labelbias

thanks for the downvotes, I'm out of here for good!

I thought that i++ ++i difference is related to the inner data structures of C++ iterators and that when using the former you need to keep 2 states compared to just one for that fraction of execution.

Yet the best NNs are deep, not wide.

I'm also interested in this. How hard would it be to remember 440Hz and then just go up in your head, searching with octaves and then going up and down when close?

Lindy effect, stuff worth reading is stuff that has been filtered by time.

Reading old books is worth it. New ones are hard to filter and many will turn out to be completely unnecessary.

Leaves you on the sides of contemporary culture but at least you're reading efficiently.

The main thing about GPT-3 is that they wanted to demonstrate one-shot fine-tuning and succeeded at it.

So the model can be transformed to output part-of-speech words, dependency grammar trees or named entities in input even if training data is sparse. Similarily, you could fine tune it to produce game lore and then see how it works for that. The model easily switches to different modes of operation and achieves state-of-the-art or close to state-of-the-art performance.

It's quite funny how NLP folks tried to solve low level tasks (POS tagging, NER, Named entity relationship extraction, dependency parsing, sentiment classification etc.) to get to higher level tasks (good summarization, machine translation, text generation, question & answering) and now a single model captures all the low level stuff for free and does high level stuff so good that finetuning it to do low level stuff is unnecessary.

I think there is a large collection of evidence showing that education does not bring prosperity. It's prosperity that brings education. The importance of education is exaggerated.

The greatest managers of capitalism are authoritarian communists.

Just recently used xq to convert and clean up a bunch of weird xml files to JSON. It was such a breeze.

I love how nested lists inside objects can expand into a particular one with the .[] operator.

For example:

{ a: [{b: 5}, {b: 3}], c: 5} can be transformed into: [{ b: 5, c: 5}, {b: 3, c: 5}] using jq '{c: .c, b: .a[].b}'

For heavily nested XMLs I can get a nice flat output.

€45,000 does sound a bit low. I lived on €40,000 (after taxes, rent was only 8-10% of the yearly income) in Croatia and lived like a king (travelling, restaurants, buying stuff without worrying about the price, massive amounts of good quality food).

Moving to any other major European city would require me to at least double the income to maintain my standards.

Masai all have atherosclerotic plaque build up and it is also present in their children. Childhood with plaque forces various adaptations to blood vessels. They get wider and more elastic to counter the reduced blood flow and rigidity.

Anyone who isn't a child won't be able to make these adaptations and anyone who isn't on a starvation mimicking diet but is trying to gain or maintain weight on diets filled with dietary cholesterol will almost certainly start building up plaque.

Yes, dietary cholesterol has a low effect on blood serum cholesterol after some amount. So yes, increasing the consumption of eggs from 5 to 10 a day won't linearly effect the blood serum cholesterol.

Problem with fasting mimicking diets is that they are not a cure and are not long-term.

Also, consuming only things that are easy to digest makes you digestion more fragile to varying stimuli in the future. Which is why people who are on these "rejuvinating" diets for a longer period of time tend to experience their disease symptoms much quicker when they go back to eating vegetables rich in antinutrients.

Here's a very nice study where all Masai participants had atherosclerosis. Some even died from heart disease being less than 20 years of age: https://academic.oup.com/aje/article-abstract/95/1/26/167903

"These pastoral people are exceptionally active and fit and they consume diets of milk and meat. The intake of animal fat exceeds that of American men. Measurements of the aorta showed extensive atherosclerosis with lipid infiltration and fibrous changes but very few complicated lesions. The coronary arteries showed intimal thickening by atherosclerosis which equaled that of old U.S. men. The Masai vessels enlarge with age to more than compensate for this disease. It is speculated that the Masai are protected from their atherosclerosis by physical fitness which causes their coronary vessels to be capacious."

I got hooked on chess because of Lichess. The interface is well done and it's so easy to play. Something about it feels much better than chess.com. Unfortunately, I got heavily addicted and had to stop.

Although, I had so much fun. I was playing mostly openings with 2nd move being the one with queen. Got to about 1700 rating on that site in bullet (3-5 min game). This was the result of about 1 year of heavy addiction (2-10 hours of chess per day).

There's absolutely no way these oil wells are leaking as much methane as Siberian plain slowly melting. I did a calculation a while ago and it ended up being equivalent to 125 years of 2019 CO2 emissions.

Once these natural stores of gas start getting into atmosphere due to 2nd-3rd-nth order effects these little gas leaks from human made oil wells are nothing.

I can't handle dark mode. If letters are bright they burn into my eyes and follow my gaze. I use black on solarized background color for years after I dropped the white on black and it feels much better.

Fortunately, apples are not candy bar. The metabolic effects of eating 1kg of apples is completely different from eating glucose and fructose extracted from those 1kg of apples.

Natural sugar is magnitudes times better than processed/added sugar.

Fructose consumed through fruits has never had a similar metabolic effect as high-fructose corn syrup.

At my previous employer we were using GA and it's very flexible. We were able to track all DOM interactions by just configuring the included js. Although, querying that later with tools engineers can use and build is a whole different story.

Hmm... 256M records is still not that much for postgres or mysql so it shouldn't be too much for kdb/q. Several gb of data also shouldn't be too much.

RBF kernel represents vectors of infinite dimension and that makes the dataset linearly separable. Increasing dimensions improves separability and going to infinity makes it always true. Representation is not infinite but is implicitly infinite.

For example when you use polynomial kernels you do not add extra dimensions to input vectors explicitly but you could use a linear kernel and add quadratic terms to the input vectors and still get the same separability.

I'm wondering if these tasks have a form of bias that decreases the performance. If the model sees only positive examples and no negatives then it is biased on the positive paths of decisions. The moment where one changes the path to be incorrect, the model can't recover from the mistake because there weren't any negative examples during pretraining. There's many words that never follow some words but the model never sees that.