3 ms·
His examples are CNN image applications, and it is true that these were spectacular new results 10 years ago. Since then, the tedious engineering has improved a
by johbjo 6y ago
His examples are CNN image applications, and it is true that these were spectacular new results 10 years ago. Since then, the tedious engineering has improved and recreating the results now are indeed just a few lines of python.
The reason they were so impressive then is because it seemed they were "more intelligent" than they actually are, and people extrapolated the same level of performance on any task. Achieving it seemed more impressive then, than now.
But "AI skills" should be broader than image classification.
- willeh 6y agoI think the many point of the article isn't necessarily that it is easy to construct models in terms of lines of code but that any tweaking of the model is dominated by increases in data-size. Why pay for a PHD when you can invest in more data instead.
- AstralStorm 6y agoExcept when you cannot: with spaces too big, such as strategies and games, or too rarified to be sampled. Or ones where generating completely new output is required, such as music and writing.