5 ms·

If it was just computer that caused the various assistants to get bad, then why haven’t they gotten good again? Google assistant came out in 2016! That is 7 y

by spott 3y ago

If it was just computer that caused the various assistants to get bad, then why haven’t they gotten good again? Google assistant came out in 2016! That is 7 years ago! Since then, GPU compute in flops has 5xed… for single precision. If you include half-floats (not a thing back then), it is 75x the single precision flops. I think the reason that assistants have “gotten bad” is a combination of edge case controls (avoiding politically problematic results), increased capabilities mean more opportunities for failures, and their failures are amplified, and their successes ignored (in other words they aren’t necessarily worse, we just expect more out of them).

assistantist 3y ago
If it was just a matter of flops then AI would already be driving cars and making door to door food deliveries. The problem is neither the energy expenditure nor the compute capacity. No one knows what the proper architecture should be for general purpose intelligence assistants so everyone has to fine-tune the existing models on their specialized use cases and even then it's a hit or miss depending on how much training data you have available. No one currently has any ideas on how to deploy auto-adaptive ML software that continuously adapts to new data so the industry has settled on deploying snapshots and then hoping the data set that was used for the snapshot is going to be good enough for most use cases. It seems to work well enough for now and since Facebook has decided to open source most of their work someone in the open source community might figure out how to continously update these models without deterioration in output quality.
- thfuran 3y ago
  The difficulty of improving AI capability towards AGI isn't really related to a purported worsening of an already-implemented system for financial reasons.
  - financialisty 3y ago
    If money is the bottleneck then we will never get AGI. A superior machine intelligence would have no use for money.
    - BlueTemplar 3y ago
      Only if it's singular.
    - astrange 3y ago
      On the contrary. If we invented an AGI, it wouldn't be superhuman because it would be too busy working a job to pay for its own AWS credits.
- spott 3y ago
  I think you are missing my point. The OP said Siri et al. got worse because their creators optimized them to reduce compute costs. I’m saying that if that were true, then they should be at least as good as they were because you can get tons more compute for cheaper these days. So it isn’t compute that makes the various assistants “worse” than when they were introduced. I made no statements about their quality in general, just their relative quality compared to when they were released.
  - verve_rat 3y ago
    That assumes that holding perceived performance steady while letting compute consumption get cheaper saves enough money, quick enough. If voice assistants were losing $10 billion a year (made up number) in 2013 and bringing in zero revenue, then it is easy to see why a decade later they would have less compute allocation now than they did back then, even in absolute terms. If the revenue is near zero, then the spend on compute will be near zero too, after everyone figured out that it wasn't going to win some AI battle for all the consumers. If they are not willing to set money on fire for no measurable benefits anymore, then compute will tend to zero over time.

by spott 3y ago

assistantist 3y ago

If it was just a matter of flops then AI would already be driving cars and making door to door food deliveries. The problem is neither the energy expenditure nor the compute capacity. No one knows what the proper architecture should be for general purpose intelligence assistants so everyone has to fine-tune the existing models on their specialized use cases and even then it's a hit or miss depending on how much training data you have available. No one currently has any ideas on how to deploy auto-adaptive ML software that continuously adapts to new data so the industry has settled on deploying snapshots and then hoping the data set that was used for the snapshot is going to be good enough for most use cases. It seems to work well enough for now and since Facebook has decided to open source most of their work someone in the open source community might figure out how to continously update these models without deterioration in output quality.

thfuran 3y ago
The difficulty of improving AI capability towards AGI isn't really related to a purported worsening of an already-implemented system for financial reasons.
- financialisty 3y ago
  If money is the bottleneck then we will never get AGI. A superior machine intelligence would have no use for money.
  - BlueTemplar 3y ago
    Only if it's singular.
  - astrange 3y ago
    On the contrary. If we invented an AGI, it wouldn't be superhuman because it would be too busy working a job to pay for its own AWS credits.
spott 3y ago
I think you are missing my point. The OP said Siri et al. got worse because their creators optimized them to reduce compute costs. I’m saying that if that were true, then they should be at least as good as they were because you can get tons more compute for cheaper these days. So it isn’t compute that makes the various assistants “worse” than when they were introduced. I made no statements about their quality in general, just their relative quality compared to when they were released.
- verve_rat 3y ago
  That assumes that holding perceived performance steady while letting compute consumption get cheaper saves enough money, quick enough. If voice assistants were losing $10 billion a year (made up number) in 2013 and bringing in zero revenue, then it is easy to see why a decade later they would have less compute allocation now than they did back then, even in absolute terms. If the revenue is near zero, then the spend on compute will be near zero too, after everyone figured out that it wasn't going to win some AI battle for all the consumers. If they are not willing to set money on fire for no measurable benefits anymore, then compute will tend to zero over time.