Tech Congestion Pricing
A new California law will fine robotaxi/AV operators if their disabled vehicle “blocks police or firefighters for more than 30 minutes.”
Believe it or not, here in San Francisco this is one of the weird things that’s actually an issue. Especially during rush hour.
I wonder if there’s a policy design lesson for frontier AI developers…
Department of Ruh-Roh!
Couple related tidbits from the always informative ChinaTalk:
at a recent Chatham House Rules event in DC, Democratic staffers indicated that should Congress flip, they are preparing to hold flat or even decrease the Pentagon budget.
Anduril said if the FY27 NDAA doesn’t pass soon, they will no longer be willing to work “at risk” for the government.
I’m sure it’ll all work out fine. No need to panic.
…Speaking of Panics
Cataloging historical techno-panics is an amusing pastime. It’s bracing to see what we collectively lost our minds about at one time or another. My friend Rob Raffety just launched a new series called Panic Patrol. The first episode focuses on “Bicycle Face” which is both as ridiculous as it sounds and conveniently only applied to women.
Revisiting these overreactions of the past can be a helpful reminder to put our present in context. Most of the time, extreme outcomes are seductive to the imagination, and social contagion gives them a popular boost, but they did not in the end reflect an evidence-based conclusion.
The “Y2K Panic” amounted to a nothing-burger, no? And yet…a categorical dismissal can sweep away justifiable concerns. There were high-value systems with demonstrated failures from the “bug,” like:
UK hospitals where an “unusual number of babies were being born with Down’s Syndrome” because women “in a high-risk category were wrongly told that they did not need further testing.”1
Both owed to Y2K-induced coding errors. The former was caught ahead of time to be fixed and the latter found only after an anomalous spike in real diagnoses. In the end it’s still fair to say the world overspent in response to a panic, but that doesn’t mean the optimal amount of concern was zero.
Remember the ozone hysteria of the 1980s? Probably not. Tl;dr basically some nerds warned of a depleting ozone layer in the atmosphere owing to proliferation of chlorofluorocarbons (mostly CFC-11, CFC-12, and CFC-113). Certainly no one’s worried about it now. Because we fixed the glitch. The Montreal Protocol (1987) led to a “nearly 99%” reduction in the offending substances.
In both cases you can argue the intervention (read: cost) was disproportionate to the actual problem. But that is not the same thing as concluding there was not an actual problem.
Which is what’s so tiresome about the reductive AI “debate.” Noting a problem — even a non-hypothetical, demonstrable risk — gets you thrown in with AI-pocalypse doomers. LOL, they’re just like those people who got all worked up about women riding bikes.
Please. I hate bikes being ridden by anyone.2
AI Models—They’re Just Like Us!
Maybe the whole solution is to train AI models to be more neurotic. A recent paper proposes that instilling conservative risk tolerances could constrain unwanted agentic behavior.
Risk aversion in resources could prevent misaligned AI agents from causing catastrophic harm. Misaligned but risk-averse agents would tend to favor safer strategies like making deals with humans over riskier strategies like rebelling. We train agents to be risk averse through character training, finding that persona traits provide a robust mechanism for instilling risk preferences.
In other words, trying to employ “do as we say, not as we do” classifiers may be exactly the opposite strategy we should be looking to.
Written in Invisible Ink
Forthcoming is a longer piece about how the lack of “a law” does not mean “permissionless innovation.” It’s just regulation through ad hoc, capricious, and compiled together separate pieces. Most of which is not fit for purpose.
Consider this: the day after the White House Super Intelligence Accord between the leading tech companies, ensuring a continued hands-off approach from the federal government, it came out that the FTC had weeks earlier initiated civil investigative demand letters against OpenAI, Anthropic, and METR. And under Section 5’s “unfair or deceptive acts” authority from a law over 100 years old to boot.
Good, bad, or ugly, that is what absence of formal law gets you. I dare you to argue it’s superior.
Etc.
From The Grand Resign:
Billing Me Softly: How AI Changes What Law Firms Charge and How
Three hypotheses on how Congressional leadership controls the agenda
and finally…
How cool was it to see Baltimore homies Turnstile on SNL? Hardcore lives!!3
It had to do with the system’s incorrect tracking of the women’s ages. See, Martyn Thomas, “What Really Happened in Y2K?” Gresham College (monograph), April 4, 2017, p. 11.
For a real doozy of the genre this see this article (City Journal Substack). Allow me to dunk on one of his premises:
…consumers and businesses are rushing to embrace AI. On the consumer front, the September release of Muse, Meta’s personal AI agent, set an all-time record for iOS downloads, with nearly 2 million in the U.S. alone in its first 12 days.
Not sure what this tells us. Indeed, “As of May [2026], 2.2% of consumers were paying for AI.” Doesn’t seem like such a rush. For comparison,
Over 50% of households in the U.S. pay for Netflix. So what?
Pokemon GO had 10 million downloads in its first week (2016), and Threads (also from Meta) had 150 million in its first week. Again, what does any of this actually tell us?
Then there’s this claim, stated as fact:
many, arguably most, truly critical systems are air-gapped, i.e., not connected to the public internet at all.
Simply not true…as he himself goes on to acknowledge
a case to be made for worrying about AI-enhanced hacking of physical, internet-connected systems, [like] power and water plants
Y’mean like when an IRGC-backed hacker group penetrated and disrupted access to some 100 municipal water systems across 11 states this summer with the help of AI?
Pick a lane my guy!