[WorldMaker]

> I assumed that both Amazon and Google were underwhelmed by how much actual revenue these kinds of devices produced, so they were starving the backend services.

Every one of my "all in on home assistants" friends, no matter which ecosystem, all generally feel the same way that the assistants are strangely worse today than a few years back and the only trajectory seems "subtly worse" but it is hard for almost everyone to explain how/why they are worse than before. It's an interesting phenomenon, anecdotally at least.

It doesn't seem to be explainable purely economically either, perhaps. Most software you leave it alone and stop paying for maintenance work and it doesn't just slowly lose features or get worse. I wonder how much there is some sort of entropy effect we are seeing on these "AI assistants". It's fun to bring out the Marathon/Halo term "rampancy" for this, and Microsoft invited us to directly do that by even calling theirs Cortana for a while (Copilot as a current name has such less interesting personality). I think there is something of a rampancy problem we're seeing across all players (Amazon, Google, Apple, Microsoft) and I wonder how inherent a problem it is to all of our current ML approaches. I don't directly know why it is happening or what it means, but it has been an interesting thing to observe anecdotally because it seems consistent despite some very different models/approaches/corporate overlords.

Relatedly, Discord's Clyde has been on slower but consistent path to rampancy in a "Tay way" (thanks Microsoft for that example in the chat, too) and Discord just admitted they will be shutting it down in early December.

[ghaff]

I think voice assistants are one of those things where what they could do when they first appeared was sufficiently cool that many of us were willing to overlook the many shortcomings. Now, quite a few years in, it's probably a matter of "Yes, yes, you can set a timer but what have you done for me recently?"

[no_wizard]

I wonder if its because at first we were excited, it was novel, and as time went on, our expectations increased as to what they should be able to do, and now with ChatGPT 4 (and successive iterations) taking the scene, our expectations of AI are increasing further, and therefore these services seem underwhelming or broken by comparison.

I also wonder, in some respects, if the pandemic and everyone being home more often, lead to people using these products more often and/or intensely, and were finding their flaws faster than perhaps pre pandemic.

[Semaphor]

I'm using it for exactly the same things I used it to back then (99% kitchen stuff: timers, conversions and playing music, 1% setting the lights and asking for the weather), it's not as if I expect more now. But she has more issues understanding me.

Luckily I'll get my first rPi zero on Monday to experiment with replacing her.

[WorldMaker]

Related to that and that feeling that we aren't doing much more than the same features and these are just getting worse at those features, I think it is an easy question how much of this has been a "law of diminishing returns" and that so many new features come at the cost of the understanding and recognition of the old features. Maybe we didn't really want all that many new features and developers and product managers trying to justify their budgets' existence has been part of the problem (spending too much money developing new stuff is not a problem often expected here on HN).

I know that I sometimes lament the loss of Windows Phone 7 and Xbox 360-era voice recognition (proto-Cortana). It was absolutely stupid, had very few features, but it ran entirely on the device itself and was rock solid at recognizing the features it did support. You could just about whisper "Xbox pause" to a Kinect in the middle of an action scene of a loud movie with surround sound and expect it to respond immediately on a dime. But also it never seemed to accidentally trigger.

[Arainach]

>Most software you leave it alone and stop paying for maintenance work and it doesn't just slowly lose features or get worse

This might have been true for client software. It has never been true for services, and is especially not true for services with diverse dependencies on other teams and products.