4 ms·

There's many levels of pattern matching, and computers have to start at the bottom. First you have to convert the real world into a form you can use, like visi

by ken 7y ago

There's many levels of pattern matching, and computers have to start at the bottom. First you have to convert the real world into a form you can use, like vision or voice recognition. Then you've got to parse that into concepts. Then you've got to analyze those concepts for patterns. Most of our software is still struggling at stage 1, and our software is so poor at pattern matching it can't detect obvious stage 1 failures at stage 2 or stage 3. We can make a program to recognize species of birds from photos, but it will also tell us that a bunch of static is (with 99.5% certainty!) a robin. I don't think any of the experts I know are at any risk of being replaced by AI in their lifetimes.

joshspankit 7y ago
I don't know if I agree with you there. A lot of AI work is around taking data and finding relationships between them. A current-gen AI might not understand what red is, for example, but it has no problem grouping red, blue, and green in to the the same cloud then mapping the relationships to objects that can be those colours. Personall it feels like the current limit on the way to an AI that can take unknown information and turn it in to a pattern matching system is how much data it can fit in it's matrix at once. If I remember correctly (and my info isn't out of date, it's hard to google) entity matching in photos is usually done internally with very low resolution files like 320x320 and lower. It's no wonder that it could easily mistake static for a robin. Take the right person's glasses off and you'll be lucky if they can tell you it's a bird.
- d1zzy 7y ago
  > I don't know if I agree with you there. A lot of AI work is around taking data and finding relationships between them. A current-gen AI might not understand what red is, for example, but it has no problem grouping red, blue, and green in to the the same cloud then mapping the relationships to objects that can be those colours. It's also worth pointing out that it doesn't matter what red is. All it matters is that the entity you talk to understand to recognize in the situations you'd expect them to. In other words, red isn't some absolute concept, for all we know every human may subjectively perceive colors differently but most of us can still recognize the color red when someone asks us to.

by ken 7y ago

joshspankit 7y ago

I don't know if I agree with you there. A lot of AI work is around taking data and finding relationships between them. A current-gen AI might not understand what red is, for example, but it has no problem grouping red, blue, and green in to the the same cloud then mapping the relationships to objects that can be those colours. Personall it feels like the current limit on the way to an AI that can take unknown information and turn it in to a pattern matching system is how much data it can fit in it's matrix at once. If I remember correctly (and my info isn't out of date, it's hard to google) entity matching in photos is usually done internally with very low resolution files like 320x320 and lower. It's no wonder that it could easily mistake static for a robin. Take the right person's glasses off and you'll be lucky if they can tell you it's a bird.

d1zzy 7y ago
> I don't know if I agree with you there. A lot of AI work is around taking data and finding relationships between them. A current-gen AI might not understand what red is, for example, but it has no problem grouping red, blue, and green in to the the same cloud then mapping the relationships to objects that can be those colours. It's also worth pointing out that it doesn't matter what red is. All it matters is that the entity you talk to understand to recognize in the situations you'd expect them to. In other words, red isn't some absolute concept, for all we know every human may subjectively perceive colors differently but most of us can still recognize the color red when someone asks us to.