However, most people who ask that do it while pointing to people who are actually quite important producers, such as financiers. Be careful not to conflate "don't produce anything of value" with "do something I don't understand the importance of".

Of course there are people in every profession who get paid a lot more than they're worth. This is less true of manual labor jobs where the output is easy to see and measure, but it's true across the board. Even in manual labor jobs you can have people whose output is negative. They may pick X apples or whatever, but they might do it while making everyone around them work slower.

Just to complete the description of the "empty chair" tactic, this is why lawsuits typically name anyone and everyone who might possibly be blamed, including many who clearly aren't culpable. It's not because the plaintiff or the plaintiff's attorney actually thinks all of those extra targets really might be liable, it's so that the culpable parties can't try to shift the blame to an empty chair, forcing the plaintiff to explain why the empty chair isn't culpable (i.e., defend them). Of course this means that those clearly non-culpable parties might have to defend themselves, which sucks for them.

So... you didn't actually look at the size of WikiMedia Foundation's bank account.

WikiMedia absolutely has enough money to run Wikipedia indefinitely if they treated their current pile of money as an endowment and just used the income from it to support the site. They don't have "billions upon billions", but they do have almost $300M, and they spend about $3M per year on hosting, and probably about that much again on technical staff to run the site, so about $6M per year. That's 2% per year. Assuming they can get a 6% average return on their assets, they can fully fund Wikipedia forever, and then some.

So, what do they do with all of the donations instead, if the money isn't needed to run Wikipedia? It funds the foundation's grant programs. Of course, you might actually like their grant programs. I think some of their grants are great, myself, and if they were honest about what they're using it for I might be inclined to give. But they're not, and the fact that they continue lying to Wikipedia's user base really pisses me off, so I don't give and I strongly discourage everyone I can from giving, at every opportunity.

Me too, but it took a lot longer and I was a lot less thorough. I would skim a half-dozen links from the search result, the LLM reads a lot more, and a lot more thoroughly.

I do. I tell the LLM to always include links to its sources, and I check them. Not all of them, but enough to make sure the LLM is accurately representing them. Granted that other people might not do this, but those other people also wouldn't check more than the first hit from the search engine, which is basically the same problem. If you only read the top hit, you're trusting the search engine's ranking algorithm.

That seems like a potential risk. I have't actually seen that happening in any of the stuff I've looked at.

Adverts pay for the web. And also clutter it up. Both of these things are true. Without advertising, there would be very little content that isn't paywalled, and there would be far less content than there is. Slashdot wouldn't exist, for example. The key is to keep advertising sufficiently profitable that it can fund the web, but not so intrusive that it make the web awful.

How do we do that? The best idea I've seen is to use adblockers that selectively block the obnoxious ads. But not enough people do it, so that doesn't work either.

This. There seems to be a widespread assumption that random genetic changes are somehow less problematic than carefully-selected ones because they're "natural" or something. It's not like cosmic rays, mutagenic chemicals, transcription errors and other sources of random genetic mutation are somehow careful not to make harmful changes. Engineered changes might not be better than random mutations, but they're clearly not worse.

Er, I guess I couldn't decide whether to say "closer" or "more closely" and munged them together. :-)

If only there were a way to preview posts before submitting them...

I'm not seeing why "launch and loiter" is beneficial. If Mars transfer windows were only hours, or even days, long, I could see that it's useful to launch early so that you don't end up missing your window because of weather or ground equipment problems, but the transfer windows are weeks to months in duration.

It seems to me that this strategy is mainly driven by lack of confidence in New Glenn, which makes sense given that it's a completely unproven platform. Over the 8+ weeks of the 2026 launch window they could certainly get to space with a reliable platform. Something like Falcon 9 might have some delays due to weather or minor technical issues, but it's extremely unlikely it would miss the window entirely. But New Glenn might have weeks of delays, so launching early might make sense.

What would make even more sense is if NASA is concerned that New Glenn might fail catastrophically. Making the attempt a full year early might provide enough time to build and launch a replacement.

Does anyone who follows this more closer have a better explanation?

Well, he renamed it and made minor changes around the edges so he could claim that it was broken and he fixed it. And then, of course, proceeded to violate the agreement he signed.

Excellent post, just a couple of comments.

First, it's interesting that Nikkos said "a previous administration", without naming it. It was, of course, Trump 1.0.

Second, international treaties on refugees don't require a country to accept every refugee and there are multiple examples where nations have made agreements that modify which county must handle asylum claims. For example, the US-Canada Safe Third Country agreement specifies that asylum seekers must make their asylum claim in whichever country they arrive in first. If the US and Mexico had a similar agreement, then refugees could not enter from Mexico at all. Trump tried to get Mexico to sign a Safe Third Country agreement, but Mexico refused -- and it probably would have been invalid anyway, since Mexico might not satisfy the requirements of a "safe" country under the US law that authorizes the signing of Safe Third Country agreements.

Instead, Trump signed the "Migrant Protection Protocols" agreement with Mexico, which was the "remain in place" agreement. You said that no other country should be willing to put up with it, but Mexico did formally agree to it, though only to avoid tariffs. Of course, Mexico has declined to renew the protocols in Trump 2.0 (though Trump announced they had, which Mexican President Claudia Sheinbaum immediately denied -- Trump's habit of unilaterally announcing that an agreement has been reached obviously doesn't really work).

Anyway, there are lots of reasons why countries might agree to various limitations on asylum processes to manage refugee volumes, and these agreements are often perfectly valid under international and national law. Trump, of course, doesn't care about legality, or humanity, only what he can get away with.

With a president who solves an existential threat by finding the best expert he can find, and using his own formidable political skills and charisma to run interference for that expert.

Seriously, Comacho was a meathead, but Iâ(TM)d vote for him.

It's also worth noting that NAFTA obliterated Mexico's smaller farmers which massively ramped up illegal migration from Mexico to the US.

No. Like any software, AI requires maintenance, and that maintenance costs money, lots of money.

It does not. Models need nothing more than the storage of some gigs of weights, and a GPU capable of running them.

If you mean "the information goes stale", one, that doesn't happen at all with RAG. And two, updating information with a finetune or even LORA is not a resource-intense task. It's making new foundations that is immensely resource intensive.

Can you integrate it into your products and work flow?

Yes, with precisely the difficulty level of any other API.

Can you train it on your own data?

With much less difficulty than trying to do that with a closed model.

I think the actual winners of this AI bubble will be whoever wastes the least money on this folly.

This.

I'm also surprised he didn't mention the housing bubble. Just because it's been running for decades and has been intentionally stabilized in an inflated state doesn't make it less of a bubble.