On April 16 2013, a team of highly skilled gunmen opened fire on the Metcalf Power Substation in San Jose California.
In just under 10 minutes, they disabled 17 transformers and caused $15m in damages.
This is the most important terror attack you've never heard of. Quick thread
The PG&E Metcalf substation provides most of the Santa Clara Valley with power-- Facebook, Stanford, etc. are all on this grid.
The attackers are still unknown, they were never caught, and the motive is still unknown.
A timeline of the attack:
12:58 a.m. – fiber-optic lines were cut not far from U.S. Route 101 just outside south San Jose. The substation loses internet and phone service.
1:07 a.m. – Some customers lost service. Cables in its vault near the Metcalf substation were also cut.
1:31 a.m. – A surveillance camera pointed along a chain-link fence around the substation recorded a streak of light that investigators from the Santa Clara County Sheriff's office think was a signal from a waved flashlight. It was followed by the muzzle flash of rifles.
1:37 a.m. – PG&E received an alarm from motion sensors at the substation, possibly from bullets grazing the fence.
1:41 a.m. – Santa Clara County Sheriff's department received a 911 call about gunfire, sent by an engineer at a nearby power plant that still had phone service.
1:45 a.m. – The first bank of transformers, riddled with bullet holes and having leaked 52,000 US gallons of oil, overheated, whereupon PG&E's control center about 90 miles (140 km) north received an equipment-failure alarm.
1:50 a.m. – Another flashlight signal, caught on film, marked the end. More than 100 expended 7.62×39mm cases were later found at the site.
1:51 a.m. – officers arrived, and found everything quiet. Unable to get past the locked fence and seeing nothing suspicious, they left.
In the subsequent investigation, it became incredibly clear how professional of an operation this was. Of the 100+ shell casings found-- all had been wiped clean of fingerprints.
there were also stacks of rocks found all over the site, commonly used to gage firing distance.
they knew exactly where to attack-- shooting directly at the cooling fans, the weakest part of the transformer.
they knew where to dig to disable fiberoptic cables, and the location of every camera.
there still exists no footage of how the gunmen entered the site.
in the months following the attack, the US gov ran a simulated attack on the electrical grid.
You can read the report below, but the results were terrifying. A group of unskilled actors could easily disabled a majority of the US grid.
1) moral alignment matters more than incentive alignment
people focus too much on aligning incentives. incentives are messy and can hardly be aligned. find people you share convictions and faith with, and keep working with them for long amounts of time. it'll work out.
2) even on long days, it should be fun
there's a difference between challenging and exhausting. the best people are extremely challenging, but never exhausting, and always in enough control of their emotions to know when to step back after a long day.
within months you will be able to buy genomics data from 14 million americans for +/- $200m?
the inevitable fire sale of this mess to an overseas PE firm is going to be a national security matter on the scale of which we haven't seen in healthcare in years
in general, hhs has left open a dangerously large hole around healthcare data sales
the reg we have now are so deeply embedded in precision oncology/2010s-RWD that they are completely unprepared to address what post-LLM healthcare data sales will actually mean
ONC should build federal data lake, incentivize state funded systems to contribute, allow companies to access these data for training/benchmarking but not directly touch/view the data.
use this as a backdoor to build a non-SAMD regulatory pathway for healthcare AI.
The Executive Order on AI issued by the White House today represents the beginning of a complete overhaul of the regulatory landscape for healthcare AI.
Here are my notes on what matters from the 100+ pages:
Let's start with a quick summary of the EO's requirements:
- Within 180 days, HHS shall publish a plan to promote responsible AI use in public benefits like Medicare/Medicaid. The plan should address access to benefits, notice, evaluation for unjust denials, etc
(Section 7.2b)
- Within 365 days, HHS shall establish an AI safety program that partners with Patient Safety Organizations to create a framework for capturing clinical errors caused by AI, analyzing data, and developing informal guidance aimed at avoiding these harms.
healthcare is having its top deck of the titanic moment
what happens in the next year will define the next century of american healthcare, and basically everyone is ignoring it.
here's the real story:
Healthcare has been defined by four factors: 1. Extremely limited supply (MDs are scarce and costly). 2. Principal-agent problem with payments (your insurance pays, not you). 3. Inelastic demand and high trust (you need it and have confidence in it). 4. Regulatory capture.
this has allowed the industry to behave in incredibly weird ways
cost can constantly grow, patient experience can constantly degrade, and clinics will still fill up with waitlists.
because patients trust it, need it, aren't paying for it, and have no choice.