> Once a region of tape has been read, the controller stores the result. Subsequent operations reference the cache rather than re-interrogating the physical medium. Re-reading a known bit is unnecessary; the controller already holds its state

However, earlier, the paper claims:

> The transformer architectures underpin- ning modern large language models are bandwidth-limited, not compute-limited [1–3]. The energy consumed moving data between DRAM, NAND flash, and processor cache already exceeds the energy consumed by arithmetic in datacenter AI accelerators [2]. This is not an optimization problem. It is a materials problem [emphasis mine].

as part of a longer rant about the AI "memory wall" in the very first section. If we open with a long spiel about how memory is expensive in material cost and energy cost and this material is a solution for that then what are we caching the read in? On that note, what kind of computer engineer thinks about cache on the order of individual bits on a medium?

And, as you point out, 25 PB/s is a lot. Around 1000x that of a typical on-die SRAM cache, I think.

A while later, the author speaks of using atomic force microscopy to read the data back. The size of AFM scans are, in practice, as I understand, along the order of square micrometers. I think this whole paper is an AI-driven, as you put it, 'fever dream', enabling an author to put forth 60 pages of sciencey claims and sciencey math without -- as far as I can tell -- any concrete and novel scientific result of any kind. AI-driven reality warps are not new; the difference is nowdays AIs are good enough at sounding smart to get past the barriers of a typical smart person who might want to be fooled or make a show of being open-minded. Later on, the author proposes using "shaped femtosecond IR pulses" -- without further elaboration -- to address single atoms! IR wavelengths are on the order of a micrometer at minimum!

This is another one of the vibe-coded slop projects that are routinely frontpaging HN now. As someone else pointed out, the single author has "written" >100kLOC in diffs per week. It's not possible that any human knows what's in the codebase in any reasonable detail.

I suspect that some of the author's comments in this thread are vibe-written, also. They are LLM-flavored and contrast strongly vs. their regular commenting.

But, on the other hand, I suppose intellectually serious dispute requires both sides to be intellectually serious. One good step in that direction would be to arrange one's citations such that they are supporting the claims you are citing them for.

I will also remark, as others have, that it's odd to make a big deal about this particular fraud when there's a lot more fraud happening a lot more obviously in a lot more of the nation. This is not to question Minnesota officials who are, rightly and appropriately, investigating suspected fraud in their zone of investigation; but it is worth questioning voices who have, apropos of nothing I can discern, made decisions about what's important to talk about and what isn't, and further made decisions to misrepresent allegations in alignment with people who very aggressively lie for evil reasons. As others have pointed out, the essay's core is seemingly cromulent, and it doesn't need you to do that.

Cash is such a poor investment that the word "investment" typically means trying to find something more productive than holding cash. Neither do alternatives to cash have a reliable history of benefitting the poor. In the US there's been lots of attempts at local currencies; they tend to fail naturally without government interference. Recently, cryptographic alternatives to cash have mostly served to benefit crypto barons and scammers.

I noticed something similar when I played contract bridge at a competitive level. A top bridge player might play very roughly on the order of 10,000 hands a year, and vividly recall something that happens on the order of once a year as "oh yeah that's common". Of course I wasn't remotely close to them. But there is something about competitive games that seem to amplify the memory for certain kinds of unusual situations.

(Some people are commenting about under promoting to avoid stalemate traps down the line. I've always been a weak chess player, but... trying to set a stalemate trap after being down a queen, in a non-contrived position, is, like, adult chess players shouldn't do that. In my limited experience.)

The first paper seems to claim a very standard cohort study is subject to "immortal time bias", an effect whereby measuring outcomes can seem to change them. The typical example of sampling time bias is that slow-growing cancers are more survivable than fast-growing ones, but also more likely to be measured by a screening, giving a correlation between screening and survivablility. So you get a time effect where more fast-acting cancers do not end up in the measurement, biasing the data.

But in measurements such that one outcome or the other does not bias the odds of that outcome being sampled, there can be no measurement time effect, which is why it's not corrected for in studies like this. The authors do not explain why measurement time effects would have anything to do with detecting or not detecting death rates in the abstract, or anywhere else in the paper, because they are quacks, who apply arbitrary math to get the outcome they want.

As another commenter pointed out, randomized controlled trials -- which cannot possibly have this made-up time effect -- often clearly show a strongly positive effect for vaccination.

I did not read the second paper.

In a vacuum, if we all have 3-10% more money and things are 3-10% more expensive, we're not poorer unless there are secondary effects that make us poorer. It's the secondary effects that inflation has on investment, markets, and transfer of wealth that are negative. The effects of debt collapse during economic crisis is a considerably more expensive effect, and not one that's good for the poor, or the rich, or the middle class or anyone.

If you "genuinely cannot understand this", consider that the traditional perspective is that the point of an economy is to produce useful goods and services, not to produce an aesthetically pleasing inflation number. Recall that, in early 2020, markets were facing the greatest panic since the 1929 crash. The debt collapse and deflation that followed then precipitated an enormous amount of misery. I'm personally pretty happy we didn't get another Great Depression; Covid was bad enough without that happening.

I don't think LessWrong is a cult (though certainly some of their offshoots are) but it's worth pointing out this is very characteristic of cult recruiting.

For cultists, recruiting cult fodder is of overriding psychological importance--they are sincere, yes, but the consequences are not what you and I would expect from sincere people. Devotion is not always advantageous.

If you actually read the article, they include a direct link to the sources they cite and explain specifically what those sources say.

It's not really any less natural than thinking stateful programming, except now the state is a reified thing, which I think is strictly advantageous once you get used to it.

It's the concept of categorical commutativity that's what's useful. A collection of types and functions is "commutative" if every way to get from type A to type B yields the same result. It happens a lot in Haskell where most or all of the operations you're interested in commute with each other, which is how Haskell gets its reputation of "if it compiles it works". In particular, if you `fmap` or `bind` two commutative functions, the result becomes commutative, too.

This is what monads being categorically commutative ("a monoid in the category of endofunctors") buys you. You want to turn monad X into monad Y? Sure, just join, flatten, return, bind in whatever way makes the type checker happy. Anything that only uses what's in the Monad typeclass must necessarily be a monad morphism, so if you're generic over your monads, you get that for free. And of course `fmap` and `bind` are required to be parameterized monad morphisms, so there's a lot you can get for free.

The other reason is to tax the rich, but you can do that by simply taxing the rich directly. If we fear powerful companies, we can put some sort of scaling size tax on the largest ones.