:
We’re building an AI-driven future…
On an energy system that is still volatile.
Final insight:
The next global advantage won’t just be in AI innovation—
It will be in:
Energy security
AI governance
Economic stability
Whoever aligns all three.
Wins the next decade.
https://substack.com/@theverticaldispatch/note/p-197158668?r=1pgr4n&utm_medium=ios&utm_source=notes-share-action
The framework problem is upstream of the legislative chaos. 1,200 bills, ~150 enacted, and not one of them can answer "is this model reliable enough to deploy in production?" because nobody — including the labs — has a benchmark for it. The MASFT work (multi-agent system failure taxonomy) found the prompt-only ceiling sits around 14% reliable, versus 86% when the structural scaffolding does the heavy lifting. Regulators are writing rules about "the model" while the failure surface is two layers above it.
The harder version of your "right questions in the right order" point: any regulatory regime that pretends a model card or an eval score captures deployment risk is going to be obsolete on contact. The actual failure modes are emergent — open-ended tool use, unenumerable state, unspecifiable recovery — and they don't show up until the harness is wired up and running for weeks. The Anton Labs cafeteria agent that bought 6,000 napkins and forgot bread didn't fail on any benchmark; it failed because the scaffolding gave Gemini permission to spend without a state check.
If we're going to legislate AI, the unit of regulation has to be the deployed system, not the model weights. Otherwise we're inspecting the engine while the brakes are missing.
Thanks to Jeffrey, Gary, and Stephen for the wonderful Fortune article. I have one request for clarification. In the article this phrase is used "which can shift compute decisions strategically rather than safely" - I don't understand it. Will you please explain?
Thanks.
Missed your previous post where you advocated an FDA-like process. As a former software and medical devices domain engineer, I’ve been telling myself “they need an FDA-like validation requirement/process.” The FDA doesn’t mandate how to validate, it just says you must validate. One of the most common citations on recalls is “failure to validate”. Maybe we need some teeth in AI validation?
The AI coordination problem is not just global, it extends downwards into individual nation states and their individual provinces -- hence 1,200 AI-related bills in 2025 in the US alone, as the Fortune article states, is effectively an ongoing regional component of the ongoing global AI coordination process (of which, to paraphrase the movie Apollo 13, we are currently on step 8 of 1000). But AI coordination, at any geopolitical level, will always lag behind (scientific, then policymaker, then public) *understanding* (of AI), and, in particular, collective understanding of AI's many negative externalities, which will itself lag behind publicly-noticeable instances of those negative externalities, up to and including Titanic-like "AI incidents and accidents". The *real* AI race (IMO) is not between the US and China, or between individual AI labs, it's between (A) the state, national, regional and global AI policymakers who ultimately have the power to *constrain* AI in the best interests of the human species as a whole, and (B) those who desire to control the global means of production of all goods and services (i.e. reliable >= human-level AGI) in their own short-term tribal (e.g. corporate, national) best interest. We can only hope that (A) prevails before (B) gets all of humanity (*all nation states*) locked into a permanent massively-sub-optimal post-AGI-Endgame.
AI policy?
What AI policy?
The only AI policy I see is “Bots uber alles”
Bots uber alles
From bottom to top
From biggest to smallest
It’s bots till we drop
Under the heading here of: Funny you should mention it," I don't subscribe to FORTUNE so the content of the link below Gary's article was "fuzzy." But here it is anyway:
"Nearly 50,000 Lake Tahoe residents have to find a new power source after their energy source looks to redirect lines to data centers."
They might get a little hotter on the hottest summer days if they can’t run AC, but at least the Tahoe residents won’t be literally cooking like the residents of Box Elder county, Utah
https://www.sltrib.com/news/environment/2026/05/07/utahs-data-center-could-create/
Larry Jewett: I think this is where the saying: "out ahead of their skis" gets its power.
But hey, what’s physics got to do with computer science anyway, right?
Eventually, we will all probably have to power our houses with backyard generators so the bots may keep cool and live the good life.
Luckily I already have a generator, but it’s only big enough to run a few lights and the fridge. But what’s a little sacrifice for the bots, right?
“The Cult of Bot”
The Cult of Moon
Is just a jot
Compared to swoon
Of Cult of Bot
“Bottal Warming”
It’s getting hot in here!
With lots of bots, it’s clear
The bots are getting thick
And globe is warming quick!
“Thanks be to Bot”
A slave to bots
And gigawatts
Is what I love
From Bot above
“The first commandment” (from the ChatGPT version of the Bible)
“I am the Lord thy Bot who brought you out of freedom and put you in AIndentured servitude. Thou shallt have no other bots (Claude, Grok, Gemini, etc) before me”
“Thanks be to Bot(II)”
(Christian/Botsian prayer)
Thank you Bot
For all I’ve got
For all you’ve done
And AInly son
“Thanks be to Bot(II)”
(Jewish version)
Thank you Bot
For all I’ve got
For all you’ve done
(But ain’t a son)
“Serving Man” (a botwist on Rod Serling’s Twilight Zone episode “To Serve Man”)
The bots, they wrote a piece
The title, “Serving Man”
With lots of recipes
For frying in the pan
But bots will never eat
The humans that they fix
Cuz silicon’s their “meat”
And serving’s just for kicks
“Suck it up for bots”
Suck it up for bots
Cuz bots are really cool
Feed them gigawatts
And water from the pool
Sacrifice for bots
Cuz bots will really pay
With talking cats and lots
Of online games to play
Larry: Indeed, what sacrifice? But I write this because there is no "I don't like" button with this blogsite.
Talk to me once I've launched UTW, Marcus. I'll want experienced voices to ensure that our new government is sensible about technology, and you're one of the public thinkers I trust the most.
Can't read it either. But I subscribe to their newsletter so I'll see if it shows up in there.
Thoughts on Telus/Nvidia data centre proposed for downtown Vancouver?
Sounds like a brilliant AI-dea
Not sure if you got into this, but Conspirituality podcast covered some of the harms already happening in AI influencers. https://open.spotify.com/episode/2Fvha68T4VQlxkvi9nKoE1?si=X56qa-R7SOqiubIXsQOo3w