GPT 4: Will We Ever Train Again? 

Thanks for the run down, great insight, fun speculation. Kind of scary how fast this is all moving!

I really appreciate this. Thank you!

Since GPT has this "real world understanding" as we humans get from walking and observing the earth for years, I expect training of new tasks to require far less examples. E.g. some factory part detector might be trained with like 10 examples (fine-tuned GPT vision model). For e.g. medical imaging it depends on how many datasets are open source. I do not expect localization to be a problem, just ask the model to output JSON with COCO format etc... I guess they will fine-tune for that capability. OpenAI might also introduce task-specific pruning etc. to reduce cost and latency. I think this might make the need of annotation platforms significantly reduced and for the fine-tuning, the money might go to the company who provides the best foundational model. Just some speculation :)

Would you guys consider doing a new video on using an existing RF model, and implementing detection (inference) for it? Last attempt, the colab failed to compile and it's just been integration hell.

Hi, you think gpt4 can do object detection and count an image based on text prompt?

yes you do need to still train CV models for sure

If models can only be accessed through an API people will continue to train and create their own model weights

GPT -4 will NOT be able to do localization. Not even GPT 5 will be able to do that...