Exactly. As technologists, we need the output of computers to be precise and accurate. LLMs might be precise, but they're very often inaccurate, and that's not acceptable to us.

The average person doesn't live in a world where accuracy matters to them. A colleague said she used AI all the time, and I asked her how. She said she often tells it the contents in her fridge and asks it for a recipe that would use those ingredients. She said, "yeah, and it's really accurate too." I don't know how you measure accuracy on a test like that, but it doesn't really matter. If you're just mixing some ingredients together in a frying pan, you probably can't go too far wrong. As long as you don't ask it for a baking recipe, it'll work out.

And I think that's what's going on. The people who love AI don't know enough to realize when it's wrong, or are just asking it open ended questions, like you would ask a fortune teller, and it spits out something generic enough that you can't disprove it anyway.

Bullshit. I live in the western US and have regularly driven through some of the least-populated areas of the country, but I've never seen an area you can go 500 miles without encountering any infrastructure. You might be able to accomplish it if you take careful note of where the truck stops are and go out of your way to avoid them, but on any realistic route you'll encounter truck stops -- if not towns -- at least every 150 miles.

As for charging infrastructure, if you stay on the interstates I don't think there's anywhere in the country you can go more than 100 miles without finding a Tesla Supercharger. Those aren't designed for truck charging, but this demonstrates that building out the infrastructure isn't that hard.

They certainly seem to have completely missed Article I. You know, the part that says that the legislature makes the laws? Even if you think restricting AI regulation to the federal government is a good idea, the right way to do it isn't with an executive order to set up a DOJ task force aimed at litigating state AI regulations out of existence based on complex legal theories about interstate commerce. The right way is for Congress to pass a law barring states from regulating AI. This is simpler, cheaper and should invoke public debate about the issue, which is how things are supposed to be done in constitutional republics.

I don't even think Trump is taking this route because he and his advisors don't believe they have the votes for it. I think they're doing it this way because they don't even consider governing through legislation rather than through executive power. Granted that Congress is fairly dysfunctional, but they actually can and do make laws... and the way to fix the dysfunction is to work the system.

I agree. Where I can find data to backup his claims, I generally accept what he says. Otherwise, I'm like, "hmm, ok, maybe."

The above was already quite long, but allow me to add a bit :-)

I spent a few minutes looking for the state of the art in C++/Rust interop for contexts that don't have a nice intermediary like binder. It turns out that the situation isn't as bad as I thought. The CXX project enables automatic generation of bi-directional definitions between Rust and C++ and is being used at scale by the Chromium project and that seems to be going pretty well.

There's also a Google-funded Rust Foundation project to define a better solution, though I don't see what, if anything, has happened since it was announced last year. Hopefully that's because there's a small group working too hard to waste a lot of time talking about it.

The reason I went to look is that my new team (I left Google a couple of months ago) might need such a thing. I've been asked to define an API that would benefit from being implemented in Rust and usable from C++ and Rust.

Telling people that vaccines are bad, ensuring that any parent who wishes to refuse to vaccinate their children is fully supported in that decision, and working to make vaccines harder to get, more expensive and more painful (RFK Jr. wants to separate the MMR vaccine into three shots, each of which will still require three injections, so kids will have to get 9 shots to be fully vaccinated instead of three).

This is similar to their plan to fight inflation by imposing tariffs and forcing the Fed to lower interest rates in spite of rising inflation (note that this last part hasn't really happened yet -- the interest rate cuts have been measured, cautious and justified by economic conditions -- but Trump is working on it). Though to be fully fair, by making the tariffs arbitrary and capricious so that business leaders are completely unable to plan, Trump is also causing a contraction in US economic activity that might eventually generate significant unemployment, which actually does reduce inflation. I see no corresponding "silver lining" in the mumps plan, though.

And, ultimately, slavery was a far less efficient and effective economic system. One might think that keeping a big chunk of the populace poor is efficient, since you're not "wasting" a lot of production on providing them with unnecessary goods and services, but it's really not, at least not since the industrial revolution. I think the core reason that it's so inefficient is the same reason that Marxist communism is inefficient: From an economic perspective, both systems value the masses only for their physical labor, and fail to cultivate and take advantage of their brains, which also actually tends to reduce their labor output. Harnessing the distributed ingenuity of your workforce requires giving your workforce some reason to exercise ingenuity and some way to benefit from doing so.

It's going to be interesting (or maybe terrifying, or maybe just sad) to see what happens when we fully automate ingenuity, too, which will mean that the system no longer depends on or benefits from distributed ingenuity because the machines are smarter and think faster, just as the machines are already stronger and indefatigable.

I don't think vibe coding is going to last long as a thing, because it's just a sort of intermediate step to telling the AI to do what you want and having it do that. Right now, people are telling the LLM to write code to accomplish a thing and then running the code to see how it works, then telling the LLM to refine it, but that's a lot of unnecessary extra steps. I'm sure that in the not-too-distant future people will just tell the LLM what they want to do, which may require creating a custom user interface to make user interaction convenient, and may require creating databases or performing network queries or whatever, and the LLM will understand what they want, and do it.

In that future, it's possible that the LLM may generate code to implement the requested functionality, but if it does so that will be a compute-saving shortcut, essentially a way to cache the LLM's work and be able to repeat it with less effort. There won't be any need to show any of the code to the user, or even tell the user that the LLM chose to generate some code.

As an aside, the whole notion of leaning "prompt engineering" is another intermediate step that will die. The whole point of natural language-capable AI is that it will be able to understand what humans want when we express ourselves as we would to other humans. As the LLMs get more capable, it will become less necessary to treat them as something different from an entity that is fully capable of understanding and acting on human communication.

One of the things Android did very right with the Rust transition was to set up a small team of people who were entirely focused on Rust support. It wasn't a large team, only 2-6 people (it varied over time) out of approximately 1500 engineers. Having that core team who either were or became deep Rust language and toolchain experts was critical to smoothing the path for everyone else. It provided a group that had the knowledge and bandwidth to solve the problems that inevitably came up, as well as to offer advice and code review support to the early adopters.

That group no longer provides code reviews and design advice because Rust knowledge is now widespread enough that teams have their own, homegrown, Rust experts (not people designated as Rust experts, just engineers who became enthusiastic and dived deep), but the group still exists to resolve complex technical problems with language integration and to work on improving tooling and performance.

I think any shop adopting Rust (or any new language or complex tool) needs to have some people who become deeply expert in it and are allowed the time and freedom to support others who are picking it up.

So does Android. Google has been a primarily-C++ shop since its inception and although I'm not sure if Android had a lot of C++ in it when Google bought Android in 2005, it definitely became a C++-based system as soon as that happened.

The highly-segmented architecture of Android really helped facilitate the transition. Most of Android is structured as a web of collaborating services that communicate through a common language-independent [*] IPC mechanism (binder). Implementing Rust binder IDL generation and support libraries was a moderately big job, but once that was done it was easy to begin writing new system components (or replacing existing system components) in pure Rust, generally without any unsafe blocks at all.

If your code runs as a monolithic process, or has a lot of different IPC mechanisms, or uses a lot of existing libraries, it will be a lot harder, and the benefits will come slower. You'll have to wrap a lot of C interfaces in Rust -- and they will have to be C, not C++, since there isn't a good way for Rust to interoperate directly with C++. People are working on that, but it's a very hard problem and at present the best option is to layer a C interface on top of your C++ code, then wrap a Rust interface around the C interface. Yuck. Or, in the alternative, insert some other language-agnostic boundary between them.

So in a lot of ways Android got lucky because of its modular architecture and single, language-agnostic IPC mechanism. OTOH, that wasn't really "luck", it was a lot of work, done for good reasons, one of which was cross-language compatibility, notably between Java and C++.

[*] Language independent-ish, maybe I should say. The binder IDL is definitively Java-based, but this maps fairly nicely onto OO languages that support common primitive types (int, char, enum), basic composite types (array, vector, class/struct, string (which is just a vector, but used enough to be worth treating as a first-class thing)) and Java-like methods (fixed argument list, single return value). Further, it's based on "old" Java, before Java acquired functional extensions, when doing things like passing method references as argument was uncommon, and therefore not supported. So it's moderately-expressive but avoids things that get weird and complicated. My one big complaint about it is that I wish it supported unsigned integer types. That's my biggest gripe with Java, too.

I watch a lot of PZ, but you need to take that guy with a grain of salt. Yes, he's generally correct about the overall trends, and he got famous by predicting the Russian invasion of Ukraine, but he's also prone to wild exaggerations, and often spouts incorrect details about countries or industries he talks about.

In the case of China, PZ's biggest argument that they're past peak and falling is demographic data. Unfortunately, just like their financial numbers, their demographic numbers are unreliable too. Most of this was systemic... instead of having a central institution do the census, China relies on data from provinces and adds them up. Unfortunately it also uses that data to determine transfer payments to provinces for things like school funding. So if you were governing a province and wanted more money for schools, you had an incentive to inflate the numbers, and this is what was happening for a long time. As China has been slowly releasing corrections to the data, massive numbers of people are just disappearing from the official counts. Like maybe one or two hundred million people.

What this means is that China's much further along the path to demographic decline than most other countries, and worse, they were actively trying to suppress birth rates with the one-child policy for a good portion of this timeframe. By the time they figured out what was going on, it was too late. You can't suddenly manufacture a bunch of 18 year-olds to fill your factories. And as a nation, they're very unfriendly to immigration, so that option is off the table too.

Now, China also has a leader who likes to shoot the messenger. Anyone under him who publishes data that Xi doesn't like finds himself on the outside, and unemployed. (Which is why we were all so aghast when Trump did the same thing with the jobs numbers.)

So yes, the reliable data we do have seems to indicate that China is in a lot of trouble. But don't take everything PZ says as gospel. He's often wrong.

When in the history of humanity has there not been a big problem coming? Or occurring, for that matter.

The problem in this story is not the wealth, but its form. Cryptocurrency transactions are generally irreversible and not subject to the layers of process and protection that have been built up around large banking transactions. Keep your money in banks and brokerages like a sensible person and you don't have much risk.

Most of the people on Slashdot have been screaming that the emperor has no clothes for a while now. Building a machine that spits out semi-plausible dialogue is very different from making an intelligent machine. I just asked Google's AI to summarize information about myself (my own name from my own town) and it rather hilariously indicated that my wife was actually my (IRL) sister. It had apparently retrieved the names from an obituary for our grandparent but didn't actually understand the relationships. We're not seeing the meaningful improvements that you would hope to see given the ludicrous capital that's been invested into LLMs. This isn't like Moore's Law in the 90's where there is constant improvement along a single axis (transistors per square mm). There have been a couple really big breakthroughs (first deep learning, and then transformers). Throwing more and more compute power at it isn't going to create the payoff that all the investors think. It's not going to get significantly better without more big breakthroughs, and those might come tomorrow, or not until long after we're all dead and gone. LLMs are a very risky bet right now.

Uh, yes, the 24-hour cancellation option is always the most expensive one for a given room (ignoring paying extra for add-ons like free breakfast or extra points). What other option would be more expensive? The one that gives the consumer the most flexibility is the one with the highest risk to the property, and that's priced in.

Yeah, TFA overstated it. Though if you're not booking through the chain directly, in many cases it is hard to get a 24-hour cancellation policy. Many of the travel aggregator services hide them.

Yep. That's all true even if AI is the most transformative technology ever invented, even if it generates trillions per year in economic output -- it might not do it soon enough to prevent another crash. You don't have to believe that AI is "sand-castle based or vapor based" (which it's really not) to see a big problem coming.