rwallace

Reasonable people can differ on whether quiet NaN propagation is good or bad, so I don't fault you for being in the latter camp, but I am surprised by your suggestion that it was particularly friendly to 1980s hardware. I was under the impression that an 8087 would have been equally happy to respond to NaN by quiet propagation, setting a global sticky flag or crashing your program, and it's modern hardware, with multiple heavily pipelined vector units running in parallel, that prefers to quietly propagate. Am I missing something?

How? I was under the impression that RISC-V was deliberately minimalist in the hope of obtaining smaller cores, and I can see how that would allow higher clock frequency, but I would've expected the trade-off to be longer programs. How did it get shorter programs than ARM? What were the cases where a single RISC-V instruction corresponded to multiple ARM instructions?

Sympathies! Now I'm curious, how did it happen? What equipment was it, that your employer preferred Apple over IBM, let alone had such a vehement preference as to actually fire you over it?

> they still pay royalties per chip, even with a perpetual license on the ISA

Why would they have to pay royalties per chip? I would've thought the whole point of a perpetual license on the ISA was that they didn't have to do that. What am I missing?

> General purpose GPU hardware is way more wasteful for matrix math, like maybe >10x waste on power and equally worse performance, than tensor cores.

I didn't realize there was that distinction; I thought GPU's were just optimized for vector arithmetic across the board. What is the difference between general purpose GPU hardware and tensor cores? What does general purpose GPU hardware do that tensor cores do not?

I'm curious about the specifics of the problems this would solve, having been doing some thinking about what an ideal file system would look like. In particular, is this something you would need primarily for files you create, or existing files created by other people? And, to be fully useful, would it need to allow the attaching of such metadata to arbitrary files/folders regardless of type?

Now that is interesting! Do you remember where he said this?

Can you summarize, or point to a discussion of, the ways in which PTX differs from a traditional RISC architecture?

> But the fact is we have now cheap existing clean technologies, with a fast decreasing cost (decreasing cost of energy, and decreasing cost of managing intermittence)

The cost of solar and wind energy when the sun shines or the wind blows has certainly decreased a lot in recent years, you're right about that.

But decreasing cost of managing intermittence? I haven't heard of any big recent advances in that (aside from the successful use in Australia of batteries to damp fluctuations on a timescale of minutes, which is great but doesn't address timescales of hours to months). What am I missing?

> From my perspective Julia is sacrifising some amount of "duct tape UX" to gain speed, and that's the wrong direction.

What particular language features of Julia make that trade-off? (Not a rhetorical question, I'm not disagreeing with you, just curious; I'm familiar with Python, not really familiar with Julia.)

What would you have considered the normal/preferred way of handling this? After all, hardware price drops were normal, that's how the market was expanded; wouldn't it be the case that no matter when the price came down, someone would've bought it the previous day?

So why exactly does anyone want that, if not as a subgoal of spotting fraud? (I mean, if the first link in the chain is "because that's what's necessary and sufficient to comply with the law" then why was the law written that way?)

I am surprised! I always thought that was precisely their job. So if it's not, what is their job instead?

Intuitively I would expect that to work even with twisted-pair; each sensor needs to transmit for maybe a millisecond every 300 seconds, which would allow for 150,000 sensors while still using only half the available bandwidth. What am I missing?

> One case where timer-based scheduling is needed is when implementing a web browser. The JavaScript standard has a function called setTimeout which asks the browser to call a JavaScript function some number of milliseconds later. Chromium uses timers (mostly WaitForSingleObject with timeouts rather than Sleep) to implement this and other functionality. This often requires raising the timer interrupt frequency.

Why does it require that? Timeouts should normally be on the order of minutes. Why does Chrome need timer interrupts to happen many times per second?

Suppose we grant that hell exists with some nonzero (albeit tiny) measure.

By the same logic, heaven (some extraordinarily joyous and worthwhile state of existence) also exists with some nonzero measure.

You probably feel that the latter cannot outweigh the former (even though it logically should). Essentially that bad outweighs good. But why do you feel that? Literally, what is the cause of your brain containing that feeling?

It's because human intuition evolved in a world of asymmetric harms and benefits. For an animal living on Earth - the state in which we evolved - a single bad event can cause death, which is a greater harm than can (for most animals) be outweighed by any single good event.

Okay. So evolution gave us intuition tuned for the tradeoffs of life as a mortal animal, that often generates valid conclusions while we are living as such. But once we start talking about wild metaphysics like dust theory, all that goes out the window. The basic premise is that we are immortal! So if we are going to accept such wild metaphysical speculations at all, the first thing we need to do when thinking about them is remember that they completely invalidate our intuitive feelings about things like bad outweighing good.

I'm curious, what do you find are the advantages of structural typing?

> Or a learning tool to make computer science & networking concepts concrete and tangible?

That seems like a perfectly good use case to me?

My safety worry wouldn't be the lack of flight attendants per se, it would be distracted pilots. The pilots have enough to keep track of already. The last thing we need is having planes start crashing because a pilot was distracted dealing with a passenger.

> back when Microsoft was developing their 32 bit Windows, they were all set to not save the x87 state during a context switch, effectively killing the x87

Do you mean '64-bit Windows'? (When 32-bit Windows was being developed, killing x87 would have meant killing all hardware floating point on x86.)

> We said we liked their CPU performance, but didn't like the Archimedes graphics hardware which was too low res and low depth, and could not be upgraded.

I'm surprised about that! Looking at https://en.wikipedia.org/wiki/Acorn_Archimedes it says the machine could do 640 × 512 with 2, 4, 16 or 256 possible colours, or 800 × 600 with 2, 4 or 16 possible colours. That's pretty spectacular for the late eighties; far superior to the Amiga (previously considered the king of graphics) and even superior to VGA, let alone the Mac.

What am I missing?

> resurgence of fixed point arithmetic in hpc

Where is fixed point arithmetic resurging in HPC? Serious question; I haven't heard about this, but if it's happening in places I haven't heard about, I would be interested in looking at them.

I think your summary is accurate. For reference, I'm willing to cut Wolfram more slack than some people are, I think his quest to discover the ultimate laws of physics from the other end is potentially interesting, but I don't think this particular article really has even any interesting conjectures.

> my friend who started a 20 million dollar software company in Chicago and then moved to Prague to try to repeat it and he knows exactly why software engineering doesn't scale in Eastern European countries

I'm curious, why doesn't it? What went wrong with your friend's enterprise in Prague?

Great! Then you're the ideal person to ask something I've been wondering: We see the occasional headline about machine learning algorithms outperforming human radiologists. What's the catch? I'm prepared to believe they aren't really outperforming humans once you take all the unstated factors into account, just wondering what the unstated factors are.

> nobody explicitly specifies "result of an addition of natural numbers should match ..."

Yes they do. In most programming languages, by default integer arithmetic is modulo 2^64 (at best). If you want arbitrary precision arithmetic, you have to explicitly specify that.

Interesting article! One thing I'm curious about: I can see what the bit shifter was for, but the article also mentions a separate byte shifter. What was that for?

Look, I completely agree with you that the housing crisis is terrible and needs to be fixed, but sniping at proposals for particle accelerators is not the way to express that view! Let's make political comments in contexts where they are appropriate.

Does Wolfram's model correctly generate the many worlds of quantum mechanics? I get the impression that it doesn't, and I've seen it claimed that this has been mathematically demonstrated, but I can't say for sure.

That includes variability in some things that are constant in our region of the universe. It's fine for everyday purposes for us to regard things like the ratio of masses of the particles or strengths of the forces as fundamental and given, but a truly fundamental theory needs to generate the entire landscape, possibly including variation in the number of extended dimensions of space.

At the same time, there is a sense in which it's easy to do that: just use a universal Turing machine. A correct fundamental theory should generate something akin to the landscape of string theory, but at the same time not generate e.g. Conway's Game of Life. Is Wolfram's theory being evaluated on achieving that delicate balance between generating too little and too much?

Excellent article! Just one thing surprised me:

> While I was writing a book about Wolfenstein 3D[1], I wanted to vividly demonstrate how much of a handicap it was to work without floating point

I would've expected fixed point to work fine for games, because that's a domain where you know the data, and in particular the dynamic range, in advance, so the 'automatically adjust to whatever dynamic range happens to be in the data' feature of floating point isn't needed. What am I missing? (If the answer is 'it would take too long to explain here, but he does actually explain it in the book', I'm prepared to accept that.)