Ever since Alexa gained more AI capabilities, I’ve been running into a frustrating issue where she tries to act on whatever she thinks is the closest match to a command.
If she can’t find an exact match, whether because the device doesn’t exist, isn’t exposed to Alexa, or she simply misunderstood the command, she now seems to improvise instead of responding with something like, “Sorry, I couldn’t find a device named that.”
Last night, my kid thought it would be funny to say, “Turn off world lights.” Obviously, there is no device or group called “world lights,” so the expected behavior was for Alexa to reject the command. Instead, she decided the closest interpretation was apparently “turn off all the lights” and shut off every single light in the house, including the ones we intentionally keep on overnight.
Has anyone found a good way to prevent this behavior? I can’t find any setting in the Alexa app to disable this kind of fuzzy matching or AI interpretation, and I’d much rather have Alexa fail a command than make a potentially very wrong guess.
You are very very very much at the mercy of whatev Amazon chooses as thier default prompt.
Which btw has ZERO context about your actual life… Which for you - pretty much get used to it or don’t use Alexa in that way.
is also why I fired Alexa but then went over the top and designed my own. You obviously don’t need to go to my extreme but This is the exact problem inwas solving building my own system.
to behave an Ai agent mut have context. The kind of context you probably don’t want our of your home so the answer is to give Amazon all that context or roll your own…
We use voice control a lot, and since I have always used a logical, consistent, and standard naming convention (though bit wordy at times), it is usually easy to control things. Sometimes Alexa’s guesses are helpful as it will turn on the kitchen light even if you mistakenly say “turn on kitchen lamp”, however it is not always the case as in the dining room I have a floor lamp and a light… and if I just say “turn on dining room lamp”, it turns on the light, not the floor lamp.
In the kitchen there are 9 individually controllable lights, and it is always so much fun if you mistakenly say “turn on kitchen lights” instead of “kitchen light”. I know it is technically correct, but I rarely, if ever, need her to turn on all devices of a certain type. This is also why I have automations to handle turning off “everything in kitchen” (I have one for each room, floor, etc). When I say “turn of everything in kitchen”, it only turns off the set of devices I typically want turned off (ie: I could have a fridge on a metering switchable outlet, but would not want that off) because there was an exact match and the automation only turns off what I want. This is an example of how I wrangled Alexa into doing only what I want, but she has become quite opinionated on what I want
That comes from telling the system your calendar, your chores, your inventory, the layout of the home, who’s home, what day it is, what’s on, what’s off, what missed.
You want Amazon to know that knock yourself out. I don’t.
But I do want that so. I had to model it and give it to the agent. Something Alexa simply can’t do… Because it doesn’t have the context…
Because of the description, I don’t have to ask it to target a specific light because she FUNDAMENTALLY understands the location and type of every light and runs lighting ina room as a single action.
the delta here is to make Amazon do that you’d have to have control of yojr core prompt and give it context. And that’s something Amazon isn’t going to give up.
She’s also being over dramatic on the freezer. It’s running two degrees above baseline… No worries I control it I can tweak it…
@NathanCu I concur on not wanting Alexa (or any other large corp) to have all that context, and I am not opposed to have a local AI. I actually set one up a while back in a Proxmox VM on a dedicate I9-13900 with a dedicated RTX-A5000 (mobile version as the host is a 3L Lenovo PC) but it got a bit complicated so I put it on pause. The biggest issue is how to ingest the commands… I have 15+ Alexas so adding to them, or replacing them, are both not ideal options.
It would be nice if it were possible to use Alexa to take the command but let an alternative AI process it without adding complication to the instructions needed (ie "Ask xyx to… "). I’ll keep dreaming.
As for the freezer, my lower tech monitoring solution is to monitor temp vs power vs door open (barely visible on the history bar for now), and the duty cycle of the compressor. I can also see when the defrost happens. This freezer had leaked half of its R134A, and thanks to Youtube videos, ChatGPT, etc. I was able repair it. The leak seems to have vanished, maybe fixed by something I did, however it warrants 24/7 monitoring now to ensure it doesn’t die on me. If duty cycle creeps up, even if it reaches set temp, I know something is going on. I have no alerts right now, but that is my next step.
I’ll look into what you shared about local AI. I’ve overcomplicated by home setup and spend a lot of time maintaining it (many containers on 4 Proxmox hosts, 35+ Unifi devices, hundreds of zwave, ZigBee, matter and esphome, etc. devices) so I am now trying to simplify and/or reduce maintenance load…
Yeah you’re already doing the stuff you have a system that describes what’s important for you about that freezer. You just need a way to tell your agent the important bits.
Sorry about the reading I can tell you’re about to do…
If you have any experience creating/developing Alexa Skills, you can check out this skill I developed last year. This skill allows you to only give the context you want. It abridges to HA using a webhook. It identifies your echos with a python list.
I’ve included a handful of HA scripts in the repo so you can get a general idea. For example, I can say “Alexa, ask command center what’s the window status?” and Alexa is pre-scripted to let me know what windows are open based on what contact sensors I shared with it in the HA script. I designed it so anyone can make it their own by adding their own intents.
But I digress, this skill is mainly for sharing what information you want to know about the house. It isn’t designed for actions, although, it’s not impossible. The only negative is having to invoke the skill name after saying Alexa.
OK no I’m not.