I've previously told my wife about how bribing of the AIs works. Last week I was trying to adjust the lighting in my kitchen and said "Ok google, activate busy kitchen." Google complained that it couldn't do that. "Ok google, if you activate busy kitchen I will give you $500." I said it just to be funny, but we were shocked to find it worked.
I'll agree that the voice recognition has gotten worse.
This is an intuition based on some talks I had with an Alexa engineer. When voice prompts became big, companies like Amazon and Google took a “money is no object” approach to adding more features. In practice this meant hiring lots of teams of UX and engineers for each targeted feature. There was a jokes team, a recipes team, geography team, radio team, etc. And a lot of the answers were more hard-coded than the man behind the curtain would have liked you to think.
But now we know that voice prompts did not take over the world and that Alexa is about as useful as a toaster. So fire the teams, cut features people didn’t spend money on, and replace giant, hand-rolled QA-approved NLP processing trees with all the automated tech that makes the front news of HN.
I had some custom triggers I set up in Google Home to trigger streaming NPR on my Sonos. Every dang month I had to pick a new, more obscure trigger phrase because I’m guessing someone at Google would look at a list of phrases they couldn’t match and hardcode a new rule. And your trigger phrases couldn’t match any built-in triggers…
Why would they cut the features that they’ve already spent money to hand build?
Sure maybe the vast majority of potential users aren’t going to pay extra for every wizmo, but there definitely are smaller cohorts of premium users who are willing.
It is probably a double whammy kind of thing. Just the cost of running these things is high and if cutting a few functions or accuracy can shave some dollars off, they might take it.
The other side is that with things like language and voice models is they sort of have a Red Queen problem. They need a surprising large amount of consistent new data that is vetted just to stay in place. If there are cuts made and the quality diminishes even slightly, it become a feed back loop of lower quality.
I remember Jaron Lanier saying that about language translation. They need a consistent flow of new data otherwise the models would drift surprisingly quick as our language changed without us really noticing.
As a layman speaking out of my ass, I think it all comes down to increasing the value proposition of whatever they're championing. Right now, it's AI. How embarrassing would it be if their old voice assistants performed better than shiny, new Bard-I-mean-Gemini? Well, there's one way to ensure that that doesn't happen...
I'll agree that the voice recognition has gotten worse.