ctoa
The thing that made these breakthroughs feel important was that they seemed to be open ended problems that the system itself made independent progress on. That the results were not primarily great retrieval into the corpus of mathematical research + stitching together.
A major point of having novel frontier problems as a benchmark itself is to try to sidestep this issue a bit, where it's hard to tell if one is evaluating reasoning vs retrieval capabilities.
But at least one of the problems was identified as definitely being plausibly mostly retrieval, since it hinges on but does not cite a 2016 paper that the author identified. Closely enough that the author describes it as plagiarism. Another problem seems to combine results from 2016 and 2019.
Ignoring the question of credit, it makes it look like OpenAI doesn't actually understand the eval well in the first place, and undercuts the notion that the results represent great leaps forward in mathematical reasoning.
~80% of acetaminophen poisoning cases severe enough to end in hospital are intentional suicide attempts.
Those patients also have better outcomes though: family members know what happened, they get them to a hospital, they can take N-acetylcysteine as a timely antidote.
In the unintentional poisoning group, ~90% is from multi-day accumulation, they don't realize they've been poisoning themselves. They have a much higher rate of acute liver failure and death.
https://doi.org/10.1186/cc1475 https://doi.org/10.1016/j.cld.2013.07.001 https://doi.org/10.1097/mcg.0b013e31818a3854
Fatalities happen basically from: product stacking combination medicines people don't realize contain acetaminophen and/or multi-day accumulation exceeding max daily limit over multiple days.
Chronic drinkers have impaired acetaminophen processing so they can't handle otherwise safe doses, but fatalities still typically occur in multi-day accumulation scenarios. Their safe daily max is ~half that of a non drinker.
The effects of one round of acute drinking don't impair the liver in the same way. People are not dropping dead because they took a normal dose of acetaminophen for a hangover. Not that I'm recommending you start doing it, but it is a myth.
OED (171k) is comparable in content to RAE + RAE Americanisms (93k + 70k = 163k) because OED is pan dialect. RAE core dictionary by itself is not the same type of coverage as OED.
The 1M+ English number is essentially a garbage number from a media tracking company called Global Language Monitor and includes things no serious dictionary would include. 300k+ is for a very comprehensive legit Spanish dictionary with technical and historical words. Those two numbers aren't comparable.
"which you admit is a far lower figure in Spanish vs other languages." no you just misunderstood, the numbers are very comparable.
Looking to your perceptions of ngrams in pop lyrics is not a great way of doing linguistics, there are in fact many alternatives in Spanish for expressing emotions about love, whether you are aware of them or not, or whether song writers over use certain words. English lyrics repeat an awful lot of "love" "heart" and "baby".
The rest is not right, the word counts in particular is just a reflection of why counting words in a language is hard:
93k comes from the number headwords listed in the core RAE dictionary. The RAE's dictionary of Americanism adds another 70k entries. When you include historical and technical terminology, more comprehensive dictionaries will have well over 300k words.
Counting "over a million" from English comes from way, way more inclusive counting methods that throw in technical jargon, acronyms, global slang, etc. The OED, which would compare to the RAE dictionary numbers, lists only 171k words.
Beyond this, counting is complicated by the fact that compounding and morphological changes work differently, English will use different words in cases where Spanish would use suffixes, and will count compounds as words that in Spanish would be phrases.
"there's only a few ways to express yourself in Spanish compared to other languages." is very wrong. Spanish word order is massively more flexible than English, grammar and morphology more nuanced, they do things in different ways but this is a deep misunderstanding.
Linear bounded automata (LBA) the halting problem is decidable. But many properties of LBA are undecidable:
Emptiness: Does an LBA reject all possible inputs? Universality: Does an LBA accept all possible inputs over its alphabet? Equivalent: Do two LBA accept the same language? Finiteness: Does an LBA accept a finite number of strings.
Neuralink is just another medical device company right now. The things it is succeeding at aren't the crazy sci-fi Elon "telepathy" vision at all, and there are other BCI companies.
Boring Company is quite stagnant and I would say what they achieved is very underwhelming. The original framing was that it would revolutionize the costs of subway and high speed rail tunnels and it has not achieved that at all. The things it currently does are nothing like the vision Musk was selling in 2017, much of the original concept was shelved.
Robotaxi is wildly off schedule and his timelines that he puts out for it have been non-stop over optimistic and wrong. This may catch up eventually, but it's currently not at the point that Musk was already claiming it was at 10 years ago.
I'd list as another big flop of the last 10 years the Tesla Solar Roof.
Also, I think by far the biggest Musk reality mismatch is the story he pushes regarding Optimus vs what anyone has actually seen of that product. That mismatch is absolutely wild. Anyone who believes the things he says about timelines for it needs to get a grip.
My casual observation is that people's perception of bean's gas causing properties is way out of line with what they do for most people and strongly influenced by people who have unusually high intolerance, who are the minority.
A part of the article that irked me is where he talks about why he's going to ignore people who say you get used to it because he thinks the microbiome change doesn't make sense misunderstands the gas/bacteria process in the gut.
There is a big mix of FODMAP eating bacteria, both inefficient generalists that can eat FODMAP but do it poorly and create a lot of gas, but also specialist bacteria like Bifidobacterium that also eat them but are much more efficient at it and produce much less gas. With more food the specialist bacteria becomes more prevalent. There are also gas consuming bacteria, not just gas producing, which also will shift in population. The idea is not that you just grow a larger population of the existing problematic bacteria.
You are correct that it doesn't rule out lipids + unknown additional factor(s). However: If there is a mystery factor, it must be close to universal. We know from autopsy studies of non-cardiac deaths that fatty streaks are present in virtually all children and that by 30 most people have advanced fibrous plaques (including soft plaque invisible to calcium score tests). The double variant people don't. So LDL is at minimum a necessary, limiting factor.
It also doesn't exclude that there is some more specific subset of the low density lipids that cause problems (this is what switching to ApoB testing is supposed to get at). Which is actually where we are at in the first place with LDL-C measurement being a refinement over previously looking at total cholesterol.
First of all, I agree with your points that you should consider the individual. My long term interest in this is also from being a very fit, low blood pressure, metabolically healthy person who always had at least somewhat elevated LDL (sometimes very elevated) that doctors would flag.
PCSK9 people are as close to a natural experiment on the effects of life time low LDL as you will get and they get near total protection, even when they have no other risk factors. People like smokers, hypertensives and diabetes have ~90% less than other high risk people, but people without any of those factors also have significantly less heart disease. People with two broken PCSK9 genes have close to zero LDL and have noticeably completely plaque free arteries as adults. I do think this does pretty fatal damage to the theory that you must have some other health issue for LDL to be bad.
It's very likely that "LDL-C" the lab measurement isn't as good as measuring ApoB, but for most people, they are concordant. And ApoB is a different way of looking at low density lipids, by particle count instead of weight. Dietary stuff like the fats in the article that lowers LDL measurements typically also lowers ApoB in most people.
So, in part, I agree that more precise biomarkers can help adjust individual risk. But most people are concordant. And the evidence that the underlying "low density lipids", no matter how you measure them, are causally part of the disease process is very strong.
And it's an area where there is legitimate nuance: we only measure LDL-C partly because it's easy and available, there are other ways of looking at blood lipid particles we could be measuring that might be more effective (like ApoB), it interacts with other things like insulin, inflammation, metabolic health, blood pressure.
But all that said, the case for high LDL being bad for heart attacks and strokes is very strong.
To me the strongest short list of evidence is simply:
- People who have familial hypercholesterolemia, from different genetic causes/pathways, all have massively increased heart disease at young ages.
- People who naturally have a disabled PCSK9 gene have extremely low LDL levels, where PCSK9 is directly involved in the liver's ability to clear LDL from blood, and these people also have incredibly low incidence of heart disease.
- Modifying cholesterol levels via PCSK9 or statins both have very strong evidence that they work on people who have heart disease, we have many RCT involving people who have already had one heart attack, and they have clear dose response curves: the amount of LDL reduction is directly proportional to risk reduction. We have less clear evidence on healthy people and from diet but those people are just a lot harder to study.
It's true that not everyone with high LDL develops plaque and we don't know why, but I feel a lot of "lipid hypothesis skeptics" tend to swim around in the gray areas and just don't interact with the more smoking gun bits of evidence that have to be explained away if you are going to say that LDL has no effect.
Backpropagation has been reinvented multiple times, because it is a basic application of the chain rule. The earliest recognizable usage of it is in control theory at NASA during the Apollo program.
It's a mistake to be dismissive of academic work which has been very important, but it's equally a mistake to think that academia is the sole source of foundational work.
I live in Mammoth where the town is significantly snowier than say Truckee or lake level Tahoe. The grocery store is open and operating normally no matter how snowy it is. Including the 22/23 winter when 695" fell in town. Lots of buildings did collapse that year though and snow removal was a constant struggle.
But A-frames or other very angled roofs are not typical here, roofs have to handle 300 lbs/sq foot, and there are requirements for where a roof is allowed to shed to. Typically they will angle in one direction to control where shedding happens. Keeping the snow on the roof also provides insulation, in a typical snow year we may do basically no removal and just have a blanket of snow on the roof the whole winter.
Adding lean meat, dairy, eggs to protein poor diets is good, I love all of those things. Trying to hit a high protein target and understanding this to mean having to eat all or mostly meat is simply over correcting in a different direction.
And meat isn't a perfect PDCAAS! Beef is 0.92 close to soy at 0.91. Milk, eggs, soy protein (isolate), whey are 1. Beans are 0.75.
There's even more nuance, beans are partly lower because they are low in methionine, the essential amino acid that adults needs far less of than any other amino acid and that you don't need more of. In the context of a whole diet, it's not 92% of every gram for beef and 75% of every gram for beans, it doesn't work that way.
Relatively small servings of animal foods add up to a lot of protein, like some other comment was freaking out about needing to eat 4 hamburgers or a 16 oz steak to hit the (very modest) goal of 90g in a day. But something like, 1 egg + 1 can tuna + 1 serving Greek yogurt == 42g, is already at half the goal, much of the rest of it will fill out just fine from a balanced set of other non-empty calorie sources.
Your diet contains many sources of protein lower quality than beans (as in the linked study with high level Dutch athletes getting 19% of their protein from bread), you do need to count those. They do add up and if you don't, you end up assuming you need way more protein than you do.