5 ms·

The last 10 years of AI is basically defined by proof of concepts like that that were 80% (or whatever) solutions and claimed there was a path to something comm

The last 10 years of AI is basically defined by proof of concepts like that that were 80% (or whatever) solutions and claimed there was a path to something commercially viable. Turns out that ~20% is always basically impossible - self driving cars being the archetypal example. I work in the field and I think it can be a great tool, but it needs to be acknowledged what its limitations are and how we don't actually know how to address them yet

jeffbee 4y ago
Now it seems like you are the one moving the goalposts. There are tons of machine-learned models in production, in translation, text segmentation, image segmentation, image search, predictive text composition, etc. It's just that people forget the novelty of all these things immediately after they were launched. You can point your phone at printed Chinese text and have it read aloud to you in English. That is alien tech compared to 10 years ago.
- ForHackernews 4y ago
  > You can point your phone at printed Chinese text and have it read aloud to you in English. Yeah, but it's not really that good. Machine translation has improved a great deal, but reading those translations actually involves bringing a lot of human intelligence to the table, "Oh I bet, 'maximum fire alarms spread' on this menu actually means 'very hot sauce'" If all you're claiming is that ML models exist and have useful commercial applications, then I don't think anyone is going to argue against that point. But a lot of these AI promoters go further: in the case of the LessWrong folks some of them are convinced that a superintelligent machine capable of enslaving humanity is right around the corner.
  - jeffbee 4y ago
    You’re saying the bear doesn’t dance all that well.
    - ForHackernews 4y ago
      I'm saying the bear was trained to dance, it didn't choose to dance.
ForHackernews 4y ago
That might just be a Google problem. Historically, they've had the good fortune to operate in search advertising, where being 80% right half the time translates into billions of dollars. Many other fields (e.g. self-driving cars) are less forgiving.

jeffbee 4y ago

Now it seems like you are the one moving the goalposts. There are tons of machine-learned models in production, in translation, text segmentation, image segmentation, image search, predictive text composition, etc. It's just that people forget the novelty of all these things immediately after they were launched. You can point your phone at printed Chinese text and have it read aloud to you in English. That is alien tech compared to 10 years ago.

ForHackernews 4y ago
> You can point your phone at printed Chinese text and have it read aloud to you in English. Yeah, but it's not really that good. Machine translation has improved a great deal, but reading those translations actually involves bringing a lot of human intelligence to the table, "Oh I bet, 'maximum fire alarms spread' on this menu actually means 'very hot sauce'" If all you're claiming is that ML models exist and have useful commercial applications, then I don't think anyone is going to argue against that point. But a lot of these AI promoters go further: in the case of the LessWrong folks some of them are convinced that a superintelligent machine capable of enslaving humanity is right around the corner.
- jeffbee 4y ago
  You’re saying the bear doesn’t dance all that well.
  - ForHackernews 4y ago
    I'm saying the bear was trained to dance, it didn't choose to dance.

ForHackernews 4y ago

That might just be a Google problem. Historically, they've had the good fortune to operate in search advertising, where being 80% right half the time translates into billions of dollars. Many other fields (e.g. self-driving cars) are less forgiving.