Re:What is GPT-3? written in layman's terms
by Causality1 on Wednesday July 22, 2020 @11:47PM
(#23923984)
"Generate text that could believably have been written by a human" is a task we're getting very good at. I'm curious how far it extends towards concrete applications. When can we start using it to fill functionality gaps between the functions of a human-coded framework?
[ Parent
| on HN ]
Re:What is GPT-3? written in layman's terms
by Krollifi on Thursday July 23, 2020 @02:00AM
(#23924562)
In about the average life span of a redwood tree.
Re:What is GPT-3? written in layman's terms
by Causality1 on Thursday July 23, 2020 @02:42AM
(#23924774)
I'm probably going to be dead before Google Assistant can correctly interpret "take me to the McDonald's beside the Target"
Re:What is GPT-3? written in layman's terms
by londons_explore on Thursday July 23, 2020 @04:26AM
(#23925317)
I reckon that's 3 years away...
Current assistant features rely on trying to match your query fuzzily to an action template. Eg. "Set alarm for 10pm" might match the template "Make alarm at time {TIME}", with $TIME=22:00. Minor transformations of parameters can occur (Converting the time from 10pm to 22:00), but those are mostly hard-coded.
Future assistants will use a neural network to do the matching and parameter conversion, with the networks big enough to encode world data. So things like "Mcdonalds beside the target" can be encoded by the neural net so your query matches "Navigate to ${ADDRESS}", with $ADDRESS="Mcdonalds, 21 Foo Avenue".
It's all do-able today, but nobody has done it yet. Probably 3 years away I'd guess.