A less cynical explanation: It's heavily trained to ask follow-up questions at the end of a response, to drive more conversation and more engagement. That's useful both for making sure you want to renew your subscription, and also probably for generating more training data for future models. That's sufficient explanation for the behavior we're seeing.

I could be wrong, but I remember that Claude models didn't really ask follow-up questions. But since GPT models are doing that, and somehow people like that (why?), Anthropic started doing it as well.

Because, Anthropic can do no wrong, correct?

With this new announcement, Anthropic is saying they can _specifically_ "do wrong" since it's in their best interests...

Yep I agree, I forgot to add /s to my comment

It's true that track maintenance is costly but it's mostly costly because there's a whole lot of track to maintain; many hundreds of miles of it, often in hard to reach areas. This looks like it'd be a few hundred meters at most, all parallel to each other in one place. So hopefully it's easy enough. The tower crane also requires maintenance.

The funding model where, if I understand correctly, the construction contractor gets paid less (maybe none?) up front but gets 0.75ca per passenger-kilometer traveled on the system, is cool and does seem like it'd help align incentives and keep costs down.

Some of the other advantages listed in this article were enormous free wins not easily replicated elsewhere though. Somebody else had already paid for the new bridge to be wider to have a transit reservation on it. The same is true for the mount royal tunnel, which was conveniently and cleverly reused for sure, but which cost a staggering amount to construct when new.

This article calls standardization a key, but this system is entirely separate from Montreal's existing subway, and therefore doesn't match any existing standard within the city. https://xkcd.com/927/

Many of the great features suggested here (platform screen doors, full automation) are only possible on a brand new system like this. Very few cities have enough leftover rights-of-way to piece together entire new networks like this for cheap, without enormous land takings or tunneling. If you do have space for a new network, then inventing a new standard and using it makes sense; but if you don't, the best you can do is incremental evolution of the one you have.

Tunneling is expensive. But didn't more cities used to elevate their metros over city streets, in the manner of Chicago's loop ? Isn't this doable for contemporary projects?

Not sure what you mean. apt-get, yum, and even things like snap act like app stores for free apps, no?

It's only a real "App-Store" if it has arbitrary restrictions and you must pay fees to a company, obviously.

I think a 'real store' generally allows you to exchange money for something. If I wanted to sell software to Linux users, Steam is probably the closest thing to an 'app store' you could expect to find. Windows has the Microsoft Store, and Macs have the Apple Store.

Some distros have literal stores. But "Linux" isn't a centralized platform the same way Windows and Mac is. Also, the stores for desktop apps just don't really seem to be as effective. Companies are used to hosting their desktop apps on web.

Games seems to be more of an exception to the rule, for historical reasons.

Even with a non-free package, simply add the repository and you're ready to install it.

Silly. You know what I mean. A way to sell an app. Contributing to open source is nice, but some people want to eat food sometimes.

When you're new it can be hard to tell what to ignore; it makes it tempting to pick a simpler framework that you can entirely grasp. Also any published examples, chatgpt etc won't be aware of the subset you've chosen to use when they're providing examples; they're gonna draw from the full set.

I feel like that's more of an issue with the examples and LLMs? Discounting a framework just because it has ever increasing, completely optional capabilities doesn't compute to me. I'm not convinced there's a real problem.

The problem comes in when the complexity is both not optional and not rational.

Hooks do not work as real functions. They are magic. Why are they magic? I don't know, they certainly don't need to be. What state do they change? I don't know. Why do they look pure but actually mutate the application states? I don't know.

Why is react not reactive? Why is it if I change state the entire website rerenders? I don't know. React has a virtual dom. It knows when I change state because I have to tell it, manually. And then... It doesn't use those.

But it's okay, because you can `useMemo`. Why do I have to do that? I don't know.

Evidently I don't have to do that, because react has a compiler that does it automatically now. Why can't react just do it? I don't know. Clearly it's possible. And also every other framework does it.

There are real functional libraries like Effect, pretend functional library like React and just honest old style library like MobX. I think I know my preferred style.

Also you may have to maintain code bases that don’t use your preferred subset.

And you may have to work with developers who have a different preferred subset.

Ideally: there's a train close enough to walk, or a bus or tram that's nearby that runs frequently, is clean, and doesn't get stuck in traffic because there's not much car traffic.

Slightly more realistic: enough people can and do cycle or walk to the train that pressure is relieved on the roads for those who cannot cycle or walk.

Scott Alexander has issued many studies in his time and is surely aware of this phenomenon. He was very cautious even in this study to calibrate for this sort of noise; see the section about Michaels you know.

Heh, both figuratively and literally quixotic. He's on a quixotic quest, but he's literally fighting against windmills.

The airline websites that I've used are remarkably bad with their UI/UX. Far from any grandma being able to use it, definitely. I can rarely get through the process without some inexplicable error or missing field that I can't find or misbehaving custom date picker.

The loaded cars can go coast to coast, but trains are broken up and reorganized often in yards, especially in places like Chicago. this is efficient because two cars from LA headed to Toronto and NYC can travel together until Chicago and only then be split into two new trains. Even within a single railroad this is commonly done.