Tests like these shouldn't be a "gotcha." They should be fair.

* Had employees already been through some basic anti-phishing training?
* Did the test include something that someone who paid attention during the training course would be reasonably expected to spot?

There is one other use for tests like these: To survey the population of employees: What % of employees clicked on the link, and what do we need to do organizationally (not "just for those who failed") to lower this number to something we can live with?

But it shouldn't be a "gotcha."

They can turn Ukraine into glass. "At all costs" means they will start and lose World War III because most of their nukes were spent turning every hectare of Ukraine into uninhabitable wasteland. But for the few minutes that Putin lives before the incoming missile vaporizes him, he can laugh maniacally and say "I beat Ukraine! Ukraine is no more!"

--
For the gallows-humor impaired, this is NOT a serious suggestion. There are no winners in a nuclear war. If he's even thinking about ordering a nuclear strike on Ukraine, even one bomb, hopefully one of his aides will be brave enough to make sure such an order is never followed.

Back in the 20th century some power companies would put your hot water heater on a "controllable" meter that could be turned off during periods of high demand. The rate on that meter was lower than the rate for the rest of the house.

Industrial customers who are willing to be "turned off" can also get cheaper rates from many utility companies.

The work this agency is or will be doing has probably been done for decades. This may just be a re-organization to bring people together under one agency.

... that wetware has vulnerabilities too?

Or to put it another way: "Oops, I didn't mean to do that, sorry."

The use case I envisioned is experts who COULD do the work by hand use AI as a tool to be more efficient. This assumes they are expert-enough that will recognize if their AI of choice is giving a false-positive rate that makes their work harder, not easier. It also assumes that since they are being paid for their time (but perhaps only a pittance or not at all if they are students), they will be paid by their employers or funders as much for reviewing the AI's output as they are for the other parts of their research.

... the usual caveats apply: An expert human, or team of humans, needs to look at what the AI is suggesting, understand what it is suggesting, test it with the same or more rigor they would give to the same suggestion if a human had come up with it, and not "sign off" on it until they are satisfied that it is what they say it is.

It's the humans who are putting their professional reputation and, in some cases, their professional licenses on the line if they get it wrong. The AI is incapable of caring if it's correct or not.

I am going to make a dangerous assumption and assume that does not mean "we think we should use the written word to train our LLM" as that isn't actually anything that requires consensus. So presumably, the "team" discussion was about how to acquire said books. That the "consensus" answer was "through shadow libraries" sounds an awful lot like "conspiracy to commit commercial copyright infringement" and appears to be a rather shortsighted argument to be making.

Wait, "some other company" can replace Falcon in "a few years" but everyone is "20 years behind" a design that has only been under active development since late 2018? Your math ain't mathing, dude.

You are young. I remember the older days when you had to set up an external modem to connect to a university mainframe that MIGHT have internet. With no pictures.

No ftp or telnet from your actual home computer. If you wanted to download files, kermit, xmodem, and the mainframe's equivalent of 'cat' were your friend.

An AI that can parse my query and give me a useful result faster than me reading through several pages of documentation is a good thing. But this means the AI and its training data have to be suited for the task. There also needs to be two "escape hatches" - one to give me the pre-AI way of doing things and another that will get me to a human being, at least during business hours.

As for how Google runs its support pages, I don't have enough experience to comment. If the existing documentation is poor, I wouldn't expect better from an AI.

In the long term - 5+ years down the road - big projects like these will want on-site or near-site power or at least on-site battery storage so they can buy power when it's the cheapest.

On-site power should also be attractive to investors: If the data center flops, you still have a power plant that can sell power back to the grid.

--

By the mid-'30s, on-site base-load power may mean small nuclear plants. There are logistical, regulatory, and political/NIMBY reasons why that's not a good solution for the 2020s.

How much of this data needed to be "live" and how much of it should've been in "near-line" storage or in offline storage to reduce the risk of bulk-exfiltration?

Which is pretty funny, since the only major difference between the two these days is "Is Google spying on you, or is Microsoft?"

If selling becomes economically less viable than leasing or renting or a library-subscription, then selling will no longer be offered by vendors.