Thanks for this! Really glad you included the sycophancy example in the end. RLHF breaks once humans can no longer tell the difference between answers that look correct and answers that are correct!
1:08 People's expectations from these models are skyrocketing. We have just achieved the milestone of making computers "talk" and now they want is super intelligence regardless of the silicon tech trying very hard to keep up with scaling. I think we would need some sort hardware evolution/innovation before we can start thinking of achieving AGI or ASI. 14:00 From the rumours, the original GPT-4 had 1.7 trillion parameters. Now GPT-4o is just ~200B which makes sense since running such a big model is quite expensive. Similar can be said about Claude 3.5 Sonnet. 19:18 LMAO. I was thinking of the same paper!!!
Thanks for this! Really glad you included the sycophancy example in the end. RLHF breaks once humans can no longer tell the difference between answers that look correct and answers that are correct!
1:08 People's expectations from these models are skyrocketing. We have just achieved the milestone of making computers "talk" and now they want is super intelligence regardless of the silicon tech trying very hard to keep up with scaling. I think we would need some sort hardware evolution/innovation before we can start thinking of achieving AGI or ASI.
14:00 From the rumours, the original GPT-4 had 1.7 trillion parameters. Now GPT-4o is just ~200B which makes sense since running such a big model is quite expensive. Similar can be said about Claude 3.5 Sonnet.
19:18 LMAO. I was thinking of the same paper!!!
I agree re the hardware actually; something I did not at all point out in the video, but super relevant. Thank you also for the other comments! :))