Shawn Wang: observation

Transcript context

…Agents still get confused if they have too many memories, to my point earlier about that. So I just am out of a call with a member of the Lama team at Meta, and we were chatting about Lindy, and we were going into the system prompt that we sent to Lindy, and all of that stuff. And he was amazed, and he was like, it's a miracle that it's working, guys. He was like, this kind of system prompt, this does not exist, either pre-training or post-training. These models were never trained to do this kind of stuff. It's a miracle that they can be agents at all. And so what I do, I actually prune the memories. You know, it's actually something I've gotten into the habit of doing from back when we had GPT 3.5, being Lindy agents. I suspect it's probably not as necessary in the Cloud 3.5 Sunette days, but I prune the memories. Yeah, okay. The reason is because I have another assistant that also is recording and trying to come up with facts about me. It comes up with a lot of trivial, useless facts that I... So I spend most of my time pruning. Actually, it's not super useful. I'd much rather have high-quality facts that it accepts. Or maybe I was even thinking, were you ever tempted to add a wake word to only memorize this when I say memorize this? And otherwise, don't even bother. I have a Lindy that does this. So this is my inbox processor, Lindy. It's kind of beefy because there's a lot of different emails. But somewhere in here,…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.