Hacker News new | ask | show | jobs
by PlunderBunny 120 days ago
Given the way current LLMs hallucinate, and given that Apple (presumably) won’t accept this behaviour in Siri, I’m skeptical that existing technology (or existing technology scaled up) can ever create the Siri Apple and its customers want.
5 comments

I'll settle for "gets voice to text right most of the time". Seriously, Apple is so far behind on the cheapest table stakes at this point I highly doubt their high standards is the issue.
Oh absolutely. The amount of times I have to pause, take a deep breathe and OVER-enunciate (still with mixed success) because my voice, pulse rise and my patience decreases with every absolute butchering (like not even "close but no cigar" but "how on earth did you come up with that?") Siri does to dictated text message in CarPlay...
I don’t even bother anymore. When it reads back the text message and asks if I want to send it I just laugh heartily and say yeah. Sometimes the recipient has to read it aloud and try to phonetically guess what the original words were.
Yeah, but isn't the voice recognition (as opposed to voice comprehension) separate from the supposedly LLM powered bit of Siri? I want better voice comprehension too, but I don't think that moving to a LLM powered Siri will solve that.
Wouldn’t it? Something like Whisper is great for recognition, and is built on a transformer architecture, like most of the SOTA voice stuff is.
I agree with the other poster and gladly converted to a paying customer of Wispr because they did this right.

Honestly, I bet your question is exactly what every team adjacent to this problem at Apple is doing. Pointing fingers at each other and saying, "This isn't my problem. This is some other team." It's so egregiously broken that obviously no one inside there considers it their problem. I think this must be rampant at Apple currently. There's just no explanation for how their software has gone so completely to shit over the last ten years.

Literally what's the difference between that and Siri now.

Siri can't understand or pronounce very well.

A few weeks ago Siri via Car Play responded to a text and sent it without me saying a word or radio on, and with the setting where it asks first before sending enabled. It responding "Why?" to a serious text was seriously inconvenient in the moment. I watched it happen in disbelief.

(Edit: Didn't see your last paragraph before writing the response below)

I think there is a distinction between Siri misunderstanding what was said (which you can see/hear), and Siri understanding what you said but hallucinating an answer. In both cases, you strictly have to check the result, but in the first case it's clear that you've been misunderstood.

The Siri experience just really really really sucks for the year being 2026. So much more frustrating than the claude and chatgpt experiences I have had in recent months.

To my Apple Watch: "Hey Siri, tell me what the time is in the central time zone right now"

"I found this on the web", watch shows a link to time.gov

The only thing I find Siri useful for is: a voice-activated timer, handy in the kitchen when my hands are full and I am juggling multiple timed process. It does that well about 80% of the time.

I don't think that's at all a safe presumption, given that AI still happily hallucinates summaries of text messages/email that is contradictory to that actual content of the message.
Unless I misunderstand your reply, I think we're agreeing.
Yeah. Apple don’t half ass things. This is why people take their products seriously.
You are in a thread complaining about the most half-assed digital assistant in the industry.