pjscott · HackerLangs

From Google's spokesperson, quoted in the article: "This is a short-term, timely agreement to ensure we have bridge capacity to meet surging customer demand for our agent platform, Gemini Enterprise, which has been even higher than we expected."

Google wants a lot of compute sooner rather than later, and they're willing to pay a premium for speedy delivery of that compute. SpaceX has the capacity already built and ready to go. Hence the high price.

pjscott

bulan laludiscuss

Musk made big investments in building AI data centers starting in 2024, continuing through the present. SpaceX got those assets from xAI, spun off from X, via an acquisition. More details:

https://en.wikipedia.org/wiki/Colossus_(supercomputer)

pjscott

bulan laludiscuss

Even then, the numbers don't even remotely work out. Consider this "Elon Bezos" person; the combined net worth of Elon Musk and Jeff Bezos is about $1.3 trillion or so, and the "Top 0.1%" category in the Fed's data has a net worth of about $25 trillion. This is not a small difference.

pjscott

2 bulan yang laludiscuss

You're right, of course, but "those in government" aren't a single entity. There's always an incentive for one corrupt part of government to take more than their fair share of the loot, and then for the next part to take more, and so on... until their combined cut is over the revenue-maximizing percentage, and they make less than they could have if they had coordinated better.

(I want to call this "the tragedy of the commons," but that phrase doesn't sound quite dark enough.)

pjscott

3 bulan yang laludiscuss

It doesn't track cost of living? The way it's calculated is all about cost of living!

In the US, the official inflation numbers are based on a "basket of goods" meant to be representative of a typical person's spending. Housing currently makes up about a third of the basket, while luxury items are a fairly small percentage. Here's a pretty well-written summary, albeit with numbers from 2022:

https://www.pewresearch.org/short-reads/2022/01/24/as-inflat...

Changes in housing prices have a large effect on the BLS's inflation figures. Downward changes in the price of luxury goods have a small (and bounded) effect. Even if all luxury goods became free, the reduction in inflation wouldn't be all that much.

pjscott

3 bulan yang laludiscuss

The wording was a bit unclear. The previous paragraph mentions wanting something cheaper than "those pesky XORs and multiplications". The multiplication is the expensive part; the (very cheap) XORs are just mildly annoying because you have to think about what they're doing.

pjscott

4 bulan yang laludiscuss

> I’d like to know the memory profile of this. The bottleneck is obviously sort which buffers everything in memory.

That's not obvious to me. I checked the manuals for sort(1) in GNU and FreeBSD, and neither of them buffer everything in memory by default. Instead they read chunks to an in-memory buffer, sort each chunk, and (if there are multiple chunks) use the filesystem as temporary storage for an external mergesort.

This sorting program was originally developed with memory-starved computers in mind, and the legacy shows.

pjscott

4 bulan yang laludiscuss

If money ever starts looking particularly illusory, try thinking in terms of the underlying resources that markets allocate.

That's 'resources' viewed as expansively as possible, everything from the specialized labor-hours of people who know how to do quality control on bulk-manufactured photovoltaics to the ore used to make ball bearings in the factory all the way to the guy in charge of managing a grain elevator that was involved in making the bread for the sandwich one of the janitors had for lunch. The web of collaboration between all these far-flung people who mostly don't know each other, too vast and intricate to fit in any living mind, is how we currently get most of our material stuff.

... And in a conventional market system, the core of how those people coordinate their efforts is money. The price that each person is willing to buy something for or sell it for sends a signal about how much they care about it relative to other things. And markets are one popular way of aggregating that information, helping guide society's cooperative efforts in the direction of what people care about.

There are various allocation systems that don't involve money, both theoretical and historical. Community-based mutual reciprocity with a reputation mechanism to discourage freeloading, for example, can be found all over the place in pre-modern history because it worked – as long as your community was small enough that you can realistically all know each other. Or, back in the 20th century, there were a number of efforts to scale up operations research toward the level of nations, since suddenly we had computers fast enough to handle e.g. non-trivial linear programming. (The successes and failures were both instructive.)

--

Coordination problems are hugely underrated in political discourse. So when I hear people say things like "The economic system is the ideology holding us back", I always have to wonder: how carefully has this person thought about a what a viable alternative would look like?

"I dislike the current system" is only the first and most trivial part of a real reform agenda; the next part has to be "... and here is how to meaningfully change it in a way that doesn't result in disaster, with a detailed discussion of mechanism design and a look at relevant historical prior attempts. [Insert essay or hyperlink here.]"

pjscott

5 bulan yang laludiscuss

So it looks to be just the latency improvement that's noteworthy, then. Thank you!

pjscott

5 bulan yang laludiscuss

If you're familiar with the technical specs, I'd be interested in hearing what size of objects the star trackers can sense and at what range. In theory the fancier star trackers can see objects around 10 cm diameter hundreds of kilometers away, without needing to worry about a pesky atmosphere [1], but I don't know how sensitive the sensors on Starlink's current generation satellites are, and this web site isn't saying.

They're mostly touting the improvement in latency over existing tracking, from delays measured in hours to ones measured in minutes. Which is very nice, of course, but the lack of other technical detail is mildly frustrating.

[1] https://www.mit.edu/~hamsa/pubs/ShtofenmakherBalakrishnan-IA...

pjscott

5 bulan yang laludiscuss

Honestly, these two paragraphs are one of the most compelling things they could possibly say in a press release:

> Stargaze already has a proven track record in its utility for space safety. In late 2025, a Starlink satellite encountered a conjunction with a third-party satellite that was performing maneuvers, but whose operator was not sharing ephemeris. Until five hours before the conjunction, the close approach was anticipated to be ~9,000 meters—considered a safe miss-distance with zero probability of collision. With just five hours to go, the third-party satellite performed a maneuver which changed its trajectory and collapsed the anticipated miss distance to just ~60 meters. Stargaze quickly detected this maneuver and published an updated trajectory to the screening platform, generating new CDMs which were immediately distributed to relevant satellites. Ultimately, the Starlink satellite was able to react within an hour of the maneuver being detected, planning an avoidance maneuver to reduce collision risk back down to zero.

> With so little time to react, this would not have been possible by relying on legacy radar systems or high-latency conjunction screening processes. If observations of the third-party satellite were less frequent, conjunction screening took longer, or the reaction required human approval, such an event might not have been successfully mitigated.

Looks like a non-trivial upgrade to previous systems, and they're making Stargaze's data available to other satellite operators free of charge. Nice!

pjscott

5 bulan yang laludiscuss

Slowing the adoption of much-safer-than-humans robotaxis, for whatever reason, has a price measured in lives. If you think that the principle you've just stated is worth all those additional dead people, okay; but you should at least be aware of the price.

Failure to acknowledge the existence of tradeoffs tends to lead to people making really lousy trades, in the same way that running around with your eyes closed tends to result in running into walls and tripping over unseen furniture.

pjscott

6 bulan yang laludiscuss

It's a tale as old as schlock journalism: an article seems interesting... until it talks about something you actually know about personally, at which point it suddenly starts saying obvious nonsense.

pjscott

7 bulan yang laludiscuss

I'm not too familiar with the JVM so perhaps I'm missing something here: how would that help? The file is tiny, just a few bytes, so I'd expect the main slowdown to come from system call overhead. With non-mmap file I/O you've got the open/read/close trio, and only one read(2) should be needed, so that's three expensive trips into kernel space. With mmap, you've got open/stat/mmap/munmap/close.

Memory-mapped I/O can be great in some circumstances, but a one-time read of a small file is one of the canonical examples for when it isn't worth the hassle and setup/teardown overhead.

pjscott

7 bulan yang laludiscuss

As with all the ahead-of-time compiled languages that I checked, the answer is that it generates non-SIMD code for the hot loop. The assembly code I see in godbolt.org isn't bad at all; the compiler just didn't do anything super clever.

pjscott

7 bulan yang laludiscuss

The common element is that they're written with the most obvious version of the code, while the ones in the faster bucket are either explicitly vectorized or written in non-obvious ways to help the compiler auto-vectorize. For example, consider the Objective C version of the loop in leibniz.m:

  for (long i = 2; i <= rounds + 2; i++) {
      x *= -1.0;
      pi += x / (2.0 * i - 1.0);
  }

With my older version of Clang, the resulting assembly at -O3 isn't vectorized. Now look at the C version in leibniz.c:

  rounds += 2u; // do this outside the loop
  for (unsigned i=2u; i < rounds; ++i) // use ++i instead of i++
  {
      double x = -1.0 + 2.0 * (i & 0x1); // allows vectorization
      pi += (x / (2u * i - 1u)); // double / unsigned = double
  }

This produces vectorized code when I compile it. When I replace the Objective C loop with that code, the compiler also produces vectorized code.

You see something similar in the other kings-of-speed languages. Zig? It's the C code ported directly to a different syntax. D? Exact same. Fortran 90? Slightly different, but still obviously written with compiler vectorization in mind.

(For what it's worth, the trunk version of Clang is able to auto-vectorize either version of the loop without help.)

pjscott

7 bulan yang laludiscuss

The delta there is because the Rust 1.92 version uses the straightforward iterative code and the 1.94-nightly version explicitly uses std::simd vectorization. Compare the source code:

https://github.com/niklas-heer/speed-comparison/blob/master/...

https://github.com/niklas-heer/speed-comparison/blob/master/...

pjscott

7 bulan yang laludiscuss

They have very limited power these days. They advise the House of Commons, as more or less a hereditary think tank. They can delay the passage of bills, though this has been limited to a maximum delay of one year since 1949 (less for some types of bills) and there are some checks on this ability. They have a few other things they can do that are (IMO) too boring to warrant much thought unless you're a member of parliament.

The idea of a House of Lords does strike me as a bit odd, but it's not really the big deal it used to be.

pjscott

9 bulan yang laludiscuss

They do, technically, allow JIT. You need a very hard-to-obtain entitlement that lets you turn writable pages into executable read-only pages, and good luck getting that entitlement if (for some reason) your name isn’t “mobilesafari”, but the capability exists.

pjscott

10 bulan yang laludiscuss

It may comfort you to hear that the Starlink satellites are tiny in comparison to the vastness of their orbit – the visualization makes them appear larger than they are, so you can see them clearly – and that they’re low enough that they’ll naturally de-orbit and burn up in the atmosphere after about 15 years even without using their maneuvering thrusters.

They’re providing worldwide rural broadband, and according to the FAA they’re doing so in a way that’s careful and responsible about space debris and collision avoidance. Is disgust truly warranted in this case?

comments