3 ms·

> Siri works great for a lot of the basic use cases that didn’t work in the video. Reminders, notes, alarms, messaging. Honestly I’m guessing it would work just

> Siri works great for a lot of the basic use cases that didn’t work in the video. Reminders, notes, alarms, messaging. Honestly I’m guessing it would work just as well for some of the knowledge questions that they asked. It just doesn't or people would be using it and people wouldn't be as excited about LLMs as they are. Comparing Alexa/Siri style, it works if you know the right trigger words and how it expects you to ask the question, also just does that one thing you can't follow on and expand on it is just a completely different world to what an LLM agent can do. Apple Watch and Siri are dead products for good reasons.

ninkendo 3y ago
> it works if you know the right trigger words and how it expects you to ask the question I don't think we should discount this. There's something to be said for a product that works in a specific way, and if you use it in that way, it actually works. I don't need Siri to divine my hidden meaning regardless of how casually or ambiguously I phrase something... but I do want to know the phrase that works 100% of the time, and then I'll use that 100% of the time, and I'll be happy. What I hate is how Siri fails at even that basic requirement a lot of the time. It's the worst for HomeKit scenes: "Set the scene 'In The Basement'" bounces between working and failing with an inscrutable error, release to release. Sometimes just saying "In the basement" works, other times it lists all the lights in the basement and asks me what it should do. I have a "Bed Time" scene that starts lullaby music in my kids' bedrooms, but sometimes Siri thinks "Bed Time" really means "Good Night", and turns off all of the lights in my house instead. Which is super great when my wife is in the kitchen trying to prepare things and suddenly the whole house goes dark. For me, the biggest problem with Siri and other assistants is the opposite of what you describe: The product is sold as "natural language", which really means that it fuzzily matches what you're saying to its discrete list of features, and that fuzziness is only partially accurate, leading to broken use cases some non-zero percent of the time. LLM's are likely to be even worse here... maybe you've used a certain phrase every time you wanted it to do something, then some model update comes along and it decides to do something else with this phrase, and you'll have to constantly re-discover what works every time it updates.
- whywhywhywhy 3y ago
  >LLM's are likely to be even worse here... We already know they’re better, even if you screw up the question in some nuanced way which would trip Siri up completely they often get the gist of what you want and provide the expected solution. Prediction engine vs trigger term syntax
- MBCook 3y ago
  Naming things in such a way Siri won’t confuse them is an important skill to make things reliable. Shouldn’t need to be, but is.

ninkendo 3y ago

> it works if you know the right trigger words and how it expects you to ask the question I don't think we should discount this. There's something to be said for a product that works in a specific way, and if you use it in that way, it actually works. I don't need Siri to divine my hidden meaning regardless of how casually or ambiguously I phrase something... but I do want to know the phrase that works 100% of the time, and then I'll use that 100% of the time, and I'll be happy. What I hate is how Siri fails at even that basic requirement a lot of the time. It's the worst for HomeKit scenes: "Set the scene 'In The Basement'" bounces between working and failing with an inscrutable error, release to release. Sometimes just saying "In the basement" works, other times it lists all the lights in the basement and asks me what it should do. I have a "Bed Time" scene that starts lullaby music in my kids' bedrooms, but sometimes Siri thinks "Bed Time" really means "Good Night", and turns off all of the lights in my house instead. Which is super great when my wife is in the kitchen trying to prepare things and suddenly the whole house goes dark. For me, the biggest problem with Siri and other assistants is the opposite of what you describe: The product is sold as "natural language", which really means that it fuzzily matches what you're saying to its discrete list of features, and that fuzziness is only partially accurate, leading to broken use cases some non-zero percent of the time. LLM's are likely to be even worse here... maybe you've used a certain phrase every time you wanted it to do something, then some model update comes along and it decides to do something else with this phrase, and you'll have to constantly re-discover what works every time it updates.

whywhywhywhy 3y ago
>LLM's are likely to be even worse here... We already know they’re better, even if you screw up the question in some nuanced way which would trip Siri up completely they often get the gist of what you want and provide the expected solution. Prediction engine vs trigger term syntax
MBCook 3y ago
Naming things in such a way Siri won’t confuse them is an important skill to make things reliable. Shouldn’t need to be, but is.