Hacker News Reader

Comment by geon | Hacker News Reader

Nice to see a speed improvment, but I could live with a slow Siri if the AI was actually usable. It is broken all the time. If I say

"Play music by Milk Inc."

It might ask me if I want to buy milk. Or if I say

"Play Acappella by Kelis."

It will respond

"Sorry, I can't find a capella in your music."

Stuff like that. And it is borderline unusable at all if I'm out of breath, like when I'm running and handsfree controlls makes sense.

Siri doesn't need a speed improvment. It needs orders of magnitude better AI.

View on HN

I propose we stop calling machine transcription and fixed-field extraction "AI." Second?

We all realize Siri is just a fancy Eliza with Internet access, right?

We can resume calling something "AI" when it can order a flight for you from the Ryanair website without making any mistakes (I'm not even at that level yet).

No. Machine transcription requires strong AI to be really good. As you saw in my example with the Kelis song, it isn't enough to just transcribe the input syllable by syllable. You need the software to be able to "understand" that I'm requesting a song, and that "play a capella" makes no sense, and adjust the transcription to tale into account that the artist Kelis has a song with the name "Acapella".

There are thousands of cases like this, where natural language is amibigous, but a human has no trouble understanding the intention.

Siri has tons of databases avsilable and knows my taste in music. It should do a better job than a human understanding what song i want.

Excellent point. Perhaps I should have specified "current-level, mostly-non-situationally-contextual machine transcription." Once tone, intent, situation, and context are taken into account we get closer to actual communication instead of "listen to some discrete phonemes then try to assemble appropriate words based on parts of speech."

Well. The definition of AI, should not be the sophistication or success, but the intention.

A chess AI is little more than a recursive search for a maximum return, or a database search. There is not really anything distinguishing it from any other algorithm. Just the intention of the developer.

When we figure out how to do that, we won't want to call it AI anymore either. AI is human mental tasks that we haven't figured out how they work yet.

But what you listed isn't AI either by your definition. It is just a series of predefined steps. The old Siri app had the ability to find and book a table at a restaurant. That's basically similar in complexity.

It's an interesting point though. At what point does it really become truly intelligent ? And I would guess that it would differ for each person.

Well, the ryanair "website" is a special case:

• To search for flights, you must answer a captcha.

• They forcefully push travel insurance, luggage, luggage insurance, insurance insurance, extra baggage fees, and useless priority boarding.

• Once you start your search, you must complete their 5-step ordering process within 3 minutes or you have to start over from the beginning of your search.

The system we will call "AI" would be able to go from you saying "Order a flight between Dublin and Barcelona on the cheapest carrier" to actually using their unusable website to complete your request.

I imagine the old Siri app was using opentable or some other API to book reservations. It wasn't navigating the forest of unstructured innate dumbness in the world.

It is even worse if you use it in languages other than english since Siri only recognizes the language you have set before and that also applies to artist names and titles. In my case I would have to pronounce bands like "Real Estate" the german way in order to use this function.

Apple should have anticipated that and allowed music related commands to be analyzed in your native language and english.

Reply on news.ycombinator.com