dynm

People are asking for evidence. FWIW, pangram classifies this post as 97% AI, 3% human.

I'm a bit skeptical of arguments of the form, "You should not use LLMs without disclosure because LLMs at bad at writing."

My reasoning is: If LLMs get better at writing—which I think is extremely likely—will you switch positions and say that now using LLMs without disclosure is A-OK?

Surely some people are willing to bite that bullet and say yes. But for most people, my guess is that the answer will remain no. Thus, I tend to think that the "real" reason most of us don't like it when people use LLMs to write without disclosure is that it's misleading: It's a sort of a claim that certain thoughts can be attributed to a human being when in fact they can't.

(The em dashes in this message were rendered using keyboard shortcuts.)

I think you're getting a lot of unfair pushback on this, when you're making some excellent points. Just to clarify, I think you're making these points:

- Existing air pollution causes a ton of harm.

- Realistically, the optimal amount of air pollution is probably not zero.

- Logically, we should treat the cost/benefit of adding this "new" air pollution the same as the cost/benefit of removing existing air pollution.

- As an absolute number, the harm caused by this facility is very large, but as a percentage it is tiny.

People seem to be interpreting your conclusion as, "therefore it's completely fine to build this facility because only percentages matter, not absolute numbers." But I think your conclusion is more like, "our society has a complex relationship with air pollution and our existing choices reveal that we are willing to accept some air pollution sometimes, so you have to do a cost/benefit analysis, not just look at costs."

I've withdrawn this claim, which I believe was overstated. Thank you for following up!

There's definitely quite a lot of scientific support, e.g. see this review paper: https://doi.org/10.1016/j.envint.2018.08.059

I'm open to the possibility that this is wrong, but as far as I can tell, the conventional scientific view is that high-CO2 environments are somewhat bad (although not catastrophic, even at 8000 ppm).

I don't want to stop anyone from debating the merits of using plants for carbon sequestration. But the ones explaining why my article did such a bad job of explaining those merits are kind of eye-opening, yes.

(Author here) I've added the word "indoor" to the title, which will hopefully make it harder to miss that this is not about climate change or carbon sequestration.

It might not be clearly worded, but 132.5 watts is the power used by 2 incandescent light bulbs, whereas 918 watts is the power present in the light produced by 765 light bulbs. The factor of 50 comes from the fact that incandescent lightbulbs are only ~2% efficient at turning power into light.

Like most people, surveillance pricing feels gross to me. But I think we need to separate two things: Surveillance pricing is a combination of (a) price discrimination and (b) surveillance.

If it were up to me, I'd probably give people much stronger legal rights to control their personal information, which would in turn decrease "surveillance".

But I haven't seen a convincing argument to ban price discrimination. There seems to be some kind of implicit assumption that price discrimination reduces consumer surplus on average. But I've never seen a convincing argument to that effect.

For example: Pharmaceuticals are often priced much low in poor countries. That's price discrimination. (You could think of it as a primitive form of surveillance pricing.) But if that were banned, pharmaceutical companies wouldn't lower prices for everyone. That would destroy almost all their revenue which comes from high prices in rich countries. Instead, they'd charge rich-country prices to everyone, and most people in poor countries would simply have to go without. That seems like it would be pure harm.

It's entirely possible that some kind of "coarse" price discrimination is OK/good but "fine-grained" price discrimination is bad. But where's the argument?

For what it's worth, I have a markdown-based website with a domain I own, and I stubbornly link to that website only, treating substack as just a temporary distribution channel.[^1] And I am constantly, endlessly asking myself if all the extra friction I've added to my life is worth it.

[^1]: I can't stand how slow and tracking-heavy substack is, but the problem of "send email and not get filtered as spam" is non-trivial.

Perhaps on brand for Meta, there appears to be no mention anywhere on this page or any of the docs pages, of any data retention policy. Maybe this is hidden in an agreement you need to sign when making an account? Has anyone actually done that?

> I'd like to know how long it takes to write a post like this.

In this case, about three weeks.

> The "nicer" sunscreens that use these ingredients often sneak in SPF boosters which are actually derivatives of other chemical sunscreens but are treated differently on the ingredients label, pretty much cheating the system.

Interesting, thank you for pointing this out. I had a little trouble understanding what the link was saying at first, but it seems to (correctly) state that many "mineral" sunscreens contain active chemical ingredients like butyloctyl salicylate. (And they're sometimes labeled as non-active ingredients?)

The FDA did (3 days ago!) finally approve a new ingredient: https://www.fda.gov/news-events/press-announcements/fda-expa...

My personal hot take is that we should all be using zinc (or titanium) oxide sunscreen which AFAICT maxes out both effectiveness and chemical safety. (And is the best for the fish?) Interestingly, these are the only ingredients that the FDA currently deems both safe and effective.

Ack, I'm sorry you had to waste your time transcribing the data! I uploaded it here: https://dynomight.net/img/coffee/temps.csv

How? By default, state governments can pass basically whatever laws they want. They don't have (theoretically) limited enumerated powers like the federal government.

Before the law, I think the state government or local governments could (by passing a law) restrict computing for any reason, even without a government interest. Now, they'd have to repeal this first.

I think the main content of this law (https://legiscan.com/MT/text/SB212/id/3212152) is just two paragraphs. I'd suggest reading them yourself rather than relying on secondary description:

"Government actions that restrict the ability to privately own or make use of computational resources for lawful purposes, which infringes on citizens' fundamental rights to property and free expression, must be limited to those demonstrably necessary and narrowly tailored to fulfill a compelling government interest."

"When critical infrastructure facilities are controlled in whole or in part by a critical artificial intelligence system, the deployer shall develop a risk management policy after deploying the system that is reasonable and considers guidance and standards in the latest version of the artificial intelligence risk management framework from the national institute of standards and technology, the ISO/IEC 4200 artificial intelligence standard from the international organization for standardization, or another nationally or internationally recognized risk management framework for artificial intelligence systems. A plan prepared under federal requirements constitutes compliance with this section."

In particular, I think the reporting is straight wrong that there's a shutdown requirement. That was in an earlier version (https://legiscan.com/MT/text/SB212/id/3078731) and remains in the title of this version, but seems to have been removed from the actual text.

As you say, it's sort of a cat-and-mouse game, and I'd rather not play it. Fortunately, there are plenty of competitors that don't require a phone number.

Not sure why you're being downvoted. It's unusual and harmful to privacy to require a phone number.

Not too interested in debating the semantics of "micropayment", but it sounds like if we swap in "news sites" in place of "LLM providers" everything should still still be possible? Consumers could pay tiny amounts of money for individual articles?

Everything you say makes sense. But can you help me understand why this doesn't also apply to the LLM service I use today? Doesn't that service, in effect, makes a "micropayment" to the LLM providers every time I make a query? Is the key difference that there are only a small-ish number of LLM providers? (Not doubting, just interested!)

There is a lot of hate for the idea of micropayments here, so I'd like to offer a counterpoint. I use a service that provides access to a bunch of different LLMs. Each time I call an LLM I, in effect, pay a $0.001 - $0.05 for the response. (Technically, this is implemented as me having to renew earlier.) Each time I make a call, I don't know if the answer will be useful. I don't even know how much it will cost! And in practice, the answers are often garbage, and I have to pay anyway. I find this annoying, but--to my surprise--only very mildly annoying. This has made me much more open-minded about micropayments for news / articles.

FWIW, the EU is working on zero-knowledge proofs: https://digital-strategy.ec.europa.eu/en/news/commission-mak...

But I strongly prefer my solution!

What I find puzzling about these proposals is that it SEEMS like they could be designed to achieve 90% of the stated goals with almost 0% of the loss of privacy.

The idea would be that devices could "opt in" to safety rather than opt out. Allow parents to purchase a locked-down device that always includes a "kids" flag whenever it requests online information, and simply require online services to not provide kid-unfriendly information if that flag is included.

I know a lot of people believe that this is just all just a secret ploy to destroy privacy. Personally, I don't think so. I think they genuinely want to protect kids, and the privacy destruction is driven by a combination of not caring and not understanding.

When you have lots of non-randomized dropouts from a randomized trial, that greatly weakens the causal link. The results are effectively non-randomized.

Meanwhile the evidence from actual drinking levels was much stronger (far fewer dropouts) and showed zero effect. Before this trial was done, you may have predicted that there would be positive results for the lab experiment but zero results in ecological conditions. But I think that prediction would be quite unusual. For anyone who expected results in ecological conditions (like me), this was disappointing.

The results for how much people actually drank in daily life were basically zero. No effect at all. The effects you're talking about are for a weird lab experiment where they sort of had people sit there in the lab and drink (or not). A huge percentage of people declined to participate in that experiment, too, which makes causality non-obvious.

I'm surprised that this article doesn't appear to mention the RCT on semaglutide and alcohol use disorder by Hendershot et al. that was published in JAMA Psychiatry in early 2025 (though it's possible I missed it) https://doi.org/10.1001/jamapsychiatry.2024.4789

This was largely portrayed as a great result in the popular press although personally I think it was a bit of a disappointment given all the amazing anecdotes https://dynomight.net/glp-1/

Incidentally, this is not a blog that makes it easy to look at the archives!

- No link to other posts

- This post is at https://relaxing.run/blag/posts/top-gun-landing/

- https://relaxing.run/blag/posts/ gives a 403

- https://relaxing.run/blag/ gives a 403

- https://relaxing.run/ gives a full-page picture of some beautiful mountains

- No Atom/RSS link hidden in source

Not a complaint! If this is an intentional choice, I respect it.

Contrary to what many replies are telling you, the link clearly states that if you don't own a smartphone, you can check in online and then obtain a boarding pass for free at the airport.

(Not sure how easy that will be or if they actually verify that you don't own a smartphone, etc.)