One thing I wonder is: why is it expected that fine-tuning will be of less importance in the future? Without finetuning how would you define what similarity means for your specific use case. Assume you have a fashion image. What search results would you expect? Maybe for some use case it is important that the color of the results match. For another use case you might only accept similar shapes and the color does not matter. A general model no matter how smart it is will always come up with the same results - independant of the use case.
This is the debate of the next 3-5 years of AI - how do we digest things like GPT-3’s Zero Shot inference abilities academically? I have no idea haha, will be interesting to follow
Thanks for the video! Great stuff.
Thanks Florian, really appreciate it!
🙏
One thing I wonder is: why is it expected that fine-tuning will be of less importance in the future?
Without finetuning how would you define what similarity means for your specific use case.
Assume you have a fashion image. What search results would you expect?
Maybe for some use case it is important that the color of the results match.
For another use case you might only accept similar shapes and the color does not matter.
A general model no matter how smart it is will always come up with the same results - independant of the use case.
This is the debate of the next 3-5 years of AI - how do we digest things like GPT-3’s Zero Shot inference abilities academically? I have no idea haha, will be interesting to follow
6:54