Stacking AI Tools and the Self-Improving Workflow

**Beth Lyons** (0:01)
One circle is giant, and one circle is like a golf ball.

**Andy Halliday** (0:04)
I'm a promiscuous AI user and payer.

**Gareth Hood** (0:07)
You will never say it's finished. There's always something. So.

**Beth Lyons** (0:14)
Hey, everybody. Good morning. Happy Friday, Friday, June 5th. It is episode 740 We are almost halfway through the 700s, Andy. And I feel like we just did 700 like yesterday.
So I just implied I'm Beth Lyons with me in the studio today is Andy Halliday with us in the chat today is Jude Simmons. Yay. And Brian Maucere. Woohoo.
No, Maucere because you say a Modi. Duh. Yes. I know how to pronounce my co-host name for sure. I never call him by his last name for reasons. Here we go.
You are watching and listening to The Daily AI Show and it is a very Friday Friday. We do expect Gareth to come in. Andy gave us a challenge yesterday. I have some things to say about it and I expect Gareth did quite a lot. So good morning, Andy. How are you doing?

**Andy Halliday** (1:22)
Good morning. I'm well. Thank you. Happy Friday.

**Beth Lyons** (1:24)
Happy Friday. Brian says he was worried that I did not know how to announce his name.
So awesome. Anything happening? What's top of your mind in terms of AI today?

**Andy Halliday** (1:38)
There's a couple of interesting things I came across this morning. Excuse me. The first is that yesterday, ChatGPT announced that they have been improving memory behind the scenes and they have version 3 of their ChatGPT Dreaming function, which updates your memory by basically reviewing all of your chat history that has accumulated since the last time that it ran, and then updating a memory file. There's a thing in your settings, which is in the personalizations thing called memory summary specifically, that it was interesting to read. I looked at it this morning, and as a full-page dossier on who I am, and what I do, and how I work. It was pretty spot-on. I didn't find anything that I would object to in there. That's like the Anthropic slash Dream Command.
But this is just automatically happening for you if you're a ChatGPT user.
Those memories are viewable in that summary that I just mentioned. Then you can also understand what ChatGPT knows about you. ChatGPT also offers a memory management UI where you can add things that you want it to know about you, like somewhat akin to specific or custom instructions. But anyway, then if you want to improve the memory around a certain thing, just chat to ChatGPT about it. It'll automatically find its way through the dreaming process, which is probably a process that's kicked off on their systems during the lowest volume use hours of the day.

**Beth Lyons** (3:45)
Well, that makes sense and also makes sense why all of the systems have funky memories about me. Because I'm often in like, hey, let's imagine this scenario, what would happen after that. And so all of the AI think that I have done things that I've just investigated. Like, no, no, no, no, I don't run seven businesses right now. Those were like asking how it could be done and then getting instructions for that does not equal that I then went and did it and have been doing it since we had that conversation.
Very cool. And I don't know, how do you feel about it being automatic versus something that you initiate?

**Andy Halliday** (4:34)
Well, because there are so many things that you have to configure and set up in the other systems. It's a convenience.
And I don't really object to it, since it's not going to be used broadly, it's not going to be published anywhere. It's really available just for my review. And as a tuning, if you will, of how ChatGPT responds to me. By the way, on their kind of blog about this, they show a bunch of side by side examples of what the prompt and response would be if you didn't have dreaming and what it would be with dreaming.
And so, and those are compelling. You can see very clearly that it becomes a much more knowledgeable AI assistant when you have context about who you are, what you have done in the past. And it actually updates. One of the examples they used in the updating process was, oh, one of the things I know about you is that you're going to be traveling to Spain in July. Well, when July goes past, it says you traveled to Spain in the past. It understands the temporal nature of those things. So it's updating that constantly with some intelligence. It's not just putting down notes that age badly or get stale.