Volume II · Episode 1 · Transcript
Context Over Capability
Full transcript of the recorded conversation.
Back to the episode · timestamps jump to the video · lightly machine-transcribed, may contain errors
0:11 Sam Rogers
Welcome to Signals and Subtractions Episode 1. I'm Sam Rogers, here with Lee Rodriguez from R&P Associates. Lee, would you tell the nice people of the internet about how and when we first met?
0:25 Lee Rodrigues
Well, the time was 2012, San Bruno headquarters at YouTube. I got brought in after having a comfy job at Apple teaching everyone how to edit video, being asked, can you come do instructional design video editing? And on my way in, one of the people I was working with said, By the way, they've hired five consultants in four weeks and fired every single one of Good luck.
And then I met Sam and we talked about ten minutes about scripting and stories and realized I think together we can do this. I think we have the right people here to get this done. And we did.
0:58 Sam Rogers
Yes, and since then we've kept in touch.
as we've both worked at the intersection of learning and performance, multimedia, IT, change management, both inside large organizations and startups. And on this show, we'll typically have a guest, but today being the very first one, it's just the two of us. I'm really glad that you're here.
with me, Lee.
1:25 Lee Rodrigues
I'm glad we can do this together. It seems like old days,
1:28 Sam Rogers
Yep. So the format for this show is one signal each, one subtraction each, some stories, hopefully, to make it stick. And before we get to all of that stuff, I think it's time for the current state of AI.
1:52 Sam Rogers
Yeah, so you might have heard about SpaceX in the news. XAI and their Colossus data centers are a core part of that. they're the ones supplying all the compute for the frontier models. Anthropic is racing to IPO. OpenAI just said that they're backing off for a moment, probably because there's literally not enough money in the world to support.
all three of these valuations all at the same time.
2:19 Lee Rodrigues
It's a lot. let's talk about the models withheld.
2:21 Sam Rogers
so that's the big news just within the last two weeks. Fable Five is actually released
Again today, July 1st. it's been released, it's been revoked, it's been a rolling restore. meanwhile, also on the open AI side, ChatGPT and their 5.6 model, it's In preview mode. It's in the same kind of mythos mess. These are the two leading AI labs that are basically receiving instructions from the government that say, yeah, you give it to us first, and then a staged release to certain approved companies, approved people, the US government is now receiving the frontier models and approving their.
Use cases.
3:08 Lee Rodrigues
Let's see that headline again.
3:10 Sam Rogers
regulations are being postponed because like the EU AI Act, which was supposed to be going into effect on August second of this year, parts of it are still going into effect, but most of it is not. It's been delayed sixteen months. So as somebody who is developing products to to help relieve some of the regulatory strain there.
oops, 16-month error. there's a lot of executive order stuff happening and there's this whole voluntary system that came out. Basically, the federal regulations that everyone agrees we should have aren't coming. The state regulations, such as just happened in Colorado, are getting gutted. And the EU AI Act is trying to figure out how it works.
3:59 Lee Rodrigues
Well what about the agent fleets that are eating software as a service?
4:02 Sam Rogers
AI agents are everywhere, and they're coming for your software. just yesterday, part of Sonnet's release was optimized for fanning out agents in this advisor pattern, which is where they take a a less intelligent model and it pops out to a more intelligent model to maybe answer some questions, get some direction, and then the bulk of the work happens on lower model. For those of us who are deep in AI, we've been doing stuff like this for months. This is how I've been doing translation.
for a long time, actually with local models popping out to more intelligent ones. But this is actually getting integrated as a pattern now into fleets of agents that are working inside Enterprise. And that is new. Also
the open weights models, the ones that you get from China or something like that, the the AI that you actually run on your device without touching the internet, those tend to be about two-quarters behind the frontier stuff. So this thing that the government has to approve, that's the 10 trillion parameter models that are coming out now.
from the Frontier Labs. They can gate it for now, but those level of models are coming by the end of this year free and open source.
Unless something huge changes, it's likely to actually catch up even more. So for all those regulated industries, things that like can't rely on frontier models to absorb all that context about their business, being able to run that completely locally is possible with the right hardware. And it is eating software.
5:45 Sam Rogers
between February and June
There's been an average of a new state-of-the-art frontier model every 11 days. I'm talking about those big models. So the ChatGPT 5.6 that was just announced last Friday, their sole like huge one, Fable 5, which
we
were talking about before, which came out three weeks ago and then kind of uncame out, but is now coming back out again, that level of model. And two weeks before that, like at the end of May, there was Opus 4.8,
These things are pushing the frontier, and something completely new coming out largely at this point, for the last quarter, just from those two labs. But previous to that we were seeing things from Google, seeing things from X as well, that were kind of pushing the frontier in new ways. The world changes. What AI can do changes.
Every two weeks or less. It's averaging every 11 days. And this is the first day of Q3 I'm guessing that the only thing that we really are sure of with what's coming in Q3 is that it won't look like Q2. It's probably gonna be 10 days. It's probably gonna be faster. You can throttle down some of the some of the releases and you can stage them, but the pace
of progress is happening.
Perhaps we could go to the signals segment here.
7:14 Lee Rodrigues
the signal for me has been finding a way to get really back to the basics. And what I mean by that, I started like developing in scripts and stuff for like a course I'm gonna produce and I would feed it into Claude and have Claude create a seven page document of how this works.
And what I found is I was getting lost in the sauce with all the formatting and all these fields and everything. And you can't remember how much of this is me, how much did it come up with, how much relates to what I do. And the big thing that changed for me was when I used to train designers, start with a one-page content outline. One page outline that clearly defines what we're talking about.
It's like, well, got all these ideas. I don't want to see a course. I don't want to see a video script. I don't want to see an exam. I want to see a one-page content outline to make sure this stuff flows and all connects with each other so subject matter experts can understand it and everything. When we get into the big complicated documents, it can be tough for everyone to navigate. And I think that over polish makes it extremely difficult.
said, hey, this is a beautiful seven-page document that I'm already lost in. Give me a one-page content outline. And going back to the basics without the polish and looking at this one page that we used to use back in the day made a whole lot of sense. That structure right there let me see where Claude's making stuff up.
It's finding things that were not a part of my content that it's bringing in from somewhere else, and I can just trim to subtract what isn't a part of it, and then I I get what I needed. So the idea is you have to get back to the basics. You have to get back to the basics because it will put it in a format that is overcomplicated. And then you end up sharing it with a client or something, and they say, Where did this come from? And then you all of a sudden you sound like a child in high school.
8:53 Sam Rogers
Yeah.
9:08 Lee Rodrigues
Reading a book report for a book they clearly have not read, which is not a great place to be. How about you? Which what's your signal?
9:11 Sam Rogers
Yeah.
well for me lately it's it's similar. So as all these models get more and more and more powerful, I'm realizing how much more it's about context.
overcapability. And it it's not that different than for people, like like so many things. This works for people as well as AI. We tend to trust when someone gets us, right? Like when they reflect back the things that are relevant to what we care about, the things that we think are obvious. You know, it's not complicated. The intelligence coming from these huge models is astonishing.
but it kind of doesn't matter how astonishing it is if the context goes wonky. Building trust up from the basics. And i it works so much better than like building it down from something very esoteric and difficult to understand, something really challenging. just what you're sharing there, Lee, about the
The one page version, like the simplified version. Remove all flashy sparkles. What's the actual thing that we're doing here that you can't bluff? your book report example. Like, you don't need to memorize what's on, you know, page 82, but you need to be able to say, like, who the characters are and why they're there in the story, you know, things like that.
again just back to people. Like you could take a genius level person and put behind fry machine at McDonald's and they'll be confused because
These two, they don't have the context for that. Are they smart enough? Absolutely. Will they burn themselves? Probably. Like will they burn your business? Absolutely. So having that kind of awareness and ability to communicate context with AI, I think at this point, is becoming more important than the intelligence itself. We've gotten there.
11:03 Lee Rodrigues
Likely.
Hundred percent.
Well, that brings up a really interesting story. when we started it, Google, I was auditing courses and I found that they had a course about how to write better emails because so much of communication was in emails. And their story was content without context is very difficult to absorb. And your job as the writer of the email is to provide the context, how this relates to your team.
what you're working on, how that ties to this, because everybody has an email from someone else's team that's working on something that has nothing to do with us and you're asking for help. But if there's context, wait a minute, you're trying to solve the same problem I am from a different angle. And if I help you with this, you may just chip away hours of development work on my side. We're friends now. Let me help you with your project, right?
In Ford, they got a bunch of AI agents to troubleshoot designs and to and to simplify and put designs out. And what happened is they laid off a whole bunch of engineers to get this code base and these AI tools to make all this better. And what just recently happened is they're now hiring back the gray beard engineers. There was a bunch of older engineers, and they said what they have is the context, none of the code does, because they're getting ready to launch this car, and one of the engineers comes over and says, Hey, I don't want to be a dick or anything, but
12:36 Lee Rodrigues
You release this turbocharger in a car like five years ago and it overheated disastrously and melted the car and everything. It needs a much larger intercooler to work, and that intercooler needs a larger radiator to transfer all the fluid around. You all are trying to compact this down and it's gonna result in overheating. I know that the AI says that this is a really good turbocharger. All the specs are good, the costs are good. Yeah, but it runs hot. It always ran hot, and you're talking about a vehicle that's going to tow.
What that's gonna mean is someone on the side of the road with a trailer with a with a radiator boiling all over the place, w under warranty, calling saying, Guess who doesn't want this truck anymore? And they learn that from the original Ford F-150s with double turbochargers, but no one remembered all the service calls. And the gray gray beard engineers are like, We run calls with dealerships trying to resolve this overheating issue when this thing came out. I remember those calls. AI wasn't on phone calls.
that experience is now being called like the gray beard engineers, the gray beard army. Not that they're going to replace all the AI, but that content needs the context of an experienced engineer with a few great hairs to look over their shoulder and go, let me give you an article you're missing.
That is probably in paper form from dealerships. But if you integrate this into the AI, the AI goes, Holy crap, we got a heating problem. No, you've always had a heating problem. You just didn't have that context of how it applies because you're thinking about performance. And by the way, the horsepower, the fuel consumption, all those things is good. But climbing over I-80 at 101 degree temperature, towing a trailer at 65 miles an hour is gonna overheat that vehicle.
14:11 Sam Rogers
it's no surprise really for those of us who have some gray, as as you do, Lee. If if this grows in it gets a little gray for me too. But yes, exactly. so we've been through some tech disruptions over the last
14:19 Lee Rodrigues
There's sparkles of wisdom.
14:27 Sam Rogers
several decades, and every time it's pretty much the same thing, which is we overinvest in the technology and we underinvest in the people.
And then as if we didn't make that mistake just five years ago. We go, oops, I guess we should bring some of them back. Like we overcompensated. especially the human context, the context that has not yet been made machine readable. we have to be able to point.
AI to that context for it to even have any chance of understanding what it is that we're trying to do.
so yeah, we're just doing the same thing that we always do, and it'll probably go similarly of cutting a little bit too deep and then oops, I guess we need to bring some actual people back who know what they're doing.
15:16 Lee Rodrigues
And
and just a quick plug, what inspired me was I was reading your article on LinkedIn vocabulary debt. And that really ties into this because you have the AI's using a different term because we called it three different things as we were developing this product. And when we identified that the platinum turbocharger is what we called it as we released it, before it was called something else, it had major heat issues.
15:22 Sam Rogers
Yeah.
15:43 Lee Rodrigues
hey, wait a minute. There's four other names for this. Call it these four things. You're gonna pull up the trouble tickets, the cases, the support issues. That context will fold in and you'll say, wait a minute, there's a couple more things attached to this.
15:54 Sam Rogers
the platinum example is a great one because what I was saying in that signals and subtractions newsletter from last week is platinum can mean more than one thing in different parts of the company. So the marketing, they're calling it the platinum one. If you say marketing platinum, it means this. But if you say engineering platinum, they're thinking it's got platinum in it? Like y
16:17 Lee Rodrigues
Got platinum in it? That that's an expensive metal.
That's unique. Where are we going with this?
16:20 Sam Rogers
And and yeah, exactly. That that's I where are we going here? Like that
being able to differentiate those things and that use terminology that is actually AI friendly. When we talk about making things machine readable, that's a core part of it is how does that
ontology breakdown of when to use which context. So it's not just like all the context all the time, it's the slice of context that you need so that within this team, when they say this, that's what that means. And over here on this team, even though it's the same word, it's going to have the same mappings in its understanding of how it is that it's you know predicting the next token, it's going to use the same platinum
to mean those different things, but it's gotta like correlate the trajectory of that in its whole map of everything. And we can make that easy and we can make it hard. So the easy way tends to work better, but it's more work up front that we're used to humans being able to compensate for. And even though it makes sense to humans too, we've just never had to go to the trouble.
You know, and now we kinda have to go to the trouble.
How about a word from our sponsor? What do you say? I I put some flashy sparkles in there just because I figured we would be talking about all the flashy sparkles. you may be familiar with the in this sponsor.
18:52 Sam Rogers
For you, yeah, I'll I'll hook ya up. No no worries. So so signals done. now on to the subtractions. What did we each decide to stop, to kill, to refuse to engage with this week and why? Take it late.
18:57 Lee Rodrigues
I love it.
My subtraction is exactly what we were just talking about. It's subtracting the fluff whenever possible, particularly for me, my design documents, my scripting, getting to a simple outline, removing all the fluff to make sure we're on the same page, and it's one page. To quote one of my favorites writers in the world, Tim Ferris, if you can't fit it on one page and understand it, you probably don't understand it.
And getting it down to one page, what I learned in grad school a 250-word paper? That's a challenge, man. Every word has to be doing something, and it has to be super simple and to the point. If you drift it all in 250 words, it's all gonna fall apart. So getting your AI message, getting the things you're working on down to a one page content outline, subtract the formatting, subtract the bars, subtract the just in case I need the context.
The content and the flow. Subtract the unnecessary.
20:07 Sam Rogers
Sounds great. And and while you're at it, you might as well d do that in what format of document? What format are you using?
20:15 Lee Rodrigues
T X T or Mark Down
20:16 Sam Rogers
Hey, hey, there we go. Now we're talking. yeah, do something that's token efficient, that's the lingua franca of all the LLMs that were trained on text. Just use text. Text is different than a Word document. So for my subtraction, it's to stop reaching for a better tool to fix the context,
and subtract down to what actually makes it reliable. So example.
I I have a whole lot of context management that I'm constantly doing anyway, and quite a few skills, agentic skills, I've been making or modifying an average of one a day since February. So like there's over a hundred of them. I've got them all in their own monorepo. I'm like coordinating across multiple LLMs to use them. but I'm realizing as these models get better and better, there's actually less use for all those custom skills.
that I used to need. But with A B testing, the results, I'm realizing I sometimes get as good or better responses from a more better informed model with the context that I'm giving it as opposed to the skill and the context that's in that.
So so I've actually been removing a number of skills that I developed several months ago that were very useful. But as soon as it becomes like almost about as useful, kill it.
Because it's just not worth the maintenance. It's not worth the the choices, the complexity. So I'm down from around 100 skills to around like 40 now. I've cut over half of them because they're I I gotta admit, like they aren't always helping. So I'm archiving way more than I ever thought I would. and I'm realizing in the process that most AI assets are very short-lived.
Like back to our news segment, when when you've got the new state of the art model every eleven days that's changing what is possible with AI, how many iterations, how many cycles of that are gonna make sense for your existing infrastructure and architecture? Like I'm expecting I will probably keep things for maybe five of those cycles? Well that's only like two months. So a skill I built in February.
is like already outdated. So that's my subtraction. Just keep track of what it is that you're building for AI. And if you still need it, 'cause you might not.
23:01 Lee Rodrigues
Amen. That that's the same for every saved editing file you have for a video. It's like how many times am I gonna revise this? How many versions do I need? Maybe the current one and get rid of all the archive, you know? But what are you keeping it for?
23:15 Sam Rogers
Yeah, exactly.
I believe we're coming to the end of our episode. For those who are enjoying this thus far, I just want to direct you to Sigsub.show, where you can sign up to see future episodes, be notified as those come out. There will be additional co-hosts that are rotating in, and we will be doing this as a live live stream. That didn't quite happen for
for
the very first episode, but it will happen for the next one. We will also have guests in addition to co-hosts. Looking forward to seeing you there and all of your comments and questions. Lee, do you have a something good to to wrap it all up?
23:59 Lee Rodrigues
Without an authentic story, we're magnifying and multiplying unnecessary content, which is a lot of what your AI can do. It can just totally extrapolate on nonsense. So getting that authentic story, that authentic business problem, the thing we're trying to do, wrap the context around it.
Turns out that works pretty good.
24:20 Sam Rogers
it's been great sharing this with you, Lee. and thanks for playing along and being the
very first co-host on the very first episode of this signals and subtractions thing which I've been doing for like a year now, over a year, every week. I'm really excited about this new format and being able to take it to an even bigger audience.
24:45 Lee Rodrigues
Happy to support you any way I can, Sam. Love working with you.
24:48 Sam Rogers
So where would people find you if they wanted to find Mr. Lee Rodriguez?
24:53 Lee Rodrigues
I I work with RP Associates. That's my wife and I's consulting company. But the new thing that's coming out that'll be coming out in the next couple of weeks here is system2focus.com. It is a smarter approach for a serious search. It's all about using all the tools available to you to find that authentic part of you, to interview better, to informational interview better, to make better resumes so you don't waste your time sending 30 applications a week that get no response.
Nothing whatso you feel like you got some work done, but you accomplished absolutely nothing. If you're ready to break that cycle, systemtofocus.com will have a link in the description.
25:31 Sam Rogers
You know, we should do a a future episode on job search in general. maybe you could find a suitable guest or something for that. let's bring on a recruiter or somebody who's used to working with all the the modern AI tool sets for doing that. You and I have been hiring managers, we've been through thousands and thousands of resumes, but it'd be good to bring on some some other external expertise. The job
25:38 Lee Rodrigues
Absolutely.
25:53 Sam Rogers
thing is quite a thing these days. I think that'd be great to explore in a future episode. Cool. Well thanks again everybody for tuning in to Signals and Subtractions.
25:59 Lee Rodrigues
I love it. Let's do it.