tlb

But in fact no modern processor/OS executes this either. Pages are marked as executable or not, and static data is loaded as non-executable pages.

I like how the pictures got more and more sloporific through the essay.

It doesn't mention an important group being harmed: the creators who make high-quality, sincere podcasts about knitting. Their genuine content gets buried under a mountain of slop. In theory, recommendation algorithms ought to surface the best stuff, but that doesn't seem to align with incentives. Sad.

Yes, the entire field of software engineering ran aground on not being able to test how well people can write software.

But I'm more optimistic about testing programming models. You can run repeated tests, and compare median performance. You can run long tests, like hundreds of hours, while getting more than a few humans to complete half-day tests is a huge project. And you can do ablation testing, where you remove some feature of the environment or tools and see how much it helps/hurts.

Field termination is necessary when the connectors are too large to pull through a conduit. But if they were USB-C sized, you could just pull fully assembled cables.

You can rotate your voice with substantial effort. Just speak differently: higher or lower pitch, a different accent. Your friends may look at you funny for the first few years.

If you have something interesting you made yourself, and not some vibe-coded AI slop project, email the mods at hn@ycombinator.com and they can enable your account.

Of course, when everyone sets min-release-age=7, supply chain attacks won't get noticed until 7 days later. So you should set min-release-age=14 and be safe forever.

I understand the appeal from AWS's perspective. Customer A pays for a 32 vCPU VM, which they run on 32-core hardware. Then they can also squeeze in customer B's 1 vCPU instance running a blog, and no one notices. Free money!

But I don't want to be either of those customers. It means the whole system has an extra layer of abstraction, so they can juggle VMs around. It's why you need slow EBS instead of just getting a flash drive in the same case as the CPU, with 0.01x the latency.

I think clouds pay a huge abstraction penalty to allow tiny VMs. I guess it helps with onboarding and $10 personal VPNs. But I have never needed a fraction of a computer. I want to rent some number of full computers of various sizes, consisting of CPU, memory, and flash disk. Hetzner is closer than AWS, and I think/hope that’s what Crawshaw is aiming for.

MuJoCo is great. I have it running in the browser for robotics simulation. See for example https://visibot.com/sheet/examples/humanoid_walking.v

Cool, but why is the most rounded-off part in the center? My wrists cover the edge at 5-25% and 75-95% when typing. When mousing, my right hand fleshy pad covers the edge at 65-80%.

You’re free to invest that way if you want. You might one day wake up and wonder why your Blockbuster Video shares did so badly. But Netflix seemed way overpriced.

Investing in future prospects encourages companies to plan for the future, rather than extract what they can from the present. The stock price is a big motivation for execs, so they can only invest in R&D if the market understands why it makes sense to spend money now in expectation of future profits.

Yes, and indeed more startups are addressing more markets every year. The total number of attempts and successes are both up, keeping the success rate fairly constant.

The fact that the survival rate of startups hasn't improved doesn't show that our knowledge hasn't improved. Startups are competitive, with only 1 or 2 VC-scale winners per market. So, the claim is like "race car technology hasn't gotten better, because there's still only one winner per race."

I think the total light output of each bulb in the pair is the same at all points in time, but the orange-blue gradient is reversed. So when one is orange at one end, the bulb beside it is blue at that end.

IIRC, the end that's negative looks orange, because the electrons emitted from the filament haven't gotten up to speed yet and can't ionize the mercury atoms at that end to the highest states.

If you didn't do this, you'd see 60 Hz strobing when you looked at one end.

You and I look with dismay at the high prices, but remember that a million hospital administrators are high-fiving themselves. So ideas like "just cut waste" are opposed by a large group with a lot more skin in the game.

Heat goes up with the square of current, so putting 10x the current to get 10x the force means 100x the heat.

Still, I think this idea is under-explored. There are probably applications for robots that move really fast, but only for a second before having to cool down.

You can get a more accurate list at https://www.ycombinator.com/companies?batch=Summer%202026&ba...

One scenario: there's a short circuit somewhere, say rats chewing through insulation. This can cause a very high current through the short. A non-inverter 4500 watt 120 volt generator might have 0.2 ohms coil resistance, so the short circuit current can hit 170 volts / 0.2 ohm = 850 amps. When the shorted branch's circuit breaker trips, the inductance in the generating windings wants to keep that 850 amps flowing for at least a few microseconds, and it gets distributed across everything else that's still connected. Depending on what else is connected (hopefully including some surge protectors) the peak voltage can get into many kilovolts.

The circuit is something like this:

  voltage source -- parasitic inductor --+- circuit breaker -- short
                                         |
                                         +- circuit breaker -- your PC

HDC/VSA.

I'm interested in this, but only passingly familiar with it from several years ago. Can you link to what you believe the current state of the art is?

Isn't that essentially what happened? In the 1930s, the channels were allocated at 6 MHz spacing, but only needed about 3 MHz of bandwidth for the luma channel. There might have been some foresight involved, but the gaps between channels also allowed for cheaper, less precise tuners. Then in the 1950s they added a 1 MHz chroma signal on a 3.58 MHz upper sideband, expanding the channel bandwidth to just under 6 MHz.

Oil prices are affected by many other things too. It's valuable to isolate individual factors.

It's easier to ban things before they become entrenched than after. Banning stock trading is hard today because most senators have been doing it for years before getting elected.

I'd be in favor of senators doing prediction markets, just limited to small dollar trades so they can get better at predicting.

Great images, OK writeup. There are some bits of bullshit, like "The proximity of microphones to processing hardware minimizes latency". No, the speed of electrical signal propagation (around 2/3 the speed of light) is not significant for microphone placement.

Many of the great hacks have involved breaking through 2 layers of supposed security. You break into the 3D printer, which lets you send packets on the local network. Then you use that to break into the exercise bike, which has a camera because it's based on a generic tablet.

Either vendor might see the flaw as low-severity. So what if someone can send packets? So what if someone already on the local network can hack the camera? But combine them and you're pwned.

The rate at which molecules of plastic sublimate off the surface of the enclosure is probably a much larger amount of mass. The rate increases with e^kT, where k is such that it doubles about every 10 degrees C. So if you get a drive and fill it with data (which warms it up significantly) the lost casing material will dominate the mass balance.

International orgs usually need a company incorporated in every country they're working in. You need it to pay employees, for instance.

Compilers only do one file at a time, but an LSP needs to parse all the files.

Also, an LSP needs to update incrementally when you edit the file, not just when you compile.

Do you clear the file and use the same name for the next commit? Or create a new directory with a plan.md for each set of changes?