Adventures in AI: Design, Detectives, & Drift with Dennis & Tom

Episode Notes

Transcript

Announcer:

Web 2.0. Innovation, trend, collaboration. Got the world turning as fast as it can? Hear how technology can help, legally speaking, with two of the top legal technology experts, authors and lawyers, Dennis Kennedy and Tom Mighell. Welcome to the Kennedy Mighell Report here on the Legal Tok Network.

Dennis Kennedy:

And welcome to episode 415 of the Kennedy Mighell Report. I’m Dennis Kennedy in Ann Arbor.

Tom Mighell:

And I’m Tom Mighell in Dallas.

Dennis Kennedy:

In our last episode, we had a fresh voices interview with Tom Martin of Law Joyd, and it was a terrific conversation about legal tech, AI, and very practical innovation. Highly recommended if you missed it. In this episode, Tom and I are going to talk about some of the AI projects each of us is actually working on right now. Tom, what’s all on our agenda for this episode?

Tom Mighell:

Well, Dennis, in this edition of the Kennedy Mighell Report, we are going to compare notes on the AI projects we’re each spending time on right now, not just talking about theory, but actually what we’re doing. We’re going to alternate on a few projects. And then in our second segment, we’re going to stay with a theme, and Dennis is going to tell us about a new way of extracting the needles from long AI sessions and smelting them into useful ingots. And as usual, we’ll finish up with our parting shots, that one tip website or observation that you can start using the second that this podcast is over. But first up, Dennis, I thought we’ve been having a lot of fresh voices episodes and I love our guests. I think they’re fantastic, but I thought it was time for us to get back and talk about some things.

I thought it would be useful for us to get really concrete for a change. We spent a lot of time talking about trends and products and things that AI is doing and what might matter next, but what are we actually doing? What are the AI projects that we’re actually working on ourselves? How are we making use of it? I thought it might be useful to get out of the abstract for a change and talk about kind of what we’re actually building and testing ourselves. So I’m going to throw it to you first. What are you working on right now and is there a theme that you’re finding?

Dennis Kennedy:

As I was thinking about this time, I realized that what I would have talked about last week is different than what I would talk about right now. And I think that is the trend that’s happening in AI is the movement within the tools. But I would say one of the things I’ve been focusing on a lot is that something I call drift, which is that the AI tools tend to move around on you, especially as you go into longer sessions. And you start to lose control over what they’re doing. It’s partially not just partially. It’s a lot because of what’s built into them, their own guidelines and their own approaches that are programmed into it, not so much the large language models themselves. So my own interest has been in what happens in longer sessions and how they start to move around on you, and then what you can do about that.

So I have this notion that I’m going to talk about in some of the things I’m doing, what I call taking charge of the control plane of AI. And that once I’m working on something and I see that the AI is doing something I don’t want, how do I preserve the signal and how do I end up with something I can trust? And I think my examples, Tom, will go toward that. And you and I talked right before we started recording about how you started something with an AI and it tells you like, “Oh, I’ve done exactly what you told me to do. ” And then you say, “No, you didn’t.” And it goes, “Oh, you’re right. I didn’t do that. ” And then you have to kind of take control of it and steer it back to getting it to focus on what you need.

And I think that’s something that people talk a lot about AI without doing hands-on are surprised by or look confused about when you and I talk about that to them.

Tom Mighell:

So I’m interested in this conversation because we both decided to come up with the things that we’re both working on. And I find that both things that we’re doing are totally different from each other, which I think is actually good. What’s interesting is that I think that my project’s the through line, if we call it, is actually the flip side of yours. You’re thinking about how to preserve signal, avoid drift, extract value from the process itself. My question isn’t how do I make the AI session better? My question is, what can I actually get out of it? I want something at the end, a deliverable, something I can use, something I can put on the table, literally. We’ll talk about that in a minute, or frankly, just make my life easier. My stuff is going to be minor. I know that there’s some lawyers out there that I’ve been following who have been doing a lot of work in talking about the claude related law practice or AI robots or things like that.

So the things I’m doing are going to kind of pale in comparison, but I really think that this is where a lot of lawyers are going to find value in finding the places where AI can just take something off of your plate entirely, where you point it at a problem, it does the work and you move on. And that’s kind of the direction I’ve been chasing with the projects that I’m going to talk about today. So let’s get started. Dennis, what is your first project? What do you want to talk

Dennis Kennedy:

About? Well, I call it the trace protocol. So as people listen to podcasts, no, I’ve done a lot with AI personas and then sort of made it more concrete even and call them protocols. I’m not going to talk about on this podcast, but I’m working toward this notion of AI appliances that are very dedicated tools. So not AI agents, something different than that. So the trace protocol comes from this thing that says, “I’ve been working on a series of questions in a chat session and I reached the end of it and I know there’s good stuff in there, but it may not all be at the end. And how can I grab the important stuff out of it? ” So you’d say, “Oh, it’s like the note taker or it’s something that’s an assistant that takes notes that says, I’m going to go through this session and I’m going to grab the stuff that’s buried in there and pull them out.

So it’s a great way to do it. And I can’t rely on AI to do that for me. ” So what I’d say in this protocol is to say, kind of map out the session, trace what’s there, which is why I call it trace. And I don’t really want you to just do this little summarization. I’m going to tell you what are the important types of things and I want you to grab those things. So what seem like insights, what seems like takeaways, what seem like action steps, what fit certain criteria? And the goal is to say, “Can I pull out of the session by using one AI prompt the 20% of that session that delivers 80% of the value?” And I just have the AI do it and I say, “If it misses stuff, so be it, because I’m not going to go back through this.

” And so I’m focused on, like I said, if there were decisions made during that session, if I have insights, if there’s questions I want to follow up on, if there’s something I want to brainstorm or start the next session with, or I have a version of it, actually it’s now in the trace protocol of poll quotes, which says, if there’s something in there that I either put into a prompt or was generated as a result of a prompt that feels like a poll quote from an article, then grab that too. And so the idea is that I’m focusing on this question that says, if I do a session in AI of any kind, short, medium, long, if I can’t extract it in reusable form, something I could take with me and start again, then I feel the session didn’t really pay off. And so that’s what this tool, the trace tool is dedicated for.

It’s just, can I in a disciplinary way pull out what’s valuable and do it with the minimum amount of work on my part because I designed the protocol that will tell it what it needs to find? All right.

Tom Mighell:

Super interesting. My first project is a work related project. I think I have mentioned before on the podcast that part of what we do at my company is we create record retention schedules, what records a company keeps and for how long they need to be kept. We also do legal research that produces a long list of citations. What are the regulations that talk about the record keeping requirements? Those lists can be quite long. And once we create that, for most of our clients, our clients want to know, well, which regulations apply to which of the categories on our retention schedule. So we have to do some linking and that linking traditionally is manual. And frankly, it’s a one to many because a regulation can apply to five or 10 categories sometimes, maybe not that many, but it can apply to multiple categories. And it’s accurate, it’s important work, but it is mind numbingly tedious.

It is so boring to do, but it needs to be accurate. So I built a system with an AI where the AI is reviewing the retention schedule, it’s reviewing the list of citations, it does the linking itself and it provides me with the reasoning for every decision that it makes. It says, I’ve linked it to these because of this and because I was doing it just within the regular chat interface, it would not do the entire list. The list was too long. I think it was, I started with a list of 550 citations and it said, “Oh no, I can’t do all this at once. It’s too much for me, so let’s do it 50 at a time.” And working in batches like that was nice because I could then catch anything that didn’t look right and I could teach it, “Oh, by the way, for your next batch, pay attention to this because these are things that you got wrong.” And it kept rewriting the rules and rewriting the rules and getting it right.

And by the end, the last maybe 200, maybe the last half of the citations were all amazingly, amazingly accurate. I used it on a real client project working in batches. At the end of it, the AI said, “I can take all these rules and create an application if you just head over to my code area. I’m happy to develop it. And I really would like to do it because in that, I will then be able to handle the entire citation list at once and it can do all of that all at one time.” But even at the batch stage, it is already saving real time on real work and I’m very impressed at how it works. It’s making a very boring job interesting again.

Dennis Kennedy:

I think that’s the key is finding a job that needs to be done where you can actually understand how AI will help. So my next one I call the detective team of teams, and this is both fun and has a purpose and a job to be done. And so the focus here is Notebook LM and say, “Can I create something in Notebook LM where I am able to upload source materials as this kind of stable environment with stable content to give me a context I can work in long enough to get solid results and avoid the tendency of the big AI programs to drift and as I said, to do that in a fun way.” So I started with this experiment last fall or last summer where I decided I wanted to upload because they’re in the public domain, all of the Sherlock Holmes stories and there’s only 60 of them.

And then I wanted to draw the lessons from those and create the principle, pull the principles out of them and pull all this other information and then ask questions based on that set of sources as if I’m talking in a way to Sherlock Holmes, but with some basis behind it. And that was really fun. And then I said, “Well, but I don’t necessarily always want to talk to Sherlock Holmes.” There’s a whole bunch of great detectives, fictional detectives. It’d be really fun to get their perspective on these things, Colombo, you can think of a whole bunch of them. And so that was my area of interest. So I looked at it and I said, “Well, I’m going to pull in the characters.” And then as I started to use AI to help me put these things together, it said, “Well, if you say what the characters are, the AI tends to kind of … ” There’s this phrase called semantic flattening, but it tends to like overly summarize and simplify what the characters are.

So he said, “The better approach is to figure out what the actual techniques and methods and principles they use, and to use those as the base.” So without going into detail with Sheryl Combs, it would be like observation and what they call abductive reasoning. And you would use that more so than the Sherlock Holmes character itself. So I started building this, and so now I have this set of detectives that I like that cover different aspects and they can look at different problems that I bring to it from a variety of perspectives that can get coordinated or like the best ones are linked up to the type of problem I have for me, or I can choose the ones that I want. And it just gives me like this kind of fun way to say, “Here’s this stable environment where I know the methodology’s being used and the approaches they’re being taken and they can be suitable to things and they may be able to pull out some insights I wouldn’t see otherwise.” So I’m experimenting it … I don’t have like a bunch of murder cases to solve, obviously, but there are certain things where I go like, “Oh, I wonder if I had this group of three or four detectives who looked at this problem that I want to raise in my law school class and what would the answers be?

” And it’s just fun and it gives you that way to kind of step beyond of like, “Oh, I’m asking Claude and Claude says this. ” I’m going like, “To me, it doesn’t matter what Claude says.” I’m giving it what it’s working from and that’s what I want. I’m not really interested in some what Anthropic or what OpenAI decides is helpful or whatever they’re programming in. I’m telling it the perspectives I want. And it’s a really fun way to do things. And I’ve gotten some great insights I don’t think I could have gotten otherwise. I’ve got some stuff that doesn’t make any sense, obviously, but it’s kind of a fun and as a technique, it’s I think potentially really useful over time, something I’m definitely going to keep experimenting with.

Tom Mighell:

So practical question, if you’re using Notebook LM to keep the context similar, are you able to point any LLM at it or do you have to use Gemini?

Dennis Kennedy:

You have to use the Gemini that’s built into Notebook LM.

Tom Mighell:

Okay. All

Dennis Kennedy:

Right. But you could do the protocols for the detectives and load them into sessions and then you’d lose that kind of stability.

Tom Mighell:

The next two projects for me are more about personal productivity and less about work. The next problem that I needed to solve, and you may all have a problem like this, is I have this folder, the folder where everything lands during the week and slowly becomes a disaster. I have a downloads folder that when I get documents in my personal email, I download it there, I get pictures, I get copies of things, I get scripts, I get screenshots. That’s my catchall place. And traditionally, I will go there periodically. Sometimes it’s every week, sometimes it’s every month. And I go and I clean it up and I delete what I don’t need and I move things to the right folders and I get it organized and it takes time and really it’s just drudge work that I do not prefer to do. So I taught Claude cowork to do it for me.

And so for those of you not familiar, Claude has a separate tool apart from its AI chat tool that’s called cowork. The downside, I mean not the downside, the security caution to be aware of is that you are giving Claude access to an area of your computer. You have to really make sure that you’re following some good security protocols to be able to do that. It’s far preferable to using something like Open Claw to get on your computer, but there are still some security issues you need to be aware of. But I gave it access to one set of folders that are on my computer. I basically told it, what’s interesting part is that I was teaching it what to do. It didn’t just figure everything out on day one. It came back to me with a report. “Here are the files I handled. Here are the ones I couldn’t figure out.

What do you want me to do with these?” It wouldn’t delete files for me, which I found reassuring actually. It created instead a to be deleted folder that I could then go in and just delete it or review it if I needed to. I told it, I retaught it, it updated its rules, and now it handles those situations on its own going forward. And that iterative loop, that AI flagging what it doesn’t know and me teaching it and it getting better is a really useful model. And Dennis just talking about that when he’s refining his process for working with any AI tool. But the result for me is a folder that stays clean without me touching it. I just say run my weekly cleanup and it runs it. I’m now comfortable to the point where I may … Well, I’m now comfortable where I may be giving Claude access to more places like email and calendar to actually become a true AI assistant.

I need to work through my issues with that, but that’s coming next probably. All right, we’ve got more projects to talk about, but we need to take a quick break for word from our sponsors.

And now let’s get back to the Kennedy Mighell Report. I’m Tom Mighell.

Dennis Kennedy:

And I’m Dennis Kennedy. We wanted to remind you to share the podcast with a friend or two that helps us out. I guess I’ll go next time. So my third project is a brand new one, which actually in some ways, believe it or not, was inspired by you, and I’m not sure exactly why, so I hope you’ll agree with me that it was. And so we see a lot of things these days that seem like they’re written by AI and it’s getting hard to tell. Sometimes it’s really easy. I had a student yesterday say that they have a professor who doesn’t allow them to use AI and the professor gave out a handout yesterday that had every indication it was ChatGPT written and it was really surprising to that student. But so you’re looking at things, you’re going like, okay, for me, I don’t really care whether it’s written by AI or not.

I just want to know whether it’s reliable, but a lot of times you want to understand like how much is human, how much is AI because it shows how much it’s been worked on. So there are a lot of ways that you can tell things that are AI, certain constructions, the MDashes, there’s like a lot of things, but they’ll change as the tools change and they start to smooth some of those things out. So I said, “If it’s a moving target, then do I really want to try to figure out something that says, here are the tells.” And I create this sort of dictionary of tells and say, “Okay, take a look at these tells.” And I would probably use Notebook LM for that and then I’ll give you something and then you tell me whether it was AI written or human written. And I realized it wouldn’t really be all that helpful to me and it would move a lot.

So I said, “What might be a different way of doing that? ” And so I started to think, and this is where I think you came into play, Tom, in my thinking, was I said, “Well, what else is out there where we try to figure out where something is AI or not, or human?” And I said, “Blade Runner.” Then they said, “The silence and Battlestar

Tom Mighell:

Galactica.” Silon detector.

Dennis Kennedy:

And then I said, “Oh, and like the poker players, they have tells.” And so I said, “Whoa, there’s like half a dozen ways that we can do this. ” And so I said, “Well, what if I start to experiment with AI giving them these types of roles and the methodologies used in these fictional situations?” Some are real. There’s some psychological things you could use as a basis. And I started having the AI work with me on that. And so the idea was to say, how likely is a text to be AI generated and how much of it is AI generated? So that’s the goal that I had. And so I started to play with this and I got to a certain point and then I couldn’t make it any better. And then sort of between the AI and me, this idea came out, because I’m not sure who it came from, that said, well, I was looking only at this synthetic piece to say, “Okay, so what’s AI?” And I was trying to nail that down, but actually there was a lot to be learned by having the AI also identify what was most likely to be human.

So there’s things like cadence, big tell with AI, right? There’s no typos ever. So you could say there’s certain things that make it human. And so I ended up with now this prompt that I can run on things and it will tell you like how likely, like how much is AI and how much is human. And so it’s looking for what it called itself a forensic reading protocol looking for machine residue and telltale patterns. Machine residue. Which is not what I would say, but you get the point. And I’m not looking for certainty. I’m just saying like, if you tell me that something is 80% AI, the signal to me of how much to trust it or how much I need to verify it changes than if you tell me it’s 20% AI, but it’s a moving target. And I thought it was really interesting because it made me kind of rethink completely what my approach to AI detection is and how I realized, especially for students where you’re looking for plagiarism, AI plagiarism and stuff like that, that we’re kind of using the wrong tools.

And so this is really interesting. And then the flip side, Tom, which I know you’re thinking about already, is that you can obviously run this on your own stuff that you’ve AI generated and clean it up so you improve the humanness of it. And so that is something that we all are going to be dealing with over the next few years is that the AI is going to get better hiding that it’s AI.

Tom Mighell:

And I think that’s the point where we need to get to because I am finding that there are certain types of responses or certain types of things that I don’t want to take the time to write. And AI is the better tool to do that. And I make minimal changes to it, but I need it to be in my voice. And I totally agree with that. All right. My final project is purely selfish. I have an app at home that I use for … I cook four nights out of the week. Monday through Thursday, I cook four nights out of the week and I have an app that has a database of maybe 350 different recipes. And what I’ve been finding is, is that I have maybe five or 10 go- tos that I tend to go back to all the time. And the rest of them, the rest of the 340, I maybe make once and then I never come back to them.

And that makes me sad because some of these are really good recipes and I really want to deal with them. And frankly, every week, trying to figure out, scrolling through the recipes, trying to remember what we had recently, making sure we’re not eating chicken four nights in a row. So I went to Claude Code, the tool that can build an application, and I asked it to build me an app that could be my recipe manager. And I gave it the database. I gave it the whole database of it. I gave it all of the ingredients and I gave it some instructions. I said I want different proteins. I want different cuisines. Every other Thursday is a taco Thursday. We don’t do taco Tuesdays here. We do taco Thursdays. So every other Thursday is a taco Thursday, no repeats within a five to 10 week period of time, so mix it up.

And so now all I do is click a button. I click a button and it automatically gives me four recipes for the week. They’re balanced, they’re varied, they follow all the rules. If I don’t like that, I can reshuffle them. I can regenerate the list. But once I plug that list in, I click another button and it generates a grocery list that I can take straight to the store and buy all the ingredients that I need. It collates them all. So if there’s garlic in all four recipes, it tells me how much total garlic I need, which is nice. So I have all of that in one place rather than five tablespoons across all of that. It keeps a history. I can edit the suggestions. I can rank recipes and it took about two hours to build it. It was one of those nice things where I just gave it the instructions and I followed it in the corner of my screen.

I was playing a computer game in the other corner of my screen and occasionally would ask permission to do something and I gave it that permission and I never have to think about what’s for dinner again. And it runs on my computer, it runs on my desktop. So I don’t have to worry about a server or anything related to that, but it runs on my desktop. And for two hours worth of work, it’s a pretty darned impressive thing to use. It’s very simple, very basic, but it gets me what I need.

Dennis Kennedy:

Classic problem to solve that AI can definitely do. And I think that’s what a lot of times when I hear lawyers talk about using AI, they’re like, as they say, trying to boil the ocean as the first project. And it’s like, no, why would you do that? So I mean, I guess time to wrap up that I realized I was listening to myself that all of my three projects for me, if I told the truth, they’re really about fun. I mean, I did them for fun, but they actually do come back to the same thing, which I’m really struggling with, which is this one question, which is, how do you keep AI useful long enough in its current forum to get you to something that you can trust? And that’s the theme that’s going through almost all the AI projects I’m doing right now.

Tom Mighell:

And frankly, I think that we’re solving us, we’re trying to approach a similar problem from opposite ends. You’re building systems to make sure that the AI stays trustworthy throughout the process. I’m skipping straight to the output and I’m asking whether I can trust what lands on the other side, but I think we’re both trying to answer the same question. Can I actually rely on this thing? And I think the answer for both of us is yes, but you have to be intentional about it. And whether that means designing against Drift like you’re doing or building in reasoning transparency like I’m building within that retention schedule linker, you have to always think about trust. You can’t assume it. So I think that there’s some similarities that we have and hopefully stuff we’ve talked about today will inspire you all to think about experiments or projects that you can take on yourself.

Happy to hear about yours. If you’ve got any that you want to talk about, you know how to find us. We’d love to hear about it. We might be able to mention it on an upcoming podcast. All right, we’ve got still got more to talk about, but before we move on to our next segment, we need to take another quick break for a word from our sponsor.

Dennis Kennedy:

And now let’s get back to the Kennedy Bio Report. I’m Dennis Kennedy.

Tom Mighell:

And I’m Tom Mighell. Dennis, before the break, we were talking about the AI projects we’re working on, but before that, you had mentioned to me that you’ve got another idea that sort of touches on your drift theme. So let’s talk about it. Tell me more about what you are thinking about.

Dennis Kennedy:

Well, this goes to this notion of semantic flattening, which is I think a Really important topic for lawyers to become familiar with. And we’ll probably hit it more on a future show. But what I’ve found is that after a really good session, the valuable material is still buried in the whole chat session. There’s residue, there’s repetition, there’s filler, there’s false starts. So even if I do my trace protocol and I start to pull them together, then if I want to extract from there, it still can feel really flattened and repetitive if I just kind of stack them all up into one document. So I said, “What if I do this? What if I just take … I grab the things from a session, then I put them together, and then I do this thing where I say, I’ll pull the needles out of the haystacks from those sessions and condense them down.” Again, giving them criteria for what qualifies as what I’m calling a needle.

And then because I was on the metallurgy theme at the time, I’ve now changed that. And then I take those needles and I smelt them down. And I say, “If I had 50,000 words of trace captures from a whole bunch of sessions, then could I get it down to the three to five most important things that I would want to work with? ” And the criteria would be that they’re really innovative insights, they’re useful for a project I want, but I would set the criteria and I’m just trying to boil things down. And then the idea is to just throw out all the stuff I’m never going to use because I won’t remember it. I could recreate it if I wanted to. And the idea is like, can I end up with the most reusable and valuable assets? So can I take those ingots and I say, “I need to write a blog post.

I need to do something for class.

I have to do a talk or other things. Can I just dig through the small collection of ingots that’s like my best ideas as AI is determined?” And that becomes the starting point rather than to say like, “Oh my God, look at all these documents of either AI transcript sessions or ways I boil them down.” So it’s like boiling down, boiling down, condensing, condensing, and then saying that I use that trace protocol, as I say, to find the signal and then the secondary process more industrial turns the signal into inventory. So most of what you do, if you’re doing a long AI session, which is what I do, it produces a lot of slag if you’re in the metals industry. And the value comes from learning how to find what’s valuable and to turn it into something useful. So in a way, Tom, your dinner tool to me is an example of this that you’re saying like, “I’ve grabbed all this stuff.

I have these great recipes and I’m never going to use them all, but what if I condense them down and I say, Here are the core ones. And I build off of that. And then these other ones can filter in, but I know I have this base and I don’t worry anymore about the 300. I’ll never look at it again. I just say, I’ve done that work. Now I’m pulling out what is the core that I need to work with. And that’s what I’m calling ingots because I was in the, like I said, I was in the metallurgy framework back then. I’m switching to another framework that probably is a topic for another day because it’s not fully developed. But that’s the idea is to say, I can produce all this stuff. And especially using Claude, which is incredibly verbose to say like, “I’ve done this stuff and I just need to keep pulling stuff out of it.

” The risk is this semantic flattening, which will keep, as it summarizes, it sort of smooths off all the edges. And so I build some of that roughness back into this process. But the idea is not that I need to try to remember this, that I need to go through a bunch of transcripts, it’s like pull this stuff out and then what I need to do is just go back and look at these distilled ingots or nuggets, whatever you want to call them, the gems of it. And that’s where I start working from because I’ve refined them in a way that’s focused on how I might want to use them. So I don’t know, Tom, you said like my analogies weren’t making that much sense to you in the script, but does it seem better now?

Tom Mighell:

It does, but I’m only upset that you’ve established this one set of metaphors only to tell me that you’re changing the metaphors. And so we’re going to have to talk about new words again later. I look forward to hearing that in a future podcast. And I think that your process works to solve a real problem, especially if you’re running as many deep AI sessions as you do. The long sessions, I think that makes a lot of sense. My main thought here to kind of applying it to what I’ve been doing is that I’m not sure how many of our listeners are at the stage where they’re generating enough slag to need a smelting operation. Most lawyers I know are still trying to figure out how to get one useful thing out of one good session. The Inget system I think that you describe is very powerful, but it might be a solution to a problem you have to earn, you have to get to.

And I think my workflow is similar to what you’re describing just without the extraction step. All of my projects started as sessions where I was figuring something out, but instead of extracting the insights and turning them into reusable language, I just said, Claude, build me something that I can use repeatedly. So I guess maybe to your point, the app is the ingot. If we’re looking at that, I just skipped the smelting part. I just went straight to the ingit. So for me, I think the question our listeners could be asking is, what’s the one thing I keep doing manually that I could just turn into a tool? To me, for many of our listeners, that’s a lower barrier to entry and you still can potentially end up with something reusable at the end of it.

Dennis Kennedy:

Right. And I think that part of it, for me especially, because I will admit to being older than you are, is that you say, I have this stuff from AI and I’m not going to remember it all. So I’d rather have the AI sort through it and figure out like what, based on my criteria, what is the stuff I should remember and then not bother with the rest of it. So that’s another way to think about what I’m trying to do. But Tom, what it’s time for, it’s time for our parting shots, that one tip website or observation you can start using the second this podcast is over. Tom,

Tom Mighell:

Take it away. Well, I actually have a non-AI parting shot and it is another purely selfish act on my part because it’s not a tip, it’s not an observation, it’s not a website. It is a story in the life of Tom and maybe a cautionary tale, but I also don’t want you to get scared about it. I use a pair of … I use sleep earbuds when I go to bed. I like to listen to certain things when I try to go to sleep. And I use sleep earbuds because I sleep on my side. I don’t want something like AirPods are not going to work because they stick right out of your ears. They would be uncomfortable. So you need something with a low profile. So I use Anchor. Anchor’s a reputable brand. Anchor has some sleep buds that I have. And a couple of weeks ago I was getting ready for bed.

I only put one in because we’ve got an elderly dog that makes noise during the night that I need to pay attention to. So I put one in my right ear. I turned on what I was going to listen to. It made a high whining noise and the earbud battery exploded in my ear. It completely exploded. The fortunate thing for me was that it exploded outward. The rubber that you insert into your ear was completely unharmed. So my inner ear, the interior of my ear was fine. According to the nurse who cleaned it, I had a blast radius around my ear of soot and other stuff. I had to throw my shirt out because it had burning metal in it. I only tell this story to make to … Because I went and looked it up. There’s been no other reported episodes of any anchor earbud exploding anywhere.

So it’s not like even a thing that’s happening. I guess I just want to say it’s something that can happen. So just be careful. I’m not sure I’m ever going to use sleep buds again when I sleep. I don’t use AirPods. I have Bose earbuds and I have my bone conduction ones, which right now seem much more safe than something that goes in your ear. But I guess I will just say beware and be careful. If there was one thing wrong about these earbuds, they were older. They were not too old, maybe two years old, because if you know me and you’ve listened to me on this podcast, I go through new earbuds and earphones all the time, but they were a little old. So I guess just be aware and maybe learn from my experience. Dennis.

Dennis Kennedy:

Okay. So I feel like I shouldn’t even do a parting shot because there’s no way for me to top this. This is the greatest parting shot ever. So my parting shot is something that I’m hearing more people talk about, which is, and I call it the drift. And so I want listeners to pay attention to this. So I think that the real skill going forward isn’t really generating output from AI. It’s building systems and developing skills that you notice when the AI is starting to drift and then figure out how you can either work with the drift or work against it. So I love long AI sessions, and I think there’s a growing concern that the big AI companies are making us optimize for the AI’s convenience and not for what we want to do. So my parting shot is to … When you have the chance, just like lean into the drift, because that’s where you can learn what the system can do, what it can’t do, and maybe how it can work a little better for you.

And if you’re a lawyer, you can definitely represent your clients better if you understand those things in a very visceral way.

Tom Mighell:

All right. So that wraps it up to this edition of the Kennedy Mall Report.

Dennis Kennedy:

Thanks to the Legal Talk Network team for producing this show. You can find the show notes and transcripts on the Legal Talk Network website.

Tom Mighell:

If you like what you hear, please subscribe in your favorite podcast app and leave us a review.

Dennis Kennedy:

You can also connect with us on LinkedIn with your questions.

Tom Mighell:

So until the next podcast, I’m Tom Mighell.

Dennis Kennedy:

And I’m Dennis Kennedy, and you’ve been listening to the Kennedy Mighell Report, a podcast on legal technology within internet focus since 2006.

Announcer:

Thanks for listening to the Kennedy Mighell Report. Check out Dennis and Tom’s book, The Lawyer’s Guide to Collaboration Tools and Technologies, Smart Ways to Work Together from ABA Books or Amazon. And join us every other week for another edition of the Kennedy Mighell Report only on the Legal Talk Network.

Continue Reading