Learning while listening to imaginary friends

I learn from peers who share as they learn. Recently, I started chatting with Simon Toon via asynchronous audio and Signal texts. Here’s one of the chats.

Listen to “While listening about Claude E732” on Spreaker.

Simon and I are upskilling with two independent tools. The first tool is Obsidian and it holds our sovereign data inside its Vault. The second tool is Claude, a very clever AI that the two of us are learning how to use more efficiently. While you can use both of these tools separately, we are asking Claude to extract more intelligence from the way we use Obsidian. Much of the chat in the audio clip covers techniques about leveraging Claude when working with tick off items on checklists.

Sometimes we share questions and big discoveries inside a vibrant back channel we have with the free encrypted Signal app. At the moment we use asynchronous interactions. Signal could be used as a active video channel too if we opted for video calling. However, I believe our lives are better when we simply share gnawing thoughts through Signal and then let the responses percolate when ready.

I am also curious about whether it’s possible for me to more efficiently bundle together our chats in another app.

So I’m looking at Marco Polo since its architecture would allow me to quickly respond as well as its service gives me the option to download and reuse some of the learning pathways that Simon via audio and I have started exploring. I reckon both of us will continue on this spontaneous learning journey. Claude listened to the Spreaker clip that is embedded above and then used ffmpeg and the faster-whisper (smaller model) to generate a transcript below.

Unedited audio transcript below

[00:47] Simon (Tune): Hello, my imaginary friends. This is Simon Tune here. I wanted to tell you about the little argument I had with Claude earlier on. It was about the refresh tasks skill, which is the one I get to manage the to-do list. And I ran this towards the end of the day, because I'd been looking at my to-do list and trying to look at what's important to do. And there was one thing that was important. And I thought, my colleague would have done that. I checked and he had indeed done it. So I then checked the box on the Obsidian note and yeah, rerun the skill. And it didn't pick it up.

[01:25] Bernie Goldbach: Hey, Simon, listen. When I want a checklist run to see how I've done on tasks I have assigned to myself, and when the checklist involves tick off lists, I have to sometimes tell Claude to look at portions of the list in a way that does a calculation, not just an annotation or a computation. And so in the case of how I run my checklist, the daily tasks I have are blocked into terms of activities that are around different kinds of things. So in your case, it might be activities that are outward facing or scrum related or sprint related or whatever. So I have tasks broken down based on money coming in, money going out, calls that to be made, calls that to be followed up, and then some generic housekeeping tasks. And I asked Claude to give me a percentage for each of the tick-offs for each category. That forces it, at least in my case, to make sure it doesn't miss anything that I've ticked off.

And it said, this self-healing thing again, it said, okay, yeah, I can see there's a problem there with my logic. I could widen my search and look for open check boxes and checked boxes and just search for all of them. It said, well, I could do that. But I think we should also look at all of the files, including the ones, all the files I've got checked boxes in already, that are checked, X's in the box. So I should search for all of the files, the ones with open check boxes along with crosses in, and also if you really want me to look in the open tasks, the list.

So Simon, when Claude talks back to me, it kind of informs me of, hey, boss, you might be able to do this better. So it tells me stuff it finds in the places I let it look that I might not be aware the file exists. So it finds anomalies. So in past runs, sometimes Claude goes and dumps the results of the previous run somewhere where I told it not to. I didn't, well, that I didn't know was going to do that. And in a subsequent run, it discovers an artifact where it puts something that it wasn't told to do into a place that it's not told to put it. It's like a log file that it's made in the form of a final report that it just kind of put in place. Now also it would have put the report that I wanted in the correct place, but it sort of leaves behind at the root normally of my obsidian, a file that's relevant to another kind of task that I ask it to run a day or two later. So it's Claude, not apologizing for doing it before as an extraneous task, but Claude pointing out to me, hey, this thing exists. You want to have a look at it? And so I do. I do cleanup on Aisle 4 from Claude when it has done something I didn't tell it to do previously.

Make sure if you're using obsidian for the storage locations, if you're using obsidian for stuff for Claude to do, occasionally scan it for stuff that may not supposed to be there. So what I do as a regular occurrence is I ask Claude to count the number of files that I have in obsidian and also annotate, call out the names of the files that were just produced in the last three days. I can spot the anomalies just on that list. That might work for you. But it was just really interesting to me as an observation that I was explaining to Claude as I would another human being in a respectful way why my idea for the logic was more efficient and more appropriate for this situation compared with the idea that it had put forward.

So yeah, it was really very much like having a conversation with a programmer. I mean I do, I do that a lot in my job. I talk to developers and they tell me about how they're doing things and I put forward other ideas and we have a debate and talk about the pros and cons and then a decision is made that we're both happy with or comfortable with. Yeah, I find that really quite inspiring. One of the things I really like about Claude is that it tells me that I could have asked for something a better way. That depends. I think it depends upon the time of the day, the model, the LLM amusing, the iteration, the previous task that I asked to do. It's like it hasn't purged its memory of stuff. Even though I'm running scheduled tests which should be operating in a new chat box, a routine that shouldn't need to purge memory, but there are certain things that are actually annotated in the log files of what Claude does that tells you it's going back to something that it previously did to see if it could actually inform its judgment and activate certain tasks in a different sequence that I would have asked for it to do. And it comes up as a result, different answers based on different models that I've asked to do a task. So I guess what I'm saying here is I need to show your log files.

But the point is Claude politely telling you, hey, you can do it this way and you say, no, I want to do it that way. You can sometimes spot like you're doing this argument that's going back and forth where Claude's politely telling you, hey, do it this way, it's better. It is fun. It's like an actual person telling you, I think we should do it the different way, use a red marker. And you say, no, blue is better. Shut up, just do the job.

There is a certain lifting that happens to me cognitively, especially in my low energy points of the day, which happens around three o'clock for me, where I invite these kind of AI arguments. Makes me think. And that's good. Sounds like it's making you think too, right Simon?

[08:26] Simon (Tune): It's become an attractive and for a certain part of the population, use a friendly way of interacting with technology. Thank you for listening to the book.

[08:47] Bernie Goldbach: Yes, Simon, you and I are the same when it comes to having some human virtual assistant. People tell me, people I respect, don't make it a person. I'm not, I'm just saying, I think that Claude has helped me think deeper. I know Claude has helped me link my thinking to thoughts I've had as far back as 50 years ago, because sometimes those thoughts are inside my obsidian as notes I kept in journals I maintained when I was learning how to run 50 miles nonstop, how I was learning how to fly different kinds of aircraft. I've flown 19 different types of aircraft, so they are in my obsidian vault, those thoughts. I'm talking to my past self, hoping that some of the stuff that I'm able to write will result in something people will enjoy and that grandchildren of mine will find interesting.

We should do more of this, Simon. I'm going to send you a link to where we could do this, both on audio, like we're doing now, but I think we can accelerate the process if you just tapped into Marco Polo. It's faster for me to respond right at the second that you're talking about something so that I'm highly relevant. I can make a 30 second response and maybe then create something that's like a 10 minute audio response.

We are laying the foundation by talking through Signal and audio clips, proof that you can learn by doing, after reading, after trying, after listening, and then the power in a lot of this is — it's right there at your fingertips that you can see with your own eyes how to do something and watch it malfunction or work and that you can rewind. As you've mentioned, go back a month and say, now I get it — because there will be people listening to this audio clip saying, I still don't know what those guys are talking about. Then a month later, they'll say, oh, God, how can I learn more? Well, there are better people to teach you how to do this. Simon and I have exchanged links and Simon, I need to get some of the places you've visited, voices you've heard that have helped you parse something that was unintelligible from me because I want to learn from what you've learned. Let's keep doing this.

Okay, I'm Topgold on Good Social Networks. Topgold Audio Clips is what you've been listening to. SM2N my Imaginary Friends is the other voice. Hey, check back in if you want to. I'll put some of the notes of this up on insideview.ie because that's what I do. I write and I link my thoughts often with my pencils in the middle of it all. Bye for now.