uniqueuid

Among my secret research ideas, this is the most dangerous and morally wrong one.

If future generations of researchers will wonder why IRB reviews became mandatory for computer science, studies like this will be the answer.

Seriously, some people don't seem to realize the point at which they are becoming Fritz Haber.

It's also about variance in the number.

Expert software engineers will still accidentally burn $500 or $5000 on tasks that don't work, or are not efficient. Amateurs will accidentally spend $100 to get something great.

So part of the change is a change in the risk structure of using frontier models. Before, you'd burn your quota; now, you can burn uncapped (less-capped) money.

I can relate to the attitude, but I also suspect it's a lot easier to live like this in one of the two richest countries in the world with the least stressful labor markets (going by the experience of my friends who emigrated there).

Whoops, misread the digit! Thanks for catching.

But the numbers are still the same (it's about rates not counts), just that we can expect the 40% figure to be more accurate than with 2M.

To put this into perspective, young children have an average amount of 16 colds per year (there's even a cool repo for incidence data, I suspect you could get more precise [1]). With an increasing number of parents both being in the labor market, and with childcare facilities increasingly (and rightfully IMO) refusing to take in sick kids, this means that parents need to take their own sick days to care for them.

It's difficult to do a back of the napkin calculation, but you can easily see how 16 colds x 3 days / 2 parents leads to 24 excess sick days for parents per kid (discounting for some overlap when multiple kids are sick) over the first few years.

[1] https://github.com/robert-koch-institut/GrippeWeb_Daten_des_...

I assume you mean there might be a self-selection issue with people who are voluntarily ("privately") insured, as compared to those who have the normal state-mandated ("gesetzlich") insurance?

This data is from AOK which is one of the state-mandated insurers. It insures around 2M people, and my gut feeling is that they are not terribly unrepresentative of the workforce as a whole.

But of course the point is, everyone with a tiny bit of true data could tell much more precise stories, and the journalists (as usual) didn't care or didn't think it would fly with readers.

To be fair, the median is just as little information as the mean. We live in a world where ink is cheap, they should just show a histogram.

That's exactly right.

This article [1] mentions 40% of sick days being from people with long-term (> 6 weeks) illnesses. That's data from one of Germany's large insurers. While I don't know the proportion of those with long-term illnesses, if we assume it's at most 10%, then the average for people with "normal" short-term illnesses is at most 12 days. So much lower.

[1] https://www.deutschlandfunk.de/sind-die-deutschen-wirklich-h...

Which is exactly (part of) why people call ZFS a "rampant layering violation"!

It's very simple, this is about the threat model.

If you are worried about big players profiling you (hard to avoid, high likelihood of happening, low likelihood of damage), then you want Sailfish.

If you are worried about apps profiling you (easy to avoid, high likelihood of happening, moderate likelihood of damage), you want Android or iOS.

Graphene and Sailfish sit on different points on that spectrum, just like OpenBSD and Linux do.

I have nothing but great things to say about typst, and this is my personal favorite from this release:

"A single document can now contain multiple bibliographies"

That was nicely put.

I think you can learn about it most by reading clever, capable people from big tech corporations. Their framing often involves tradeoffs against a slow but inevitable societal pressure that is helped by compromising on freedom.

So I don't believe they are ignorant of all your points; it's rather that they don't see a realistic way how tech, corporations, and perhaps even ordinary people can go forward (being better, or richer, or more sophisticated or whatever) without making that compromise. It's as if they saw the forking paths of the future, and none will end up without technical restraints, regardless of whether they do it or whether things just get worse and someone else then does them.

The causal loop you mentioned makes social science hard, but I’d argue that falsification and hypothesis-driven research can still work. Otherwise all the behavioral targeting Meta and Google and co are doing would not work.

The link is only to the marketing fluff piece, the paper is here and has three authors [1]

[1] https://www.frontiersin.org/journals/developmental-psycholog...

Thanks, that's a great point. I know a huge amount of the literature exactly in this area, and I'm honestly still pretty skeptical of the theoretical idea of the diversity principle. Mostly because there doesn't seem to be a good consensus what actually is a good amount of diversity.

But in practical terms, while there are some very obvious gaps and biases, I think the German media system is still pretty good in a lot of ways and everybody does have a lot of freedom.

I'd even go so far as to say - the typical accusations of censorship are mostly from right-wing actors who are not censored, but who demand more views and make a spectacle out of claiming to be censored.

I mean, that's a very intuitive question if you're not used to the system, and a bit surprising if you are.

The historic precedent is that Germany's first Chancellor wanted to establish a state-funded and state-directed TV station. This was explicitly shut down by the constitutional court.

As a result, a system was established to (1) ensure funding is not decided upon by state institutions but instead by an independent body of experts (KEF). (2) control over the meta-level content decisions is exercised by a body (essentially like a parliament) of representatives from societal groups (e.g. including politicians, doctors, churches etc etc.)

Now the "gross intervention in the media" is a very recent American idea - up until pretty recently the US did have the fairness doctrine, it has licensing and so on, all of these are gross interventions in media. And so are libel laws etc. So the German insight that underpins its media regulation is: You cannot have functional mass media without enabling them through some form of state action, you can only try to be light-handed and implement checks and balances.

Before this blows up:

- The world has very diverse ideas about how media regulation works. Germany has a strong public service broadcast tradition that is constitutionally anchored. You may not like it, but that's the cultural and legal tradition and changing it is not easily done. Please respect that especially coming from a country with a more liberal tradition (and perhaps a less functional media system, e.g. in the US).

- The source is a fringe right-wing outlet that most Germans would consider a bit suspicious (not necessarily factually wrong, but tasteless and a bit hysteric perhaps?)

- Public service broadcasting in Germany is not influenced by the state (note this says state, not politics or politicians) by design.

- Prioritizing public service broadcasters is a pretty logical conclusion from a certain tradition of media regulation and has precedent, e.g. in must carry rules for cable, EPGs etc.

So sure, debatable whether this is sensible, but it's at least neither surprising nor evidently nonsensical.

Source: Was a tiny bit of an expert on this for a short while.

The problem space that ducklake solves is smaller, but it helped me to get a working metabase dashboard quickly on ~1tb of data with 128gb ram. Queries were much, much faster than all alternatives.

Some downsides are: No unique constraints with indexes (can accidentally shoot yourself in the foot with double ingestion), writing is a bit cumbersome if you already have parquet files.

yes especially because homebrew will also setup deno for you, which is required for youtube itself.

It's interesting how the sentiment around Apple has turned, for the first time in what feels like 20 years.

The true reason is, as the recent norwegian report quipped: We love our tech, but it betrays us - that's an abusive relationship.

Consent prompts are a band-aid for users being exploited: They are not fixing the root but covering it with legal painkillers.

But the only true remedy is actually feeling in control of and empowered by your device - a vision that Apple once at least promised, but now has less and less legitimacy of heralding.

Publishers have the final say in appointing editors in chief (EIC) and editors. So they bear the ultimate responsibility for holding editors accountable.

A lot of people are to blame here, but Elsevier is definitely among them.

The dataset contains ~80% of food sold and inclusion in it is very probably skewed towards large volume. So the lower bound is something like 56% (if the 20% rest are not ultraprocessed)

Here it actually means 70%, but the paper is in a paper from mdpi which have been under criticism for predatory (i.e. fraudulent, junk-science enabling) practices.

From TFA:

"We report results of a cross-sectional assessment of the 2018 US packaged food and beverage supply by nutritional composition and indicators of healthfulness and level of processing. Data were obtained through Label Insight’s Open Data database, which represents >80% of all food and beverage products sold in the US over the past three years. Healthfulness and the level of processing, measured by the Health Star Rating (HSR) system and the NOVA classification framework, respectively, were compared across product categories and leading manufacturers. Among 230,156 food and beverage products, the mean HSR was 2.7 (standard deviation (SD) 1.4) from a possible maximum rating of 5.0, and 71% of products were classified as ultra-processed. "

While your assessment may be true in many contexts, this is not one of them.

American hyper polarization does not permeate other countries in the same degree and German academia is actually full of sober, level-headed, nuanced people.

Nitpick: The Kiel Institute is not a think tank in the sense that people would understand the word.

It is a federally funded research organization (part of the family of Leibniz institutes) similar to a university but without teaching. Here's a list of the others [1].

These are independent, high-quality research institutions without political money or a designated political agenda.

[1] https://www.leibniz-gemeinschaft.de/en/institutes/leibniz-in...

That's pretty easy.

If you have a randomized controlled trial, the sugar dose is varied and other confounding variables are controlled by randomization. So you measure the causal impact of sugar only. There are studies showing that.

With observational studies, if you have a dose-dependent effect, then that's good evidence (although not completely conclusive) of a causal relationship. This is what many studies do.

If you have a meta analysis covering many primary studies, and if those vary a lot of context (i.e. countries, year, composition of the population), and you still get a consistent effect, then that's another piece of support for a causal relationship.

The few studies that I've looked at seem to show a pretty robust picture of sugar being a cause, but there might be selection bias - i.e. we'd need an umbrella / meta meta study (which ideally accounts for publication bias) to get the best estimate possible.

It's really good to ask these questions.

I'm not a medial researcher, but my impression is that many fields find it difficult to produce the robust high-level risk comparisons that you ask about. I.e. if you're looking at blood fats, even there you'll find many complicated contextual factors (age, sex, ethnicity, type of lipids i.e. LDL or lp(a) or ...?). The same might be the case for sugar. So it's not really easy/cheap to combine detailed state-of-the-art measurements of different causes into one randomized controlled trial.

As for the effects of sugar, I think there's some evidence that's not too hard to find, e.g. some meta analyses showing something around 10% increase in dose-dependent risk (RR ~ 1.10) [1,2]. A lot of the literature seems to be focused on beverages, e.g. this comparative cross-national study [3].

[1] https://jamanetwork.com/journals/jamainternalmedicine/fullar...

[2] https://www.sciencedirect.com/science/article/abs/pii/S08999...

[3] https://www.nature.com/articles/s41591-024-03345-4

Point taken, but I'd prefer 200 lines of rocket science that I understand and control over 60k lines of (cleanly written and documented) rocket science.

(Although admittedly both plain typst and this project are still way less complex than LaTeX.)

I get the motivation, but it honestly feels a bit weird to use tens of thousands of lines of python code to do something that you can just directly do in typst.

I mean, a CV is not really rocket science and there are quite a few great typst templates out there.

Thank you for that work. I hope it's asymmetric meaning one hour of your work wastes thousands of hours for bad actors.