Comment by NoLinkToMe | Hacker News Reader

Comment by NoLinkToMe | Hacker News Reader

Parent Full thread

> the responses to the various voice assistance was almost eerie in that it could infer what you wanted/needed. Then as things cooled off, they got gradually dumber/worse every year since then. They're legitimately bad today (Siri in particular

It's a compelling argument but I find that difficult to believe given the timespan. Siri launched in 2011.

The FLOP/s per dollar rates for GPUs involved in ML doubled every 2 years across this peripd [0]. Meaning in the past 12 years, you're talking about 64x GPU performance per dollar spent. Combine that with 36% general inflation + some increased spending on mobile computing, plus a decade of time on making algorithms for a big consumer product more efficient, and you get to wildly more favourable economics, as in, two orders of magnitude more favourable than when Siri launched in 2011.

I'm not surprised this could well be happening for ChatGPT (would be cool to see the first public audit standards of AI products emerge, to test this hypothesis). But I don't think it holds true for the 12y period for Siri and co.

https://epochai.org/blog/trends-in-gpu-price-performance

I do remember meeting the Siri team before Apple bought them (I was writing for a tech blog at the time) and having them show me a bunch of really cool queries. I was pretty impressed. Shortly after, Apple bought them, and significantly after that, actually launched it as a product. I was not at all impressed by what launched.

Keeping in mind that the pre-sale demo I got was in pretty controlled conditions (i.e. I could have been duped):

After Apple's launch of Siri, someone else who had been working with Siri told me that he was really unhappy with what Apple had done; according to him, they gutted a bunch of capabilities, mainly because A) they couldn't be launched to all customers or B) just didn't fit with Apple's product vision. To give a specific example, the pre-Apple Siri could make reservations at a hotel in SF / NYC / Austin for you but not in Dubuque or Tokyo, so that had to go.

Of course, that original Siri was very different from ChatGPT: it was based on structured queries and the "cool stuff" it could do was basically a ton of API linkages to outside services. At that point, there'd already been a wave of "natural language" companies like Powerset and Semantic Web that had promised the world and then flamed out, so everyone was aware that unstructured queries were basically impossible. Funny how fast things change.

I don’t think Siri et al got worse, we just had time for the hype of the excellent voice to text and back again to wear off, and realized that they were too limited at figuring out actions to be very useful.

The hardware being cheaper only matters if they truly want to improve the product, which it's not clear that they do. Lowering the number of FLOPs needed will always reduce the cost, as the hardware gets cheaper the cost is just reduced faster.

The question is, would improving Siri sell more IPhones? If not, then why bother?

Reply on news.ycombinator.com