WEBVTT
00:00:01.760 --> 00:00:13.679
The saying of keeping up with AI is like trying to drink water from a fire hose, though it's tired and cliche is obviously 100% true.
00:00:13.919 --> 00:00:20.000
I mean, with all the nonstop updates, your head is probably spinning trying to keep up.
00:00:20.239 --> 00:00:25.120
So you're thinking, with all these new releases, how do I use ChatGPT?
00:00:25.199 --> 00:00:29.280
Or is using Claude kind of the same as using Gemini on the desktop?
00:00:29.440 --> 00:00:33.679
Or can I prompt Copilot kind of like I'd prompt Codex?
00:00:34.079 --> 00:00:42.240
Or maybe you've had to start the race sprinting and you never really got the proper 101 on how all of these AI chatbots work.
00:00:42.479 --> 00:00:48.159
Regardless, you definitely aren't alone in the fire hose updates drowning you out.
00:00:48.479 --> 00:00:54.960
That's the new struggle for every enterprise, because it's not like there's a cheat code, right?
00:00:55.359 --> 00:01:02.320
Because as the AI models are changing almost daily, so too does the input required to get the best output.
00:01:02.399 --> 00:01:03.280
And I get it.
00:01:03.520 --> 00:01:09.840
As someone that covers AI daily, I understand the struggle of trying to play a game, yet the field dimensions change without notice.
00:01:10.000 --> 00:01:16.560
And you were using a softball yesterday, but cricket gear today and tomorrow you might kind of play it out like rugby.
00:01:16.719 --> 00:01:20.560
The input rules and the output capabilities are moving targets.
00:01:20.799 --> 00:01:25.040
But there's one thing I've found out over the last three or four months.
00:01:25.280 --> 00:01:31.200
The big players have all just kind of started to copy each other, which is actually a good thing for you.
00:01:31.439 --> 00:01:42.719
So I think it's actually been over the last two months that we've finally stumbled on a set of best practices for getting the best outputs out of any large language model.
00:01:42.959 --> 00:01:53.840
So that means, well, there is a cheat code, or at least a set of somewhat concrete rules that, when followed, will give you stellar outputs, really, no matter what model you're using.
00:01:54.000 --> 00:01:58.319
And that's exactly what we're going to be giving you today on Everyday AI.
00:01:58.480 --> 00:02:01.200
Welcome to the Start Here series.
00:02:01.359 --> 00:02:10.879
Uh, if you're new here, the Start Here series is your essential guide to getting caught up and getting ahead with AI.
00:02:10.960 --> 00:02:14.479
But first, let's talk about the big picture.
00:02:15.360 --> 00:02:20.719
So the capabilities of these models are very, very real.
00:02:21.199 --> 00:02:23.840
And it is hard to keep up.
00:02:24.000 --> 00:02:24.400
Right.
00:02:24.560 --> 00:02:28.319
So right now, you probably see examples all over the place.
00:02:28.479 --> 00:02:36.560
You see people sharing examples, whether it's online, whether it's media articles that are so poorly written, and you're like, well, these AI models are really bad.
00:02:36.719 --> 00:02:37.680
Look at all these mistakes.
00:02:37.759 --> 00:02:53.120
But then you also see these benchmarks and these tests, and you know, these big companies laying people off in lieu of spending billions of dollars on AI, and you're confused because you can't even know or don't even really understand what button to click or which model should I use.
00:02:53.280 --> 00:02:58.159
And up until recently, it's been too hard to follow because there's been too many different paths.
00:02:58.479 --> 00:03:02.879
And now I think essentially we have cookie cutters, right?
00:03:02.960 --> 00:03:08.560
We have the McMansions of models because they're all kind of the same.
00:03:08.719 --> 00:03:12.719
Obviously, the capabilities and the harnessing of the tool is all completely different and unique.
00:03:12.960 --> 00:03:28.719
But I think for the first time, maybe ever, at least when it comes to you know, the three to five handful of big AI players, we have a set of rules that can apply unilaterally to really any of them.
00:03:29.039 --> 00:03:33.120
So that's what we're gonna be going over on today's show.
00:03:33.280 --> 00:03:34.000
Stick with me.
00:03:34.159 --> 00:03:36.719
This one, trust me, is gonna be 30 minutes.
00:03:36.879 --> 00:03:37.199
All right.
00:03:37.280 --> 00:03:42.159
I know sometimes I say 25 to 30, and then you're like 45 minutes in, but stick with me on today's show.
00:03:42.240 --> 00:03:43.199
Here's what you're gonna learn.
00:03:43.280 --> 00:03:48.800
You're gonna know why the cookie cutter nature of today's large language models is actually both a good and a bad thing.
00:03:48.960 --> 00:04:08.479
You're gonna know the up-to-date capabilities, features, pros, and cons of the big players, specifically focusing on Chat GPT, Claude, Gemini, technically applicable, the co-pilot, but we're not gonna go into that one as deep, as well as I mean, you can apply it to perplexity, grok, open models, etc., because they are kind of all getting the same.
00:04:08.560 --> 00:04:14.479
And you're gonna leave today's show with the 10 essential steps to get the most out of any AI chatbot.
00:04:14.639 --> 00:04:17.439
All right, this is the Start Here Series.
00:04:17.600 --> 00:04:23.360
This is the essential podcast series to both learn the AI basics and to double down on your AI knowledge.
00:04:23.519 --> 00:04:32.000
Because after doing this for three and a half years, nearly almost 800 episodes now, I never had a good answer when people ask, Jordan, I'm new to the podcast.
00:04:32.079 --> 00:04:32.720
Where do I start?
00:04:32.879 --> 00:04:37.040
Well, you start here with the start here series and you go to startheseries.com.
00:04:37.120 --> 00:04:40.399
That is going to give you free access to our inner circle community.
00:04:40.560 --> 00:04:41.279
It's exclusive.
00:04:41.360 --> 00:04:48.560
This is the only way you can actually get access in the Start Here series uh space inside of our free community.
00:04:48.720 --> 00:04:54.079
Uh, you can go find every single Start Here Series episode all in one easy-to-find place.
00:04:54.240 --> 00:04:58.399
Uh, a Spotify playlist that's updated, all the newsletters, go read it.
00:04:58.560 --> 00:04:59.600
It's awesome.
00:04:59.839 --> 00:05:00.160
All right.
00:05:00.319 --> 00:05:10.399
So uh if you miss our last Start Here Series episode, we covered Bill by Partner or Wait, the four-layered AI stack decision framework for 2026.
00:05:10.560 --> 00:05:17.839
And today we are giving you the LLM cheat code 10 essential steps to get the most out of any AI chatbot.
00:05:18.319 --> 00:05:18.879
All right.
00:05:19.040 --> 00:05:22.079
Um, and I'm just gonna give you those steps now.
00:05:22.160 --> 00:05:26.800
And we're obviously going to break them down as we go along.
00:05:27.040 --> 00:05:28.560
But here they are.
00:05:28.959 --> 00:05:34.319
Um kind of steps, best practices, rules to live by, whatever.
00:05:34.480 --> 00:05:38.879
But number one, you have to understand what a large language model is and what it is not.
00:05:39.120 --> 00:05:50.160
Number two, you have to choose your AI operating system, like ChatGPT, Claude, Gemini, or Copilot, and really stick to those as much as you can for the majority of your day-to-day knowledge work.
00:05:50.319 --> 00:05:52.959
Number three, you have to choose the right surface.
00:05:53.040 --> 00:05:57.279
And the surfaces are unfortunately changing, but I think it's actually for the good.
00:05:57.519 --> 00:06:01.519
Number four, you have to choose the right account plan and model.
00:06:01.680 --> 00:06:03.759
I'm gonna break down those best practices as well.
00:06:03.920 --> 00:06:07.040
Number five, you have to understand the context layer.
00:06:07.120 --> 00:06:12.240
First, you have to understand it before you get to number six, context engineering basics.
00:06:12.319 --> 00:06:20.800
I'm gonna tell you prime prompt polish, refine queue, five, five, five, a lot of the secrets that I've kind of held close to the best throughout the years.
00:06:20.959 --> 00:06:30.560
Uh, number seven, I'm gonna tell you how and why, and risks on working with files, apps, and company data and why that matters.
00:06:30.720 --> 00:06:38.079
Number eight, we're gonna talk about privacy, permissions, and governance and how that actually plays out in these AI operating systems.
00:06:38.240 --> 00:06:42.079
Yeah, I call AI chatbots now AI operating system because that's what they are.
00:06:42.319 --> 00:06:46.399
Number nine, transparency, observability, and reasoning artifacts.
00:06:46.480 --> 00:06:51.759
I'm gonna show you how to use those things to your advantage to get better outputs.
00:06:51.920 --> 00:06:56.639
And last but not least, verification, iteration, and workflow design.
00:06:56.800 --> 00:06:57.439
There you go.
00:06:57.600 --> 00:07:04.240
You can stop here if you want, but trust me, this is like thousands of hours of conversations over the past three years.
00:07:04.319 --> 00:07:09.360
And I'm gonna do my dangedest uh to get them to you in 30 minutes or less.
00:07:09.600 --> 00:07:15.519
So uh here's what I want to talk about, why this is both a good and a bad thing.
00:07:15.680 --> 00:07:19.680
And I think this really became evident when Google at their I.O.
00:07:19.839 --> 00:07:23.600
conference, uh uh like last week, and I know, right?
00:07:23.759 --> 00:07:25.839
So right now it's the end of May 2026.
00:07:25.920 --> 00:07:31.680
I should put this out there because I know you know a lot of you might be listening to this in you know, July or December.
00:07:31.839 --> 00:07:37.680
So obviously, some of the things are gonna change here when I'm talking about certain surfaces, models, modes, etc.
00:07:37.920 --> 00:07:44.560
But hopefully these concrete-ish uh 10 steps will still hold.
00:07:44.800 --> 00:07:48.800
But I noticed at Google's I.O.
00:07:48.959 --> 00:07:52.959
conference when they announced their new anti-gravity 2.0 desktop app, right?
00:07:53.120 --> 00:08:01.680
And not gonna get into it, but I'm like, wait, this just looks like codex, it functions like codex, and then I'm like, wait, cursor kind of does too.
00:08:01.759 --> 00:08:12.000
And that's when I started to notice that for the most part, especially compared to 18 months ago, when all of the different, whether we're talking about the web interfaces, right?
00:08:12.160 --> 00:08:19.360
Gemini.google.com, claude.ai, chat gpt.com, grok.com, whatever you're using, 18 months ago, they all looked really different.
00:08:19.439 --> 00:08:21.600
It was almost like speaking different languages.
00:08:21.759 --> 00:08:25.360
Now it's all speaking the same language with slightly different dialects, right?
00:08:25.439 --> 00:08:31.439
Like anti-gravity literally looks like it is codex light, right?
00:08:32.000 --> 00:08:35.200
Color schemes, fonts all look the same, layouts.
00:08:35.360 --> 00:08:36.799
So it's good and bad.
00:08:37.039 --> 00:08:45.039
From a bad perspective, it might seem to you, the non-technical user, that the front-facing innovation is kind of slow, right?
00:08:45.200 --> 00:08:56.159
As an example, go back to you know, 2024, front-end innovation for um users and maybe early in 2025 was everywhere, right?
00:08:56.240 --> 00:09:12.000
All of a sudden we had artifacts from Claude, and then we had, you know, Canvas mode, and you know, we had these agents that worked inside, and all of a sudden, you know, large language models went from simple, you know, next token prediction transformers to models that could think and reason and call tools, right?
00:09:12.080 --> 00:09:19.120
So I think there is this period of fast innovation in terms of what buttons you would click, all of these modes, deep research, et cetera.
00:09:19.279 --> 00:09:22.720
But if I'm being honest, in 2026, we haven't seen that same thing.
00:09:22.799 --> 00:09:42.720
And I think that's actually good because what happened is it, I think it gave us number one, now we can kind of apply these best practice rules, but more than uh anything else, I think it allows enterprises to be able to modular modularly uh pivot as needed, right?
00:09:42.879 --> 00:09:53.600
Because 18 months ago, if you were a heavy chat GPT Teams user back when it was called before it turned into Chat GPT business, and then you try to go to, you know, uh Gemini, right?
00:09:53.759 --> 00:09:57.440
There was no Gemini business at the time, but it was just Gemini, you'd be confused.
00:09:57.600 --> 00:10:02.879
But now they've kind of all copied each other with certain features, projects, GPTs, gems, right?
00:10:02.960 --> 00:10:05.200
They're all kind of just the same thing now.
00:10:05.360 --> 00:10:11.600
Uh but the actual race is smarter, faster models with autonomous desktop harnesses.
00:10:11.679 --> 00:10:13.279
So the surface is changing a little bit.
00:10:13.360 --> 00:10:21.360
But ultimately, the thing that you have to keep in mind if you don't get anything else from today's show, AI models are smarter than all of us.
00:10:21.440 --> 00:10:23.519
If you're using them the right way, and that's a big if.
00:10:23.600 --> 00:10:26.879
And when I talk about jagged, jagged capabilities, right?
00:10:27.120 --> 00:10:30.480
People will share something where it's like, oh, AI models are dumb.
00:10:30.559 --> 00:10:34.879
And then you have this false sense of security, like, oh, I don't have to worry about keeping up with these day-to-day.
00:10:35.120 --> 00:10:35.679
Yes, you do.
00:10:35.919 --> 00:10:39.039
Anyone that shares those things, that's a skill issue.
00:10:39.200 --> 00:10:43.840
You know, uh, send them this video and then they'll be like, oh, I've been using everything wrong.
00:10:44.000 --> 00:10:51.840
Um, but AI models by default, studies show this, even judged blindly by experts.
00:10:52.320 --> 00:10:58.720
Today's AI models, with all of this harnessing built around them, they can produce artifacts one shot, right?
00:10:58.879 --> 00:11:05.840
Uh, you know, complete websites, complete apps, uh, spreadsheets, PowerPoints, Word docs, PDFs, etc.
00:11:06.080 --> 00:11:06.639
Uh, right.
00:11:06.720 --> 00:11:09.440
So they're no longer just a little chat bot to talk to.
00:11:09.519 --> 00:11:13.600
And we've covered that uh in depth in the Start Here series uh so far.
00:11:13.759 --> 00:11:15.519
So let's dive in.
00:11:15.600 --> 00:11:25.919
We have some upgraded visuals uh today, but let's get into our uh kind of the 10 new rules of work, I'll say.
00:11:26.720 --> 00:11:32.240
So number one, what is a large language model and what is it not?
00:11:32.639 --> 00:11:46.159
Well, if you're only using Chad GPT or Claude or whatever as a version of Google, as a smarter, faster, better source version of Google, you're missing out, right?
00:11:46.320 --> 00:11:52.000
If you're just treating it back and forth like talking to a smart friend, you're also missing out.
00:11:52.159 --> 00:11:55.759
So by default, large language models are generative.
00:11:56.000 --> 00:12:03.679
So that means they are non-deterministic, where a search engine like Google, aside from localization and personalization, is deterministic, right?
00:12:03.840 --> 00:12:19.360
If we put in the exact same query, chances are, aside from that personalization and localization, but if we all did an incognito Google search, right, uh five years ago before AI overviews in AI mode, we for the most part all would have gotten the exact same outputs, right?
00:12:19.440 --> 00:12:26.240
Even if there's, you know, 20,000 people listening to this episode, and we all did it, we would for the most part get the exact same results.
00:12:26.480 --> 00:12:28.080
It's not the same with generative AI.
00:12:28.159 --> 00:12:28.960
It is generative.
00:12:29.120 --> 00:12:33.200
There is an element of uh next token prediction.
00:12:33.279 --> 00:12:46.639
And uh essentially, large language models have been trained on the entirety of the internet, copyrighted works, uh, you know, open source coding projects, videos, photos, everything, uh, just terabytes of data.
00:12:46.799 --> 00:12:56.000
Um, so the old quote unquote, I always say that there is a line in the sand that came with open AI with the first um widely available reasoning model.
00:12:56.080 --> 00:13:02.480
So previously, when you first started using ChatGPT, or maybe you're listening to this and it's been a while, right?
00:13:02.639 --> 00:13:07.440
Um, the older models were transformers, and that's really all they were is next token predictors.
00:13:07.519 --> 00:13:10.240
Um, and but still they're generative.
00:13:10.399 --> 00:13:20.879
So the whole point is if the 20,000 people listening to this show all put the same prompt in ChatGPT or Claude or Copilot or whatever, you're probably going to get very different results, right?
00:13:20.960 --> 00:13:26.159
There might be 2,000 very different results, there might be 20,000 very different results, uh, right?
00:13:26.240 --> 00:13:39.840
It's it kind of goes through this next token prediction process, but um, that is why you have to really understand how these work and build up a uh context and good practices to get the best output.
00:13:40.240 --> 00:13:48.240
So um there's also in this, without going too deep, uh, there's a lot of nuance to this, right?
00:13:48.320 --> 00:13:52.639
Uh, because now these models can think uh very much like a human would.
00:13:52.799 --> 00:13:54.639
And we're gonna get to that a little bit later.
00:13:54.799 --> 00:13:58.879
But for the most part, a large language model is not a search engine.
00:13:58.960 --> 00:14:11.840
It is not a chatbot, it is a reasoning engine that combined with your data and your human knowledge can output economically viable work at a rate better and faster than humans.
00:14:12.000 --> 00:14:18.559
All right, so that is number one, what a large language model is and what it isn't.
00:14:18.720 --> 00:14:22.559
So now that we know what it is, where do we use it?
00:14:22.720 --> 00:14:35.840
All right, so uh also in the start here series, I had an entire uh episode on this concept of an AI operating system, but I think that's important to talk about uh because now they can all connect with your data.
00:14:36.000 --> 00:14:48.960
All right, and depending on what you go and uh you know what route you go down, I do think it's best to move to move most of your organization, most of your enterprise into one AI operating system, right?
00:14:49.120 --> 00:14:55.279
So uh personally, I'm using Chat GPT via codex more than anything.
00:14:55.440 --> 00:14:59.519
Uh, a lot of people are using Claude, whether that's on the web or the desktop.
00:14:59.600 --> 00:15:00.720
We're gonna get to that in a minute.
00:15:00.879 --> 00:15:09.840
There's also Gemini for those uh people who are uh really ingrained um in the uh kind of Google workspace uh ecosphere.
00:15:10.000 --> 00:15:19.840
And then obviously with Copilot, it's a little more difficult to explain because there's so many hoops to jump through in terms of permissions, role base, access control, all these things.
00:15:19.919 --> 00:15:30.639
It's a little convoluted, but um, I always tell you, I tell companies, try to move the majority of your day-to-day operations inside one of these AI operating systems.
00:15:30.799 --> 00:15:33.440
Uh, obviously, you should be building modularly.
00:15:33.679 --> 00:15:43.440
So when there are you know new kind of frontier breakthroughs uh or huge advantages, maybe a section of your team can take that context over to another one.
00:15:43.679 --> 00:15:48.720
Also, now every single big player makes it fairly easy to bring context over.
00:15:48.879 --> 00:15:57.120
So, whether that's things like skills and plugins, uh, your context, memory, chat history, etc., there's easy ways to port this over.
00:15:57.279 --> 00:16:02.799
It is more or less just a very um in-depth and detailed prompt, FYI.
00:16:02.960 --> 00:16:06.320
All right, so number two is you have to choose the operating system.
00:16:06.399 --> 00:16:08.960
Uh, number three, you have to choose the right surface.
00:16:09.120 --> 00:16:30.000
And this, at least in 2026, is where we've seen the most movement on, not necessarily what we saw in you know, mid-2024 to mid-2025, which is the features and the modes, the deep research is right, the you know, agentic capabilities with these reasoning models by default, the connectors, the apps, the surface is what's changing.
00:16:30.159 --> 00:16:31.360
So, what do I mean by that?
00:16:31.519 --> 00:16:34.879
In 2022, with the advent of ChatGPT, it was the web.
00:16:34.960 --> 00:16:36.639
We went and chatted, right?
00:16:36.799 --> 00:16:42.240
Then we started to bring our teams on board and you could collaborate with others and kind of a shared workspace.
00:16:42.399 --> 00:16:54.080
Uh, you know, and now in you know, late 2025 with Claude Code, uh, then Claude Co work in early 2026, and now Codex and everyone else, right?
00:16:54.159 --> 00:17:00.240
Google got on board with their Gemini desktop app with their anti-gravity, which I don't think is that great, but hopefully it'll get better, right?
00:17:00.399 --> 00:17:02.000
Now everything is moving to the desktop.
00:17:02.080 --> 00:17:11.519
Uh, and one of the reasons is is well, it can use um kind of your local machine to do a lot more uh faster and access your data.
00:17:11.680 --> 00:17:14.559
It can read and write on the desktop surface, which is huge.
00:17:14.640 --> 00:17:18.160
And then obviously, everything is agentic by default.
00:17:18.400 --> 00:17:23.599
So there's no easy answer to say, okay, well, should I still use the web version?
00:17:23.759 --> 00:17:25.599
Should I use the desktop version?
00:17:25.680 --> 00:17:38.160
Uh, you know, to get to that answer, you have to sit down and talk about what's important uh in terms of access, uh, in terms of automations, in terms of you know what you want in the cloud versus running locally.
00:17:38.319 --> 00:17:49.599
Actually, you know, although the that there's a lot of risk uh reward trade-off with a desktop version, in some instances it might be better for some companies that are uh, you know, privacy first.
00:17:50.079 --> 00:17:56.079
So, number three, uh, you know, is understanding that that surface dictates access, right?
00:17:56.400 --> 00:18:04.240
Um, so uh, because it's important to know that the agent tools, I mean, they can use your actual browsers, uh, right.
00:18:04.319 --> 00:18:12.480
And that's one of the biggest advantages, I think, to the desktop surface, aside from being able to read and write to your local folders, right?
00:18:12.559 --> 00:18:18.240
But to be able to use your actual browsers, I think Codec's built-in browser is so far ahead.
00:18:18.400 --> 00:18:23.440
Uh, computer use technology where it can literally, you know, use your computer, launch different programs.
00:18:23.599 --> 00:18:33.200
I have mine, you know, can go through and read my iMessage on my computer and you know, open up all these different desktop programs if there isn't a built-in app or connector.
00:18:33.440 --> 00:18:33.920
All right.
00:18:34.079 --> 00:18:36.799
Number four, choose the right model, people.
00:18:37.119 --> 00:18:39.119
Stop using a free plan.
00:18:39.680 --> 00:18:40.319
Period.
00:18:40.559 --> 00:18:44.559
If your company is like, yeah, use the free plan until we get this approved, don't do that.
00:18:44.720 --> 00:18:46.160
The risks are too high.
00:18:46.319 --> 00:18:46.640
All right.
00:18:46.720 --> 00:18:51.200
I don't care if you're using ChatGPT, Quad, Gemini, it doesn't matter.
00:18:51.359 --> 00:19:02.079
For the most part, you might get like one or two quote unquote prompts or outputs, uh, you know, with a decent enough uh plan if you're on uh a decent enough model.
00:19:02.160 --> 00:19:03.839
If you're on a free plan, don't do it.
00:19:03.920 --> 00:19:04.960
It's not worth it.
00:19:05.200 --> 00:19:11.200
Um, if you want real outputs, you have to use a paid model, right?
00:19:11.279 --> 00:19:14.640
All these people that are sharing things online, and you're like, oh, AI is so dumb, right?
00:19:14.720 --> 00:19:22.000
They're usually using an old model, a free model, um, and they really just don't understand uh kind of the underlying harness.
00:19:22.480 --> 00:19:31.920
So this is like, you know, you could have a convertible body, but oh, the engine is actually, you know, a dude on a bike, right?
00:19:32.079 --> 00:19:34.240
That's the equivalent of using a free model.
00:19:34.400 --> 00:19:37.839
You can't hide it in the fact that, like, oh, I'm using Claw.
00:19:38.400 --> 00:19:46.319
No, if you're using the free model, uh, right, for the most part, you probably don't even know because companies don't always tell you what version it is, right?
00:19:46.400 --> 00:19:50.000
Like, did you know that there's a GPD 5.4 latest?
00:19:50.240 --> 00:19:54.079
Like, no, you don't, because unless you're like me, you're not out there testing these things, right?
00:19:54.160 --> 00:19:56.640
You just think, oh, I'm using the best in the world.
00:19:56.799 --> 00:19:59.039
And you probably aren't if you're on a free plant.
00:19:59.119 --> 00:20:02.400
Uh, and you also need to be using thinking models.
00:20:02.559 --> 00:20:05.599
Do not use these models that don't think, don't reason.
00:20:05.759 --> 00:20:12.400
All right, you need to be using those always, even if that means waiting an extra 10, 20, 30 seconds.
00:20:12.559 --> 00:20:21.920
All right, I'm gonna get to this later, but read the chain of thought, do some exercise, have multiple you know, prompts running at the same time and you know, dictate, you know, move between.
00:20:22.000 --> 00:20:25.440
I do think the the job of the future is agentic orchestration.
00:20:25.599 --> 00:20:26.000
All right.
00:20:26.160 --> 00:20:32.079
So don't just, you know, go and chat with one, you know, free chatbot thread.
00:20:32.160 --> 00:20:33.519
At that point, you're wasting your time.
00:20:33.599 --> 00:20:39.680
And if you're making any business decisions based on using a free or a non-thinking model, I'm sorry.
00:20:39.839 --> 00:20:44.720
Uh, as the kids say, NGM, I'm not gonna make it.
00:20:44.880 --> 00:20:49.920
You, your department, your company, if you're doing that, if that's part of your actual strategy, don't.
00:20:50.160 --> 00:20:50.559
Right.
00:20:50.720 --> 00:20:51.759
Just just please don't.
00:20:51.839 --> 00:20:56.079
Unless you're just trying to learn like where all the buttons are, like how do I work this thing?
00:20:56.319 --> 00:20:58.319
That's the only time you should be doing that.
00:20:58.640 --> 00:20:59.279
All right.
00:20:59.519 --> 00:21:03.359
Uh, so number five, understanding the context layer.
00:21:03.599 --> 00:21:10.079
And this part is important before we even dive into the you know best practices of context engineering.
00:21:10.240 --> 00:21:15.599
But context is essentially everything that an AI model can see or not see, right?
00:21:15.759 --> 00:21:27.359
Uh, the the way that I've you know, we've taught uh you used to do like two live uh, you know, kind of prompt engineering courses a week for like two and a half years.
00:21:27.440 --> 00:21:29.680
Uh, and kind of the analogy I've always used.
00:21:29.759 --> 00:21:34.640
If if if you think back to the Star Wars credits from the original Star Wars, right?
00:21:34.799 --> 00:21:40.079
And it comes in and the words come in slightly slanted, and it says in a galaxy far, far away, right?
00:21:40.240 --> 00:21:45.359
And eventually the words they get smaller, smaller, smaller, and then they're off the screen.
00:21:45.599 --> 00:21:47.519
Think of that as a context window.
00:21:47.680 --> 00:21:50.240
Different models have different context windows.
00:21:50.400 --> 00:21:59.599
So essentially, you might be using a large language model, or you might have a skill saved, or a plugin, or a workflow saved, and you're using it repeatedly, and all of a sudden, you know.
00:22:00.160 --> 00:22:03.039
Started out great and all of a sudden it's like, wait, this stinks now?
00:22:03.200 --> 00:22:06.720
You know, oh, you know, this company must have nerfed the model.
00:22:06.880 --> 00:22:08.160
Maybe, but probably not.
00:22:08.319 --> 00:22:10.240
Probably you just ran over your context window.
00:22:10.400 --> 00:22:16.799
So that's essentially the amount of information like a computer's hard drive that a large language model can retain.
00:22:16.960 --> 00:22:17.279
All right.
00:22:17.359 --> 00:22:22.720
And a lot of times the first prompts, the first pieces of information you give a model are oftentimes the most important.
00:22:22.880 --> 00:22:24.319
It's where you give it direction.
00:22:24.559 --> 00:22:37.359
So the difference between a context window or understanding how context works in a hard drive is if you try to save a two gigabyte file on a full hard drive, it's just gonna say, nope, can't do it, no room.
00:22:37.519 --> 00:22:37.839
All right.
00:22:37.920 --> 00:22:42.880
A large language model is gonna do it and it's just gonna kick off something and you have no clue.
00:22:43.039 --> 00:22:54.960
So working with relevant, accurate business context is one of the most important things to do when working with a large large language model, uh, because essentially large language models get data from one of three places.
00:22:55.119 --> 00:22:59.519
Number one, their internal training data, which for the most part is very, very old.
00:22:59.680 --> 00:23:02.880
Uh, number two, any of your company data that you share.
00:23:02.960 --> 00:23:06.880
That can be, you know, apps, connectors, your file memory, chat history.
00:23:07.039 --> 00:23:16.240
You know, there's essentially a layer of your uh, you know, personal context, company context, you know, your all your SaaS apps that you use, right?
00:23:16.400 --> 00:23:17.599
And then there's the web.
00:23:17.920 --> 00:23:18.160
All right.
00:23:18.240 --> 00:23:23.200
So you have to understand that there's now this context layer that didn't really exist before, right?
00:23:23.279 --> 00:23:24.960
It used to just be the training data.
00:23:25.039 --> 00:23:28.400
And keep in mind, training data is usually really old.
00:23:28.559 --> 00:23:41.200
There's a knowledge cutoff date, but I would venture to guess the overwhelming majority, like 99% of the data that actually ends up in today's large language models is nowhere close to that date.
00:23:41.279 --> 00:23:52.880
So, you know, let's just say if there was a way to turn off web search and to turn off your company's context, and if you were to only rely uh on a model's training data, it would be very bad, right?
00:23:52.960 --> 00:23:59.839
It's actually crazy to think it took anthropic so long to add web search, and which is why I was like, companies should never use it, right?
00:23:59.920 --> 00:24:08.720
Because without web search, you know, you're probably playing with data at minimum that's probably on average 15 to 18 months old on the good side.
00:24:08.880 --> 00:24:14.559
And think of how quickly your company, uh, your competitive landscape, your sector moves.
00:24:14.720 --> 00:24:18.160
And imagine if you only had access to data that was 15 to 18 months old.
00:24:18.319 --> 00:24:19.759
Yeah, recipe for disaster.
00:24:19.920 --> 00:24:30.640
And that's why also you get these, you know, out outputs sometimes, especially, you know, pre, you know, 2024 reasoning models that could agentically call the web and pull your data.
00:24:30.799 --> 00:24:34.079
Why you'd get these outputs that looked like they were absolutely terrible?
00:24:34.240 --> 00:24:35.119
Well, because they were.
00:24:35.279 --> 00:24:40.319
It was relying on old data, and large language models are trained to be helpful assistance.
00:24:40.480 --> 00:24:40.799
All right.
00:24:40.960 --> 00:24:53.839
So make sure you go back and listen to the start here series going over the, I think we called it the seven deadly sins of AI, uh, where we kind of talked about some of these things, but that's why sometimes outputs from large language models seem generic.
00:24:53.920 --> 00:24:57.279
And you're like, is this actually truthful or is it kind of lying to me?
00:24:57.359 --> 00:25:07.200
Well, the context helps steer that away from just giving you these general, it sounds helpful, but is it true, to being pinpoint specific and valuable for your company.
00:25:07.440 --> 00:25:09.680
Number six, context engineering.
00:25:09.839 --> 00:25:18.640
All right, this obviously plays in line with the context layer, but this is where you insert right one of those three layers of data into the context window.
00:25:18.720 --> 00:25:22.559
Uh, so this can be things like your traditional prompt engineering.
00:25:22.640 --> 00:25:26.160
So you can share that context via your prompting, uh, right.
00:25:26.240 --> 00:25:31.279
But the important thing uh really is knowing that it works in layers.
00:25:31.440 --> 00:25:33.839
So I'm gonna give you a couple hints, right?
00:25:34.000 --> 00:25:36.000
So we taught prime prompt polish.
00:25:36.160 --> 00:25:44.559
All right, the 30-second version of that is you should always prime a model, work with it a lot before you prompt it and ask it for an output.
00:25:44.640 --> 00:25:54.640
And then polish, we're actually gonna get to that technically in number nine and 10, because the first output that a model gives you is usually garbage, even if you do everything else correctly.
00:25:54.880 --> 00:25:57.119
Um, all right, refine queue.
00:25:57.200 --> 00:26:04.720
That's where we talk about, you know, role, examples, fetch, insights, narration, um, explanation, and questions.
00:26:04.880 --> 00:26:05.279
All right.
00:26:05.359 --> 00:26:09.759
And if you want to go more on the, you know, prime prompt polish, refine queue, 555.
00:26:10.079 --> 00:26:16.880
This is like a decision tree on when you should use um a large language model versus when you should do it, quote unquote, manually without AI.
00:26:17.200 --> 00:26:23.599
Actually, when you do go to starthirseries.com uh and sign up for our community, uh, you will also get access.
00:26:23.759 --> 00:26:25.119
Yeah, a lot of people don't know this.
00:26:25.200 --> 00:26:26.640
Uh, we recorded that.
00:26:26.720 --> 00:26:27.599
It's on demand, right?
00:26:27.680 --> 00:26:28.400
I used to do it live.
00:26:28.480 --> 00:26:31.359
I can't do it live anymore, but I keep the videos pretty updated.
00:26:31.440 --> 00:26:34.880
So you can go take our updated prime prompt polish course.
00:26:35.039 --> 00:26:38.000
I think it was updated as of like GPT 5.4.
00:26:38.079 --> 00:26:41.440
So I'll probably update it once we get like five, six or something.
00:26:41.599 --> 00:26:46.960
Uh, but the uh the basic, this is basic context engineering 101.
00:26:47.119 --> 00:26:47.440
All right.
00:26:47.519 --> 00:27:06.319
So it can be as simple as you know, doing the role, goal, sources, constraints, examples, whatever it is, but it's making sure that through your words that you're sharing, you are steering the model in the right direction and telling it what context to use, what output you need, and working with it before you just expect it to give you amazing outputs.
00:27:06.559 --> 00:27:07.039
All right.
00:27:07.279 --> 00:27:10.720
Number seven, working with files, apps, and company data.
00:27:10.880 --> 00:27:12.480
This one is huge.
00:27:12.720 --> 00:27:17.359
So files make AI far more specific to the actual task.
00:27:17.519 --> 00:27:17.759
All right.
00:27:17.839 --> 00:27:40.799
And a lot of people think oh, if I just dump all of my context into one, one of these large language models and tell it, you know, okay, now go, you know, write, you know, next quarter's press releases and update all our job descriptions based on the SOPs and right, all you're doing in this instance without getting too technical and accidentally turning this again into a 90-minute show, um, these think of them as books on the bookshelf, right?
00:27:40.960 --> 00:27:54.319
You still have to tell the model when, via whether that's through a skill, uh a plug-in, uh, a certain workflow, automation, etc., you still have to tell it when to use the book and sometimes direct it to look into what chapter.
00:27:54.559 --> 00:28:07.839
But uh creating these um, you know, app connections are huge because depending on if you're talking about, you know, Chat GPT, Claude, or Gemini, as an example, they all treat these connectors a little bit differently.
00:28:08.000 --> 00:28:15.279
As an example, some of them might index or cache some of your most important connectors, like your Gmail, right?
00:28:15.359 --> 00:28:17.839
So you don't even have to wait, it just knows.
00:28:17.920 --> 00:28:25.440
And it can actually actually uh, you know, proactively go out and find these things and surface them to you, which is actually huge.
00:28:25.680 --> 00:28:41.680
So, you know, whether we're talking about uploading files, um, you know, connecting uh a connector or an app, right, like a Google Drive or a Microsoft um SharePoint, OneDrive, Box, etc., your CRM, right?
00:28:41.759 --> 00:28:50.000
Uh, most of the big three, I don't know if Google does, but I think mostly everyone else has like HubSpot and some of the big names and CRMs, right?
00:28:50.240 --> 00:28:53.759
All of these things now, for the most part, have read and write.
00:28:53.839 --> 00:29:00.240
And that is big, and that's much different than where we were at, you know, in the early days of connectors, like nine months ago.
00:29:00.319 --> 00:29:00.880
Uh, right.
00:29:00.960 --> 00:29:04.319
So now these can, well, run actions for you.
00:29:04.480 --> 00:29:23.039
So, you know, let's just say you get some information off of a Google Sheet, um, and you have to go, you know, update uh something in a database, and then you have to go, you know, change some options on a CRM, and then you have to update your project management tool, click up as an example, right?
00:29:23.200 --> 00:29:27.440
These are all processes that normally I call the human duct tape that you would have to do.
00:29:27.599 --> 00:29:31.599
Now these connectors can talk to each other and carry that context over.
00:29:31.839 --> 00:29:35.680
We did a show called Agentic Context Carry in the Start Here series.
00:29:35.759 --> 00:29:38.640
So go back and listen to that if you want to know more about how that works.
00:29:38.799 --> 00:29:54.480
But essentially, this is why these models with the context um you know layer in this in agent being able to carry that context and keep it in the context window is huge because this is the majority of what so many knowledge workers spend their time doing.
00:29:54.640 --> 00:30:00.480
You know, you have the five to 15 probably different SaaS applications that you spend your time in.
00:30:00.640 --> 00:30:05.440
You use your brain, your role, your requirements, your KPIs.
00:30:05.680 --> 00:30:10.640
You uh grab all of that important context from all of these different, you know, apps.
00:30:10.880 --> 00:30:15.920
You, you know, personalize it, create some value out of it, and then you piece it together in all of these other apps.
00:30:16.079 --> 00:30:17.839
This is what agents do now, right?
00:30:17.920 --> 00:30:22.400
So working with files, apps, connectors, and your company data is huge.
00:30:22.559 --> 00:30:23.200
All right.
00:30:23.599 --> 00:30:26.799
Number eight, and this is important, especially following up.
00:30:26.880 --> 00:30:28.960
Uh, maybe it should have been number seven, right?
00:30:29.200 --> 00:30:32.880
Uh don't do the shadow IT, right?
00:30:33.039 --> 00:30:39.680
Shadow IT is when you don't have permission or you know, shadow AI or shadow agents, whatever you want to call it.
00:30:39.759 --> 00:30:43.279
When you don't have permission to use your company data, you shouldn't be uploading it, right?
00:30:43.359 --> 00:30:55.599
But the good thing is, whether we're talking about Claude Enterprise, Gemini Enterprise, Chad GPT Enterprise, they have the exact same data security and privacy uh that you would have from your cloud provider.
00:30:55.759 --> 00:31:05.119
So if you upload your company's uh data to a cloud, it's the exact same thing.
00:31:05.359 --> 00:31:09.039
It is the exact same thing as using these connectors in these apps.
00:31:09.119 --> 00:31:18.079
As long as you're going through, you know, turning off model training, doing the you know, one-on-one best practices, um, you're not running any additional risk.
00:31:18.240 --> 00:31:28.640
Um, again, that's a big if as long as you're you know using these models correctly, turning off uh, you know, training data, which most of the paid uh business or enterprise versions do by default.
00:31:28.960 --> 00:31:32.000
You're not running any additional risk with your company's data.
00:31:32.160 --> 00:31:40.480
It's not like I'm gonna upload, you know, or connect my data, turn off model training, and my competitor is gonna be like, oh, what's everyday AI doing?
00:31:40.559 --> 00:31:41.440
That's not how it works.
00:31:41.519 --> 00:31:45.039
But people, smart people with brains still think that today.
00:31:45.200 --> 00:31:46.799
That's not how it works, right?
00:31:46.960 --> 00:31:50.000
Uh, governance is also extremely important.
00:31:50.160 --> 00:31:52.559
That is permission design, right?
00:31:52.640 --> 00:31:53.759
This is not paperwork.
00:31:53.839 --> 00:31:57.839
This is not a one-time 20-minute training because the systems are constantly changing.
00:31:58.079 --> 00:31:58.400
All right.
00:31:58.480 --> 00:32:03.359
Also, personal accounts, those aren't like unofficial company AI systems.
00:32:03.440 --> 00:32:07.279
Like I said, uh, you know, you should only be connecting your company's data.
00:32:07.440 --> 00:32:17.680
Number one, if you've gone through the guardrails, the governance, uh, the privacy permissions, getting the sign-off from everyone involved, but you should only also only be using it through the proper channels, right?
00:32:17.839 --> 00:32:20.400
I can't tell you the number of companies when they're telling the truth.
00:32:20.480 --> 00:32:27.279
They're like, okay, well, yeah, we we got our, you know, copilot approved, but you know, no one can get the right access to it.
00:32:27.519 --> 00:32:32.880
But well, if we're using SharePoint and copilot, we should just be able to use it in ChatGPT, right?
00:32:33.039 --> 00:32:37.279
Well, in theory, kind of, but also absolutely not, right?
00:32:37.440 --> 00:32:40.480
That is going against governance 101.
00:32:40.799 --> 00:32:54.079
So, you know, expert-driven loops are so important when we talk about, you know, properly setting up these guardrails, especially as these large language models, these quote unquote AI chatbots by default, can now take actions for us.
00:32:54.160 --> 00:33:03.359
And we can schedule agents to go take actions, right, with right W-R-I-TE, right capabilities, right?
00:33:03.599 --> 00:33:07.279
It can go off and email customers and update your CRM.
00:33:07.440 --> 00:33:16.240
And, you know, maybe if you're not, you know, keeping an expert-driven loop, it could potentially make a huge mistake that can cost your organization a lot of money.
00:33:16.400 --> 00:33:26.000
So there's obviously a downside to productivity and efficiency uh if you skip over step number eight, which is proper privacy, permissions, and governance.
00:33:26.480 --> 00:33:27.039
Yeah.
00:33:27.200 --> 00:33:32.880
Uh so after permissions, though, leaders need visibility into what actually happened.
00:33:32.960 --> 00:33:39.279
And that's where we have to talk about the importance of transparency, observability, and reasoning artifacts.
00:33:39.599 --> 00:33:49.200
So this is where uh people skip over showing your work uh and being able to observe that as a team.
00:33:49.279 --> 00:33:49.680
Uh, right.
00:33:49.759 --> 00:34:01.440
So I can't go into all the specifics because it's gonna take a while, but especially when you get up to enterprise accounts, not only can you see, you know, usage and things like that, you can see uh you're right, I think Copilot actually leads in this.
00:34:01.519 --> 00:34:07.920
Uh, you know, I think with their uh intra ID, you can see every action certain uh, you know, agents make.
00:34:08.000 --> 00:34:15.599
I think ChatGPT's new workspace agents uh for uh business and enterprise uh teams do a great job of doing that as well.
00:34:15.760 --> 00:34:24.639
But you have to be able to understand and show the work, especially when agents are uh or sorry, AI chatbots are agentic by default.
00:34:24.800 --> 00:34:29.920
You have to be able to uh observe them and understand the reasoning artifacts.
00:34:30.000 --> 00:34:30.719
What does that mean?
00:34:30.880 --> 00:34:36.400
You know, there's usually a little thing you can click, accept in Gemini, Gemini, can we get a little better at that, please?
00:34:36.559 --> 00:34:37.199
Uh right.
00:34:37.360 --> 00:34:55.039
I know there's a tiny step that we got uh with 3.5 flash and the you know updated version of Gemini, but we need complete transparency when it comes to seeing how models get from context engineering and great setups and great workflows to creating this great economically valuable work.
00:34:55.119 --> 00:34:58.400
And the reason being is because models change all the time.
00:34:58.480 --> 00:35:17.280
So if you can't properly see every single step in every tool call, in every website that a model or an agent went to in order to deliver that, you know, first draft deliverable for your team that you ultimately sent off to a client or the RFP or whatever it is, then you don't have the observability.
00:35:17.440 --> 00:35:22.400
Then you don't actually own that asset because you don't know what's happening, right?
00:35:22.480 --> 00:35:34.719
So I like to talk about the steps in the middle are sometimes the most overlooked, but they're the most important because if you don't understand what's going on, you don't own it, right?
00:35:34.800 --> 00:35:35.599
So what does that mean?
00:35:35.760 --> 00:35:44.559
If something happens in the harnessing or a model changes, or you know, there's a drastic switch up in how a model uses different tools.
00:35:44.719 --> 00:35:53.199
If you don't own and understand that observability, the the reasoning, uh, you know, artifacts, then you may not be able to replicate those same things.
00:35:53.280 --> 00:36:10.400
And if they if these things are running on loops scheduled, which is very easy to do, and one little thing changes, it can break your entire business if you become overly reliant on passive, lazy, human-in-the-loop workflows versus expert-driven loops.
00:36:10.559 --> 00:36:15.119
Expert-driven loops means that you are looking at the transparency and observability.
00:36:15.360 --> 00:36:16.000
All right.
00:36:16.480 --> 00:36:21.760
Last but not least is verification, iteration, and creating those workflows, right?
00:36:21.840 --> 00:36:23.280
So, what do I mean by this?
00:36:23.519 --> 00:36:26.480
Uh iteration is huge.
00:36:26.639 --> 00:36:29.519
So I talked about, you know, the prime prompt polish.
00:36:29.599 --> 00:36:30.639
This is the polish.
00:36:30.719 --> 00:36:42.800
You know, your first output, even if you do everything else, steps one through nine correctly, your first output for the most part from a large language model, even a great one, is at best case, it's gonna be generic garbage, right?
00:36:42.880 --> 00:36:44.079
But it's not gonna be very well.
00:36:44.239 --> 00:36:45.199
It's not gonna sound like you.
00:36:45.280 --> 00:36:48.000
It's even when you give it the proper examples, right?
00:36:48.239 --> 00:36:52.719
Um, you have to really iterate on the output.
00:36:52.880 --> 00:36:57.920
You have to understand how it got to that output and then run it over and over and make it better.
00:36:58.079 --> 00:37:03.039
That is this iteration loop where you turn, you know, your first output, right?
00:37:03.199 --> 00:37:08.480
My old journalism days, the first version of the story you turned in was never really good or the best, right?
00:37:08.559 --> 00:37:14.159
You always had to go through the multiple editors and all the the red, all the red lines to make it better.
00:37:14.320 --> 00:37:17.920
And then once you get to that point, you need to then turn it into a workflow.
00:37:18.000 --> 00:37:21.920
And this looks a little bit different in Chat GPT versus Claude versus Gemini, right?
00:37:22.000 --> 00:37:26.480
But for the most part, we are getting these universally um applicable skills.
00:37:26.639 --> 00:37:31.119
Another thing, Anthropic uh led the way on this, and everyone else is adopting them.
00:37:31.280 --> 00:37:40.000
But skills are essentially ways that at the end of iterating, you can go through and essentially, in more words than this, say turn this into a skill, right?
00:37:40.079 --> 00:37:48.000
And then once you turn it into a skill, then you can schedule it because uh most of these systems have scheduling by default.
00:37:48.239 --> 00:37:51.119
So um it's all about verify that it works.
00:37:51.280 --> 00:38:06.079
You know, make sure you're using the right model, the right mode for the right tool, you're getting the right outputs, you're verifying it, you're going through the transparency, observability, the reasoning artifacts, but then you don't stop when you get that first, you know, hey, this draft is good enough.
00:38:06.239 --> 00:38:07.519
No, keep working.
00:38:07.599 --> 00:38:09.039
That's where you polish.
00:38:09.199 --> 00:38:14.559
You turn it into a skill, you turn it into a plugin, and then you turn it to an automated workflow.
00:38:14.639 --> 00:38:21.199
And that's how we go from working with an AI chatbot to commanding AI agents.
00:38:21.360 --> 00:38:21.760
All right.
00:38:22.000 --> 00:38:24.719
I hope this one was helpful, y'all.
00:38:24.800 --> 00:38:25.599
If so, let me know.
00:38:25.679 --> 00:38:28.079
If you're listening on the podcast, please do me a favor.
00:38:28.320 --> 00:38:31.360
All right, uh, this one has been a long time in the waiting.
00:38:31.440 --> 00:38:36.239
So, you know, I hope more people can hear this because I want to cut through the BS.
00:38:36.400 --> 00:38:44.159
I want to tell people exactly how these models work, uh, and I want to be able to do it for as for free as long as I can, right?
00:38:44.320 --> 00:38:47.519
Uh, so that only happens if you're number one sharing this with others.
00:38:47.599 --> 00:38:52.719
So if you are listening um, you know, on LinkedIn as an example, please repost this if this was helpful.
00:38:52.800 --> 00:38:56.239
If you are listening on the podcast, please subscribe to the show.
00:38:56.400 --> 00:38:57.679
That would mean a lot.
00:38:57.920 --> 00:38:59.440
Leave us rating too, if you can.
00:38:59.599 --> 00:39:00.719
So I hope this was helpful.
00:39:00.880 --> 00:39:08.400
Like I said, make sure to go to starthirseries.com to get exclusive access to all of the episodes in this series.
00:39:08.559 --> 00:39:09.360
Thank you for tuning in.
00:39:09.440 --> 00:39:12.320
I hope to see you back tomorrow in everyday for more everyday AI.
00:39:12.400 --> 00:39:13.440
Thanks, y'all.
00:00:01.760 --> 00:00:13.679
The saying of keeping up with AI is like trying to drink water from a fire hose, though it's tired and cliche is obviously 100% true.
00:00:13.919 --> 00:00:20.000
I mean, with all the nonstop updates, your head is probably spinning trying to keep up.
00:00:20.239 --> 00:00:25.120
So you're thinking, with all these new releases, how do I use ChatGPT?
00:00:25.199 --> 00:00:29.280
Or is using Claude kind of the same as using Gemini on the desktop?
00:00:29.440 --> 00:00:33.679
Or can I prompt Copilot kind of like I'd prompt Codex?
00:00:34.079 --> 00:00:42.240
Or maybe you've had to start the race sprinting and you never really got the proper 101 on how all of these AI chatbots work.
00:00:42.479 --> 00:00:48.159
Regardless, you definitely aren't alone in the fire hose updates drowning you out.
00:00:48.479 --> 00:00:54.960
That's the new struggle for every enterprise, because it's not like there's a cheat code, right?
00:00:55.359 --> 00:01:02.320
Because as the AI models are changing almost daily, so too does the input required to get the best output.
00:01:02.399 --> 00:01:03.280
And I get it.
00:01:03.520 --> 00:01:09.840
As someone that covers AI daily, I understand the struggle of trying to play a game, yet the field dimensions change without notice.
00:01:10.000 --> 00:01:16.560
And you were using a softball yesterday, but cricket gear today and tomorrow you might kind of play it out like rugby.
00:01:16.719 --> 00:01:20.560
The input rules and the output capabilities are moving targets.
00:01:20.799 --> 00:01:25.040
But there's one thing I've found out over the last three or four months.
00:01:25.280 --> 00:01:31.200
The big players have all just kind of started to copy each other, which is actually a good thing for you.
00:01:31.439 --> 00:01:42.719
So I think it's actually been over the last two months that we've finally stumbled on a set of best practices for getting the best outputs out of any large language model.
00:01:42.959 --> 00:01:53.840
So that means, well, there is a cheat code, or at least a set of somewhat concrete rules that, when followed, will give you stellar outputs, really, no matter what model you're using.
00:01:54.000 --> 00:01:58.319
And that's exactly what we're going to be giving you today on Everyday AI.
00:01:58.480 --> 00:02:01.200
Welcome to the Start Here series.
00:02:01.359 --> 00:02:10.879
Uh, if you're new here, the Start Here series is your essential guide to getting caught up and getting ahead with AI.
00:02:10.960 --> 00:02:14.479
But first, let's talk about the big picture.
00:02:15.360 --> 00:02:20.719
So the capabilities of these models are very, very real.
00:02:21.199 --> 00:02:23.840
And it is hard to keep up.
00:02:24.000 --> 00:02:24.400
Right.
00:02:24.560 --> 00:02:28.319
So right now, you probably see examples all over the place.
00:02:28.479 --> 00:02:36.560
You see people sharing examples, whether it's online, whether it's media articles that are so poorly written, and you're like, well, these AI models are really bad.
00:02:36.719 --> 00:02:37.680
Look at all these mistakes.
00:02:37.759 --> 00:02:53.120
But then you also see these benchmarks and these tests, and you know, these big companies laying people off in lieu of spending billions of dollars on AI, and you're confused because you can't even know or don't even really understand what button to click or which model should I use.
00:02:53.280 --> 00:02:58.159
And up until recently, it's been too hard to follow because there's been too many different paths.
00:02:58.479 --> 00:03:02.879
And now I think essentially we have cookie cutters, right?
00:03:02.960 --> 00:03:08.560
We have the McMansions of models because they're all kind of the same.
00:03:08.719 --> 00:03:12.719
Obviously, the capabilities and the harnessing of the tool is all completely different and unique.
00:03:12.960 --> 00:03:28.719
But I think for the first time, maybe ever, at least when it comes to you know, the three to five handful of big AI players, we have a set of rules that can apply unilaterally to really any of them.
00:03:29.039 --> 00:03:33.120
So that's what we're gonna be going over on today's show.
00:03:33.280 --> 00:03:34.000
Stick with me.
00:03:34.159 --> 00:03:36.719
This one, trust me, is gonna be 30 minutes.
00:03:36.879 --> 00:03:37.199
All right.
00:03:37.280 --> 00:03:42.159
I know sometimes I say 25 to 30, and then you're like 45 minutes in, but stick with me on today's show.
00:03:42.240 --> 00:03:43.199
Here's what you're gonna learn.
00:03:43.280 --> 00:03:48.800
You're gonna know why the cookie cutter nature of today's large language models is actually both a good and a bad thing.
00:03:48.960 --> 00:04:08.479
You're gonna know the up-to-date capabilities, features, pros, and cons of the big players, specifically focusing on Chat GPT, Claude, Gemini, technically applicable, the co-pilot, but we're not gonna go into that one as deep, as well as I mean, you can apply it to perplexity, grok, open models, etc., because they are kind of all getting the same.
00:04:08.560 --> 00:04:14.479
And you're gonna leave today's show with the 10 essential steps to get the most out of any AI chatbot.
00:04:14.639 --> 00:04:17.439
All right, this is the Start Here Series.
00:04:17.600 --> 00:04:23.360
This is the essential podcast series to both learn the AI basics and to double down on your AI knowledge.
00:04:23.519 --> 00:04:32.000
Because after doing this for three and a half years, nearly almost 800 episodes now, I never had a good answer when people ask, Jordan, I'm new to the podcast.
00:04:32.079 --> 00:04:32.720
Where do I start?
00:04:32.879 --> 00:04:37.040
Well, you start here with the start here series and you go to startheseries.com.
00:04:37.120 --> 00:04:40.399
That is going to give you free access to our inner circle community.
00:04:40.560 --> 00:04:41.279
It's exclusive.
00:04:41.360 --> 00:04:48.560
This is the only way you can actually get access in the Start Here series uh space inside of our free community.
00:04:48.720 --> 00:04:54.079
Uh, you can go find every single Start Here Series episode all in one easy-to-find place.
00:04:54.240 --> 00:04:58.399
Uh, a Spotify playlist that's updated, all the newsletters, go read it.
00:04:58.560 --> 00:04:59.600
It's awesome.
00:04:59.839 --> 00:05:00.160
All right.
00:05:00.319 --> 00:05:10.399
So uh if you miss our last Start Here Series episode, we covered Bill by Partner or Wait, the four-layered AI stack decision framework for 2026.
00:05:10.560 --> 00:05:17.839
And today we are giving you the LLM cheat code 10 essential steps to get the most out of any AI chatbot.
00:05:18.319 --> 00:05:18.879
All right.
00:05:19.040 --> 00:05:22.079
Um, and I'm just gonna give you those steps now.
00:05:22.160 --> 00:05:26.800
And we're obviously going to break them down as we go along.
00:05:27.040 --> 00:05:28.560
But here they are.
00:05:28.959 --> 00:05:34.319
Um kind of steps, best practices, rules to live by, whatever.
00:05:34.480 --> 00:05:38.879
But number one, you have to understand what a large language model is and what it is not.
00:05:39.120 --> 00:05:50.160
Number two, you have to choose your AI operating system, like ChatGPT, Claude, Gemini, or Copilot, and really stick to those as much as you can for the majority of your day-to-day knowledge work.
00:05:50.319 --> 00:05:52.959
Number three, you have to choose the right surface.
00:05:53.040 --> 00:05:57.279
And the surfaces are unfortunately changing, but I think it's actually for the good.
00:05:57.519 --> 00:06:01.519
Number four, you have to choose the right account plan and model.
00:06:01.680 --> 00:06:03.759
I'm gonna break down those best practices as well.
00:06:03.920 --> 00:06:07.040
Number five, you have to understand the context layer.
00:06:07.120 --> 00:06:12.240
First, you have to understand it before you get to number six, context engineering basics.
00:06:12.319 --> 00:06:20.800
I'm gonna tell you prime prompt polish, refine queue, five, five, five, a lot of the secrets that I've kind of held close to the best throughout the years.
00:06:20.959 --> 00:06:30.560
Uh, number seven, I'm gonna tell you how and why, and risks on working with files, apps, and company data and why that matters.
00:06:30.720 --> 00:06:38.079
Number eight, we're gonna talk about privacy, permissions, and governance and how that actually plays out in these AI operating systems.
00:06:38.240 --> 00:06:42.079
Yeah, I call AI chatbots now AI operating system because that's what they are.
00:06:42.319 --> 00:06:46.399
Number nine, transparency, observability, and reasoning artifacts.
00:06:46.480 --> 00:06:51.759
I'm gonna show you how to use those things to your advantage to get better outputs.
00:06:51.920 --> 00:06:56.639
And last but not least, verification, iteration, and workflow design.
00:06:56.800 --> 00:06:57.439
There you go.
00:06:57.600 --> 00:07:04.240
You can stop here if you want, but trust me, this is like thousands of hours of conversations over the past three years.
00:07:04.319 --> 00:07:09.360
And I'm gonna do my dangedest uh to get them to you in 30 minutes or less.
00:07:09.600 --> 00:07:15.519
So uh here's what I want to talk about, why this is both a good and a bad thing.
00:07:15.680 --> 00:07:19.680
And I think this really became evident when Google at their I.O.
00:07:19.839 --> 00:07:23.600
conference, uh uh like last week, and I know, right?
00:07:23.759 --> 00:07:25.839
So right now it's the end of May 2026.
00:07:25.920 --> 00:07:31.680
I should put this out there because I know you know a lot of you might be listening to this in you know, July or December.
00:07:31.839 --> 00:07:37.680
So obviously, some of the things are gonna change here when I'm talking about certain surfaces, models, modes, etc.
00:07:37.920 --> 00:07:44.560
But hopefully these concrete-ish uh 10 steps will still hold.
00:07:44.800 --> 00:07:48.800
But I noticed at Google's I.O.
00:07:48.959 --> 00:07:52.959
conference when they announced their new anti-gravity 2.0 desktop app, right?
00:07:53.120 --> 00:08:01.680
And not gonna get into it, but I'm like, wait, this just looks like codex, it functions like codex, and then I'm like, wait, cursor kind of does too.
00:08:01.759 --> 00:08:12.000
And that's when I started to notice that for the most part, especially compared to 18 months ago, when all of the different, whether we're talking about the web interfaces, right?
00:08:12.160 --> 00:08:19.360
Gemini.google.com, claude.ai, chat gpt.com, grok.com, whatever you're using, 18 months ago, they all looked really different.
00:08:19.439 --> 00:08:21.600
It was almost like speaking different languages.
00:08:21.759 --> 00:08:25.360
Now it's all speaking the same language with slightly different dialects, right?
00:08:25.439 --> 00:08:31.439
Like anti-gravity literally looks like it is codex light, right?
00:08:32.000 --> 00:08:35.200
Color schemes, fonts all look the same, layouts.
00:08:35.360 --> 00:08:36.799
So it's good and bad.
00:08:37.039 --> 00:08:45.039
From a bad perspective, it might seem to you, the non-technical user, that the front-facing innovation is kind of slow, right?
00:08:45.200 --> 00:08:56.159
As an example, go back to you know, 2024, front-end innovation for um users and maybe early in 2025 was everywhere, right?
00:08:56.240 --> 00:09:12.000
All of a sudden we had artifacts from Claude, and then we had, you know, Canvas mode, and you know, we had these agents that worked inside, and all of a sudden, you know, large language models went from simple, you know, next token prediction transformers to models that could think and reason and call tools, right?
00:09:12.080 --> 00:09:19.120
So I think there is this period of fast innovation in terms of what buttons you would click, all of these modes, deep research, et cetera.
00:09:19.279 --> 00:09:22.720
But if I'm being honest, in 2026, we haven't seen that same thing.
00:09:22.799 --> 00:09:42.720
And I think that's actually good because what happened is it, I think it gave us number one, now we can kind of apply these best practice rules, but more than uh anything else, I think it allows enterprises to be able to modular modularly uh pivot as needed, right?
00:09:42.879 --> 00:09:53.600
Because 18 months ago, if you were a heavy chat GPT Teams user back when it was called before it turned into Chat GPT business, and then you try to go to, you know, uh Gemini, right?
00:09:53.759 --> 00:09:57.440
There was no Gemini business at the time, but it was just Gemini, you'd be confused.
00:09:57.600 --> 00:10:02.879
But now they've kind of all copied each other with certain features, projects, GPTs, gems, right?
00:10:02.960 --> 00:10:05.200
They're all kind of just the same thing now.
00:10:05.360 --> 00:10:11.600
Uh but the actual race is smarter, faster models with autonomous desktop harnesses.
00:10:11.679 --> 00:10:13.279
So the surface is changing a little bit.
00:10:13.360 --> 00:10:21.360
But ultimately, the thing that you have to keep in mind if you don't get anything else from today's show, AI models are smarter than all of us.
00:10:21.440 --> 00:10:23.519
If you're using them the right way, and that's a big if.
00:10:23.600 --> 00:10:26.879
And when I talk about jagged, jagged capabilities, right?
00:10:27.120 --> 00:10:30.480
People will share something where it's like, oh, AI models are dumb.
00:10:30.559 --> 00:10:34.879
And then you have this false sense of security, like, oh, I don't have to worry about keeping up with these day-to-day.
00:10:35.120 --> 00:10:35.679
Yes, you do.
00:10:35.919 --> 00:10:39.039
Anyone that shares those things, that's a skill issue.
00:10:39.200 --> 00:10:43.840
You know, uh, send them this video and then they'll be like, oh, I've been using everything wrong.
00:10:44.000 --> 00:10:51.840
Um, but AI models by default, studies show this, even judged blindly by experts.
00:10:52.320 --> 00:10:58.720
Today's AI models, with all of this harnessing built around them, they can produce artifacts one shot, right?
00:10:58.879 --> 00:11:05.840
Uh, you know, complete websites, complete apps, uh, spreadsheets, PowerPoints, Word docs, PDFs, etc.
00:11:06.080 --> 00:11:06.639
Uh, right.
00:11:06.720 --> 00:11:09.440
So they're no longer just a little chat bot to talk to.
00:11:09.519 --> 00:11:13.600
And we've covered that uh in depth in the Start Here series uh so far.
00:11:13.759 --> 00:11:15.519
So let's dive in.
00:11:15.600 --> 00:11:25.919
We have some upgraded visuals uh today, but let's get into our uh kind of the 10 new rules of work, I'll say.
00:11:26.720 --> 00:11:32.240
So number one, what is a large language model and what is it not?
00:11:32.639 --> 00:11:46.159
Well, if you're only using Chad GPT or Claude or whatever as a version of Google, as a smarter, faster, better source version of Google, you're missing out, right?
00:11:46.320 --> 00:11:52.000
If you're just treating it back and forth like talking to a smart friend, you're also missing out.
00:11:52.159 --> 00:11:55.759
So by default, large language models are generative.
00:11:56.000 --> 00:12:03.679
So that means they are non-deterministic, where a search engine like Google, aside from localization and personalization, is deterministic, right?
00:12:03.840 --> 00:12:19.360
If we put in the exact same query, chances are, aside from that personalization and localization, but if we all did an incognito Google search, right, uh five years ago before AI overviews in AI mode, we for the most part all would have gotten the exact same outputs, right?
00:12:19.440 --> 00:12:26.240
Even if there's, you know, 20,000 people listening to this episode, and we all did it, we would for the most part get the exact same results.
00:12:26.480 --> 00:12:28.080
It's not the same with generative AI.
00:12:28.159 --> 00:12:28.960
It is generative.
00:12:29.120 --> 00:12:33.200
There is an element of uh next token prediction.
00:12:33.279 --> 00:12:46.639
And uh essentially, large language models have been trained on the entirety of the internet, copyrighted works, uh, you know, open source coding projects, videos, photos, everything, uh, just terabytes of data.
00:12:46.799 --> 00:12:56.000
Um, so the old quote unquote, I always say that there is a line in the sand that came with open AI with the first um widely available reasoning model.
00:12:56.080 --> 00:13:02.480
So previously, when you first started using ChatGPT, or maybe you're listening to this and it's been a while, right?
00:13:02.639 --> 00:13:07.440
Um, the older models were transformers, and that's really all they were is next token predictors.
00:13:07.519 --> 00:13:10.240
Um, and but still they're generative.
00:13:10.399 --> 00:13:20.879
So the whole point is if the 20,000 people listening to this show all put the same prompt in ChatGPT or Claude or Copilot or whatever, you're probably going to get very different results, right?
00:13:20.960 --> 00:13:26.159
There might be 2,000 very different results, there might be 20,000 very different results, uh, right?
00:13:26.240 --> 00:13:39.840
It's it kind of goes through this next token prediction process, but um, that is why you have to really understand how these work and build up a uh context and good practices to get the best output.
00:13:40.240 --> 00:13:48.240
So um there's also in this, without going too deep, uh, there's a lot of nuance to this, right?
00:13:48.320 --> 00:13:52.639
Uh, because now these models can think uh very much like a human would.
00:13:52.799 --> 00:13:54.639
And we're gonna get to that a little bit later.
00:13:54.799 --> 00:13:58.879
But for the most part, a large language model is not a search engine.
00:13:58.960 --> 00:14:11.840
It is not a chatbot, it is a reasoning engine that combined with your data and your human knowledge can output economically viable work at a rate better and faster than humans.
00:14:12.000 --> 00:14:18.559
All right, so that is number one, what a large language model is and what it isn't.
00:14:18.720 --> 00:14:22.559
So now that we know what it is, where do we use it?
00:14:22.720 --> 00:14:35.840
All right, so uh also in the start here series, I had an entire uh episode on this concept of an AI operating system, but I think that's important to talk about uh because now they can all connect with your data.
00:14:36.000 --> 00:14:48.960
All right, and depending on what you go and uh you know what route you go down, I do think it's best to move to move most of your organization, most of your enterprise into one AI operating system, right?
00:14:49.120 --> 00:14:55.279
So uh personally, I'm using Chat GPT via codex more than anything.
00:14:55.440 --> 00:14:59.519
Uh, a lot of people are using Claude, whether that's on the web or the desktop.
00:14:59.600 --> 00:15:00.720
We're gonna get to that in a minute.
00:15:00.879 --> 00:15:09.840
There's also Gemini for those uh people who are uh really ingrained um in the uh kind of Google workspace uh ecosphere.
00:15:10.000 --> 00:15:19.840
And then obviously with Copilot, it's a little more difficult to explain because there's so many hoops to jump through in terms of permissions, role base, access control, all these things.
00:15:19.919 --> 00:15:30.639
It's a little convoluted, but um, I always tell you, I tell companies, try to move the majority of your day-to-day operations inside one of these AI operating systems.
00:15:30.799 --> 00:15:33.440
Uh, obviously, you should be building modularly.
00:15:33.679 --> 00:15:43.440
So when there are you know new kind of frontier breakthroughs uh or huge advantages, maybe a section of your team can take that context over to another one.
00:15:43.679 --> 00:15:48.720
Also, now every single big player makes it fairly easy to bring context over.
00:15:48.879 --> 00:15:57.120
So, whether that's things like skills and plugins, uh, your context, memory, chat history, etc., there's easy ways to port this over.
00:15:57.279 --> 00:16:02.799
It is more or less just a very um in-depth and detailed prompt, FYI.
00:16:02.960 --> 00:16:06.320
All right, so number two is you have to choose the operating system.
00:16:06.399 --> 00:16:08.960
Uh, number three, you have to choose the right surface.
00:16:09.120 --> 00:16:30.000
And this, at least in 2026, is where we've seen the most movement on, not necessarily what we saw in you know, mid-2024 to mid-2025, which is the features and the modes, the deep research is right, the you know, agentic capabilities with these reasoning models by default, the connectors, the apps, the surface is what's changing.
00:16:30.159 --> 00:16:31.360
So, what do I mean by that?
00:16:31.519 --> 00:16:34.879
In 2022, with the advent of ChatGPT, it was the web.
00:16:34.960 --> 00:16:36.639
We went and chatted, right?
00:16:36.799 --> 00:16:42.240
Then we started to bring our teams on board and you could collaborate with others and kind of a shared workspace.
00:16:42.399 --> 00:16:54.080
Uh, you know, and now in you know, late 2025 with Claude Code, uh, then Claude Co work in early 2026, and now Codex and everyone else, right?
00:16:54.159 --> 00:17:00.240
Google got on board with their Gemini desktop app with their anti-gravity, which I don't think is that great, but hopefully it'll get better, right?
00:17:00.399 --> 00:17:02.000
Now everything is moving to the desktop.
00:17:02.080 --> 00:17:11.519
Uh, and one of the reasons is is well, it can use um kind of your local machine to do a lot more uh faster and access your data.
00:17:11.680 --> 00:17:14.559
It can read and write on the desktop surface, which is huge.
00:17:14.640 --> 00:17:18.160
And then obviously, everything is agentic by default.
00:17:18.400 --> 00:17:23.599
So there's no easy answer to say, okay, well, should I still use the web version?
00:17:23.759 --> 00:17:25.599
Should I use the desktop version?
00:17:25.680 --> 00:17:38.160
Uh, you know, to get to that answer, you have to sit down and talk about what's important uh in terms of access, uh, in terms of automations, in terms of you know what you want in the cloud versus running locally.
00:17:38.319 --> 00:17:49.599
Actually, you know, although the that there's a lot of risk uh reward trade-off with a desktop version, in some instances it might be better for some companies that are uh, you know, privacy first.
00:17:50.079 --> 00:17:56.079
So, number three, uh, you know, is understanding that that surface dictates access, right?
00:17:56.400 --> 00:18:04.240
Um, so uh, because it's important to know that the agent tools, I mean, they can use your actual browsers, uh, right.
00:18:04.319 --> 00:18:12.480
And that's one of the biggest advantages, I think, to the desktop surface, aside from being able to read and write to your local folders, right?
00:18:12.559 --> 00:18:18.240
But to be able to use your actual browsers, I think Codec's built-in browser is so far ahead.
00:18:18.400 --> 00:18:23.440
Uh, computer use technology where it can literally, you know, use your computer, launch different programs.
00:18:23.599 --> 00:18:33.200
I have mine, you know, can go through and read my iMessage on my computer and you know, open up all these different desktop programs if there isn't a built-in app or connector.
00:18:33.440 --> 00:18:33.920
All right.
00:18:34.079 --> 00:18:36.799
Number four, choose the right model, people.
00:18:37.119 --> 00:18:39.119
Stop using a free plan.
00:18:39.680 --> 00:18:40.319
Period.
00:18:40.559 --> 00:18:44.559
If your company is like, yeah, use the free plan until we get this approved, don't do that.
00:18:44.720 --> 00:18:46.160
The risks are too high.
00:18:46.319 --> 00:18:46.640
All right.
00:18:46.720 --> 00:18:51.200
I don't care if you're using ChatGPT, Quad, Gemini, it doesn't matter.
00:18:51.359 --> 00:19:02.079
For the most part, you might get like one or two quote unquote prompts or outputs, uh, you know, with a decent enough uh plan if you're on uh a decent enough model.
00:19:02.160 --> 00:19:03.839
If you're on a free plan, don't do it.
00:19:03.920 --> 00:19:04.960
It's not worth it.
00:19:05.200 --> 00:19:11.200
Um, if you want real outputs, you have to use a paid model, right?
00:19:11.279 --> 00:19:14.640
All these people that are sharing things online, and you're like, oh, AI is so dumb, right?
00:19:14.720 --> 00:19:22.000
They're usually using an old model, a free model, um, and they really just don't understand uh kind of the underlying harness.
00:19:22.480 --> 00:19:31.920
So this is like, you know, you could have a convertible body, but oh, the engine is actually, you know, a dude on a bike, right?
00:19:32.079 --> 00:19:34.240
That's the equivalent of using a free model.
00:19:34.400 --> 00:19:37.839
You can't hide it in the fact that, like, oh, I'm using Claw.
00:19:38.400 --> 00:19:46.319
No, if you're using the free model, uh, right, for the most part, you probably don't even know because companies don't always tell you what version it is, right?
00:19:46.400 --> 00:19:50.000
Like, did you know that there's a GPD 5.4 latest?
00:19:50.240 --> 00:19:54.079
Like, no, you don't, because unless you're like me, you're not out there testing these things, right?
00:19:54.160 --> 00:19:56.640
You just think, oh, I'm using the best in the world.
00:19:56.799 --> 00:19:59.039
And you probably aren't if you're on a free plant.
00:19:59.119 --> 00:20:02.400
Uh, and you also need to be using thinking models.
00:20:02.559 --> 00:20:05.599
Do not use these models that don't think, don't reason.
00:20:05.759 --> 00:20:12.400
All right, you need to be using those always, even if that means waiting an extra 10, 20, 30 seconds.
00:20:12.559 --> 00:20:21.920
All right, I'm gonna get to this later, but read the chain of thought, do some exercise, have multiple you know, prompts running at the same time and you know, dictate, you know, move between.
00:20:22.000 --> 00:20:25.440
I do think the the job of the future is agentic orchestration.
00:20:25.599 --> 00:20:26.000
All right.
00:20:26.160 --> 00:20:32.079
So don't just, you know, go and chat with one, you know, free chatbot thread.
00:20:32.160 --> 00:20:33.519
At that point, you're wasting your time.
00:20:33.599 --> 00:20:39.680
And if you're making any business decisions based on using a free or a non-thinking model, I'm sorry.
00:20:39.839 --> 00:20:44.720
Uh, as the kids say, NGM, I'm not gonna make it.
00:20:44.880 --> 00:20:49.920
You, your department, your company, if you're doing that, if that's part of your actual strategy, don't.
00:20:50.160 --> 00:20:50.559
Right.
00:20:50.720 --> 00:20:51.759
Just just please don't.
00:20:51.839 --> 00:20:56.079
Unless you're just trying to learn like where all the buttons are, like how do I work this thing?
00:20:56.319 --> 00:20:58.319
That's the only time you should be doing that.
00:20:58.640 --> 00:20:59.279
All right.
00:20:59.519 --> 00:21:03.359
Uh, so number five, understanding the context layer.
00:21:03.599 --> 00:21:10.079
And this part is important before we even dive into the you know best practices of context engineering.
00:21:10.240 --> 00:21:15.599
But context is essentially everything that an AI model can see or not see, right?
00:21:15.759 --> 00:21:27.359
Uh, the the way that I've you know, we've taught uh you used to do like two live uh, you know, kind of prompt engineering courses a week for like two and a half years.
00:21:27.440 --> 00:21:29.680
Uh, and kind of the analogy I've always used.
00:21:29.759 --> 00:21:34.640
If if if you think back to the Star Wars credits from the original Star Wars, right?
00:21:34.799 --> 00:21:40.079
And it comes in and the words come in slightly slanted, and it says in a galaxy far, far away, right?
00:21:40.240 --> 00:21:45.359
And eventually the words they get smaller, smaller, smaller, and then they're off the screen.
00:21:45.599 --> 00:21:47.519
Think of that as a context window.
00:21:47.680 --> 00:21:50.240
Different models have different context windows.
00:21:50.400 --> 00:21:59.599
So essentially, you might be using a large language model, or you might have a skill saved, or a plugin, or a workflow saved, and you're using it repeatedly, and all of a sudden, you know.
00:22:00.160 --> 00:22:03.039
Started out great and all of a sudden it's like, wait, this stinks now?
00:22:03.200 --> 00:22:06.720
You know, oh, you know, this company must have nerfed the model.
00:22:06.880 --> 00:22:08.160
Maybe, but probably not.
00:22:08.319 --> 00:22:10.240
Probably you just ran over your context window.
00:22:10.400 --> 00:22:16.799
So that's essentially the amount of information like a computer's hard drive that a large language model can retain.
00:22:16.960 --> 00:22:17.279
All right.
00:22:17.359 --> 00:22:22.720
And a lot of times the first prompts, the first pieces of information you give a model are oftentimes the most important.
00:22:22.880 --> 00:22:24.319
It's where you give it direction.
00:22:24.559 --> 00:22:37.359
So the difference between a context window or understanding how context works in a hard drive is if you try to save a two gigabyte file on a full hard drive, it's just gonna say, nope, can't do it, no room.
00:22:37.519 --> 00:22:37.839
All right.
00:22:37.920 --> 00:22:42.880
A large language model is gonna do it and it's just gonna kick off something and you have no clue.
00:22:43.039 --> 00:22:54.960
So working with relevant, accurate business context is one of the most important things to do when working with a large large language model, uh, because essentially large language models get data from one of three places.
00:22:55.119 --> 00:22:59.519
Number one, their internal training data, which for the most part is very, very old.
00:22:59.680 --> 00:23:02.880
Uh, number two, any of your company data that you share.
00:23:02.960 --> 00:23:06.880
That can be, you know, apps, connectors, your file memory, chat history.
00:23:07.039 --> 00:23:16.240
You know, there's essentially a layer of your uh, you know, personal context, company context, you know, your all your SaaS apps that you use, right?
00:23:16.400 --> 00:23:17.599
And then there's the web.
00:23:17.920 --> 00:23:18.160
All right.
00:23:18.240 --> 00:23:23.200
So you have to understand that there's now this context layer that didn't really exist before, right?
00:23:23.279 --> 00:23:24.960
It used to just be the training data.
00:23:25.039 --> 00:23:28.400
And keep in mind, training data is usually really old.
00:23:28.559 --> 00:23:41.200
There's a knowledge cutoff date, but I would venture to guess the overwhelming majority, like 99% of the data that actually ends up in today's large language models is nowhere close to that date.
00:23:41.279 --> 00:23:52.880
So, you know, let's just say if there was a way to turn off web search and to turn off your company's context, and if you were to only rely uh on a model's training data, it would be very bad, right?
00:23:52.960 --> 00:23:59.839
It's actually crazy to think it took anthropic so long to add web search, and which is why I was like, companies should never use it, right?
00:23:59.920 --> 00:24:08.720
Because without web search, you know, you're probably playing with data at minimum that's probably on average 15 to 18 months old on the good side.
00:24:08.880 --> 00:24:14.559
And think of how quickly your company, uh, your competitive landscape, your sector moves.
00:24:14.720 --> 00:24:18.160
And imagine if you only had access to data that was 15 to 18 months old.
00:24:18.319 --> 00:24:19.759
Yeah, recipe for disaster.
00:24:19.920 --> 00:24:30.640
And that's why also you get these, you know, out outputs sometimes, especially, you know, pre, you know, 2024 reasoning models that could agentically call the web and pull your data.
00:24:30.799 --> 00:24:34.079
Why you'd get these outputs that looked like they were absolutely terrible?
00:24:34.240 --> 00:24:35.119
Well, because they were.
00:24:35.279 --> 00:24:40.319
It was relying on old data, and large language models are trained to be helpful assistance.
00:24:40.480 --> 00:24:40.799
All right.
00:24:40.960 --> 00:24:53.839
So make sure you go back and listen to the start here series going over the, I think we called it the seven deadly sins of AI, uh, where we kind of talked about some of these things, but that's why sometimes outputs from large language models seem generic.
00:24:53.920 --> 00:24:57.279
And you're like, is this actually truthful or is it kind of lying to me?
00:24:57.359 --> 00:25:07.200
Well, the context helps steer that away from just giving you these general, it sounds helpful, but is it true, to being pinpoint specific and valuable for your company.
00:25:07.440 --> 00:25:09.680
Number six, context engineering.
00:25:09.839 --> 00:25:18.640
All right, this obviously plays in line with the context layer, but this is where you insert right one of those three layers of data into the context window.
00:25:18.720 --> 00:25:22.559
Uh, so this can be things like your traditional prompt engineering.
00:25:22.640 --> 00:25:26.160
So you can share that context via your prompting, uh, right.
00:25:26.240 --> 00:25:31.279
But the important thing uh really is knowing that it works in layers.
00:25:31.440 --> 00:25:33.839
So I'm gonna give you a couple hints, right?
00:25:34.000 --> 00:25:36.000
So we taught prime prompt polish.
00:25:36.160 --> 00:25:44.559
All right, the 30-second version of that is you should always prime a model, work with it a lot before you prompt it and ask it for an output.
00:25:44.640 --> 00:25:54.640
And then polish, we're actually gonna get to that technically in number nine and 10, because the first output that a model gives you is usually garbage, even if you do everything else correctly.
00:25:54.880 --> 00:25:57.119
Um, all right, refine queue.
00:25:57.200 --> 00:26:04.720
That's where we talk about, you know, role, examples, fetch, insights, narration, um, explanation, and questions.
00:26:04.880 --> 00:26:05.279
All right.
00:26:05.359 --> 00:26:09.759
And if you want to go more on the, you know, prime prompt polish, refine queue, 555.
00:26:10.079 --> 00:26:16.880
This is like a decision tree on when you should use um a large language model versus when you should do it, quote unquote, manually without AI.
00:26:17.200 --> 00:26:23.599
Actually, when you do go to starthirseries.com uh and sign up for our community, uh, you will also get access.
00:26:23.759 --> 00:26:25.119
Yeah, a lot of people don't know this.
00:26:25.200 --> 00:26:26.640
Uh, we recorded that.
00:26:26.720 --> 00:26:27.599
It's on demand, right?
00:26:27.680 --> 00:26:28.400
I used to do it live.
00:26:28.480 --> 00:26:31.359
I can't do it live anymore, but I keep the videos pretty updated.
00:26:31.440 --> 00:26:34.880
So you can go take our updated prime prompt polish course.
00:26:35.039 --> 00:26:38.000
I think it was updated as of like GPT 5.4.
00:26:38.079 --> 00:26:41.440
So I'll probably update it once we get like five, six or something.
00:26:41.599 --> 00:26:46.960
Uh, but the uh the basic, this is basic context engineering 101.
00:26:47.119 --> 00:26:47.440
All right.
00:26:47.519 --> 00:27:06.319
So it can be as simple as you know, doing the role, goal, sources, constraints, examples, whatever it is, but it's making sure that through your words that you're sharing, you are steering the model in the right direction and telling it what context to use, what output you need, and working with it before you just expect it to give you amazing outputs.
00:27:06.559 --> 00:27:07.039
All right.
00:27:07.279 --> 00:27:10.720
Number seven, working with files, apps, and company data.
00:27:10.880 --> 00:27:12.480
This one is huge.
00:27:12.720 --> 00:27:17.359
So files make AI far more specific to the actual task.
00:27:17.519 --> 00:27:17.759
All right.
00:27:17.839 --> 00:27:40.799
And a lot of people think oh, if I just dump all of my context into one, one of these large language models and tell it, you know, okay, now go, you know, write, you know, next quarter's press releases and update all our job descriptions based on the SOPs and right, all you're doing in this instance without getting too technical and accidentally turning this again into a 90-minute show, um, these think of them as books on the bookshelf, right?
00:27:40.960 --> 00:27:54.319
You still have to tell the model when, via whether that's through a skill, uh a plug-in, uh, a certain workflow, automation, etc., you still have to tell it when to use the book and sometimes direct it to look into what chapter.
00:27:54.559 --> 00:28:07.839
But uh creating these um, you know, app connections are huge because depending on if you're talking about, you know, Chat GPT, Claude, or Gemini, as an example, they all treat these connectors a little bit differently.
00:28:08.000 --> 00:28:15.279
As an example, some of them might index or cache some of your most important connectors, like your Gmail, right?
00:28:15.359 --> 00:28:17.839
So you don't even have to wait, it just knows.
00:28:17.920 --> 00:28:25.440
And it can actually actually uh, you know, proactively go out and find these things and surface them to you, which is actually huge.
00:28:25.680 --> 00:28:41.680
So, you know, whether we're talking about uploading files, um, you know, connecting uh a connector or an app, right, like a Google Drive or a Microsoft um SharePoint, OneDrive, Box, etc., your CRM, right?
00:28:41.759 --> 00:28:50.000
Uh, most of the big three, I don't know if Google does, but I think mostly everyone else has like HubSpot and some of the big names and CRMs, right?
00:28:50.240 --> 00:28:53.759
All of these things now, for the most part, have read and write.
00:28:53.839 --> 00:29:00.240
And that is big, and that's much different than where we were at, you know, in the early days of connectors, like nine months ago.
00:29:00.319 --> 00:29:00.880
Uh, right.
00:29:00.960 --> 00:29:04.319
So now these can, well, run actions for you.
00:29:04.480 --> 00:29:23.039
So, you know, let's just say you get some information off of a Google Sheet, um, and you have to go, you know, update uh something in a database, and then you have to go, you know, change some options on a CRM, and then you have to update your project management tool, click up as an example, right?
00:29:23.200 --> 00:29:27.440
These are all processes that normally I call the human duct tape that you would have to do.
00:29:27.599 --> 00:29:31.599
Now these connectors can talk to each other and carry that context over.
00:29:31.839 --> 00:29:35.680
We did a show called Agentic Context Carry in the Start Here series.
00:29:35.759 --> 00:29:38.640
So go back and listen to that if you want to know more about how that works.
00:29:38.799 --> 00:29:54.480
But essentially, this is why these models with the context um you know layer in this in agent being able to carry that context and keep it in the context window is huge because this is the majority of what so many knowledge workers spend their time doing.
00:29:54.640 --> 00:30:00.480
You know, you have the five to 15 probably different SaaS applications that you spend your time in.
00:30:00.640 --> 00:30:05.440
You use your brain, your role, your requirements, your KPIs.
00:30:05.680 --> 00:30:10.640
You uh grab all of that important context from all of these different, you know, apps.
00:30:10.880 --> 00:30:15.920
You, you know, personalize it, create some value out of it, and then you piece it together in all of these other apps.
00:30:16.079 --> 00:30:17.839
This is what agents do now, right?
00:30:17.920 --> 00:30:22.400
So working with files, apps, connectors, and your company data is huge.
00:30:22.559 --> 00:30:23.200
All right.
00:30:23.599 --> 00:30:26.799
Number eight, and this is important, especially following up.
00:30:26.880 --> 00:30:28.960
Uh, maybe it should have been number seven, right?
00:30:29.200 --> 00:30:32.880
Uh don't do the shadow IT, right?
00:30:33.039 --> 00:30:39.680
Shadow IT is when you don't have permission or you know, shadow AI or shadow agents, whatever you want to call it.
00:30:39.759 --> 00:30:43.279
When you don't have permission to use your company data, you shouldn't be uploading it, right?
00:30:43.359 --> 00:30:55.599
But the good thing is, whether we're talking about Claude Enterprise, Gemini Enterprise, Chad GPT Enterprise, they have the exact same data security and privacy uh that you would have from your cloud provider.
00:30:55.759 --> 00:31:05.119
So if you upload your company's uh data to a cloud, it's the exact same thing.
00:31:05.359 --> 00:31:09.039
It is the exact same thing as using these connectors in these apps.
00:31:09.119 --> 00:31:18.079
As long as you're going through, you know, turning off model training, doing the you know, one-on-one best practices, um, you're not running any additional risk.
00:31:18.240 --> 00:31:28.640
Um, again, that's a big if as long as you're you know using these models correctly, turning off uh, you know, training data, which most of the paid uh business or enterprise versions do by default.
00:31:28.960 --> 00:31:32.000
You're not running any additional risk with your company's data.
00:31:32.160 --> 00:31:40.480
It's not like I'm gonna upload, you know, or connect my data, turn off model training, and my competitor is gonna be like, oh, what's everyday AI doing?
00:31:40.559 --> 00:31:41.440
That's not how it works.
00:31:41.519 --> 00:31:45.039
But people, smart people with brains still think that today.
00:31:45.200 --> 00:31:46.799
That's not how it works, right?
00:31:46.960 --> 00:31:50.000
Uh, governance is also extremely important.
00:31:50.160 --> 00:31:52.559
That is permission design, right?
00:31:52.640 --> 00:31:53.759
This is not paperwork.
00:31:53.839 --> 00:31:57.839
This is not a one-time 20-minute training because the systems are constantly changing.
00:31:58.079 --> 00:31:58.400
All right.
00:31:58.480 --> 00:32:03.359
Also, personal accounts, those aren't like unofficial company AI systems.
00:32:03.440 --> 00:32:07.279
Like I said, uh, you know, you should only be connecting your company's data.
00:32:07.440 --> 00:32:17.680
Number one, if you've gone through the guardrails, the governance, uh, the privacy permissions, getting the sign-off from everyone involved, but you should only also only be using it through the proper channels, right?
00:32:17.839 --> 00:32:20.400
I can't tell you the number of companies when they're telling the truth.
00:32:20.480 --> 00:32:27.279
They're like, okay, well, yeah, we we got our, you know, copilot approved, but you know, no one can get the right access to it.
00:32:27.519 --> 00:32:32.880
But well, if we're using SharePoint and copilot, we should just be able to use it in ChatGPT, right?
00:32:33.039 --> 00:32:37.279
Well, in theory, kind of, but also absolutely not, right?
00:32:37.440 --> 00:32:40.480
That is going against governance 101.
00:32:40.799 --> 00:32:54.079
So, you know, expert-driven loops are so important when we talk about, you know, properly setting up these guardrails, especially as these large language models, these quote unquote AI chatbots by default, can now take actions for us.
00:32:54.160 --> 00:33:03.359
And we can schedule agents to go take actions, right, with right W-R-I-TE, right capabilities, right?
00:33:03.599 --> 00:33:07.279
It can go off and email customers and update your CRM.
00:33:07.440 --> 00:33:16.240
And, you know, maybe if you're not, you know, keeping an expert-driven loop, it could potentially make a huge mistake that can cost your organization a lot of money.
00:33:16.400 --> 00:33:26.000
So there's obviously a downside to productivity and efficiency uh if you skip over step number eight, which is proper privacy, permissions, and governance.
00:33:26.480 --> 00:33:27.039
Yeah.
00:33:27.200 --> 00:33:32.880
Uh so after permissions, though, leaders need visibility into what actually happened.
00:33:32.960 --> 00:33:39.279
And that's where we have to talk about the importance of transparency, observability, and reasoning artifacts.
00:33:39.599 --> 00:33:49.200
So this is where uh people skip over showing your work uh and being able to observe that as a team.
00:33:49.279 --> 00:33:49.680
Uh, right.
00:33:49.759 --> 00:34:01.440
So I can't go into all the specifics because it's gonna take a while, but especially when you get up to enterprise accounts, not only can you see, you know, usage and things like that, you can see uh you're right, I think Copilot actually leads in this.
00:34:01.519 --> 00:34:07.920
Uh, you know, I think with their uh intra ID, you can see every action certain uh, you know, agents make.
00:34:08.000 --> 00:34:15.599
I think ChatGPT's new workspace agents uh for uh business and enterprise uh teams do a great job of doing that as well.
00:34:15.760 --> 00:34:24.639
But you have to be able to understand and show the work, especially when agents are uh or sorry, AI chatbots are agentic by default.
00:34:24.800 --> 00:34:29.920
You have to be able to uh observe them and understand the reasoning artifacts.
00:34:30.000 --> 00:34:30.719
What does that mean?
00:34:30.880 --> 00:34:36.400
You know, there's usually a little thing you can click, accept in Gemini, Gemini, can we get a little better at that, please?
00:34:36.559 --> 00:34:37.199
Uh right.
00:34:37.360 --> 00:34:55.039
I know there's a tiny step that we got uh with 3.5 flash and the you know updated version of Gemini, but we need complete transparency when it comes to seeing how models get from context engineering and great setups and great workflows to creating this great economically valuable work.
00:34:55.119 --> 00:34:58.400
And the reason being is because models change all the time.
00:34:58.480 --> 00:35:17.280
So if you can't properly see every single step in every tool call, in every website that a model or an agent went to in order to deliver that, you know, first draft deliverable for your team that you ultimately sent off to a client or the RFP or whatever it is, then you don't have the observability.
00:35:17.440 --> 00:35:22.400
Then you don't actually own that asset because you don't know what's happening, right?
00:35:22.480 --> 00:35:34.719
So I like to talk about the steps in the middle are sometimes the most overlooked, but they're the most important because if you don't understand what's going on, you don't own it, right?
00:35:34.800 --> 00:35:35.599
So what does that mean?
00:35:35.760 --> 00:35:44.559
If something happens in the harnessing or a model changes, or you know, there's a drastic switch up in how a model uses different tools.
00:35:44.719 --> 00:35:53.199
If you don't own and understand that observability, the the reasoning, uh, you know, artifacts, then you may not be able to replicate those same things.
00:35:53.280 --> 00:36:10.400
And if they if these things are running on loops scheduled, which is very easy to do, and one little thing changes, it can break your entire business if you become overly reliant on passive, lazy, human-in-the-loop workflows versus expert-driven loops.
00:36:10.559 --> 00:36:15.119
Expert-driven loops means that you are looking at the transparency and observability.
00:36:15.360 --> 00:36:16.000
All right.
00:36:16.480 --> 00:36:21.760
Last but not least is verification, iteration, and creating those workflows, right?
00:36:21.840 --> 00:36:23.280
So, what do I mean by this?
00:36:23.519 --> 00:36:26.480
Uh iteration is huge.
00:36:26.639 --> 00:36:29.519
So I talked about, you know, the prime prompt polish.
00:36:29.599 --> 00:36:30.639
This is the polish.
00:36:30.719 --> 00:36:42.800
You know, your first output, even if you do everything else, steps one through nine correctly, your first output for the most part from a large language model, even a great one, is at best case, it's gonna be generic garbage, right?
00:36:42.880 --> 00:36:44.079
But it's not gonna be very well.
00:36:44.239 --> 00:36:45.199
It's not gonna sound like you.
00:36:45.280 --> 00:36:48.000
It's even when you give it the proper examples, right?
00:36:48.239 --> 00:36:52.719
Um, you have to really iterate on the output.
00:36:52.880 --> 00:36:57.920
You have to understand how it got to that output and then run it over and over and make it better.
00:36:58.079 --> 00:37:03.039
That is this iteration loop where you turn, you know, your first output, right?
00:37:03.199 --> 00:37:08.480
My old journalism days, the first version of the story you turned in was never really good or the best, right?
00:37:08.559 --> 00:37:14.159
You always had to go through the multiple editors and all the the red, all the red lines to make it better.
00:37:14.320 --> 00:37:17.920
And then once you get to that point, you need to then turn it into a workflow.
00:37:18.000 --> 00:37:21.920
And this looks a little bit different in Chat GPT versus Claude versus Gemini, right?
00:37:22.000 --> 00:37:26.480
But for the most part, we are getting these universally um applicable skills.
00:37:26.639 --> 00:37:31.119
Another thing, Anthropic uh led the way on this, and everyone else is adopting them.
00:37:31.280 --> 00:37:40.000
But skills are essentially ways that at the end of iterating, you can go through and essentially, in more words than this, say turn this into a skill, right?
00:37:40.079 --> 00:37:48.000
And then once you turn it into a skill, then you can schedule it because uh most of these systems have scheduling by default.
00:37:48.239 --> 00:37:51.119
So um it's all about verify that it works.
00:37:51.280 --> 00:38:06.079
You know, make sure you're using the right model, the right mode for the right tool, you're getting the right outputs, you're verifying it, you're going through the transparency, observability, the reasoning artifacts, but then you don't stop when you get that first, you know, hey, this draft is good enough.
00:38:06.239 --> 00:38:07.519
No, keep working.
00:38:07.599 --> 00:38:09.039
That's where you polish.
00:38:09.199 --> 00:38:14.559
You turn it into a skill, you turn it into a plugin, and then you turn it to an automated workflow.
00:38:14.639 --> 00:38:21.199
And that's how we go from working with an AI chatbot to commanding AI agents.
00:38:21.360 --> 00:38:21.760
All right.
00:38:22.000 --> 00:38:24.719
I hope this one was helpful, y'all.
00:38:24.800 --> 00:38:25.599
If so, let me know.
00:38:25.679 --> 00:38:28.079
If you're listening on the podcast, please do me a favor.
00:38:28.320 --> 00:38:31.360
All right, uh, this one has been a long time in the waiting.
00:38:31.440 --> 00:38:36.239
So, you know, I hope more people can hear this because I want to cut through the BS.
00:38:36.400 --> 00:38:44.159
I want to tell people exactly how these models work, uh, and I want to be able to do it for as for free as long as I can, right?
00:38:44.320 --> 00:38:47.519
Uh, so that only happens if you're number one sharing this with others.
00:38:47.599 --> 00:38:52.719
So if you are listening um, you know, on LinkedIn as an example, please repost this if this was helpful.
00:38:52.800 --> 00:38:56.239
If you are listening on the podcast, please subscribe to the show.
00:38:56.400 --> 00:38:57.679
That would mean a lot.
00:38:57.920 --> 00:38:59.440
Leave us rating too, if you can.
00:38:59.599 --> 00:39:00.719
So I hope this was helpful.
00:39:00.880 --> 00:39:08.400
Like I said, make sure to go to starthirseries.com to get exclusive access to all of the episodes in this series.
00:39:08.559 --> 00:39:09.360
Thank you for tuning in.
00:39:09.440 --> 00:39:12.320
I hope to see you back tomorrow in everyday for more everyday AI.
00:39:12.400 --> 00:39:13.440
Thanks, y'all.
