I publish a monthly email newsletter with personal updates and interesting things I read or learned that month. The latter is archived below. If you’d like to be added to the newsletter, email me.
Alberta established a 17,000 square kilometer "control zone" on their
south-eastern border. Their other borders are too cold (north) mountainous
(west) or sparsely populated (Montana) for rat populations to survive. The
invention of warfarin (a pesticide that carries limited health risk for
humans) helped get local populations on board.
"Albertans have grown so unaccustomed to rats that they frequently mistake
squirrels, gophers, and other small animals for them: of the 875 reported
sightings in 2025, only 47 turned out to be actual rats."
French citizens are legally required to reserve 50-75% of their estate for
their children depending on the number of children. This has the unfortunate
effect of both discouraging philanthropy and having more children. France's
fertility rate is 1.59.
John Arnold suspects
similar laws across Europe make it less dynamic.
Jack Clark (Co-founder, Head of Policy at Anthropic) thinks there is a 60%
chance that by the end of 2028 a frontier model will be able to autonomously
train its successor. He believes the majority (if not all) of AI engineering
can be automated by today's systems. Whether these systems can automate AI
research is less clear. He cites Gemini solving 2 open Erdos problems and
says if you squint you can see signs of field-advancing creative intuitions
(this was written before an OpenAI model solved the unit distance problem).
"mostly the field of AI moves forward through humans methodically going
through some loop of taking a well performing system, scaling up some
aspect of it (e.g, the amount of data and compute it is trained on),
seeing what breaks when you scale it up, figuring out the engineering fix
to allow it to scale, then scaling it again. Very little of this requires
extremely out-of-leftfield insights... As the public data above shows, AI
has got extremely good at performing many of the essential schlep
components of AI development...This means even if AI systems are
relatively uncreative, it feels safe to bet they can push themselves
forward - albeit at a slower rate than if they’re able to generate novel
insight"
In response to the federal government's blocking of Fable's release
"Here are a few of the constraints on the U.S. government, not the only
ones I might add:
It needs for the main companies to stay in business. On top of that,
it wants their IPOs to go reasonably well. And it is now much harder
for the top companies to recruit foreigners, which is a significant
share of their highest quality workforce (Demis, Ilya, Andrej for a
start). It is also much harder for the main companies to drum up
foreign business in a credible and sustainable manner. 1b. How are
American multinationals operating abroad supposed to use top systems,
moving forward?
It wants to use model access as a tool of both hard and soft power, so
model access has to be possible at some level. But it is very hard to
control what foreign agents will do with their partial model access,
when they get it in the future.
The U.S. needs to stay ahead of China in the AI race.
The U.S. needs to issue restrictions that are actually enforceable,
and “U.S. citizens only” does not fit that bill. Furthermore (markets
in everything!) it is easy enough to hire a traitorous American to
access tools of wrongdoing, or for matter it is not difficult to fake
citizenship in various ways.
USG cannot nationalize these companies and then proceed to run them
effectively.
Chinese and other open source models do in fact improve at some
reasonable pace, even if they are right now considerably behind the
best proprietary models."
Dean Ball's (OpenAI, former White House staffer) lessons from the federal
government's recent actions and how it should address AI safety moving
forward. He wants the government to certify private auditors which monitor
the frontier AI companies. They would hold these companies accountable to
their own safety policies until an industry-wide policy is agreed upon.
Ever wonder why, despite the popularity of Calvin and Hobbes, there are no
themed plush toys, cartoons, or coffee-mugs? Bill Watterson fought for a
decade with his print syndicate against licensing of his characters. A fun
story of Watterson and that history.
F.D.A. approves bemotrizinol for use in sunscreen in the U.S. It provides
better UVA coverage than existing ingredients and allows for a
lighter-feeling formula. The F.D.A. is 25 years behind Europe in approving
the ingredient because it classifies sunscreen as a drug rather than a
cosmetic, which must pass a higher bar.
An interactive financial model comparing the cost of space-based to
terrestrial data centers. It will be really hard. Only SpaceX is positioned
to do it, and will require Starship's success and a vertically integrated
supply chain.
