Wiki Contributions


Complete newbie question: is it possible to construct a version of these models that uses a 3 dimensional vector, instead of the 768 dimensional vector?

From the sound of it, the 768 dimensional vector is basically a constant linear transform of the three PCA components. Can we just declare the linear transform to be a constant array, and only train up the three components that appear to be the most needed? Eg generate the 768 from the PCA?

My personal bloodwork did not. It shows effectively no change over the past two and a half years - including zero change between late last year (when I was having serious difficulties), and last month (when I had been supplementing protein by about a hundred grams per day.)

So I don't know what that test is measuring, but it doesn't appear to be related to protein levels.

(It's also worth noting that not all protein is created equal. I had a fair amount of plant protein in my diet, but the amino acid profile used to make plant cells is not the same as the the amino acid profile needed to build muscle and repair sinew. I believe I was short on specific key amino acids.)

Congrats! I went through this thought process as well, and one of your three hypotheses above seems like the right one. Vitamin D isn't the issue (I have tests for it and have heavily supplemented for years), and sulfur itself isn't an issue (onions and broccolo are both pretty big in my diet). However, the lack of sulfur amino acids is the lead hypothesis.

Over the years, I had slowly shifted my diet more and more plant based: lots of vegetables, with occasional meat and a piece of fish every couple of days. As you mentioned, not all protein is created equal. While both vegetables and meat both contain all 20 amino acids, the ratios matter. Bodybuilders eat animal protein instead of vegetable protein for very good reason. I try to be active and try to keep muscle on my frame, and the plant based sources were just not providing enough of the key unsynthesizable amino acids.

Funny you should mention this; it made me check my records. It turns out that none of the doctors actually requested bloodwork. However, I do have my own bloodwork, which I do every 4-6 months on my own. Looking through that, what I see for heptatic protein level is 7.2+-0.2 for the past two and a half years.

This includes my most recent test, where I had been taking massive amounts of protein for months. So whatever that test is measuring, it doesn't actually seem related to the amount of protein the body has available or needs.

Yeah, a big part of the problem was expert ping-pong. I only saw one doctor twice. It's part of the global dysfunction I observed. Because of that, I have been seriously considering signing up for concierge medicine, which has a similar business model to what you describe above.

FYI, structural conformation diseases are actually quite common in humans, we just call them something different: amyloidoses. For example, TTR amyloidosis kills a good fraction of our oldest humans.


"Prions were originally defined as a unique class of infectious agents, whose infectivity relates solely to protein. In mammals, they cause fatal neurodegenerative diseases, such as Creutzfeldt-Jacob disease of man, sheep scrapie and bovine spongiform encephalopathy (reviewed in refs. 1 and 2). All these diseases are related to the PrP protein, whose conformationally altered form (PrPSc) is able to convert the normal host-encoded protein (PrPC) into this altered prion form. While only one prion protein is known in mammals, the prions appear to represent just a part of a much wider phenomenon, amyloidoses.

Amyloid diseases represent a group of more than 30 human diseases, which are characterized by deposition in different tissues of fibrous aggregates of conformationally altered proteins."

The thing that makes prion diseases scary is that they're potentially transmissible; but the reality is that even if you never catch a prion disease, a protein aggregation disease will eventually kill you if you live long enough.

Yep, that pretty much handles it. Thanks for the update!

I don't find it surprising; 0.1% is a fairly low bar here on LW. I'm not considered that unusual here, and my calibrated guess is that I'm in the 0.3% category. There's a million people in the USA alone at that level, and three hundred thousand at 0.1%. That's a wide pool to select from.

Personally I wouldn't be surprised if Musk was substantially above the top 0.1%. I've seen a number of technical interviews with him; he and I have similar backgrounds and technical field strengths; and we are approximately the same age. I feel able to properly evaluate his competence, and I do not find it lacking.

For a really good example of what I would consider a 'dumb' way for AGI misalignment to be problematic, I recommend "accelerando" by charles stross. It's available in text/html form for free from his web site. Even now, after 20 years, it's still very full of ideas.

(FYI, sections are about ten years apart in the book, but last I read it it seemed like the dates are off by a factor of two or so. Eg. 2010 in the book corresponds loosely to 2020 in real life, 2020 in the book corresponds loosely to 2040, etc.)

In that book, the badness largely comes from increasingly competent / sentient corporate management and legal software.

I totally understand where you're coming from, and I apologize for straddling the line on Norm One. I saw that it was heavily downvoted without comment, and that struck me as unhelpful.

Regarding the post itself, it wasn't a matter of being unable to understand. It seemed likely to me that there were insights there, and that if I spent enough time on it I could pull them out. It was more about the other demands on my time, which I suspect isn't a unique circumstance.

Regarding probability of mistake, I think that's an unhelpful way of looking at it. IMO it's not so much mistake, as 'interface incompatibility'. A particular presentation style will be varying levels of compatible with different readers, with different message acceptance and error rates at different receivers. The total transmission of information is across the distribution. Minor changes to the presentation style sometimes have outsized effects on total information transfer. Words are hard.

The general point is that if your goal is information transfer, it's less about 'mistake' than getting the best integrated product across the distribution. If you have a different goal ("just getting the words out" is common for me), then optimizing for delivery may be irrelevant.

Load More