Episode 14 · Transcript
My Agents Can't Delete
Full transcript of the recorded conversation.
Back to the episode · timestamps jump to the video · lightly machine-transcribed, may contain errors
0:00 Limited Edition Jonathan
1.2 million documents were processed in under five minutes with Jev. extraordinarily fast, like between 70 and 150 milliseconds per answer.
Sonnet did better than Sol. On low.
Like Sonnet Low with the minimum reasoning.
it was way cheaper. I just let it do the rest of the 30,000 documents and I think I use 40% of my usage limit
0:32 Sam Rogers
Welcome to Signals and Subtractions. I'm your host, Sam Rogers. What processes in your organization only exist because the actual work used to be impossible? Not the hard stuff, the impossible stuff. somebody once needed a team or a budget or six different specialists to do it. So a process grew up around that, or some sort of queue or sign-off or handoff. And now one person.
And a capable model can do the whole thing in an afternoon or less. Who would notice if that process just went away? If we met at Moonshots Live last week, welcome. Yes, this is the show I told you about, and I'll do my very best to roll all the Moonshot mates into this one episode with today's special guest. Who's that? Why, it's the return of Jonathan Edwards, aka Limited Edition Jonathan.
He's a filmmaker who teaches AI fluency by often documenting his own failures in public. he was on the show back in episode five talking about Substacks AI detector. Jonathan, welcome back. So glad you're here.
1:39 Limited Edition Jonathan
What is up?
1:40 Sam Rogers
Why is Jonathan here? So, this timing is really perfect. as we record this, OpenAI's dev day was yesterday, and Anthropic just released another new model on Monday, and Jonathan had early access before that. But this is not like launch review. I don't think either one of us really cares about anything that lands on a leaderboard. what we care about is what we can do now that we couldn't before, and what he did with.
Sonnet 5.5 recently with Jev, as we mentioned on the show last week. I'll let him tell you. there are some wonderful signals that Jonathan always brings from his mad science creative brain. getting down to the minute details of things and coming back with.
really simple and useful stuff as a result. So get your geek on. I'd love to dig in with the kind of signals that you're watching lately. just before the show, you were telling me like, well, you know, since last night, it's all different, actually, what I want to focus on. And like totally apropos, thank you. Bring it.
2:46 Limited Edition Jonathan
what I've been thinking about lately a lot is
how like one thing leads to another when you take a few minutes to learn some of the little things in the smaller,
2:55 Sam Rogers
Mm.
2:56 Limited Edition Jonathan
seemingly inconsequential things.
the foundational pieces, the primitives.
like when we were talking about Jev,
If Jev doesn't absolutely floor you, then you don't understand what it is yet.
that's honest. Like I I spent the first few days after Jev was announced. Like, what is the big deal? Like I get it, it's kind of cool. And then all of a sudden, when it clicked, I'm like, my God. Like the ideas just started flowing.
3:19 Sam Rogers
so let's let's pause on that for a minute there's been so much coverage about Jev. like you were talking, primitives. Like it is a a different kind of building block I love Nate's framing of it of like the semi-deterministic kind of thing, which is a great way to think about it. and for those at home who don't know, I'll put it in the show notes for Nate B. Jones, which is the community through which Jonathan and I met.
there were classifiers that existed for a long, long time before Jev existed. but I think there's been so many people who just weren't.
Really aware of that as a concept or have only had experience with using LLMs as classifiers, that it seems like an
3:55 Limited Edition Jonathan
Yeah, and ask yourself
3:57 Sam Rogers
even bigger deal.
3:58 Limited Edition Jonathan
so ask yourself just one of the three question types you can answer. you know, just a boolean yes or no. Right? how many files in your local directory, like in your laptop right now, how many files
Could you benefit from just having a yes or no question answered about those files? and just one question, right?
pick one where you could add up to 255 responses So pick one. Like, is this about one of up to 255 things? is this X yes or no? And then, you know, probabilities and all that kind of stuff.
Let's talk about email because that's a that's an easy one people could wrap their head around. Is this email important? Yes or no? I mean, that question alone awesome. you can look at an email and determine yes or no pretty quickly whether or not it's important, right? But can you look at a hundred thousand emails and answer that question
in under a month.
Jev can do it in thirty seconds.
4:50 Sam Rogers
the system one kind of kind of thinking where it's just like reflexive. It doesn't necessarily take a lot of reasoning. It's basically multiple choice, you know, and you can specify what those choices are and like up to 255 different choices, which is a lot, and it happens so fast and so cheap.
That things that have never been possible
5:10 Limited Edition Jonathan
Yeah.
5:12 Sam Rogers
before. Because, like you're saying, it's like cost and time prohibitive for humans. Like we're just, it's not worth it. Like we're not gonna do that. it makes a lot of sense to be able to put those kind of choices on the front end to gate what then is worthy of human effort, human time, or reasoning models and the the tokens that.
We might need that higher level reasoning from the frontier. one of the things I was using it for, is this something that really should be routed to a local model as opposed to a frontier model? how much reasoning does it need? How much privacy and security does it need? being able to make those kind of it's like semi-deterministic. so those those deterministic
like snap judgments.
that do
involve a little bit more than just a script to do it.
6:02 Limited Edition Jonathan
so
Once you understand those bare primitives, right? you start compounding on them. So, like, for example, the real estate company I was telling you about before. it was like 1.2 million files like old files, throughout the last 20 years. And
To go through all of those files and identify which ones are important, which one needs intention, what they're actually about, like whether they're duplicates, whether they're stale, what town, county, region they're from,
where they were listed, how they were listed, how many people are in them, all that kind of stuff. Jev can answer all those questions, right? It can't name the person in there. You can't say, like, list the people in this. but you could say how many people are in this, you know, probability of whether or not this is important,
is this from pick one? And then you have a list of counties or towns if you don't have enough room in that two hundred fifty-five, well, create a second list with another two hundred and fifty-five,
6:53 Sam Rogers
I actually used to work at a mortgage company and I remember trying to do that deterministically.
6:58 Limited Edition Jonathan
in
the laws are different between you know in the legislation and like it gets
7:02 Sam Rogers
Yes.
7:02 Limited Edition Jonathan
super complicated extraordinarily fast. So what you do is so let's say you want to categorize this backlog of documentation that hasn't been classified or separated out yet or you just want to answer a bunch of questions about your corpus, you work on those questions. Like you spend a few hours, like not a couple minutes, you spend a few hours figuring out the set of questions you want to ask every document.
And then
Then you set it up. Only takes a couple minutes to set it up and send out those questions to ask against every document. that's what I did with this real estate company. 1.2 million documents were processed in under five minutes with Jev. extraordinarily fast, like between 70 and 150 milliseconds per answer.
7:46 Sam Rogers
Dang.
7:47 Limited Edition Jonathan
And you could run things in parallel. the only blocker was how many requests you could send the API at the same time. but even that was fine because running, a hundred in parallel and waiting for that to finish and then running the next hundred was like two hundred milliseconds, two hundred milliseconds. Like you know, like so like even e even running it
8:06 Sam Rogers
Yeah. It's practically concurrent, even though it's not
technically concurrent. Yeah.
8:11 Limited Edition Jonathan
Exactly,
exactly. it's insane. But then one of the one of the main questions I asked was like, hey, the question set was designed to promote documents to the level of like, should I have an LLM, like a standard LLM, take a look at this document and extract
the important information from it. so that's what I did. I came up with all the documents that we needed to actually send an LLM out to. And then we had that list. out of 1.2 million, there was still 40,000 of them. that's where, you know, using traditional LLMs get a little harder. you're burning
tokens like crazy these agents over engineer the crap out of everything. that that's a whole other story
8:52 Sam Rogers
a it's a story worth pausing on as well because the knee-jerk response for people who are maybe a little newer to this is, well, I'll use the biggest, most powerful thing because of course that's gonna figure it out, right? But that wasn't exactly the experience you had.
9:07 Limited Edition Jonathan
think about Jev being the smallest and least powerful thing I have, but the fastest and most accurate, relatively speaking.
to approach those forty thousand documents, I had started out with GPT six Sol right? I started on high. That was dumb. I burned through like half of my token usage in just like an hour. and I processed like maybe 40 documents
this is not sustainable. so I'm like, Well, why don't I have Codex manage a bunch of Luna agents? So I did some comparisons and Luna was giving pretty decent results, within ninety-eight percent of the same accuracy as Sol High for this particular job type.
I tried cursor with grok and everything, and that just that that was a no-go. not that grok did it a bad job, it's just didn't really have the architecture I needed to in order to process this. after burning through
7482 documents with codecs slash Sol that burned through so you know how on openai you have your token reset
10:01 Sam Rogers
Yeah, yeah,
10:01 Limited Edition Jonathan
I burned through two of them so
10:03 Sam Rogers
Yeah.
10:04 Limited Edition Jonathan
two of them in an hour to get seven thousand and I needed to get to
forty seven thousand documents. so I had thirty thousand documents remaining. I'm trying to be cheap right now and I can totally use the API and burn five hundred bucks to finish this job, but I don't want to right now. let me try it on Sonnet
five point five
Sonnet did better than Sol. On low.
Like Sonnet Low with the minimum reasoning.
so
10:25 Sam Rogers
Yeah.
10:26 Limited Edition Jonathan
it did better than GPT six Sol and it was way cheaper. I just let it do the rest of the 30,000 documents and I think I use 40% of my usage limit on anthropic.
but yeah
10:37 Sam Rogers
Well, well, you are you are very good at
it. And thank you as always for sharing on your sub stack, in conversations like this one what it is that you're doing and making those tools available in many cases. I'd like to go over to my signal as well because I had plans for what I was gonna talk about today.
but then Jonathan gave me something that just totally changed it all.
11:03 Limited Edition Jonathan
yeah, y'all are gonna like this.
11:04 Sam Rogers
so fresh from last night, built on Jonathan's own work. a video production guide for every page of a website to make a video, like a two-minute explainer video.
live now on signals and subtractions.
what I'm showing is the the takeaways page. Every episode has a specific takeaway so that you can make everything that we talk about on the podcast actionable.
11:27 Limited Edition Jonathan
this is a good idea.
11:28 Sam Rogers
there's a voice, there's background music, it basically walks you through
Everything about what takeaways are and how to use them and why you should care. And it's a completely self-contained two minute video that I did not make. I simply directed leveraging what Jonathan has created with his AI video guides. So what I did do was download this repo that Jonathan makes available on his GitHub.
And I pointed it towards the takeaways page, and I just said go. Now, in order to configure it, I did, of course, have to have some things in place. So DaVinci Resolve happens to be my default go-to editor anyway, which Jonathan also uses. 11 Labs, you may have heard of for generating voices. Now they also generate music. I've
Been a member of that since it was in beta, but I didn't have a paid account anymore. So I had to go like pay the right level to do that. There were a couple things that I ran into along the way that were specifically around some unique things about my computer there were some things that had to be worked out, but the end result.
Is that I now have on the Signals and Subtractions YouTube some walkthroughs There's also a private and unlisted one that's specifically for guests to be able to do the prep for this show with a two-minute video walkthrough that shows how to find your own signal and subtraction specific to guests. why everything on
Signals and subtractions is machine readable by default and the whole MCP setup and how all that works for the podcast, which is kind of a challenge for me to explain. It just takes time for me to explain. I just pointed this repo at the existing resources, the existing accounts, let it go. I made a few small edits to the script in each case, very minor, mostly for tone, and let it build in the video editor.
And in order to do that previously, as someone who's made probably thousands of video walkthroughs,
for all sorts of companies, internally, externally, marketing uses, all that. to be able to just assign the work, monitor the work, keep up with the work, but basically just let it happen and then review it when it was finished,
It allows me to make videos in places where I just simply wouldn't have before. It allows me to create value where I simply just would never have done it. Like this isn't taking jobs away. This is somebody who does jobs like this, like for a living. I don't want to do this kind of work. If I can have agents do this work, probably better than me. And to make an edit is just a simple.
Text edit and regeneration, as opposed to doing the voiceover all over again. It's really incredible. and it's just an illustration of the kinds of work like I opened with the show with, the kinds of work that used to take a team, that used to take a sizable budget, that used to just not be possible for someone like me to point this tool at this thing, let it run, come back.
Review, yes, good, post. Like that is all it took for me. And in about a half an hour, I made four videos, including the edits, and posted them to Signals and Subtractions to be able to assist anyone who wants to go there as a fan, as a guest, as a machine, even, to be able to have the resources that they need.
Just being able to think that way is fairly new to be able to do it at this level. And thank you again, Jonathan, for making that possible. it's been possible technically for a while, but to being able to integrate it and taking the little things like you're talking about and just understand enough about how they work across domains to stitch them together.
and have them like find their own errors and repair them. it's really a new way to think about what work is and how work happens, where, like you're saying earlier, every developer is now the team lead.
15:57 Limited Edition Jonathan
when it comes to the videos there's a mountain of things that are important about that. like it it is assembling all the little things, but also where my particular domain expertise came into with building that is like one thing I don't like about ninety-nine percent of video generation products out there is that
has nothing to do with deterministic or probabilistic or how they're generated and stuff like that. It is it's the same reason I don't like creating like a menu with AI like Chat GPT make you make me a menu for this and then well it's missing this and this. Can you make it like this? And then everything about it changes. I prefer using Photoshop and Lightroom and and professional tools where we have layers.
And timelines and you know, assets are separated. Like if I don't like the logo on something, I could remove the logo and replace it with the one I like or change how it works, change one element at a time or all the elements at once. so the way this works is it plugs directly into DaVinci Resolve so that you can do that and you can make those tweaks on your own. If the intro
16:56 Sam Rogers
Yeah. Yeah.
16:59 Limited Edition Jonathan
I noticed one thing that's bothering me is the intro to the videos are lasting for freaking ever. Like, you know, we're we're like
17:06 Limited Edition Jonathan
Yeah, we're we're like 15, 20 seconds into the video and it's still like telling you what the video is gonna be about. And I'm like, okay, let's get on with it. So like
17:11 Sam Rogers
Right. Yeah.
17:13 Limited Edition Jonathan
I can go into DaVinci Resolve and just manually do that myself. Or I can manually change the whole workflow. if you're doing client work, they want you to change little tiny things. And if you used AI to generate that video, you're totally screwed.
17:27 Sam Rogers
Yeah, so so well it depends on, of course, the approach to AI. So I I liken it back to the days of like vector graphics versus raster graphics, like if that helps people understand.
17:37 Limited Edition Jonathan
Yeah, yeah. It might help like four people understand, but yeah.
17:42 Sam Rogers
thinking in terms of composability and the the changes that you're making really being like code changes or algorithmic changes as opposed to
changing the where the pixels would be, you know, like to make one little change for a client to a logo or something like that. I've lived that nightmare of, the company just rebranded. We have to change everything for every internal training for every employee because they might forget what company they work for, or if we had a rebrand. Like, you know, somebody decided it was important. It wasn't really, but there goes a couple months of work.
Of finding
18:19 Limited Edition Jonathan
Yeah.
18:19 Sam Rogers
the individual sized graphic and placing it across thousands of other assets. When really, had it been built with composability in mind, it would just be one change, no different than like having one style sheet that everything looks to, and you want to change one color, you change it in one place, and everything cascades in the style sheet down. That same kind of
18:44 Limited Edition Jonathan
Did you notice
18:45 Sam Rogers
idea is what we're talking about.
18:46 Limited Edition Jonathan
did you notice some of the other opinions that I in injected in there, like the way I organize all my projects and bins inside of Resolve, the way I organize my external folder structure and stuff like that? That
18:57 Sam Rogers
I could tell I was dealing with a pro, yes.
19:00 Limited Edition Jonathan
it could just grep it real quick and it doesn't care. It doesn't care about the humans. And I wish it would care about us a little more.
19:08 Sam Rogers
It just has
to know how. It has to know what's important to the human. And then it absolutely will. But we're not used to making that explicit, right?
19:17 Limited Edition Jonathan
I wrote a post about this yesterday, called Send Messages, which is just this little tool Claude calls itself and it allows you to send messages
between sessions. it is this really small thing. But knowing about it allows you to do some interesting things. you could ask Claude to, hey, we're talking about this right now. Why don't you send a message explaining the decisions we made here over to the other session?
it enables a bunch more things.
You could have a bunch of agents working at the same time for you. And you could have agents with sub agents and all that kind of stuff. but sometimes the only thing stopping you from doing more is not knowing what the status of like if you have 10 threads open, not knowing what the status is, not knowing what kind
19:55 Sam Rogers
Yep. Yep.
19:57 Limited Edition Jonathan
of questions you have to answer, you have to go through each and every thread and make sure that, okay.
You know, this one's going good. Now let me check the next thread. And you have to read through all the traces and the answers and responses and all that kind of stuff. That was driving me crazy. Like absolutely insane. the only real thing that makes me visibly angry is having to look for something. That I I am
20:19 Sam Rogers
huh, yeah. Yeah.
Which one of these dozen threads was the one that's the thing that I'm thinking of right now? And I'm getting distracted
20:28 Limited Edition Jonathan
Yeah, yeah.
20:28 Sam Rogers
on thread number eight along the way to twelve, yeah.
20:31 Limited Edition Jonathan
yeah, and how many times i is it like, Okay, I had this idea, I know I sent it over to an agent, I know it's working on it, but damn it, which which session is it? Like
20:39 Sam Rogers
Yeah. Especially once you
start forking the sessions and all of that and the the labeling gets weird. I can totally relate,
20:46 Limited Edition Jonathan
I asked the one session, hey, look through all these other sessions, tell me what's going on, what the status is, anything I've missed, flag any conversation that I need to go and respond to. And it did.
then for the next like hour, I would keep going back to that thread and being like, All right, give me a status update because I have a lot of features to push, I have a lot of bugs to fix, I have a lot of random things that I need to take care of because I'm building this really extraordinarily complicated thing and I don't want one thread doing everything because that's just a way to destroy your life. And
21:11 Sam Rogers
Yeah. Right.
21:14 Limited Edition Jonathan
And I just started thinking, like, okay, well, we have send messages. This is really working. Claude does these live artifacts. so I asked him, I'm like, hey, you know what? Why don't you take over as the manager here? And by manager, I mean you don't do anything but communicate with me. and you communicate what I want to the other sessions, and don't you open up an artifact
And if you listen to this, you could develop it on your own. you don't need to follow my guide or use any of my stuff. You could just like follow the steps that I'm laying out here. And I'm like, why don't you just create a live artifact for me and update it every time we talk? and it did, and it was a really sloppy, AI slop looking job. And I was like, okay, well, listen, you have the context up here, and then you have questions down here.
Then you have more context. these questions need to have the context attached to them. So why don't you like each thread gets its own little section? Context questions right there.
why don't you put their suggestions into a copy paste in each one of the modules? over time I refined it a little bit, instead of just like copy the recommendations, I should be able to edit it and then copy it and then paste it.
22:10 Sam Rogers
Mm-hmm.
22:12 Limited Edition Jonathan
And then I thought, wait a second,
why don't you just append what the question is and what the session it belongs to is, and then my answer when I hit the copy button. and that's what it did. all of a sudden I realized I'm already working faster. And then I started closing tickets out faster. I started
you know, pushing PRs faster. Like I I'm talking about like three or four PRs an hour. And
22:32 Sam Rogers
Dang.
22:33 Limited Edition Jonathan
i insane. fully tested, implemented, merged on green is solid. so then I thought, well, I have a to do list
take a look at that to-do list and at the very bottom of this manager, I want you to go through and create starter prompts for each item in the to do list.
So that you can continue managing it. that brought me up to like five or six PRs an hour. granted, you can't just do this if you don't have a good idea If you're just trying to do one little task, it's kind of worthless. But if you're working in a big repo, you're a solo dev, even if you're a team dev, honestly, like you know, there there's the distinction between solo and team is becoming very
not there anymore.
23:11 Sam Rogers
Yeah.
It well it's becoming more fluid as as our definition of managing changes, you
23:18 Sam Rogers
Yeah, exactly.
23:20 Limited Edition Jonathan
but all of that started from understanding the function of the send message tool. And if I had not taken the time yesterday morning to really look into how it worked and spend twenty minutes on it, none of the other stuff would have happened.
Take the time to learn the little things. learn how something works.
23:38 Sam Rogers
your your pet peeve of looking for things, right? That's what he ain't doing.
23:42 Limited Edition Jonathan
Exactly. Yeah. Yeah. I
I have
hundreds of terabytes of storage of video projects from over the years. And probably 15 years ago, I learned to be hyper precise. never throw things onto my desktop, naming conventions, folder structure. I could find a clip from five years ago from a very specific client from a very specific shoot, like.
in a minute, just myself without having to like use any kind of search tools or anything like that. I can just click over to it I hate losing things. I hate having to find things I don't delete things either.
every photo, every frame of every video I've ever shot they're all organized
my agents don't delete. They they can't they're allowed to create new things, but they're not allowed to delete even what they created. they're allowed to version. You know, that's awesome. Like, you know,
24:25 Sam Rogers
Yeah. Yeah.
24:27 Limited Edition Jonathan
I want a new version. I don't want you to overwrite the old version. I want you to create a new version.
For example, three D models. I've been I've been experimenting and learning and building some skills for AI to make three D models for my printer. in fact Anthropic just shared some of my work on X and on Instagram.
they gave me early access to Sonnet five point five. I used it for like a week they loved it so much they decided to share it on their socials, which is super cool.
24:50 Sam Rogers
It is super cool and I will add that to the show notes as well so people can applaud your Instagram share by Claude.
24:57 Limited Edition Jonathan
The reason I brought that up is because each thing gets organized in a project. each project has internal versions for every asset because the first time rarely works very well. with Sonnet, with this one it did,
but yeah, my agents can't delete things.
25:12 Sam Rogers
Well, I think that that covers the subtraction element of the show really well. basically
delete the deleting. you're to a place where not only can agents not delete anything, which is of course a good safety measure that you can enforce with your architecture rather than just a request
It's a mistake a lot of people make is I'll just put it in the AGENTS MD file. You know, it's it's basically a request when you're doing that. You're asking for something to happen. And it's no different than with people. You can ask for anything you want, but that's not a guarantee that that's what's gonna happen. So there's architectural ways to control that, which is what Jonathan's referring to. But being able to just remove the concept of even deleting anything to work with composability as
the core fundamentals and then just version that along the way.
26:02 Limited Edition Jonathan
When it comes to working files, you should always version. Period. And storage should not matter. If you get to the point where you're filling up your storage, well, now it's time to get more storage.
26:10 Sam Rogers
Building up from the fundamentals, as Jonathan's been talking about, you know, throughout this whole episode, your knowledge of the little things and how it is that they actually work. Not just buying a tool off the shelf and doing this because somebody said so, but your knowledge of how it is that the systems that you rely on actually operate really does make a difference in how it is that you're able to scale.
26:30 Limited Edition Jonathan
Yeah. my God.
you'd be surprised how many little things that you'll learn along the way that make it so that you don't have to buy the product off the shelf anymore. it's all about the foundations, the primitives, the little things that build up and stack up to become important things
26:48 Sam Rogers
as we come to a close, for everyone who's listening, one thing to do this week, pick one handoff in your organization, a place where work passes from one specialist to the next. And write down what that handoff protects. Then ask whether the person with one capable model.
Could do the whole chain. And if they could, where that would live? Like who would check that? How can you think about that? What is actually worth protecting now? And I hope that you found today's episode to be inspiring for how to think about that, where to start. for today, that's our show. Thank you so much, Jonathan. I see a finger. I see a finger. There's a finger.
27:36 Limited Edition Jonathan
Just got an email. Google's finally embracing skills. Finally.
27:44 Limited Edition Jonathan
Ha ha ha.
27:45 Sam Rogers
skills, which have now been around for roughly a year, agentic skills, as skills are coming, now GPTs are going away, gems are going away, all of these other artifacts that we've had for longer as ways to approach how to work at a higher level with AI, we're iterating, we're finding our way forward together.
the giants on Mount Olympus will fight their fights on what standard is what for payments and for processes. but you and I here on the ground level can pay attention to our signals and what it is that we subtract in order to lighten the load as we all proceed exponentially up the curve together. thanks so much, Jonathan, for coming back to the show. you are most welcome.
to come back anytime
28:33 Limited Edition Jonathan
Before we went live, Sam was like, you and I can go and talk about just about anything And he was telling me how he wanted to keep this kind of constrained of maybe maybe two topics, three most. and
28:48 Limited Edition Jonathan
I just came up with a one liner to end the show with, you know
28:50 Sam Rogers
Please, please.
28:51 Limited Edition Jonathan
while you're subtracting things, subtract a few dollars from your bank account, just subscribe to both signals and subtractions and limited edition Jonathan on Substack. There you go. How's that for a one
29:03 Sam Rogers
Yes.
29:03 Limited Edition Jonathan
liner?
29:04 Sam Rogers
Yes, that's good. Thanks for bringing it back to Signals and Subtractions on Substack. new episode out every Friday, new newsletter every Sunday. please go to the podcast app and click follow so that they just magically show up for you. Sunday will be issue 71 of the newsletter.
One signal, one subtraction, one analogy. It's about five minutes to read. And everything on the show, including the takeaways that we discussed in the new videos, are at sigsub.show. One signal worth watching, one subtraction worth making. Now go find yours.