dynm

I think this is the main content of the law. (Everything below is quoted.)

---

Section 3. Right to compute

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 in public health or safety.

---

Section 4. Infrastructure controlled by artificial intelligence system -- shutdown.

(1) When critical infrastructure facilities are controlled in whole or in part by an artificial intelligence system, the deployer shall ensure the capability to disable the artificial intelligence system's control over the infrastructure and revert to human control within a reasonable amount of time.

(2) When enacting a full shutdown, the deployer shall consider, as appropriate, disruptions to critical infrastructure that may result from a shutdown.

(3) Deployers shall implement, annually review, and test a risk management policy that includes a fallback mechanism and a redundancy and mitigation plan to ensure the deployer can continue operations and maintain control of the critical infrastructure facility without the use of the artificial intelligence system.

Strange, right? Take a look at figure 4 here: https://jamanetwork.com/journals/jamapsychiatry/fullarticle/...

This is as in the abstract:

1. drinks/day declined in both groups and somewhat more in the treatment group but wasn't statistically significant 2. number of drinks/day basically wasn't different at all 3. drinks/drinking day didn't change in the placebo group but did decline in the treatment group

(These are all actually regression coefficients computed on non-random samples but nevermind.) Somehow it seems like what's happening is that 3 rises to statistical significance even though 1 doesn't.

OK looking at the original abstract:

> Semaglutide treatment did not affect average drinks per calendar day or number of drinking days, but significantly reduced drinks per drinking day (β, −0.41; 95% CI, −0.73 to −0.09; P = .04)

So they didn't find any reduction in (1) drinking, or (2) in the number of days that people drank, but they did technically find (3) a reduction in the number of drinks that people consumed on the days that they drank. So I guess what they said is technically correct... but I still think it's very odd not to mention the headline result that there was no actual reduction in drinking!

> A more recent RCT showed that low-dose semaglutide reduced laboratory alcohol self-administration, as well as drinks per drinking days and craving, in people with AUD [72].

I think this quote is... wrong? Or at least extremely misleading? Here, citation 72 refers to a paper by Henderson et al. That paper did (sorta) reduced laboratory alcohol self-administration, but did not find any reduction in the amount that people drank. https://dynomight.net/glp-1/

This paper states:

> The study protocol was approved by the Institutional Ethics Committee and registered with the Clinical Trials Registry of India

As far as I can tell, that registry is here: https://www.ctri.nic.in/Clinicaltrials/pubview.php

Doing a keyword search for the first author's last name reveals zero hits. (It's possible I'm missing—that search does not inspire confidence.)

It seems like if anything the current model would end up paying obscure artists more? (If you assume that people who listen to obscure music tend to listen to more music overall, which would be my guess.)

Hmmm, I use https://njal.la/ for DNS. Could spamhaus really just auto-mark every njalla user as suspicious?

This is a bit of a tangent, the whole concept of "domain reputation" can be infuriating. For example, my blog has been marked as suspicious by spamhaus.org: https://check.spamhaus.org/results?query=dynomight.net

As a result, some ISPs apparently block the domain. Why is it listed? I have no idea. There are no ads, there is no user content, and I've never sent any email from the domain. I've tried contacting spamhaus, but they instantly closed the ticket with a nonsensical response to "contact my IT department" and then blocked further communication. (Oddly enough, my personal blog does not have an IT department.)

Just like it's slowly become quasi-impossible for an individual to host their own email, I fear the same may happen with independent websites.

I think there is a "correct" answer to this question! The best book on Monte Carlo methods is this one: https://artowen.su.domains/mc/

It's not published yet, but already a classic. (Might be more intermediate than beginner, though.)

For something a bit more gentle, I also recommend chapter 29 of this book: https://www.inference.org.uk/mackay/itila/book.html

You seem to be responding as if I claimed that population-level correlational studies are bad, or that I claimed that air pollution is not bad for you. But I did not claim either of those things.

What I claimed is that this press release takes a population-level correlational study and presents it in a misleading way that implies causation was established. Which this press release most certainly does.

This article repeats the common mistake of conflating correlations and causality. The main results are (1) that PM2.5 exposure is correlated with dementia in humans, (2) some experimental results with mice. This does not establish causality in humans. The paper is careful to stay juuuust on the right side of the line by carefully using "associated" in the right places. But the press release discards that pretense at rigor and jumps straight to full-on claims of causality in people:

> Long-term exposure accelerates the development of Lewy body dementia and Parkinson’s disease with dementia in people who are predisposed to the conditions.

I think it's entirely possible (perhaps even likely) that this is true. But the paper does not show it.

FWIW, I think the wirecutter's quality for air purifiers is pretty bad, and likely more influenced by affiliate payments than science: https://dynomight.net/ikea-purifier/

If you're wondering the most obvious thing:

- Cost per mile: $4.72

- Minimum charge: $2296

There are also a huge number of other fees that I can't tell if you'd need to pay in practice, e.g.:

- Additional Locomotive Fee (per loco mile): $7.54

- Amtrak Locomotive Daily Charge: $2513

- Head End Power Daily Charge: $3433

- Annual Administrative Fee: $574

https://www.amtrak.com/content/dam/projects/dotcom/english/p...

I'm confused. Usually a "natural experiment" is a chance event that affects some random subset of a population. Here, they seem to be using "natural experiment" to refer to the event that someone decides to move to a different city. But obviously the subset of people in Amarilllo, TX who decide to move to New York, NY are going to be somewhat different than the subset who don't. So isn't this confounded?

It's really strange that they just jump into the paper and keep saying "natural experiment" over and over again without any justification that they actually have one. They do eventually get to this in the "Selection effects in relocation and mobile app usage" section, but I think they really downplay the seriousness of the issue.

Thanks for pointing this out. I submitted a request to spamhaus but it's an auto-responder black hole that tells me to contact my "IT department". (My personal blog, oddly enough, does not have one.) They don't explain why this domain which has no ads, sells nothing, and never sends email would be listed.

Just to be clear, these are hidden prompts put in papers by authors meant to be triggered only if a reviewer (unethically) uses AI to generate their review. I guess this is wrong, but I find it hard not to have some sympathy for the authors. Mostly, it seems like an indictment of the whole peer-review system.

Maybe they have some light to show if they're on. But is everyone supposed to just know that? Pointing a video camera at everyone you talk to is an... interesting social choice. Glasses like these should be designed with a physical mechanisms to cover the cameras.

(author here) This wraps JAX and JAX's version of NumPy. It would surely require some development to keep up, although it's quite short and simple (only 700 lines), so I don't think it would be a big burden. That said, I should be clear that my goal here is just to show that this is possible/easy, and possibly inspire existing array packages to consider adding this kind of syntax.

I like your alternatives! I agree that having to write

  X = dp.Slot()

before assigning to X is unfortunate. I settled on the current syntax mostly just because I thought it was the choice that made it most "obvious" what was happening under the hood. If you really want to, you could use the walrus operator and write

  (X := dp.Slot())['i','j'] = ...

but this cure seems worse than the disease...

Actually, doing something like

  X = new['i','j'] = ...

could simplify the implementation. Currently, a Slot has to act both like an Array and a Slot, which is slightly awkward. If new is a special object, it could just return an Array, which would be a bit nicer.

From the title, I read this expecting another lame observational study which I would probably distrust on the basis that it doesn't show anything causal. It's not that! Rather, if I understand it, they (1) took mice and introduced leukemia cells and (2) took human leukemia cell lines. In both cases, they found biomarkers related to leukemia growth.

(I welcome corrections to that understanding from experts!)

Personally, this seems far from convincing evidence that taurine in energy drinks is actually causing cancer. But it is suggestive and it seems like one might reasonably avoid taurine out of an "abundance of caution".

By complicated and unstated rules, I meant sort of norms about how you talk about previous work, how you caption figures, what order you have for the sections, how self-congratulatory vs. humble you are, etc. Theoretically, there are no rules for these things. But in practice, insiders seem to adopt a (deliberately?) complex ruleset and use it to signal to each other that they insiders.

I think the example of how to "correctly" cite a paper actually makes this issue seem smaller than it is. In reality, these conferences have very complicated (and unstated) "rules" for how a paper is supposed to look. If an "outsider" wanders in and submits a paper with new ideas, it will be very obvious that they are not a "member of the community" and their paper will usually be treated much more harshly as a result. This adds a huge amount of friction to research.

And what's particularly frustrating is that many organizers will try to combat this by writing papers saying they "particularly encourage" papers that are interdisciplinary, or focused on less fashionable topics, etc. It's good that they are trying to change things, but I think the main effect in practice is to encourage people to spend their time writing papers that have little chance of being accepted.

This issue isn't at all unique to computer science, though. Try publishing a paper in a top economics journal as an outsider!

The amount of text in books is surprisingly finite. My best estimate was that there are ~10¹³ tokens available in all books (https://dynomight.net/scaling/#scaling-data), which is less than frontier models are already being trained on. On the other hand, book tokens are probably much "better" than random internet tokens. Wikipedia for example seems to get much higher weight than other sources, and it's only ~3×10¹⁰ tokens.

There's absolutely no contradiction here.

Currently less than 20% of Americans work in factories. All those 80% need to want is that the 20% of people who want to work in factories can do so.

Yeah, it's quite disturbing that the differential entropy (unlike the discrete entropy) depends on the units. Even worse, the differential entropy can be negative!

Interestingly, the differential KL-divergence (differential cross-entropy - differential entropy) doesn't seem to have any of these problems.

It's tricky when you think of a continuous system because the "differential entropy" is different (and more subtle) than the "entropy". Even if a system is time-reversible, the "measure" of a set of states can change.

For example: Say I'm at some distance from you, between 0 and 1 km (all equiprobable). Now I switch to being 10x as far away. This is time-reversible, but because the volume of the set of states changed, the differential entropy changes. This is the kind of thing that happens in time-reversible continuous systems that can't happen in time-reversible discrete systems.

In the notation of this page, the entropy H(P) is best thought of as:

"The mean number of bits to encode a member of P, assuming an optimal code."

And the KL divergence KL(P,Q) is probably best thought of as:

"The mean number of WASTED bits if you encode members of P assuming that they had come from Q."

Most definitely different types of particles cause different levels of harm. (Extreme example: asbestos.) However, we don't really have good data to quantify this. It intuitive from evolutionary perspective that "natural" sources would be less harmful since we've been breathing sea spray and dust for millions of years. Yet, smoke from wood fires is clearly still extremely harmful. So... my instinct is that the harm is probably less, but I find it very hard to be confident that any particular source is totally safe.

Perhaps the most surprising sources of particular matter is... sea spray. As water crashes around, stuff in the water (e.g. salt) often ends up suspended in the air. This apparently contributes a non-negligible percentage of PM2.5 matter in coastal areas, though exact percentages are hard to come by.

I reckon if you wanted to choose just a single rule it should be, "Write something that you yourself would actually read." The problem is that our brains are designed to sort of lie to us and tell us that what we've created is amazing when in fact we'd never actually read it if someone else had written it. If you can find a way to be objective and see your own writing as the far-from-perfect mess it actually is.

(In principle, you could use "write something that someone else would actually read", but I think this is much harder, because it's much harder to know how other people would react! If you yourself would read it, well, we aren't that unique, lots of other people would read it too.)

Also, props for this stark picture of reality: https://refactoringenglish.com/chapters/write-blog-posts-dev...

Fun fact: Americans are seemingly allowed to own up to 1.5 kg of yellowcake. (That's pure uranium, refined from uranium ore, but not enriched to extract the 0.7% U-235 from the 99.3% U-238.)

Citation: https://www.nrc.gov/reading-rm/doc-collections/cfr/part040/p...