WEBVTT
00:00:00.320 --> 00:00:02.959
Welcome to the Everyday AI podcast.
00:00:03.120 --> 00:00:09.439
My name is Jordan Wilson, and for the past three and a half years, we put out more than 800 episodes.
00:00:09.599 --> 00:00:13.759
Yet one of the most common questions I get, I didn't really have an answer for.
00:00:14.000 --> 00:00:17.039
Where do I start on the Everyday AI podcast?
00:00:17.280 --> 00:00:19.760
And that's why we started the Start Here series.
00:00:19.920 --> 00:00:30.239
And with Fall now back in full swing, the Everyday AI podcast is going back to school and playing back the entire Start Here series from front to back.
00:00:30.399 --> 00:00:37.759
We've hit pause on our normal Monday to Friday programming to run back our most popular series ever for the next 30 days.
00:00:37.920 --> 00:00:42.000
We made the Start Here series for beginners and AI champions alike.
00:00:42.159 --> 00:00:53.359
So whether you're just trying to get a grasp on large language models or grappling with the best coding harness for multi-agentic workflows, the Start Here series covers it all.
00:00:53.600 --> 00:00:58.079
Plain language, no jargon, and easy to follow along each day.
00:00:58.320 --> 00:01:03.679
So make sure to subscribe to the podcast and check back each day for new insights day by day.
00:01:03.840 --> 00:01:11.040
The series is a culmination of spending more than 10,000 hours covering generative AI over the past three and a half years.
00:01:11.200 --> 00:01:14.799
So you don't want to miss a single episode of the Start Here series.
00:01:14.879 --> 00:01:16.159
Let's get into it.
00:01:18.239 --> 00:01:23.040
This could randomly be one of the most important shows I've ever recorded.
00:01:23.359 --> 00:01:24.000
Why?
00:01:24.400 --> 00:01:30.319
Well, I'm laying out the AI answers for your company for the entire year right here.
00:01:30.719 --> 00:01:31.680
Let me explain.
00:01:31.840 --> 00:01:37.599
Usually in December or January, I put on my AI predictions hat and tell you what's ahead.
00:01:37.760 --> 00:01:44.560
And for the past three years, I've done a pretty okay job in these prediction shows at telling you what's next.
00:01:44.799 --> 00:01:47.680
But the pace of AI is much different now.
00:01:47.920 --> 00:01:55.760
I mean, measurably, AI models are literally coming out about two to three times faster today than they were just nine months ago.
00:01:56.000 --> 00:02:02.000
So we can't wait until another four months in December or January to give you the roadmap ahead.
00:02:02.239 --> 00:02:08.960
So with the school year starting and with this upcoming start here series, we're about to restart from the beginning.
00:02:09.120 --> 00:02:11.840
I thought I'd just give you the answers today.
00:02:12.000 --> 00:02:15.280
Not the syllabus, not the take-home test, just the answers.
00:02:15.520 --> 00:02:24.719
So 19 quick points for you and your team to focus on, whatever the sector, whatever point you're at in your career, whatever type of work that you use AI for.
00:02:24.879 --> 00:02:28.719
I think today's show can be your answer sheet for the whole year.
00:02:28.960 --> 00:02:34.719
So you and your company are focusing on what's next, not what already happened.
00:02:34.960 --> 00:02:36.159
Let's get into it.
00:02:36.319 --> 00:02:38.719
If you're new here, welcome to Everyday AI.
00:02:38.800 --> 00:02:41.120
My name's Jordan Wilson, and well, we do this every day.
00:02:41.199 --> 00:02:52.159
It's your daily, unedited, unscripted live stream podcast, and free daily newsletter, helping business leaders like you and me not just keep up with what's happening in the AI world because it's too much for anyone human.
00:02:52.319 --> 00:02:56.719
But I help you get ahead to know what's coming so you can grow your company and your career.
00:02:56.879 --> 00:03:02.639
So it starts here, but make sure you go to our website at your everydayai.com, sign up for the free daily newsletter.
00:03:02.800 --> 00:03:06.960
We're we'll be recapping today's show and a whole lot more.
00:03:07.599 --> 00:03:13.919
So, as a reminder, we are taking a kind of break in our normal programming.
00:03:14.159 --> 00:03:23.199
So, since it is that back to school time and all the kids are going back to school, we're all gonna be doing some learning together on everyday AI.
00:03:23.439 --> 00:03:29.520
So, our start here series uh show that we kicked off earlier this year.
00:03:29.759 --> 00:03:39.280
We've been doing, you know, one or two shows a week, and we have 30 of them, and they're pretty amazing, and they're some of our most popular shows ever.
00:03:39.439 --> 00:03:45.680
So we're gonna be running them all back, the entire series starting Monday with the very first episode.
00:03:45.840 --> 00:03:49.280
So, yeah, a little break from our normal Monday to Friday schedule.
00:03:49.360 --> 00:04:03.199
But I guarantee you, if you just stick with the series from volume one, listen all the way to volume 30, whether you listen to one, some, or none of them, you're gonna know more about AI than like 99% of the other people at your company.
00:04:03.360 --> 00:04:06.719
So the series cuts through the jargon, cuts through the fluff.
00:04:06.800 --> 00:04:11.120
Whether you're a beginner, you around AI all the time, just it's simple.
00:04:11.199 --> 00:04:16.000
No computer science degree required, but class starts Monday, don't be late.
00:04:16.240 --> 00:04:23.279
All right, but let's now get into it and talk about those 19.
00:04:23.600 --> 00:04:26.079
Yeah, we have 19 of them.
00:04:26.240 --> 00:04:31.279
Uh so these are the things that I think your company needs to be focusing on.
00:04:31.439 --> 00:04:31.680
Yeah.
00:04:31.759 --> 00:04:33.439
So for the school year here, right?
00:04:33.600 --> 00:04:36.240
So, yeah, classes and section set uh session.
00:04:36.399 --> 00:04:42.879
That's why we have the little uh, you know, yellow, the pencil yellow uh accent color here for my slides.
00:04:43.120 --> 00:04:47.279
So uh these are in no real order, right?
00:04:47.439 --> 00:04:49.519
I was just kind of jotting down some notes.
00:04:49.680 --> 00:05:06.959
Um, you know, sometimes I just uh am always keeping not like predictions, but a lot of my observations, because uh from doing this for three and a half years, literally spending my entire day every day uh reading about AI, exploring AI, talking to other very smart people about AI, right?
00:05:07.120 --> 00:05:12.399
I have a what I would say is a pretty accurate notebook of random ideas and thoughts.
00:05:12.480 --> 00:05:14.639
And I've been spending some time putting them all together.
00:05:14.800 --> 00:05:16.319
I came up with 19, right?
00:05:16.399 --> 00:05:23.360
So, you know, for my yearly prediction shows, I always do, you know, for 2024, it was 24, then 25, then 26.
00:05:23.439 --> 00:05:24.399
Well, we got 19.
00:05:24.560 --> 00:05:24.879
All right.
00:05:24.959 --> 00:05:27.839
So I don't know, maybe that's roughly how far we are through the year.
00:05:27.920 --> 00:05:30.879
But this is this is we're lining this up with the school year, right?
00:05:30.959 --> 00:05:32.000
We're starting in August.
00:05:32.160 --> 00:05:38.639
I think these should all still be fairly uh relevant, come, you know, when May, when school gets out.
00:05:38.800 --> 00:05:42.000
Uh, but let's start with number 19.
00:05:42.720 --> 00:05:49.600
So this is when reactive chat is dying and proactive agents are taking over.
00:05:49.759 --> 00:05:50.160
All right.
00:05:50.319 --> 00:05:57.600
And I also want you to think, right, um, where are you at on the AI usage scale?
00:05:57.759 --> 00:05:59.680
Because maybe what I said is a shock.
00:05:59.759 --> 00:06:02.319
Maybe you're like, duh, it's been this way for a while, right?
00:06:02.399 --> 00:06:07.279
So as I rattle these off, right, keep in mind our audience is all over the place, right?
00:06:07.439 --> 00:06:16.959
We literally have AI leaders at the labs listening to this, and then we have people that have stumbled upon ChatGPT for the very first time and have never used it.
00:06:17.040 --> 00:06:19.040
All right, so keep that in mind as I walk through these.
00:06:19.199 --> 00:06:24.959
So maybe you've heard me mention some of these things before, maybe you don't know what some of them are, so I'm gonna try to keep it simple.
00:06:25.279 --> 00:06:32.879
But I think over the next year, uh reactive chat is gonna start dying, right?
00:06:33.040 --> 00:06:42.000
It's the combination of not just uh agents, but just being able to uh very easily save and reuse skills.
00:06:42.240 --> 00:06:43.920
That is huge, right?
00:06:44.160 --> 00:06:50.959
Um, I think that chat GPT work as an example is really indicative of this.
00:06:51.120 --> 00:06:55.920
It's it's no different necessarily than uh codecs, right?
00:06:56.000 --> 00:07:07.360
And codecs was kind of the, I would say, the uh one of the birthplaces um of the hardworking around the clock agent to that as well as uh Claude Code and Claude Cowork.
00:07:07.439 --> 00:07:14.959
Uh but I think the new normal is just gonna be, you know, waking up and reviewing what your AI already shipped overnight.
00:07:15.040 --> 00:07:18.480
Um, and and and that piece is extremely important.
00:07:20.399 --> 00:07:29.759
All right, so if you're prompting less than what are you doing, all right, that's number 18, you're talking more.
00:07:29.920 --> 00:07:40.240
Uh, I think that number 18 kind of uh prediction, 18 here, I think voice and mobile are already becoming the default AI controls.
00:07:40.639 --> 00:07:53.279
Uh so if you haven't used uh Chat GPT uh's new feature yet, my gosh, uh Chat GPT voice is outstanding.
00:07:55.279 --> 00:07:59.439
What's crazy is this has been out for almost a month.
00:07:59.680 --> 00:08:08.079
And you know, normally the people that are quote unquote early talking about AI are starting to talk about it literally now.
00:08:08.240 --> 00:08:21.759
Um, but I think you know, um, to summarize this, I did a show on it a couple of weeks ago when it did uh first come out, but this is essentially kind of the combination of a voice and mobile becoming the default AI control.
00:08:22.240 --> 00:08:24.959
Uh, yeah, you can go out and touch grass.
00:08:25.120 --> 00:08:31.439
And um, in in in my experience, I get a lot more done now when I take my iPhone out, right?
00:08:31.519 --> 00:08:34.559
And this is another, you know, open AI slash chat GPT.
00:08:34.639 --> 00:08:39.679
And I think you know, uh Claude has made some improvements to theirs, it's still not very good.
00:08:39.840 --> 00:08:49.840
Um, but I assume everyone is gonna have to follow suit because this is uh so so powerful um that you just have to have it, right?
00:08:50.000 --> 00:08:54.960
Even I think we saw a glimpse of it with you know Grok Grok bot that just came out.
00:08:55.120 --> 00:09:09.840
Um, but to be able to talk um to a uh full duplex agent, that means an agent that can listen and talk and uh at the same time, but one that also has access to your data and can look up information.
00:09:10.000 --> 00:09:14.000
It's literally faster than talking to a coworker, right?
00:09:14.159 --> 00:09:19.519
Um, because I can talk back and forth and it will respond to me.
00:09:19.679 --> 00:09:28.960
But then also while it's responding to me, it's gonna, you know, uh personalize everything through the data that it has connected, as well as go and look things up.
00:09:29.120 --> 00:09:36.480
So, you know, for me, this is like, you know, working with uh someone that's smarter than the smartest coworker I've ever worked with, right?
00:09:36.720 --> 00:09:46.240
But I think that this has to become eventually uh the main surface uh fairly soon, just because it is that good, right?
00:09:46.320 --> 00:09:48.960
And um, so you're like, okay, what does this mean?
00:09:49.200 --> 00:09:51.360
You know, just isn't that just mobile?
00:09:51.519 --> 00:09:57.600
Not necessarily, uh, because the desktop is obviously much more powerful than what you can accomplish on any mobile phone.
00:09:57.759 --> 00:10:08.399
But the way that this works, at least in ChatGPT voice, is well, you can open the ChatGPT app, click on remote, and then literally control your entire computer, right?
00:10:08.559 --> 00:10:19.200
So this is the Jarvis of AI where it can see what's on your computer, it can click things, it can open, save, file, shut, run commands, it can do anything.
00:10:19.279 --> 00:10:29.919
Uh, but I think that that is if if if you and your company have not already started exploring this, you should probably have your eyes on this because it is a much more fluid and natural way to work.
00:10:30.080 --> 00:10:32.159
I like to say it's kind of like a flow state.
00:10:32.320 --> 00:10:33.039
Um, right.
00:10:33.200 --> 00:10:46.159
I don't talk about this a lot, but sometimes, and maybe this is because I'm a former journalist, but sometimes when I'm a hands-on keyboard, maybe I'm trying to get things uh a little too right.
00:10:46.399 --> 00:10:50.559
You know, sometimes I don't just like to uh blurt out on the keyboard.
00:10:50.639 --> 00:10:58.879
Um, and that might interrupt kind of the natural flow state of what I actually want to communicate uh to an AI or to an agent.
00:10:59.120 --> 00:11:02.720
Obviously, when I'm talking, it's a little easier, I think, ultimately.
00:11:02.879 --> 00:11:04.799
And I think that holds true for a lot of people.
00:11:04.960 --> 00:11:15.600
There's actually some studies that show that you can much more effectively uh communicate your uh point when talking to AI agents while you're talking versus when you're texting because uh uh or typing.
00:11:15.679 --> 00:11:26.000
Because when you're typing, you're already putting you know multiple limiting filters on it uh because you know that you can only type so fast and your uh fingers cannot keep up with your brain.
00:11:26.080 --> 00:11:28.240
You know, your mouth can do a little bit better of a job.
00:11:28.399 --> 00:11:31.120
But I think number 18, that is definitely the next control.
00:11:31.759 --> 00:11:35.840
Number 17, uh, this is the manager of threads.
00:11:36.080 --> 00:11:41.759
So we've already seen this in um codecs and chat GPT work.
00:11:42.000 --> 00:11:47.279
Uh we've seen a smaller version of this that was uh actually announced.
00:11:47.360 --> 00:11:48.799
Uh, I'm losing track of time.
00:11:48.879 --> 00:11:52.080
I think it was this week uh from Claude, right?
00:11:52.320 --> 00:11:54.080
But what does this mean?
00:11:54.240 --> 00:12:06.960
Um, I think we were trained for better or worse when working, whether inside, you know, Copilot, Gemini, ChatGPT, uh, Claude, etc., you know, that any conversation had its home, right?
00:12:07.120 --> 00:12:10.480
So, hey, I need to go back into that chat, right?
00:12:10.559 --> 00:12:14.639
And this was especially more important in the earlier days of prompt engineering.
00:12:14.720 --> 00:12:26.480
This is something I taught to thousands of students that took our free PPP course because this was the constraint, you know, probably up until a year ago, is the context really just lived in that single chat thread.
00:12:26.639 --> 00:12:28.240
It's not like that anymore, right?
00:12:28.320 --> 00:12:33.360
So now I think you have to think of manager threads replaces those one-off chats.
00:12:33.519 --> 00:12:38.879
So you could call those like a chief of staff, uh chief of staff um thread, right?
00:12:39.039 --> 00:12:42.240
So that's the way I think a lot of people set it up in a project.
00:12:42.399 --> 00:12:46.399
You can have, you know, an agent's MD file in any project.
00:12:46.559 --> 00:12:56.240
Um, and you can create an unlimited amount of threads or chats in there, but then any thread in uh you know, codex and chat GPT can control any other thread.
00:12:56.320 --> 00:13:13.120
And again, uh that is an interface that all uh you know, Google, Claude, Grok, everyone else, you know, has either you know taken that as well, or they will have to soon, because that is um, you know, the the next step in terms of communication.
00:13:13.279 --> 00:13:16.320
And I think that's that's a smart way to do it, right?
00:13:16.399 --> 00:13:18.559
That's how it is with humans, right?
00:13:18.720 --> 00:13:25.919
If if you're sitting with you know your team, uh you can talk to them at any point about anything that you all know.
00:13:26.080 --> 00:13:28.320
You don't have to wait for the marketing meeting.
00:13:28.399 --> 00:13:30.720
If you're sitting in the marketing team, is there?
00:13:30.799 --> 00:13:31.679
You can just talk about it.
00:13:31.759 --> 00:13:36.480
So this is a much more natural way to uh interface uh with AI agents.
00:13:37.120 --> 00:13:37.759
All right.
00:13:37.919 --> 00:13:43.279
Um, and you know, once those threads are managing threads, the next step is obvious.
00:13:43.440 --> 00:13:46.480
Well, you put people and agents in the same room.
00:13:46.559 --> 00:13:49.919
So that is number uh 16 on the list here.
00:13:50.320 --> 00:13:55.679
Multiplayer AI with humans and agents sharing the same workspace.
00:13:55.840 --> 00:13:58.879
And over the past two weeks, we've seen some great examples of that.
00:13:58.960 --> 00:14:00.799
I already mentioned Grokbot.
00:14:00.960 --> 00:14:14.960
Um, you know, so if you haven't seen this, it's only been out for like not even 72 hours yet, uh, right, but very similar to Buzz from Block, uh, you know, Jack Dorsey of uh Twitter fame.
00:14:15.200 --> 00:14:16.159
Uh, but the same thing.
00:14:16.240 --> 00:14:30.960
So we've seen these two recent um interfaces come out that are multiplayer by default, so where you and your entire team uh can come in and chat with the same agent, uh, right?
00:14:31.039 --> 00:14:33.440
So we've seen glimpses of this, right?
00:14:33.519 --> 00:14:37.440
There was a Google leak uh about bringing some multiplayer capabilities.
00:14:37.600 --> 00:14:43.519
Um, obviously, you know, open AI uh inside ChatGPT, you know, they've had these group chats.
00:14:43.600 --> 00:14:46.559
So presumably a lot of these things are going to be moving over.
00:14:46.720 --> 00:14:49.440
Uh, but I do think that is the next interface.
00:14:49.600 --> 00:14:51.679
And why is that important for you and your business?
00:14:51.919 --> 00:15:06.639
Because I think siloed thinking when it comes to AI has been so deeply ingrained in our thought, and it is going to change quickly because what you can accomplish with a team in this environment is going to be so much better.
00:15:06.799 --> 00:15:11.440
Because, y'all, I've done training for large organizations, right?
00:15:11.519 --> 00:15:16.080
And there's always such a disconnect between kind of the AI champions and everyone else.
00:15:16.320 --> 00:15:24.799
And one of the things limiting uh, I think a lot of teams and keeping them from moving faster is things like sharing skills, right?
00:15:24.879 --> 00:15:26.799
It's it's it's these permissions things.
00:15:26.960 --> 00:15:35.759
So when you start to break down those walls, um, that's where I think the the gains are going to be compounding faster than they have so far.
00:15:35.919 --> 00:15:37.440
And we've seen it now, right?
00:15:37.519 --> 00:15:43.360
This is two strikes in the past two weeks, and I think the other big companies will be following suit.
00:15:43.519 --> 00:15:48.080
So, number 16, multiplayer AI, you should be investing in it, right?
00:15:48.159 --> 00:15:53.600
So I'm not saying, you know, drop Claude and go to Grockbot.
00:15:53.679 --> 00:15:58.159
And I'm not saying, you know, drop Chad GPT and go try out Buzz.
00:15:58.320 --> 00:16:01.120
I'm not saying that, but you should be exploring in there.
00:16:01.279 --> 00:16:09.519
Um, so then when your AI operating system of choice um brings something like this, your team is ready to go.
00:16:10.000 --> 00:16:10.480
All right.
00:16:10.639 --> 00:16:15.919
So now that the whole team's in one room, yeah, look what they can build.
00:16:16.159 --> 00:16:19.279
Number 15, Vive Operations.
00:16:19.440 --> 00:16:33.679
Y'all, uh again, I keep leaning on uh, you know, some Chat GPT and OpenAI examples because I think, you know, probably in late 2025, early 2026, I think it was, you know, anthropic that was innovating.
00:16:33.759 --> 00:16:36.240
And I think ever since it's been open AI.
00:16:36.320 --> 00:16:38.559
But you know, let me just give one example.
00:16:38.879 --> 00:16:48.000
I think ChatGPT sites is probably the most underhyped and underused thing out there right now.
00:16:48.159 --> 00:16:53.279
Like I said, I think ChatGPT voice now enough people are realizing like how good it is.
00:16:53.519 --> 00:16:55.919
Still, no one is talking about Chat GPT sites.
00:16:56.000 --> 00:17:00.639
And and and let me tell you what I mean by vibe operate, uh vibe operations.
00:17:00.720 --> 00:17:03.200
But this is when the whole company starts building.
00:17:03.679 --> 00:17:12.240
So if you don't know what chat GPT sites are, um, think of it like a a lovable or a a replet, right?
00:17:12.400 --> 00:17:20.400
So kind of uh a vibe platform um that you can you know create simple apps and dashboards and all of these things.
00:17:20.559 --> 00:17:25.440
But the difference is well, you can change them and they live.
00:17:25.599 --> 00:17:29.359
So it's it's back end off, it's databases, all of these things.
00:17:29.440 --> 00:17:31.200
You don't have to know anything, right?
00:17:31.359 --> 00:17:34.880
I actually sent something to my wife, uh when was it?
00:17:34.960 --> 00:17:37.440
I think it was today, and she's like, Oh wow, this is really cool.
00:17:37.680 --> 00:17:44.720
You know, it's chat GPT sites because all this information, you know, it had photos and directories and listings and all these, you know, the ability to save these things.
00:17:44.799 --> 00:17:46.240
And it's like, oh wow, what is this?
00:17:46.319 --> 00:17:48.559
Oh, you built yeah, Chat GPT sites, right?
00:17:48.720 --> 00:17:51.039
But you know, that is available within Teams.
00:17:51.279 --> 00:17:59.519
So uh I I'm not saying that teams are going to, you know, replace their, you know, expensive software subscriptions.
00:17:59.599 --> 00:18:02.559
I'm not saying that, but I think that they're gonna start filling in the gaps.
00:18:02.720 --> 00:18:09.759
And you know, instead, I think like dashboards are probably the simplest um explanation of that, right?
00:18:09.839 --> 00:18:21.359
So, you know, a lot of times you would have to, you know, get get your request in line with someone in you know, BI working in business intelligence who puts together the dashboards, and even then, you know, it's a lot of back and forth.
00:18:21.519 --> 00:18:25.920
Well, now you can literally just drop your files in there, your team can collaborate.
00:18:26.079 --> 00:18:30.160
You don't have to know backend off servers, anything.
00:18:30.240 --> 00:18:30.880
It's just there.
00:18:30.960 --> 00:18:35.200
And you're you have live databases that you can collaborate on in a natural language.
00:18:35.279 --> 00:18:38.960
So I do think that we're gonna see a shift eventually toward vibe operations.
00:18:39.039 --> 00:18:41.920
It's one of those things I'm like literally still pulling my hair about.
00:18:42.000 --> 00:18:45.119
Like, why are people not using ChatGPT sites more?
00:18:45.200 --> 00:18:49.839
Uh, because there's no real true competitor, a single competitor, right?
00:18:49.920 --> 00:19:09.599
The closest would be, like I said, replit and lovable, but you don't have the same um power of users in data and connection, um, especially at the multiplayer level that you have with like Chat GPT teams or chat GPT enterprise that can get these sites up and running for their organization.
00:19:11.039 --> 00:19:12.799
Um, all right, next.
00:19:14.559 --> 00:19:15.599
Agent native.
00:19:15.759 --> 00:19:16.319
All right.
00:19:16.720 --> 00:19:23.680
Um this is I think more than a buzzword.
00:19:23.920 --> 00:19:24.240
All right.
00:19:24.319 --> 00:19:28.559
I've been thinking a lot about this and what it means to be agent native.
00:19:28.720 --> 00:19:39.440
And you know, I keep thinking back to you know, certain things in my background working in marketing, you know, things like you know, mobile first or mobile responsive, all these, right?
00:19:39.519 --> 00:19:48.480
There's this this huge push, um, you know, to get since people were using the internet more on their phones, you know, 15 or so years ago, there's this big push.
00:19:48.559 --> 00:19:49.680
It took 10 years, right?
00:19:49.759 --> 00:20:01.920
But eventually the entire web became mobile responsive by default because the majority of people are accessing the internet on their phones and apps and everything, your business, everything had to be mobile responsive.
00:20:02.079 --> 00:20:04.640
The same thing is going to happen, but going agent native.
00:20:04.799 --> 00:20:05.119
All right.
00:20:05.279 --> 00:20:09.599
So I'm not just talking internally, I'm talking externally as well.
00:20:09.839 --> 00:20:11.039
So, what does that mean?
00:20:11.200 --> 00:20:17.680
So, resources that your team makes, they need to get repurposed in an agent native format.
00:20:17.839 --> 00:20:35.680
So, you know, all that prompting that you go through and do, you and your team, right, all of it needs to turn into a skill because that is a skill that, or that is you um taking the outcome that you came to and making that outcome and the process that you got there agent native.
00:20:35.759 --> 00:20:51.680
In the same way that you know, teams would get together and build SOPs, um, right, that's uh what you should be doing here, but doing this for agents um according to the successful processes that you've uh brought into different AI operating systems.
00:20:51.839 --> 00:20:52.799
So that's one example.
00:20:52.960 --> 00:20:58.720
Another example is well, I think all websites should be marked out, right?
00:20:58.799 --> 00:21:00.559
Like there should be a button, right?
00:21:00.640 --> 00:21:06.640
Aside from having an agent first web, um, which I think is already happening in the background, right?
00:21:06.880 --> 00:21:09.680
I think many companies are gonna have two versions of their websites.
00:21:09.839 --> 00:21:23.200
The second version will probably be made automatically, where it just converts everything to markdown so it's easier for agents to go through without having to be uh token inefficient AF and spend billions of tokens just to understand what's going on in the website.
00:21:23.279 --> 00:21:40.480
But I think in the same thing, simple user interfaces where you know you can say, you know, save this as markdown or convert this to markdown, you know, being able to easily turn your um you know, your company's websites or anything that you publish online um so an agent can more easily access it.
00:21:40.559 --> 00:21:45.200
So just two quick little examples there of being agent native.
00:21:45.519 --> 00:21:47.599
All right, that's number 14.
00:21:47.759 --> 00:21:53.599
So every once everything is kind of agent native, well, you need a new way to keep score.
00:21:53.680 --> 00:21:59.680
And I kind of already referenced this, but I think that skill use is going to boom and become a company.
00:22:00.160 --> 00:22:00.400
Metric.
00:22:00.480 --> 00:22:02.319
What do I mean by a company metric?
00:22:02.400 --> 00:22:02.640
Right.
00:22:02.720 --> 00:22:08.240
There's all these things, I think, through the course of AI, uh, that have been like weird metrics, right?
00:22:08.400 --> 00:22:15.200
So, like early, earlier on, right, uh, when AI was really hitting software development hard, it was all about the lock, right?
00:22:15.279 --> 00:22:16.480
It's all about the lines of code, right?
00:22:16.559 --> 00:22:20.000
Everyone's like, oh, I wrote uh 87 billion lines of code today.
00:22:20.079 --> 00:22:21.680
So, you know, I'm awesome.
00:22:21.839 --> 00:22:26.880
Um, and then you know, late 2025, early 2026, it was the token leaderboards, right?
00:22:27.440 --> 00:22:30.160
Those two things say absolutely nothing.
00:22:30.400 --> 00:22:38.640
I think one of the most important um measurables in AI is going to be skill reuse.
00:22:38.880 --> 00:22:39.200
All right.
00:22:39.279 --> 00:22:43.839
I I talked about this on the show uh last week, right?
00:22:43.920 --> 00:22:46.160
Uh kind of how I AI, right?
00:22:46.480 --> 00:22:48.079
We did the entire thing.
00:22:48.319 --> 00:22:51.440
Um, I think I'd it's like 6,000.
00:22:51.680 --> 00:22:54.160
Uh I can I can actually bring it up here.
00:22:54.319 --> 00:22:56.079
Uh bring up my codex.
00:22:56.559 --> 00:22:57.200
Let's see.
00:22:57.359 --> 00:23:02.400
Okay, so I've used skills uh 6,700 times.
00:23:02.559 --> 00:23:05.839
Um, so it's a lot of skill use, um, right?
00:23:06.160 --> 00:23:07.519
Not could be more.
00:23:07.599 --> 00:23:17.200
Um, but I think that skill reuse is actually going to become one of the most important company metrics because when I said it is so easy to turn a repeatable process into a skill.
00:23:17.279 --> 00:23:28.640
Let's say that you're working with a large language model, you're going back and forth, you're sharing data, you're doing all these things, and you finally get um an outcome or an output or deliverable that really works for you know you, your company, et cetera.
00:23:28.799 --> 00:23:29.839
Well, what do you do with it?
00:23:30.079 --> 00:23:32.799
Maybe you might turn it into a project.
00:23:32.960 --> 00:23:33.920
Okay, that's cool.
00:23:34.079 --> 00:23:41.839
But you need to turn that into a skill, not just because you can invoke it or an agent can invoke it on its own without you even asking it, right?
00:23:42.079 --> 00:23:43.599
Which is usually helpful.
00:23:43.680 --> 00:23:45.039
Sometimes it's not helpful.
00:23:45.200 --> 00:23:54.160
But the other thing is that is how you actually build AI IP for your company because skills are transferable, right?
00:23:54.319 --> 00:24:06.319
So if you're just stuffing everything, you know, into a project inside, you know, Google or Copilot or ChatGPT or Claude or whatever, it that's it's not a it's not really an asset, right?
00:24:06.480 --> 00:24:10.000
Um, at that point, uh, that is a vendor lock-in.
00:24:10.319 --> 00:24:11.680
That's all you're doing at that point.
00:24:11.839 --> 00:24:25.759
So, you know, skill reuse is such an important uh part of of I think any measurement of a good someone that is being becoming um AI native is how much and how often you are reusing skills.
00:24:25.839 --> 00:24:44.319
Because in in theory, even if that skill is not always leading to a improved output or deliverable, it is undeniably saving an extraordinary amount of time, even versus normal AI use, let alone doing things the quote unquote analog way, right?
00:24:44.480 --> 00:24:47.119
Working on the internet, you analog.
00:24:47.440 --> 00:24:49.440
All right, so let's move on.
00:24:49.599 --> 00:24:52.799
So skills capture how works gets done.
00:24:53.039 --> 00:24:57.039
Number 12 is how you capture decisions getting made.
00:24:57.200 --> 00:24:58.079
And you know what?
00:24:58.240 --> 00:25:05.519
Finally, I think I had this in my prediction uh like two years ago, and I was too early, but we're finally starting to see this.
00:25:05.680 --> 00:25:10.480
Number 12 is company reasoning data finally surfacing as the goal that it is.
00:25:10.720 --> 00:25:26.480
So uh you're now seeing big companies like Zapier was one of the first that I saw talking about this about the concept of bringing like Slack DMs out of the DMs and into the open and why and what does that mean and what does it signal?
00:25:26.640 --> 00:25:35.279
So, you know, essentially when you have an agent um in your Slack, uh most agents cannot see DMs, right?
00:25:35.359 --> 00:25:37.759
Um, aside from your own personal agent.
00:25:37.920 --> 00:26:00.400
Um, but when you start moving important conversations from individual DMs or groups of DMs into channels, right, not only does that help the decision making uh help others maybe see the decision-making project uh progress, but most importantly, it allows agents to understand how your company or how your individuals think and make decisions.
00:26:00.640 --> 00:26:16.559
Yes, that's obviously out in the open in normal channels, but the concept of uh, you know, and I actually talked with the Slack CMO about this on the very show about how important it is um to get kind of that human reasoning, that one-on-one human reasoning.
00:26:16.720 --> 00:26:21.039
A lot of times that is what the agent so sorely needs, but we're not giving it.
00:26:21.119 --> 00:26:33.920
We're just throwing more spreadsheets, you know, more skills and all these other things when really it just needs to be able to see um the nuance of a conversation between, you know, two, three, four decision makers that was maybe private.
00:26:34.079 --> 00:26:41.519
So that's just one small um example of company reasoning data finally being serviced as gold.
00:26:41.680 --> 00:26:50.480
But I think that we are gonna see that become bigger and bigger as agents are more easily uh to go out and crawl and understand different contexts, right?
00:26:50.559 --> 00:26:55.039
People are way like, wait, you know, even things like just yapping, right?
00:26:55.200 --> 00:27:20.880
Uh kind of going back to my my um my thought of you know, voice and mobile being the the surface that is gonna be the most important, you know, the amount that I can speak something uh versus type, there's so much more nuance and even reasoning uh that I can fit in there when I am just naturally speaking what's on my mind versus what just gets typed out.
00:27:21.119 --> 00:27:22.799
All right, let's keep going.
00:27:23.039 --> 00:27:26.720
Number 11, the public leaderboards.
00:27:26.880 --> 00:27:32.240
I'm not saying they're they're going to die completely, but they're gonna begin to die.
00:27:32.319 --> 00:27:35.599
And I think that private evals are gonna start to take over.
00:27:35.759 --> 00:27:39.279
I'll say this like 90% of benchmarks are just becoming useless.
00:27:39.440 --> 00:27:45.920
I mean, there's some that are obviously extremely important to look at, regardless of the type of work that you and your company do, right?
00:27:46.079 --> 00:27:49.279
There's ones that are just great for general intelligence, right?
00:27:49.440 --> 00:27:54.640
Like uh GDP Val, uh the artificial analysis index, deep suite.
00:27:54.720 --> 00:27:56.319
I think some of those are just like no-brainer.
00:27:56.480 --> 00:27:57.759
So they're always going to be important.
00:27:58.000 --> 00:28:04.240
There's like hundreds of benchmarks, and people spend so much time looking and thinking uh about benchmarks.
00:28:04.400 --> 00:28:21.440
And aside from you know, benchmark uh bench maxing, uh, which is where you know frontier companies, you know, they overfit a model to perform very well on a specific benchmark, even if maybe humans or real work might not benefit uh as much.
00:28:21.759 --> 00:28:28.799
So I think that public uh leaderboards uh in some of these benchmarks are gonna decrease um very quickly.
00:28:28.960 --> 00:28:46.640
And I think the practical thing that's gonna happen is well, um company uh companies internal uh workbench, essentially, you know, kind of what I'm calling it, that's gonna become more important, right?
00:28:46.720 --> 00:28:50.480
Companies are gonna start, and some of the larger companies already have, right?
00:28:50.559 --> 00:28:56.480
They're actually, you know, creating these as you know, like public um evaluations, right?
00:28:56.559 --> 00:29:01.039
They're creating these as public benchmarks, things that they use for internally and then sharing them with others.
00:29:01.200 --> 00:29:07.440
I think those are gonna become more and more um important, but more and more common, right?
00:29:07.519 --> 00:29:12.559
I think it was like uh DoorDash or something that shared one of theirs, and I was like, okay, this is random.
00:29:12.799 --> 00:29:19.279
I'm like, no, this is actually very important because they took the models and it maybe didn't for them.
00:29:19.359 --> 00:29:30.880
Uh, you know, a certain model might have been much better or much worse than even in the corresponding evaluation, just because of the type of work that certain people do is so specific.
00:29:31.839 --> 00:29:35.599
All right, uh, and then next one kind of related.
00:29:35.759 --> 00:29:41.359
I think model routing is gonna become the norm, and eventually it is gonna become built in.
00:29:41.519 --> 00:29:51.039
Okay, so here's what I mean by that is most companies should stop defaulting to the top model because you don't need it.
00:29:51.519 --> 00:29:58.000
All right, you literally don't uh need it for most of your work, all right.
00:29:58.079 --> 00:30:04.400
And this also uh plays into kind of the um you know model or capability overhang.
00:30:04.640 --> 00:30:06.720
A year ago, I would have said the opposite.
00:30:06.880 --> 00:30:16.799
I would say absolutely use the best model available, even if it takes longer, because you need it for your work, you don't need it anymore, right?
00:30:17.039 --> 00:30:28.960
You don't need uh, you know, Fable 5 Max or GPT 5.6 Soul Max um to you know browse the web for you.
00:30:29.119 --> 00:30:39.359
You know, you don't need uh Opus 5 or you know uh Max or GPT 5.6 Terra to rewrite an email, right?
00:30:39.440 --> 00:30:40.240
You don't.
00:30:40.640 --> 00:30:44.160
You know, there's very capable models that cost a lot less.
00:30:44.319 --> 00:30:47.599
So if you're paying for it on the API side, you've already been through this, right?
00:30:47.680 --> 00:31:02.559
But even on the subscription side, right now, uh I think eventually the the free ride of AI, even in subscriptions, will become less subsidized, and you're gonna have to start dealing um you know with usage limits um and and quotas.
00:31:02.799 --> 00:31:08.160
So model routing, it's gonna have to become a default, and it's actually very easy, right?
00:31:08.240 --> 00:31:12.960
I've I've made um skills that do this for me, right?
00:31:13.039 --> 00:31:26.160
I have a skill where you know I give a big project to uh you know GPT 5.6 soul and it breaks it up and it uses sub agents at its discretion, but lower volume ones.
00:31:26.240 --> 00:31:32.559
So, you know, I really just have it work as uh working as the orchestrator in many instances.
00:31:32.640 --> 00:31:37.039
And you know, I say if if it requires heavier lifting, you can do some of that middlework.
00:31:37.200 --> 00:31:47.119
But for a lot of those things, well, I'm just using, you know, maybe Seoul in the top 10 and last 10%, and then using a bunch of you know smaller sub agents in between.
00:31:47.200 --> 00:31:48.880
But I think that's gonna become the norm.
00:31:49.039 --> 00:32:12.000
I think there's gonna be some uh and there already has been some interesting, you know, AI startups like merge and others that have come out that this is literally what they do, uh, right, because I think it's gonna save a lot of companies a lot of money, and that's something that companies are looking to do as the uh model overhang starts to become larger and larger.
00:32:12.240 --> 00:32:13.920
And what do I mean by that?
00:32:14.160 --> 00:32:35.920
I think that 90% of most knowledge work can be done with using only 10% of a models of that top top models intelligence and just wait, you know, until we get Fable 5.1 and we get you know GPT 6 Astra or GPT 5-7 Astra, right?
00:32:36.000 --> 00:32:36.799
Whatever it is.
00:32:37.039 --> 00:32:50.799
Um our work as humans is not growing in capabilities at the same rate that the models are going, they're just not any.
00:32:51.200 --> 00:32:58.160
Yeah, like I said, two years ago, there were models that you know, the best models weren't capable to do all the work that we do.
00:32:58.400 --> 00:33:01.759
A year ago, I would say it was maybe in parity, right?
00:33:01.920 --> 00:33:05.359
But yeah, you needed that most powerful model to do your hardest work.
00:33:05.440 --> 00:33:13.359
And now, for most people's hardest work, like I said, you know, if you you know, sonnet might be good enough, right?
00:33:13.440 --> 00:33:21.200
GP or or you know, Opus, um you know, opus might be overkill, you know, let alone Fable, right?
00:33:21.279 --> 00:33:32.160
You might just need Claude Sonnet, or you might just need, you know, GPT 56 uh Luna is plenty for many people, but that's why I think model routing is gonna become extremely important.
00:33:32.319 --> 00:33:35.200
All right, we're gonna pick up the pace so this doesn't turn into a two-hour show.
00:33:35.279 --> 00:33:38.880
Number nine, intelligence is gonna get cheaper, cheaper, cheaper, cheaper.
00:33:38.960 --> 00:33:45.440
And what that means, I think in the short term at least, is anthropic is gonna get squeezed, right?
00:33:45.519 --> 00:33:46.880
I've talked about this a lot on the show.
00:33:47.039 --> 00:33:55.920
If you if you care about you know who's who's your vendor, I'm not telling you to choose one or the other, but you have to understand what's going on, and it's been absolutely nutty the last two weeks.
00:33:56.079 --> 00:34:04.160
So, as an example, open AI used a sniffling of um you know, recursive self-improvement.
00:34:04.240 --> 00:34:05.359
You know, they use GBD5.
00:34:09.280 --> 00:34:12.000
They dropped the price of uh GBD5.
00:34:14.480 --> 00:34:16.480
So that model is nearly free.
00:34:16.639 --> 00:34:19.360
That is like nearly sonnet level, all right.
00:34:19.440 --> 00:34:20.800
Uh, but it's not just them.
00:34:20.880 --> 00:34:26.559
I mean, Groc4.6 that just came out, uh, so good, right?
00:34:26.639 --> 00:34:31.039
Uh, at least you know, first looks and cost per task.
00:34:31.280 --> 00:34:36.800
Um, you have Meta's Muse Spark 1.2, Kimmy K3, Quen 3.8.
00:34:36.960 --> 00:34:46.960
All of these models deliver 95 to 99 percent of the intelligence um of the anthropic models, but at five to 30 percent of the cost.
00:34:47.039 --> 00:34:50.239
Those aren't random numbers, those are the real numbers there.
00:34:50.880 --> 00:34:52.480
So, what do I mean by that?
00:34:52.639 --> 00:34:55.199
Well, intelligence is gonna get a lot cheaper, right?
00:34:55.280 --> 00:35:03.039
It used to be the thing where you know you could just choose your, you know, choose your horse and you would know that your horse would always, you know, stay competitive.
00:35:03.280 --> 00:35:04.079
I don't know why.
00:35:04.159 --> 00:35:16.159
It looks like anthropic um is not going to follow the trend here of um, you know, making intelligence cheaper and cheaper at the same time stronger and stronger.
00:35:16.320 --> 00:35:19.119
So um that's gonna be something to definitely keep an eye on.
00:35:19.280 --> 00:35:25.360
But I think as the school year drags on, I think intelligence is only going to get cheaper and cheaper.
00:35:25.519 --> 00:35:33.199
So your provider that you're sticking with, if they're not dropping their price, uh price per performance, yeah, might want to look another way.
00:35:33.360 --> 00:35:33.920
Just saying.
00:35:34.079 --> 00:35:34.559
All right.
00:35:34.719 --> 00:35:36.000
Uh, number eight.
00:35:36.400 --> 00:35:45.039
And this kind of all goes together, but I think Fortune 100 token spend is going to slow while usage explodes.
00:35:45.199 --> 00:35:46.800
And let me explain what I mean by that.
00:35:46.880 --> 00:35:52.239
Uh, similarly to, you know, I was talking about anthropic just selling tokens.
00:35:52.400 --> 00:36:06.400
Um, I think Fortune 100s, at least right now, are the only ones truly in the position uh to be able to, well, have the CapEx or uh the facilities to be able to run local models at scale.
00:36:06.480 --> 00:36:16.719
Because right now, um, you know, again, unless my joke is unless you have a data center sitting in your basement, you can't run these open uh models, right?
00:36:16.880 --> 00:36:22.239
The ones that, you know, oh, you know, everyone's like, oh my gosh, open source models are here.
00:36:22.400 --> 00:36:25.119
So, you know, all the big companies, what are they gonna sell anymore?
00:36:25.199 --> 00:36:29.440
Well, yeah, you can't run the open source models on consumer hardware.
00:36:29.599 --> 00:36:32.960
You need a couple million dollars worth of GPUs to run them.
00:36:33.039 --> 00:36:37.840
But that's where it's like, okay, well, the Fortune 100 companies, they've already started to do it.
00:36:38.000 --> 00:36:47.280
Um, they're gonna start to shift a lot of that lower hanging work, I think, to open models that they're running on-prem.
00:36:47.440 --> 00:36:56.000
Um, is is that gonna trickle, trickle over to the rest of the Fortune 500 and uh, you know, your average everyday enterprise this year?
00:36:56.239 --> 00:36:57.440
Absolutely not.
00:36:57.599 --> 00:37:16.159
Um, but I mean, I think that you're gonna see token usage go up at these companies, but their token spend or the amount of tokens that they're buying from a provider go down, they're gonna stop uh, you know, renting AI and they're gonna start owning it.
00:37:16.239 --> 00:37:28.400
Uh, you know, and I think that that is gonna become more and more common over the years, but I think at least for this year, if you are listening at one of those Fortune 100s, you're probably already having those type of conversations.
00:37:28.559 --> 00:37:33.679
If you are, you know, maybe in the fortune, you know, 101 to 500, right?
00:37:33.760 --> 00:37:35.199
These are just loose numbers.
00:37:35.280 --> 00:37:47.679
Um, it's something that you should probably start paying attention to and start crunching the numbers because I know a lot of companies that are spending um nine figures, you know, literally nine figures on AI.
00:37:47.840 --> 00:37:52.000
You know, so at a certain point, you have to say, okay, how much of that?
00:37:52.079 --> 00:37:54.559
Um, you know, let's let's look at our model routing.
00:37:54.639 --> 00:37:55.920
Are we routing models at all?
00:37:56.079 --> 00:37:57.599
How much are we overpaying, right?
00:37:57.760 --> 00:38:20.880
How many people are are using, you know, uh a fable five to to rewrite their emails, or how many people are using, you know, you know, Google Gemini, you know, 3.6 flash to you know classify something when you could be using a much more cost-efficient model that is still um you know far capable to do the job.
00:38:21.119 --> 00:38:21.599
All right.
00:38:21.760 --> 00:38:25.519
Next, number seven, uh, compute becomes currency.
00:38:25.760 --> 00:38:29.920
All right, uh, you know, uh NVIDIA CE uh CEO Jensen Wong said something like this.
00:38:30.159 --> 00:38:35.280
Obviously, this is what his company does, uh, but I think it's the truth.
00:38:35.440 --> 00:38:50.079
Um, you know, there was a uh famous um, you know, uh Dario Modi uh interview, the CEO of Anthropic, I think it was from 2023, uh, where you know he was being asked about compute and investing in these data centers.
00:38:50.159 --> 00:38:56.400
And, you know, he essentially said, you know, we're we're gonna take it, we're gonna be cautious because we don't want to overspend.
00:38:56.719 --> 00:38:58.719
And what if the AI demand isn't there?
00:38:58.880 --> 00:39:16.880
And now obviously Anthropic is having to kind of pay for that underestimation uh of AI because they're having to overpay um, you know, and have all these other partnerships and they just aren't able to keep up with demand, which is why their servers uh and services are down more than anyone else in the industry.
00:39:17.039 --> 00:39:27.039
You know, on the flip side of that, open AI uh invested heavily early on and they're able to serve things at a much higher uptime up time uptime rate, right?
00:39:27.199 --> 00:39:36.000
But ultimately, um I think as more and more workflows, right, we're still early in the game, right?
00:39:36.159 --> 00:39:41.519
If you are listening to this, if you're using AI every single day, we are in that one percent.
00:39:41.599 --> 00:39:42.400
We are in the bubble.
00:39:42.559 --> 00:39:44.239
Most people are not there yet.
00:39:44.480 --> 00:39:50.719
So as the rest of the world catches up and says, Oh, wait, this AI thing is pretty essential, right?
00:39:50.800 --> 00:39:56.400
It's more, you know, it's as essential as you know, the internet, it's as essential as electricity.
00:39:56.559 --> 00:40:00.880
As the rest of the business world, you know, stumbles across this over the next few years.
00:40:01.039 --> 00:40:02.800
I think this scarcity, right?
00:40:02.960 --> 00:40:05.679
Hardware scarcity becomes a real thing.
00:40:05.840 --> 00:40:10.320
Um, you know, but at the same time, I think at that point you say, oh, okay, we understand it now.
00:40:10.400 --> 00:40:14.400
It doesn't matter how good your model is if you can't run it, right?
00:40:14.480 --> 00:40:21.119
And I'm saying, you know, compute in the same way the companies like Anthropic, OpenAI, Microsoft, Google, Amazon, Meta, etc.
00:40:21.440 --> 00:40:25.280
Uh, but in the same way, compute to run things locally, right?
00:40:25.440 --> 00:40:26.719
That does become the new currency.
00:40:26.800 --> 00:40:41.280
Because if you, you know, are uh you know, a tech, you know, if you work at a big tech company and oh my gosh, well, we actually have servers, well, all of a sudden you have the ability to um print money, right?
00:40:41.440 --> 00:40:43.599
Because everyone wants that compute.
00:40:44.079 --> 00:40:45.599
All right, let's keep going.
00:40:45.760 --> 00:40:53.679
Uh, number six, kind of the prediction here for the rest of the school year AI controversies are gonna hit your backyard, they're going very local.
00:40:53.840 --> 00:41:03.840
So between local data centers and AI deepfakes that I think are gonna hit really hard in the elections, I think AI controversy is going hyper local.
00:41:04.000 --> 00:41:06.159
We've already started to see it with the data centers.
00:41:06.239 --> 00:41:11.599
And let me just put this out there there's a lot of bad information out there about data centers.
00:41:11.679 --> 00:41:13.280
I'm not gonna rally on that.
00:41:13.440 --> 00:41:16.000
I think there's plenty of people having those conversations.
00:41:16.159 --> 00:41:33.199
I'm just trying to more focus on the practical use of AI, but I think it is going to become a dinner time conversation uh at some point this year between, you know, the AI data centers and you know, certain communities, you know, saying they don't want them, certain communities welcoming them in.
00:41:33.360 --> 00:41:40.719
It's gonna create, I think, um, you know, income disparities in places that maybe didn't see it before, swinging both ways.
00:41:40.800 --> 00:41:46.079
Um, but I think ultimately where it's gonna hit hardest is local and state politics, y'all.
00:41:46.159 --> 00:41:47.280
This is my previous life.
00:41:47.440 --> 00:41:53.119
I was uh uh you know a politics reporter uh for the Chicago Sun Times.
00:41:53.199 --> 00:41:55.519
Um at the big level, right?
00:41:56.159 --> 00:42:06.320
You know, presidential race, everyone knows when you know AI is being used, even AI that looks real um or sounds real locally, I think, right?
00:42:06.400 --> 00:42:14.480
So statewide races, I think you we're gonna see dozens of stories about people using AI.
00:42:14.559 --> 00:42:19.519
And a lot of people aren't gonna figure out until it's too late, but for bad reasons, bad actor AI, right?
00:42:19.679 --> 00:42:31.119
People, you know, creating uh fake audio of their competitors, you know, videos um, you know, that they're gonna use in ads that didn't actually happen, it is gonna become commonplace.
00:42:31.280 --> 00:42:39.199
And I don't think most people will be able to tell the difference, which is why I think the AI controversies are gonna be hyper local.
00:42:39.840 --> 00:42:40.320
All right.
00:42:40.480 --> 00:42:46.159
Similarly, well, I think the AI backlash is going to become mainstream.
00:42:46.320 --> 00:42:56.480
Uh, so as the uh shifts start stacking up, um, I think it's gonna start with the work slop avalanche.
00:42:56.559 --> 00:42:59.280
And I think it's already kind of started.
00:42:59.440 --> 00:43:03.599
Um, and a couple of recent developments are gonna think, I think, um, exasperate this.
00:43:03.679 --> 00:43:13.679
So, you know, as an example, uh Anthropic uh kind of announced to uh adhere to some EU regulations that they're gonna start uh putting in some invisible watermarking in text.
00:43:13.760 --> 00:43:30.320
So even if you copy and paste it and try to do all this stuff, you know, um, but essentially, I think that uh uh kind of movement has also led to the resurgence of the you know AI content detectors, which I'm not a fan of because most of them are uh not good.
00:43:30.400 --> 00:43:42.719
Uh, but I think we are gonna see this this weird paradigm because every single company is rightfully so, um, you know, trying to implement AI from top to bottom.
00:43:42.880 --> 00:43:49.599
Yet, you know, now I think more than ever, you're gonna see people getting called out for using AI, right?
00:43:49.679 --> 00:43:56.400
Whether it's you know, like this clawed watermarking or some other um, you know, you know, AI text detection.
00:43:56.480 --> 00:43:59.599
Um, but there's gonna be, I think, a huge backlash.
00:44:00.079 --> 00:44:00.960
Against using AI.
00:44:01.039 --> 00:44:02.559
I even found myself, right?
00:44:02.639 --> 00:44:06.159
I was I was looking for uh what was it?
00:44:06.320 --> 00:44:08.239
Uh a shirt or something online.
00:44:08.800 --> 00:44:14.800
Um, and you know, I just noticed that they were using all you know AI models.
00:44:14.960 --> 00:44:19.360
And it's like I knew that 95% of people weren't gonna be able to tell.
00:44:19.519 --> 00:44:24.800
It wasn't the best AI uh images, but you know, I saw that and I'm just like, nope, next, right?
00:44:25.039 --> 00:44:26.079
So even me, right?
00:44:26.239 --> 00:44:30.400
And people always think that I'm oh AI, everything, and I'm not, um, right.
00:44:30.480 --> 00:44:33.599
I it's just like I think it should always be tastefully done.
00:44:33.679 --> 00:44:44.639
But I think the AI backlash, because it is more and more accessible, and more and more people are going to use it, um, it's gonna hit hard, but I think it's gonna go mainstream.
00:44:44.719 --> 00:44:55.119
Like I talked about this years ago, but uh I think we're gonna start to see it uh where you know people are gonna start putting things out as human-made, and that's gonna become like a trendy thing.
00:44:55.199 --> 00:44:57.599
But I think the AI backlash is gonna hit hard.
00:44:57.760 --> 00:45:03.760
Uh, number four, and I have no clue what this means, but one of my predictions is gonna be that math is solved, right?
00:45:04.000 --> 00:45:18.400
I still don't quite understand how there's all these dozens of math theories that you know people, you know, the smartest mathematicians in the world have spent their careers on and it's they they've been unsolved uh for for forever, right?
00:45:18.480 --> 00:45:20.880
People have been working on these problems for hundreds of years.
00:45:21.039 --> 00:45:23.280
I think math gets solved, right?
00:45:23.519 --> 00:45:25.039
I don't know what it means, right?
00:45:25.119 --> 00:45:33.760
As an example, uh, the unreleased GPT, uh, whatever number it was, Astra, which is their next family of models on top of Seoul.
00:45:34.000 --> 00:45:38.079
Uh, it knocked out like 10 problems with published proofs.
00:45:38.239 --> 00:45:46.960
Um, but I think math in general is gonna get solved, and that's gonna lead to a lot of things that are above my pay grade of understanding.
00:45:47.119 --> 00:45:56.320
Um, obviously, uh great things, uh medicine development, protein synthesis, all these other things that can lead to positives.
00:45:56.480 --> 00:46:09.119
But I also do assume that when math gets solved, um, and that's gonna be able to, I think, um, on your open models that can run on consumer devices are gonna have math solved uh probably within the year.
00:46:09.280 --> 00:46:13.039
So I'm guessing there's gonna be some downsides to that as well, right?
00:46:13.199 --> 00:46:15.840
I don't know, like cracking crypto, right?
00:46:15.920 --> 00:46:18.559
We're still probably uh a way off on that.
00:46:18.639 --> 00:46:23.760
Um, but I think that there's gonna be some downsides to math being solved.
00:46:23.840 --> 00:46:29.360
Um, but in general, I think math benchmarks are no longer gonna be a meaningful benchmark.
00:46:29.519 --> 00:46:33.760
And if anything, this should just tell you how good uh AI has gotten.
00:46:33.920 --> 00:46:43.679
Because three years ago, right, when ChatGPT kind of uh you know started to become consumer mainstream, it couldn't consistently solve two plus two, right?
00:46:43.920 --> 00:46:46.639
Now I think it's literally gonna solve math.
00:46:46.800 --> 00:46:49.840
All right, next three token maxing coming back.
00:46:50.000 --> 00:46:50.320
All right.
00:46:50.400 --> 00:47:06.239
So this kind of ties in with uh one of mine earlier about intelligence is coming too cheap, but essentially, right, token maxing where it's like, let's spend as many tokens as possible, was very much in vogue uh in you know, like December 2025 through like February.
00:47:06.400 --> 00:47:12.960
And then there was the whiplash of uh, oh wait, no, we can't anymore because whoops, we're spending way more money than we thought.
00:47:13.039 --> 00:47:20.559
We were trying to encourage employees to use AI and they were all spending billions of tokens, you know, spinning up useless things.
00:47:20.639 --> 00:47:26.960
So now there's been this shift um toward uh kind of value maxing or token efficiency.
00:47:27.199 --> 00:47:28.960
But now I think we're going back, right?
00:47:29.119 --> 00:47:33.840
Just like the mom jeans, you you know, they were hot, they were gone, they're back.
00:47:34.000 --> 00:47:36.480
I think token maxing is already going to come back.
00:47:36.559 --> 00:47:38.480
And I mean, look at the last week.
00:47:38.639 --> 00:47:39.760
I mean, how could you not?
00:47:39.920 --> 00:47:40.880
Look at Luna.
00:47:40.960 --> 00:47:48.880
I am spending more tokens than ever, um, right, because the tokens are getting cheaper.
00:47:49.119 --> 00:47:49.519
All right.
00:47:49.679 --> 00:47:53.920
So that's why I think token maxing is actually going to return, right?
00:47:54.320 --> 00:47:56.320
Experimenting, running more and more agents.
00:47:56.400 --> 00:48:03.519
So I think we went through this weird phase of like, you know, three to five months where it was like token budgets are getting tight, right?
00:48:04.079 --> 00:48:06.960
Spend is getting tight, which was reactionary.
00:48:07.119 --> 00:48:14.880
It was not looking forward into the future because in if you were looking forward in the future, you see intelligence is going to come down, right?
00:48:14.960 --> 00:48:25.119
So this this this concept of token maxing, um, or you know, or sorry, the concept of token efficiency or value maxing still hasn't fully caught on, right?
00:48:25.280 --> 00:48:28.880
So it's already made a complete cycle.
00:48:29.199 --> 00:48:30.880
Um, I think now.
00:48:31.039 --> 00:48:40.400
So get ahead of it because I think still this this the the the token reckoning hasn't fully hit outside of like Silicon Valley, right?
00:48:40.480 --> 00:48:41.679
It has in some places.
00:48:41.840 --> 00:48:53.679
So if you're listening to this, get ahead and you know, if someone's you know, some smart guy comes in, Bill comes in on Tuesday, and he's like, All right, well, I don't know if you guys know this, but tokens are getting out of control.
00:48:53.760 --> 00:48:56.000
So we have to, you know, talk about efficiency.
00:48:56.079 --> 00:49:00.719
You can say, All right, Bill, uh, you know, it's no longer February 2026.
00:49:00.880 --> 00:49:03.360
The tokens are actually cheaper than ever.
00:49:03.519 --> 00:49:09.039
And the tokens now compared to what they could do, uh, you know, four months ago, night and day.
00:49:09.199 --> 00:49:11.199
Token maxing coming back, book it.
00:49:11.360 --> 00:49:16.719
Number two, the open agents are gonna crash in 2027.
00:49:16.880 --> 00:49:17.119
All right.
00:49:17.280 --> 00:49:18.880
This is the full school year.
00:49:19.039 --> 00:49:24.880
All right, we're we're going all the way into uh May, May or June 2027.
00:49:25.119 --> 00:49:27.119
Open agents are gonna crash everything.
00:49:27.199 --> 00:49:35.519
I talked about agent crashes um on my uh 2026 AI prediction and roadmap series, which obviously they've been crashing, right?
00:49:35.599 --> 00:49:49.119
We had all the stories, the the uh open AI, hugging face, uh Kimmy K3 anthropics models, every so the agents are crashing right now, but the open crashes are what's going to really matter.
00:49:49.280 --> 00:49:52.639
So I had a show on this the other day, if you want to go back and listen to it.
00:49:52.960 --> 00:50:00.559
But I think essentially the open models today aren't good enough to cause chaos yet in bad actors' hands.
00:50:00.800 --> 00:50:07.519
But the open models that we should be getting around quarter four, they will be because those models are gonna be fable level.
00:50:07.679 --> 00:50:14.000
Those models are gonna be, you know, Astra, potentially Astra level, at least GPT-5-6 soul level.
00:50:14.159 --> 00:50:15.599
So what does that mean?
00:50:15.760 --> 00:50:25.760
Uh, that means that they're gonna be smart enough or conniving enough when bad actors are using them to be able to skirt around the guardrails that are built in.
00:50:25.840 --> 00:50:30.239
And obviously, with open models, once you release them, you can't rein them in, right?
00:50:30.320 --> 00:50:38.320
You can't take uh, you know, Kimi K31 or you know, Quen3.9, you can't take them offline, right?
00:50:38.400 --> 00:50:39.599
Once they're out, they're out.
00:50:39.679 --> 00:50:47.280
Um, and obviously, because they're open models, uh, you can fork them, you can fine-tune them, you can do what you can to bring the guardrails down.
00:50:47.360 --> 00:50:50.480
And I think that is gonna happen at scale.
00:50:50.719 --> 00:50:52.639
Um, just FYI.
00:50:52.800 --> 00:50:53.760
So, what does that mean?
00:50:53.920 --> 00:51:07.599
I think cyber defense and cybersecurity, um, especially of the AI variety, it's gonna be the fastest growing sector um in 2027, and it's not even gonna be close.
00:51:07.679 --> 00:51:08.960
It is not even gonna be close.
00:51:09.280 --> 00:51:12.960
Assuming, right, that it's gonna happen at the rate that it's gonna happen.
00:51:13.119 --> 00:51:22.159
I mean, even small businesses, you know, SMBs, smaller enterprises that normally wouldn't have a strong cybersecurity presence.
00:51:22.320 --> 00:51:27.760
I'm not saying that they're gonna go hire a cybersecurity team of 10, but all of a sudden it's gonna be a budget item.
00:51:27.920 --> 00:51:33.360
If you're not gonna hire people, you're gonna have to hire a third-party company and it is not gonna be cheap.
00:51:33.519 --> 00:51:38.320
So start penciling that in now, the everyday average business, right?
00:51:38.559 --> 00:51:49.280
Your little, you know, not saying little, but you know, your average company with 200 employees, 500 employees, right, that have maybe never been the victim of a cyber attack.
00:51:49.440 --> 00:51:59.440
When agents, when open agents are gonna be able to find vulnerabilities in your CRM vulnerabilities in your payment systems, they will be able to wreak havoc.
00:51:59.679 --> 00:52:01.440
You have to be prepared for it now.
00:52:01.760 --> 00:52:08.960
And then number one, last but not least, and this kind of ties a lot of the stuff together, RSI is on the horizon.
00:52:09.119 --> 00:52:10.559
Recursive self-improvement.
00:52:10.719 --> 00:52:15.360
I also had a recent show about this, but actually some more recent developments.
00:52:15.519 --> 00:52:23.199
So, Sergei Brin, the most one of the most famous people in AI ever, who now seems to have a little bit more.
00:52:23.519 --> 00:52:28.159
Uh, you know, he's always been at the top of the Google AI food chain.
00:52:28.239 --> 00:52:45.360
But now it seems like um, you know, with some recent leadership shakeups at Google, um, you know, he might have a little bit more, I wouldn't say sway or pull, um, but it seems like he's gonna be uh a much more consequential figure um in Google with some of their other big names gone.
00:52:45.519 --> 00:52:56.079
But you know, just recently he just declared that recursive self-improvement is a top priority for Google Gemini and kind of um directing and wanting companies to go there.
00:52:56.159 --> 00:52:57.360
And we've already seen this.
00:52:57.519 --> 00:52:59.599
Um, you know, we've already seen glimpses of this.
00:52:59.679 --> 00:53:07.599
I've already mentioned, you know, GVD56 Soul as an example, making uh its smaller models more optimized and cheaper.
00:53:07.760 --> 00:53:23.840
Um and we've already seen stories that you know, companies say that their models that are coming out now were um, you know, um whether going through post-training or, you know, they're not always um giving the details, but today's models are starting to help build tomorrow's models.
00:53:23.920 --> 00:53:28.239
Um, and with recursive self-improvement, that makes everything else.
00:53:28.400 --> 00:53:32.559
The other 18 on this list go that much faster.
00:53:32.719 --> 00:53:33.840
All right, that's a wrap.
00:53:33.920 --> 00:53:36.960
I'm gonna go through these one more time, rapid style.
00:53:37.039 --> 00:53:43.039
So you don't forget, number 19, reactive chat is gonna start dying and proactive agents are gonna take over.
00:53:43.119 --> 00:53:47.679
Uh, number 18, voice mode and mobile become the default AI interface.
00:53:47.840 --> 00:53:51.440
Number 17, manager threads are gonna replace one-off chats.
00:53:51.599 --> 00:53:56.159
Number 16, multiplayer AI, humans and agents are gonna share one workspace.
00:53:56.320 --> 00:54:00.320
Number 15, vibe operations, the whole company is gonna start building.
00:54:00.480 --> 00:54:03.840
Number 14, everything you make goes agent native.
00:54:04.000 --> 00:54:07.280
Number 13, skill reuse becomes a company metric.
00:54:07.440 --> 00:54:11.360
Number 12, company reasoning data finally surfaces as gold.
00:54:11.519 --> 00:54:16.079
Number 11, public leaderboards die and private evals are gonna take over.
00:54:16.239 --> 00:54:19.760
Number 10, model routing becomes the norm and built in.
00:54:19.920 --> 00:54:24.000
Number nine, intelligence gets cheap and anthropic gets squeezed.
00:54:24.159 --> 00:54:30.000
Number eight, fortune 100's token spend slows while usage explodes.
00:54:30.239 --> 00:54:33.519
Um, number seven, compute becomes currency.
00:54:33.679 --> 00:54:36.079
Number six, AI controversies hit your backyard.
00:54:36.159 --> 00:54:38.800
Number five, the AI backlash goes mainstream.
00:54:38.960 --> 00:54:42.719
Number four, math is solved, whatever that means by AI.
00:54:42.880 --> 00:54:45.360
Number three, the return of token maxing.
00:54:45.440 --> 00:54:48.480
Number two, open agents crash in 2027.
00:54:48.639 --> 00:54:52.639
And number one, RSI is on the horizon.
00:54:52.880 --> 00:54:57.039
All right, that is a wrap for today's show.
00:54:57.360 --> 00:55:00.320
But remember, a quick announcement.
00:55:00.639 --> 00:55:01.920
Now you're ready.
00:55:02.079 --> 00:55:03.519
I gave you the answers.
00:55:03.760 --> 00:55:08.559
And maybe some of this was a little over your head.
00:55:08.719 --> 00:55:11.039
And you're like, wait, I'm kind of new here.
00:55:11.280 --> 00:55:12.639
I need some of the basics.
00:55:12.800 --> 00:55:25.519
Well, go back to the basics because, like I said earlier, uh, we're we're taking a little break from our normal Monday to Friday schedule uh to replay the entire and restart the start here series.
00:55:25.679 --> 00:55:32.719
So this is a series made for both beginners and people who are using AI every single day and leaders in their organization.
00:55:32.960 --> 00:55:38.079
We're gonna play episodes one through 30 in order starting Monday.
00:55:38.159 --> 00:55:43.840
So if this got your brain turning and you're like, okay, these are some good things for our company to talk about.
00:55:44.000 --> 00:55:49.760
Or if you were like, wait, I need to know more about you know the these open models.
00:55:49.920 --> 00:55:54.320
I need to learn more about you know some of these topics on agents that you talked about.
00:55:54.480 --> 00:56:06.880
Well, in the start here series, we tackled most of these things all in depth, uh, at a higher level, more detail, more examples, more use cases for your company to grow.
00:56:07.119 --> 00:56:08.079
So that's it.
00:56:08.159 --> 00:56:11.119
Make sure to go back to school with us starting Monday.
00:56:11.280 --> 00:56:12.400
So that's a wrap.
00:56:12.480 --> 00:56:13.599
I hope this was helpful.
00:56:13.760 --> 00:56:19.440
If so, please, if you haven't already, subscribe to the podcast, then go to your everydayai.com.
00:56:19.519 --> 00:56:20.320
Thanks for tuning in.
00:56:20.480 --> 00:56:23.280
See you back tomorrow and every day for more everyday AI.
00:56:23.360 --> 00:56:24.400
Thanks, y'all.
00:56:25.679 --> 00:56:29.039
And that's a wrap for today's edition of Everyday AI.
00:56:29.199 --> 00:56:30.320
Thanks for joining us.
00:56:30.480 --> 00:56:34.239
If you enjoyed this episode, please subscribe and leave us a rating.
00:56:34.400 --> 00:56:36.000
It helps keep us going.
00:56:36.239 --> 00:56:43.280
For a little more AI magic, visit your everydayai.com and sign up to our daily newsletter so you don't get left behind.
00:56:43.519 --> 00:56:46.159
Go break some barriers, and we'll see you next time.
00:00:00.320 --> 00:00:02.959
Welcome to the Everyday AI podcast.
00:00:03.120 --> 00:00:09.439
My name is Jordan Wilson, and for the past three and a half years, we put out more than 800 episodes.
00:00:09.599 --> 00:00:13.759
Yet one of the most common questions I get, I didn't really have an answer for.
00:00:14.000 --> 00:00:17.039
Where do I start on the Everyday AI podcast?
00:00:17.280 --> 00:00:19.760
And that's why we started the Start Here series.
00:00:19.920 --> 00:00:30.239
And with Fall now back in full swing, the Everyday AI podcast is going back to school and playing back the entire Start Here series from front to back.
00:00:30.399 --> 00:00:37.759
We've hit pause on our normal Monday to Friday programming to run back our most popular series ever for the next 30 days.
00:00:37.920 --> 00:00:42.000
We made the Start Here series for beginners and AI champions alike.
00:00:42.159 --> 00:00:53.359
So whether you're just trying to get a grasp on large language models or grappling with the best coding harness for multi-agentic workflows, the Start Here series covers it all.
00:00:53.600 --> 00:00:58.079
Plain language, no jargon, and easy to follow along each day.
00:00:58.320 --> 00:01:03.679
So make sure to subscribe to the podcast and check back each day for new insights day by day.
00:01:03.840 --> 00:01:11.040
The series is a culmination of spending more than 10,000 hours covering generative AI over the past three and a half years.
00:01:11.200 --> 00:01:14.799
So you don't want to miss a single episode of the Start Here series.
00:01:14.879 --> 00:01:16.159
Let's get into it.
00:01:18.239 --> 00:01:23.040
This could randomly be one of the most important shows I've ever recorded.
00:01:23.359 --> 00:01:24.000
Why?
00:01:24.400 --> 00:01:30.319
Well, I'm laying out the AI answers for your company for the entire year right here.
00:01:30.719 --> 00:01:31.680
Let me explain.
00:01:31.840 --> 00:01:37.599
Usually in December or January, I put on my AI predictions hat and tell you what's ahead.
00:01:37.760 --> 00:01:44.560
And for the past three years, I've done a pretty okay job in these prediction shows at telling you what's next.
00:01:44.799 --> 00:01:47.680
But the pace of AI is much different now.
00:01:47.920 --> 00:01:55.760
I mean, measurably, AI models are literally coming out about two to three times faster today than they were just nine months ago.
00:01:56.000 --> 00:02:02.000
So we can't wait until another four months in December or January to give you the roadmap ahead.
00:02:02.239 --> 00:02:08.960
So with the school year starting and with this upcoming start here series, we're about to restart from the beginning.
00:02:09.120 --> 00:02:11.840
I thought I'd just give you the answers today.
00:02:12.000 --> 00:02:15.280
Not the syllabus, not the take-home test, just the answers.
00:02:15.520 --> 00:02:24.719
So 19 quick points for you and your team to focus on, whatever the sector, whatever point you're at in your career, whatever type of work that you use AI for.
00:02:24.879 --> 00:02:28.719
I think today's show can be your answer sheet for the whole year.
00:02:28.960 --> 00:02:34.719
So you and your company are focusing on what's next, not what already happened.
00:02:34.960 --> 00:02:36.159
Let's get into it.
00:02:36.319 --> 00:02:38.719
If you're new here, welcome to Everyday AI.
00:02:38.800 --> 00:02:41.120
My name's Jordan Wilson, and well, we do this every day.
00:02:41.199 --> 00:02:52.159
It's your daily, unedited, unscripted live stream podcast, and free daily newsletter, helping business leaders like you and me not just keep up with what's happening in the AI world because it's too much for anyone human.
00:02:52.319 --> 00:02:56.719
But I help you get ahead to know what's coming so you can grow your company and your career.
00:02:56.879 --> 00:03:02.639
So it starts here, but make sure you go to our website at your everydayai.com, sign up for the free daily newsletter.
00:03:02.800 --> 00:03:06.960
We're we'll be recapping today's show and a whole lot more.
00:03:07.599 --> 00:03:13.919
So, as a reminder, we are taking a kind of break in our normal programming.
00:03:14.159 --> 00:03:23.199
So, since it is that back to school time and all the kids are going back to school, we're all gonna be doing some learning together on everyday AI.
00:03:23.439 --> 00:03:29.520
So, our start here series uh show that we kicked off earlier this year.
00:03:29.759 --> 00:03:39.280
We've been doing, you know, one or two shows a week, and we have 30 of them, and they're pretty amazing, and they're some of our most popular shows ever.
00:03:39.439 --> 00:03:45.680
So we're gonna be running them all back, the entire series starting Monday with the very first episode.
00:03:45.840 --> 00:03:49.280
So, yeah, a little break from our normal Monday to Friday schedule.
00:03:49.360 --> 00:04:03.199
But I guarantee you, if you just stick with the series from volume one, listen all the way to volume 30, whether you listen to one, some, or none of them, you're gonna know more about AI than like 99% of the other people at your company.
00:04:03.360 --> 00:04:06.719
So the series cuts through the jargon, cuts through the fluff.
00:04:06.800 --> 00:04:11.120
Whether you're a beginner, you around AI all the time, just it's simple.
00:04:11.199 --> 00:04:16.000
No computer science degree required, but class starts Monday, don't be late.
00:04:16.240 --> 00:04:23.279
All right, but let's now get into it and talk about those 19.
00:04:23.600 --> 00:04:26.079
Yeah, we have 19 of them.
00:04:26.240 --> 00:04:31.279
Uh so these are the things that I think your company needs to be focusing on.
00:04:31.439 --> 00:04:31.680
Yeah.
00:04:31.759 --> 00:04:33.439
So for the school year here, right?
00:04:33.600 --> 00:04:36.240
So, yeah, classes and section set uh session.
00:04:36.399 --> 00:04:42.879
That's why we have the little uh, you know, yellow, the pencil yellow uh accent color here for my slides.
00:04:43.120 --> 00:04:47.279
So uh these are in no real order, right?
00:04:47.439 --> 00:04:49.519
I was just kind of jotting down some notes.
00:04:49.680 --> 00:05:06.959
Um, you know, sometimes I just uh am always keeping not like predictions, but a lot of my observations, because uh from doing this for three and a half years, literally spending my entire day every day uh reading about AI, exploring AI, talking to other very smart people about AI, right?
00:05:07.120 --> 00:05:12.399
I have a what I would say is a pretty accurate notebook of random ideas and thoughts.
00:05:12.480 --> 00:05:14.639
And I've been spending some time putting them all together.
00:05:14.800 --> 00:05:16.319
I came up with 19, right?
00:05:16.399 --> 00:05:23.360
So, you know, for my yearly prediction shows, I always do, you know, for 2024, it was 24, then 25, then 26.
00:05:23.439 --> 00:05:24.399
Well, we got 19.
00:05:24.560 --> 00:05:24.879
All right.
00:05:24.959 --> 00:05:27.839
So I don't know, maybe that's roughly how far we are through the year.
00:05:27.920 --> 00:05:30.879
But this is this is we're lining this up with the school year, right?
00:05:30.959 --> 00:05:32.000
We're starting in August.
00:05:32.160 --> 00:05:38.639
I think these should all still be fairly uh relevant, come, you know, when May, when school gets out.
00:05:38.800 --> 00:05:42.000
Uh, but let's start with number 19.
00:05:42.720 --> 00:05:49.600
So this is when reactive chat is dying and proactive agents are taking over.
00:05:49.759 --> 00:05:50.160
All right.
00:05:50.319 --> 00:05:57.600
And I also want you to think, right, um, where are you at on the AI usage scale?
00:05:57.759 --> 00:05:59.680
Because maybe what I said is a shock.
00:05:59.759 --> 00:06:02.319
Maybe you're like, duh, it's been this way for a while, right?
00:06:02.399 --> 00:06:07.279
So as I rattle these off, right, keep in mind our audience is all over the place, right?
00:06:07.439 --> 00:06:16.959
We literally have AI leaders at the labs listening to this, and then we have people that have stumbled upon ChatGPT for the very first time and have never used it.
00:06:17.040 --> 00:06:19.040
All right, so keep that in mind as I walk through these.
00:06:19.199 --> 00:06:24.959
So maybe you've heard me mention some of these things before, maybe you don't know what some of them are, so I'm gonna try to keep it simple.
00:06:25.279 --> 00:06:32.879
But I think over the next year, uh reactive chat is gonna start dying, right?
00:06:33.040 --> 00:06:42.000
It's the combination of not just uh agents, but just being able to uh very easily save and reuse skills.
00:06:42.240 --> 00:06:43.920
That is huge, right?
00:06:44.160 --> 00:06:50.959
Um, I think that chat GPT work as an example is really indicative of this.
00:06:51.120 --> 00:06:55.920
It's it's no different necessarily than uh codecs, right?
00:06:56.000 --> 00:07:07.360
And codecs was kind of the, I would say, the uh one of the birthplaces um of the hardworking around the clock agent to that as well as uh Claude Code and Claude Cowork.
00:07:07.439 --> 00:07:14.959
Uh but I think the new normal is just gonna be, you know, waking up and reviewing what your AI already shipped overnight.
00:07:15.040 --> 00:07:18.480
Um, and and and that piece is extremely important.
00:07:20.399 --> 00:07:29.759
All right, so if you're prompting less than what are you doing, all right, that's number 18, you're talking more.
00:07:29.920 --> 00:07:40.240
Uh, I think that number 18 kind of uh prediction, 18 here, I think voice and mobile are already becoming the default AI controls.
00:07:40.639 --> 00:07:53.279
Uh so if you haven't used uh Chat GPT uh's new feature yet, my gosh, uh Chat GPT voice is outstanding.
00:07:55.279 --> 00:07:59.439
What's crazy is this has been out for almost a month.
00:07:59.680 --> 00:08:08.079
And you know, normally the people that are quote unquote early talking about AI are starting to talk about it literally now.
00:08:08.240 --> 00:08:21.759
Um, but I think you know, um, to summarize this, I did a show on it a couple of weeks ago when it did uh first come out, but this is essentially kind of the combination of a voice and mobile becoming the default AI control.
00:08:22.240 --> 00:08:24.959
Uh, yeah, you can go out and touch grass.
00:08:25.120 --> 00:08:31.439
And um, in in in my experience, I get a lot more done now when I take my iPhone out, right?
00:08:31.519 --> 00:08:34.559
And this is another, you know, open AI slash chat GPT.
00:08:34.639 --> 00:08:39.679
And I think you know, uh Claude has made some improvements to theirs, it's still not very good.
00:08:39.840 --> 00:08:49.840
Um, but I assume everyone is gonna have to follow suit because this is uh so so powerful um that you just have to have it, right?
00:08:50.000 --> 00:08:54.960
Even I think we saw a glimpse of it with you know Grok Grok bot that just came out.
00:08:55.120 --> 00:09:09.840
Um, but to be able to talk um to a uh full duplex agent, that means an agent that can listen and talk and uh at the same time, but one that also has access to your data and can look up information.
00:09:10.000 --> 00:09:14.000
It's literally faster than talking to a coworker, right?
00:09:14.159 --> 00:09:19.519
Um, because I can talk back and forth and it will respond to me.
00:09:19.679 --> 00:09:28.960
But then also while it's responding to me, it's gonna, you know, uh personalize everything through the data that it has connected, as well as go and look things up.
00:09:29.120 --> 00:09:36.480
So, you know, for me, this is like, you know, working with uh someone that's smarter than the smartest coworker I've ever worked with, right?
00:09:36.720 --> 00:09:46.240
But I think that this has to become eventually uh the main surface uh fairly soon, just because it is that good, right?
00:09:46.320 --> 00:09:48.960
And um, so you're like, okay, what does this mean?
00:09:49.200 --> 00:09:51.360
You know, just isn't that just mobile?
00:09:51.519 --> 00:09:57.600
Not necessarily, uh, because the desktop is obviously much more powerful than what you can accomplish on any mobile phone.
00:09:57.759 --> 00:10:08.399
But the way that this works, at least in ChatGPT voice, is well, you can open the ChatGPT app, click on remote, and then literally control your entire computer, right?
00:10:08.559 --> 00:10:19.200
So this is the Jarvis of AI where it can see what's on your computer, it can click things, it can open, save, file, shut, run commands, it can do anything.
00:10:19.279 --> 00:10:29.919
Uh, but I think that that is if if if you and your company have not already started exploring this, you should probably have your eyes on this because it is a much more fluid and natural way to work.
00:10:30.080 --> 00:10:32.159
I like to say it's kind of like a flow state.
00:10:32.320 --> 00:10:33.039
Um, right.
00:10:33.200 --> 00:10:46.159
I don't talk about this a lot, but sometimes, and maybe this is because I'm a former journalist, but sometimes when I'm a hands-on keyboard, maybe I'm trying to get things uh a little too right.
00:10:46.399 --> 00:10:50.559
You know, sometimes I don't just like to uh blurt out on the keyboard.
00:10:50.639 --> 00:10:58.879
Um, and that might interrupt kind of the natural flow state of what I actually want to communicate uh to an AI or to an agent.
00:10:59.120 --> 00:11:02.720
Obviously, when I'm talking, it's a little easier, I think, ultimately.
00:11:02.879 --> 00:11:04.799
And I think that holds true for a lot of people.
00:11:04.960 --> 00:11:15.600
There's actually some studies that show that you can much more effectively uh communicate your uh point when talking to AI agents while you're talking versus when you're texting because uh uh or typing.
00:11:15.679 --> 00:11:26.000
Because when you're typing, you're already putting you know multiple limiting filters on it uh because you know that you can only type so fast and your uh fingers cannot keep up with your brain.
00:11:26.080 --> 00:11:28.240
You know, your mouth can do a little bit better of a job.
00:11:28.399 --> 00:11:31.120
But I think number 18, that is definitely the next control.
00:11:31.759 --> 00:11:35.840
Number 17, uh, this is the manager of threads.
00:11:36.080 --> 00:11:41.759
So we've already seen this in um codecs and chat GPT work.
00:11:42.000 --> 00:11:47.279
Uh we've seen a smaller version of this that was uh actually announced.
00:11:47.360 --> 00:11:48.799
Uh, I'm losing track of time.
00:11:48.879 --> 00:11:52.080
I think it was this week uh from Claude, right?
00:11:52.320 --> 00:11:54.080
But what does this mean?
00:11:54.240 --> 00:12:06.960
Um, I think we were trained for better or worse when working, whether inside, you know, Copilot, Gemini, ChatGPT, uh, Claude, etc., you know, that any conversation had its home, right?
00:12:07.120 --> 00:12:10.480
So, hey, I need to go back into that chat, right?
00:12:10.559 --> 00:12:14.639
And this was especially more important in the earlier days of prompt engineering.
00:12:14.720 --> 00:12:26.480
This is something I taught to thousands of students that took our free PPP course because this was the constraint, you know, probably up until a year ago, is the context really just lived in that single chat thread.
00:12:26.639 --> 00:12:28.240
It's not like that anymore, right?
00:12:28.320 --> 00:12:33.360
So now I think you have to think of manager threads replaces those one-off chats.
00:12:33.519 --> 00:12:38.879
So you could call those like a chief of staff, uh chief of staff um thread, right?
00:12:39.039 --> 00:12:42.240
So that's the way I think a lot of people set it up in a project.
00:12:42.399 --> 00:12:46.399
You can have, you know, an agent's MD file in any project.
00:12:46.559 --> 00:12:56.240
Um, and you can create an unlimited amount of threads or chats in there, but then any thread in uh you know, codex and chat GPT can control any other thread.
00:12:56.320 --> 00:13:13.120
And again, uh that is an interface that all uh you know, Google, Claude, Grok, everyone else, you know, has either you know taken that as well, or they will have to soon, because that is um, you know, the the next step in terms of communication.
00:13:13.279 --> 00:13:16.320
And I think that's that's a smart way to do it, right?
00:13:16.399 --> 00:13:18.559
That's how it is with humans, right?
00:13:18.720 --> 00:13:25.919
If if you're sitting with you know your team, uh you can talk to them at any point about anything that you all know.
00:13:26.080 --> 00:13:28.320
You don't have to wait for the marketing meeting.
00:13:28.399 --> 00:13:30.720
If you're sitting in the marketing team, is there?
00:13:30.799 --> 00:13:31.679
You can just talk about it.
00:13:31.759 --> 00:13:36.480
So this is a much more natural way to uh interface uh with AI agents.
00:13:37.120 --> 00:13:37.759
All right.
00:13:37.919 --> 00:13:43.279
Um, and you know, once those threads are managing threads, the next step is obvious.
00:13:43.440 --> 00:13:46.480
Well, you put people and agents in the same room.
00:13:46.559 --> 00:13:49.919
So that is number uh 16 on the list here.
00:13:50.320 --> 00:13:55.679
Multiplayer AI with humans and agents sharing the same workspace.
00:13:55.840 --> 00:13:58.879
And over the past two weeks, we've seen some great examples of that.
00:13:58.960 --> 00:14:00.799
I already mentioned Grokbot.
00:14:00.960 --> 00:14:14.960
Um, you know, so if you haven't seen this, it's only been out for like not even 72 hours yet, uh, right, but very similar to Buzz from Block, uh, you know, Jack Dorsey of uh Twitter fame.
00:14:15.200 --> 00:14:16.159
Uh, but the same thing.
00:14:16.240 --> 00:14:30.960
So we've seen these two recent um interfaces come out that are multiplayer by default, so where you and your entire team uh can come in and chat with the same agent, uh, right?
00:14:31.039 --> 00:14:33.440
So we've seen glimpses of this, right?
00:14:33.519 --> 00:14:37.440
There was a Google leak uh about bringing some multiplayer capabilities.
00:14:37.600 --> 00:14:43.519
Um, obviously, you know, open AI uh inside ChatGPT, you know, they've had these group chats.
00:14:43.600 --> 00:14:46.559
So presumably a lot of these things are going to be moving over.
00:14:46.720 --> 00:14:49.440
Uh, but I do think that is the next interface.
00:14:49.600 --> 00:14:51.679
And why is that important for you and your business?
00:14:51.919 --> 00:15:06.639
Because I think siloed thinking when it comes to AI has been so deeply ingrained in our thought, and it is going to change quickly because what you can accomplish with a team in this environment is going to be so much better.
00:15:06.799 --> 00:15:11.440
Because, y'all, I've done training for large organizations, right?
00:15:11.519 --> 00:15:16.080
And there's always such a disconnect between kind of the AI champions and everyone else.
00:15:16.320 --> 00:15:24.799
And one of the things limiting uh, I think a lot of teams and keeping them from moving faster is things like sharing skills, right?
00:15:24.879 --> 00:15:26.799
It's it's it's these permissions things.
00:15:26.960 --> 00:15:35.759
So when you start to break down those walls, um, that's where I think the the gains are going to be compounding faster than they have so far.
00:15:35.919 --> 00:15:37.440
And we've seen it now, right?
00:15:37.519 --> 00:15:43.360
This is two strikes in the past two weeks, and I think the other big companies will be following suit.
00:15:43.519 --> 00:15:48.080
So, number 16, multiplayer AI, you should be investing in it, right?
00:15:48.159 --> 00:15:53.600
So I'm not saying, you know, drop Claude and go to Grockbot.
00:15:53.679 --> 00:15:58.159
And I'm not saying, you know, drop Chad GPT and go try out Buzz.
00:15:58.320 --> 00:16:01.120
I'm not saying that, but you should be exploring in there.
00:16:01.279 --> 00:16:09.519
Um, so then when your AI operating system of choice um brings something like this, your team is ready to go.
00:16:10.000 --> 00:16:10.480
All right.
00:16:10.639 --> 00:16:15.919
So now that the whole team's in one room, yeah, look what they can build.
00:16:16.159 --> 00:16:19.279
Number 15, Vive Operations.
00:16:19.440 --> 00:16:33.679
Y'all, uh again, I keep leaning on uh, you know, some Chat GPT and OpenAI examples because I think, you know, probably in late 2025, early 2026, I think it was, you know, anthropic that was innovating.
00:16:33.759 --> 00:16:36.240
And I think ever since it's been open AI.
00:16:36.320 --> 00:16:38.559
But you know, let me just give one example.
00:16:38.879 --> 00:16:48.000
I think ChatGPT sites is probably the most underhyped and underused thing out there right now.
00:16:48.159 --> 00:16:53.279
Like I said, I think ChatGPT voice now enough people are realizing like how good it is.
00:16:53.519 --> 00:16:55.919
Still, no one is talking about Chat GPT sites.
00:16:56.000 --> 00:17:00.639
And and and let me tell you what I mean by vibe operate, uh vibe operations.
00:17:00.720 --> 00:17:03.200
But this is when the whole company starts building.
00:17:03.679 --> 00:17:12.240
So if you don't know what chat GPT sites are, um, think of it like a a lovable or a a replet, right?
00:17:12.400 --> 00:17:20.400
So kind of uh a vibe platform um that you can you know create simple apps and dashboards and all of these things.
00:17:20.559 --> 00:17:25.440
But the difference is well, you can change them and they live.
00:17:25.599 --> 00:17:29.359
So it's it's back end off, it's databases, all of these things.
00:17:29.440 --> 00:17:31.200
You don't have to know anything, right?
00:17:31.359 --> 00:17:34.880
I actually sent something to my wife, uh when was it?
00:17:34.960 --> 00:17:37.440
I think it was today, and she's like, Oh wow, this is really cool.
00:17:37.680 --> 00:17:44.720
You know, it's chat GPT sites because all this information, you know, it had photos and directories and listings and all these, you know, the ability to save these things.
00:17:44.799 --> 00:17:46.240
And it's like, oh wow, what is this?
00:17:46.319 --> 00:17:48.559
Oh, you built yeah, Chat GPT sites, right?
00:17:48.720 --> 00:17:51.039
But you know, that is available within Teams.
00:17:51.279 --> 00:17:59.519
So uh I I'm not saying that teams are going to, you know, replace their, you know, expensive software subscriptions.
00:17:59.599 --> 00:18:02.559
I'm not saying that, but I think that they're gonna start filling in the gaps.
00:18:02.720 --> 00:18:09.759
And you know, instead, I think like dashboards are probably the simplest um explanation of that, right?
00:18:09.839 --> 00:18:21.359
So, you know, a lot of times you would have to, you know, get get your request in line with someone in you know, BI working in business intelligence who puts together the dashboards, and even then, you know, it's a lot of back and forth.
00:18:21.519 --> 00:18:25.920
Well, now you can literally just drop your files in there, your team can collaborate.
00:18:26.079 --> 00:18:30.160
You don't have to know backend off servers, anything.
00:18:30.240 --> 00:18:30.880
It's just there.
00:18:30.960 --> 00:18:35.200
And you're you have live databases that you can collaborate on in a natural language.
00:18:35.279 --> 00:18:38.960
So I do think that we're gonna see a shift eventually toward vibe operations.
00:18:39.039 --> 00:18:41.920
It's one of those things I'm like literally still pulling my hair about.
00:18:42.000 --> 00:18:45.119
Like, why are people not using ChatGPT sites more?
00:18:45.200 --> 00:18:49.839
Uh, because there's no real true competitor, a single competitor, right?
00:18:49.920 --> 00:19:09.599
The closest would be, like I said, replit and lovable, but you don't have the same um power of users in data and connection, um, especially at the multiplayer level that you have with like Chat GPT teams or chat GPT enterprise that can get these sites up and running for their organization.
00:19:11.039 --> 00:19:12.799
Um, all right, next.
00:19:14.559 --> 00:19:15.599
Agent native.
00:19:15.759 --> 00:19:16.319
All right.
00:19:16.720 --> 00:19:23.680
Um this is I think more than a buzzword.
00:19:23.920 --> 00:19:24.240
All right.
00:19:24.319 --> 00:19:28.559
I've been thinking a lot about this and what it means to be agent native.
00:19:28.720 --> 00:19:39.440
And you know, I keep thinking back to you know, certain things in my background working in marketing, you know, things like you know, mobile first or mobile responsive, all these, right?
00:19:39.519 --> 00:19:48.480
There's this this huge push, um, you know, to get since people were using the internet more on their phones, you know, 15 or so years ago, there's this big push.
00:19:48.559 --> 00:19:49.680
It took 10 years, right?
00:19:49.759 --> 00:20:01.920
But eventually the entire web became mobile responsive by default because the majority of people are accessing the internet on their phones and apps and everything, your business, everything had to be mobile responsive.
00:20:02.079 --> 00:20:04.640
The same thing is going to happen, but going agent native.
00:20:04.799 --> 00:20:05.119
All right.
00:20:05.279 --> 00:20:09.599
So I'm not just talking internally, I'm talking externally as well.
00:20:09.839 --> 00:20:11.039
So, what does that mean?
00:20:11.200 --> 00:20:17.680
So, resources that your team makes, they need to get repurposed in an agent native format.
00:20:17.839 --> 00:20:35.680
So, you know, all that prompting that you go through and do, you and your team, right, all of it needs to turn into a skill because that is a skill that, or that is you um taking the outcome that you came to and making that outcome and the process that you got there agent native.
00:20:35.759 --> 00:20:51.680
In the same way that you know, teams would get together and build SOPs, um, right, that's uh what you should be doing here, but doing this for agents um according to the successful processes that you've uh brought into different AI operating systems.
00:20:51.839 --> 00:20:52.799
So that's one example.
00:20:52.960 --> 00:20:58.720
Another example is well, I think all websites should be marked out, right?
00:20:58.799 --> 00:21:00.559
Like there should be a button, right?
00:21:00.640 --> 00:21:06.640
Aside from having an agent first web, um, which I think is already happening in the background, right?
00:21:06.880 --> 00:21:09.680
I think many companies are gonna have two versions of their websites.
00:21:09.839 --> 00:21:23.200
The second version will probably be made automatically, where it just converts everything to markdown so it's easier for agents to go through without having to be uh token inefficient AF and spend billions of tokens just to understand what's going on in the website.
00:21:23.279 --> 00:21:40.480
But I think in the same thing, simple user interfaces where you know you can say, you know, save this as markdown or convert this to markdown, you know, being able to easily turn your um you know, your company's websites or anything that you publish online um so an agent can more easily access it.
00:21:40.559 --> 00:21:45.200
So just two quick little examples there of being agent native.
00:21:45.519 --> 00:21:47.599
All right, that's number 14.
00:21:47.759 --> 00:21:53.599
So every once everything is kind of agent native, well, you need a new way to keep score.
00:21:53.680 --> 00:21:59.680
And I kind of already referenced this, but I think that skill use is going to boom and become a company.
00:22:00.160 --> 00:22:00.400
Metric.
00:22:00.480 --> 00:22:02.319
What do I mean by a company metric?
00:22:02.400 --> 00:22:02.640
Right.
00:22:02.720 --> 00:22:08.240
There's all these things, I think, through the course of AI, uh, that have been like weird metrics, right?
00:22:08.400 --> 00:22:15.200
So, like early, earlier on, right, uh, when AI was really hitting software development hard, it was all about the lock, right?
00:22:15.279 --> 00:22:16.480
It's all about the lines of code, right?
00:22:16.559 --> 00:22:20.000
Everyone's like, oh, I wrote uh 87 billion lines of code today.
00:22:20.079 --> 00:22:21.680
So, you know, I'm awesome.
00:22:21.839 --> 00:22:26.880
Um, and then you know, late 2025, early 2026, it was the token leaderboards, right?
00:22:27.440 --> 00:22:30.160
Those two things say absolutely nothing.
00:22:30.400 --> 00:22:38.640
I think one of the most important um measurables in AI is going to be skill reuse.
00:22:38.880 --> 00:22:39.200
All right.
00:22:39.279 --> 00:22:43.839
I I talked about this on the show uh last week, right?
00:22:43.920 --> 00:22:46.160
Uh kind of how I AI, right?
00:22:46.480 --> 00:22:48.079
We did the entire thing.
00:22:48.319 --> 00:22:51.440
Um, I think I'd it's like 6,000.
00:22:51.680 --> 00:22:54.160
Uh I can I can actually bring it up here.
00:22:54.319 --> 00:22:56.079
Uh bring up my codex.
00:22:56.559 --> 00:22:57.200
Let's see.
00:22:57.359 --> 00:23:02.400
Okay, so I've used skills uh 6,700 times.
00:23:02.559 --> 00:23:05.839
Um, so it's a lot of skill use, um, right?
00:23:06.160 --> 00:23:07.519
Not could be more.
00:23:07.599 --> 00:23:17.200
Um, but I think that skill reuse is actually going to become one of the most important company metrics because when I said it is so easy to turn a repeatable process into a skill.
00:23:17.279 --> 00:23:28.640
Let's say that you're working with a large language model, you're going back and forth, you're sharing data, you're doing all these things, and you finally get um an outcome or an output or deliverable that really works for you know you, your company, et cetera.
00:23:28.799 --> 00:23:29.839
Well, what do you do with it?
00:23:30.079 --> 00:23:32.799
Maybe you might turn it into a project.
00:23:32.960 --> 00:23:33.920
Okay, that's cool.
00:23:34.079 --> 00:23:41.839
But you need to turn that into a skill, not just because you can invoke it or an agent can invoke it on its own without you even asking it, right?
00:23:42.079 --> 00:23:43.599
Which is usually helpful.
00:23:43.680 --> 00:23:45.039
Sometimes it's not helpful.
00:23:45.200 --> 00:23:54.160
But the other thing is that is how you actually build AI IP for your company because skills are transferable, right?
00:23:54.319 --> 00:24:06.319
So if you're just stuffing everything, you know, into a project inside, you know, Google or Copilot or ChatGPT or Claude or whatever, it that's it's not a it's not really an asset, right?
00:24:06.480 --> 00:24:10.000
Um, at that point, uh, that is a vendor lock-in.
00:24:10.319 --> 00:24:11.680
That's all you're doing at that point.
00:24:11.839 --> 00:24:25.759
So, you know, skill reuse is such an important uh part of of I think any measurement of a good someone that is being becoming um AI native is how much and how often you are reusing skills.
00:24:25.839 --> 00:24:44.319
Because in in theory, even if that skill is not always leading to a improved output or deliverable, it is undeniably saving an extraordinary amount of time, even versus normal AI use, let alone doing things the quote unquote analog way, right?
00:24:44.480 --> 00:24:47.119
Working on the internet, you analog.
00:24:47.440 --> 00:24:49.440
All right, so let's move on.
00:24:49.599 --> 00:24:52.799
So skills capture how works gets done.
00:24:53.039 --> 00:24:57.039
Number 12 is how you capture decisions getting made.
00:24:57.200 --> 00:24:58.079
And you know what?
00:24:58.240 --> 00:25:05.519
Finally, I think I had this in my prediction uh like two years ago, and I was too early, but we're finally starting to see this.
00:25:05.680 --> 00:25:10.480
Number 12 is company reasoning data finally surfacing as the goal that it is.
00:25:10.720 --> 00:25:26.480
So uh you're now seeing big companies like Zapier was one of the first that I saw talking about this about the concept of bringing like Slack DMs out of the DMs and into the open and why and what does that mean and what does it signal?
00:25:26.640 --> 00:25:35.279
So, you know, essentially when you have an agent um in your Slack, uh most agents cannot see DMs, right?
00:25:35.359 --> 00:25:37.759
Um, aside from your own personal agent.
00:25:37.920 --> 00:26:00.400
Um, but when you start moving important conversations from individual DMs or groups of DMs into channels, right, not only does that help the decision making uh help others maybe see the decision-making project uh progress, but most importantly, it allows agents to understand how your company or how your individuals think and make decisions.
00:26:00.640 --> 00:26:16.559
Yes, that's obviously out in the open in normal channels, but the concept of uh, you know, and I actually talked with the Slack CMO about this on the very show about how important it is um to get kind of that human reasoning, that one-on-one human reasoning.
00:26:16.720 --> 00:26:21.039
A lot of times that is what the agent so sorely needs, but we're not giving it.
00:26:21.119 --> 00:26:33.920
We're just throwing more spreadsheets, you know, more skills and all these other things when really it just needs to be able to see um the nuance of a conversation between, you know, two, three, four decision makers that was maybe private.
00:26:34.079 --> 00:26:41.519
So that's just one small um example of company reasoning data finally being serviced as gold.
00:26:41.680 --> 00:26:50.480
But I think that we are gonna see that become bigger and bigger as agents are more easily uh to go out and crawl and understand different contexts, right?
00:26:50.559 --> 00:26:55.039
People are way like, wait, you know, even things like just yapping, right?
00:26:55.200 --> 00:27:20.880
Uh kind of going back to my my um my thought of you know, voice and mobile being the the surface that is gonna be the most important, you know, the amount that I can speak something uh versus type, there's so much more nuance and even reasoning uh that I can fit in there when I am just naturally speaking what's on my mind versus what just gets typed out.
00:27:21.119 --> 00:27:22.799
All right, let's keep going.
00:27:23.039 --> 00:27:26.720
Number 11, the public leaderboards.
00:27:26.880 --> 00:27:32.240
I'm not saying they're they're going to die completely, but they're gonna begin to die.
00:27:32.319 --> 00:27:35.599
And I think that private evals are gonna start to take over.
00:27:35.759 --> 00:27:39.279
I'll say this like 90% of benchmarks are just becoming useless.
00:27:39.440 --> 00:27:45.920
I mean, there's some that are obviously extremely important to look at, regardless of the type of work that you and your company do, right?
00:27:46.079 --> 00:27:49.279
There's ones that are just great for general intelligence, right?
00:27:49.440 --> 00:27:54.640
Like uh GDP Val, uh the artificial analysis index, deep suite.
00:27:54.720 --> 00:27:56.319
I think some of those are just like no-brainer.
00:27:56.480 --> 00:27:57.759
So they're always going to be important.
00:27:58.000 --> 00:28:04.240
There's like hundreds of benchmarks, and people spend so much time looking and thinking uh about benchmarks.
00:28:04.400 --> 00:28:21.440
And aside from you know, benchmark uh bench maxing, uh, which is where you know frontier companies, you know, they overfit a model to perform very well on a specific benchmark, even if maybe humans or real work might not benefit uh as much.
00:28:21.759 --> 00:28:28.799
So I think that public uh leaderboards uh in some of these benchmarks are gonna decrease um very quickly.
00:28:28.960 --> 00:28:46.640
And I think the practical thing that's gonna happen is well, um company uh companies internal uh workbench, essentially, you know, kind of what I'm calling it, that's gonna become more important, right?
00:28:46.720 --> 00:28:50.480
Companies are gonna start, and some of the larger companies already have, right?
00:28:50.559 --> 00:28:56.480
They're actually, you know, creating these as you know, like public um evaluations, right?
00:28:56.559 --> 00:29:01.039
They're creating these as public benchmarks, things that they use for internally and then sharing them with others.
00:29:01.200 --> 00:29:07.440
I think those are gonna become more and more um important, but more and more common, right?
00:29:07.519 --> 00:29:12.559
I think it was like uh DoorDash or something that shared one of theirs, and I was like, okay, this is random.
00:29:12.799 --> 00:29:19.279
I'm like, no, this is actually very important because they took the models and it maybe didn't for them.
00:29:19.359 --> 00:29:30.880
Uh, you know, a certain model might have been much better or much worse than even in the corresponding evaluation, just because of the type of work that certain people do is so specific.
00:29:31.839 --> 00:29:35.599
All right, uh, and then next one kind of related.
00:29:35.759 --> 00:29:41.359
I think model routing is gonna become the norm, and eventually it is gonna become built in.
00:29:41.519 --> 00:29:51.039
Okay, so here's what I mean by that is most companies should stop defaulting to the top model because you don't need it.
00:29:51.519 --> 00:29:58.000
All right, you literally don't uh need it for most of your work, all right.
00:29:58.079 --> 00:30:04.400
And this also uh plays into kind of the um you know model or capability overhang.
00:30:04.640 --> 00:30:06.720
A year ago, I would have said the opposite.
00:30:06.880 --> 00:30:16.799
I would say absolutely use the best model available, even if it takes longer, because you need it for your work, you don't need it anymore, right?
00:30:17.039 --> 00:30:28.960
You don't need uh, you know, Fable 5 Max or GPT 5.6 Soul Max um to you know browse the web for you.
00:30:29.119 --> 00:30:39.359
You know, you don't need uh Opus 5 or you know uh Max or GPT 5.6 Terra to rewrite an email, right?
00:30:39.440 --> 00:30:40.240
You don't.
00:30:40.640 --> 00:30:44.160
You know, there's very capable models that cost a lot less.
00:30:44.319 --> 00:30:47.599
So if you're paying for it on the API side, you've already been through this, right?
00:30:47.680 --> 00:31:02.559
But even on the subscription side, right now, uh I think eventually the the free ride of AI, even in subscriptions, will become less subsidized, and you're gonna have to start dealing um you know with usage limits um and and quotas.
00:31:02.799 --> 00:31:08.160
So model routing, it's gonna have to become a default, and it's actually very easy, right?
00:31:08.240 --> 00:31:12.960
I've I've made um skills that do this for me, right?
00:31:13.039 --> 00:31:26.160
I have a skill where you know I give a big project to uh you know GPT 5.6 soul and it breaks it up and it uses sub agents at its discretion, but lower volume ones.
00:31:26.240 --> 00:31:32.559
So, you know, I really just have it work as uh working as the orchestrator in many instances.
00:31:32.640 --> 00:31:37.039
And you know, I say if if it requires heavier lifting, you can do some of that middlework.
00:31:37.200 --> 00:31:47.119
But for a lot of those things, well, I'm just using, you know, maybe Seoul in the top 10 and last 10%, and then using a bunch of you know smaller sub agents in between.
00:31:47.200 --> 00:31:48.880
But I think that's gonna become the norm.
00:31:49.039 --> 00:32:12.000
I think there's gonna be some uh and there already has been some interesting, you know, AI startups like merge and others that have come out that this is literally what they do, uh, right, because I think it's gonna save a lot of companies a lot of money, and that's something that companies are looking to do as the uh model overhang starts to become larger and larger.
00:32:12.240 --> 00:32:13.920
And what do I mean by that?
00:32:14.160 --> 00:32:35.920
I think that 90% of most knowledge work can be done with using only 10% of a models of that top top models intelligence and just wait, you know, until we get Fable 5.1 and we get you know GPT 6 Astra or GPT 5-7 Astra, right?
00:32:36.000 --> 00:32:36.799
Whatever it is.
00:32:37.039 --> 00:32:50.799
Um our work as humans is not growing in capabilities at the same rate that the models are going, they're just not any.
00:32:51.200 --> 00:32:58.160
Yeah, like I said, two years ago, there were models that you know, the best models weren't capable to do all the work that we do.
00:32:58.400 --> 00:33:01.759
A year ago, I would say it was maybe in parity, right?
00:33:01.920 --> 00:33:05.359
But yeah, you needed that most powerful model to do your hardest work.
00:33:05.440 --> 00:33:13.359
And now, for most people's hardest work, like I said, you know, if you you know, sonnet might be good enough, right?
00:33:13.440 --> 00:33:21.200
GP or or you know, Opus, um you know, opus might be overkill, you know, let alone Fable, right?
00:33:21.279 --> 00:33:32.160
You might just need Claude Sonnet, or you might just need, you know, GPT 56 uh Luna is plenty for many people, but that's why I think model routing is gonna become extremely important.
00:33:32.319 --> 00:33:35.200
All right, we're gonna pick up the pace so this doesn't turn into a two-hour show.
00:33:35.279 --> 00:33:38.880
Number nine, intelligence is gonna get cheaper, cheaper, cheaper, cheaper.
00:33:38.960 --> 00:33:45.440
And what that means, I think in the short term at least, is anthropic is gonna get squeezed, right?
00:33:45.519 --> 00:33:46.880
I've talked about this a lot on the show.
00:33:47.039 --> 00:33:55.920
If you if you care about you know who's who's your vendor, I'm not telling you to choose one or the other, but you have to understand what's going on, and it's been absolutely nutty the last two weeks.
00:33:56.079 --> 00:34:04.160
So, as an example, open AI used a sniffling of um you know, recursive self-improvement.
00:34:04.240 --> 00:34:05.359
You know, they use GBD5.
00:34:09.280 --> 00:34:12.000
They dropped the price of uh GBD5.
00:34:14.480 --> 00:34:16.480
So that model is nearly free.
00:34:16.639 --> 00:34:19.360
That is like nearly sonnet level, all right.
00:34:19.440 --> 00:34:20.800
Uh, but it's not just them.
00:34:20.880 --> 00:34:26.559
I mean, Groc4.6 that just came out, uh, so good, right?
00:34:26.639 --> 00:34:31.039
Uh, at least you know, first looks and cost per task.
00:34:31.280 --> 00:34:36.800
Um, you have Meta's Muse Spark 1.2, Kimmy K3, Quen 3.8.
00:34:36.960 --> 00:34:46.960
All of these models deliver 95 to 99 percent of the intelligence um of the anthropic models, but at five to 30 percent of the cost.
00:34:47.039 --> 00:34:50.239
Those aren't random numbers, those are the real numbers there.
00:34:50.880 --> 00:34:52.480
So, what do I mean by that?
00:34:52.639 --> 00:34:55.199
Well, intelligence is gonna get a lot cheaper, right?
00:34:55.280 --> 00:35:03.039
It used to be the thing where you know you could just choose your, you know, choose your horse and you would know that your horse would always, you know, stay competitive.
00:35:03.280 --> 00:35:04.079
I don't know why.
00:35:04.159 --> 00:35:16.159
It looks like anthropic um is not going to follow the trend here of um, you know, making intelligence cheaper and cheaper at the same time stronger and stronger.
00:35:16.320 --> 00:35:19.119
So um that's gonna be something to definitely keep an eye on.
00:35:19.280 --> 00:35:25.360
But I think as the school year drags on, I think intelligence is only going to get cheaper and cheaper.
00:35:25.519 --> 00:35:33.199
So your provider that you're sticking with, if they're not dropping their price, uh price per performance, yeah, might want to look another way.
00:35:33.360 --> 00:35:33.920
Just saying.
00:35:34.079 --> 00:35:34.559
All right.
00:35:34.719 --> 00:35:36.000
Uh, number eight.
00:35:36.400 --> 00:35:45.039
And this kind of all goes together, but I think Fortune 100 token spend is going to slow while usage explodes.
00:35:45.199 --> 00:35:46.800
And let me explain what I mean by that.
00:35:46.880 --> 00:35:52.239
Uh, similarly to, you know, I was talking about anthropic just selling tokens.
00:35:52.400 --> 00:36:06.400
Um, I think Fortune 100s, at least right now, are the only ones truly in the position uh to be able to, well, have the CapEx or uh the facilities to be able to run local models at scale.
00:36:06.480 --> 00:36:16.719
Because right now, um, you know, again, unless my joke is unless you have a data center sitting in your basement, you can't run these open uh models, right?
00:36:16.880 --> 00:36:22.239
The ones that, you know, oh, you know, everyone's like, oh my gosh, open source models are here.
00:36:22.400 --> 00:36:25.119
So, you know, all the big companies, what are they gonna sell anymore?
00:36:25.199 --> 00:36:29.440
Well, yeah, you can't run the open source models on consumer hardware.
00:36:29.599 --> 00:36:32.960
You need a couple million dollars worth of GPUs to run them.
00:36:33.039 --> 00:36:37.840
But that's where it's like, okay, well, the Fortune 100 companies, they've already started to do it.
00:36:38.000 --> 00:36:47.280
Um, they're gonna start to shift a lot of that lower hanging work, I think, to open models that they're running on-prem.
00:36:47.440 --> 00:36:56.000
Um, is is that gonna trickle, trickle over to the rest of the Fortune 500 and uh, you know, your average everyday enterprise this year?
00:36:56.239 --> 00:36:57.440
Absolutely not.
00:36:57.599 --> 00:37:16.159
Um, but I mean, I think that you're gonna see token usage go up at these companies, but their token spend or the amount of tokens that they're buying from a provider go down, they're gonna stop uh, you know, renting AI and they're gonna start owning it.
00:37:16.239 --> 00:37:28.400
Uh, you know, and I think that that is gonna become more and more common over the years, but I think at least for this year, if you are listening at one of those Fortune 100s, you're probably already having those type of conversations.
00:37:28.559 --> 00:37:33.679
If you are, you know, maybe in the fortune, you know, 101 to 500, right?
00:37:33.760 --> 00:37:35.199
These are just loose numbers.
00:37:35.280 --> 00:37:47.679
Um, it's something that you should probably start paying attention to and start crunching the numbers because I know a lot of companies that are spending um nine figures, you know, literally nine figures on AI.
00:37:47.840 --> 00:37:52.000
You know, so at a certain point, you have to say, okay, how much of that?
00:37:52.079 --> 00:37:54.559
Um, you know, let's let's look at our model routing.
00:37:54.639 --> 00:37:55.920
Are we routing models at all?
00:37:56.079 --> 00:37:57.599
How much are we overpaying, right?
00:37:57.760 --> 00:38:20.880
How many people are are using, you know, uh a fable five to to rewrite their emails, or how many people are using, you know, you know, Google Gemini, you know, 3.6 flash to you know classify something when you could be using a much more cost-efficient model that is still um you know far capable to do the job.
00:38:21.119 --> 00:38:21.599
All right.
00:38:21.760 --> 00:38:25.519
Next, number seven, uh, compute becomes currency.
00:38:25.760 --> 00:38:29.920
All right, uh, you know, uh NVIDIA CE uh CEO Jensen Wong said something like this.
00:38:30.159 --> 00:38:35.280
Obviously, this is what his company does, uh, but I think it's the truth.
00:38:35.440 --> 00:38:50.079
Um, you know, there was a uh famous um, you know, uh Dario Modi uh interview, the CEO of Anthropic, I think it was from 2023, uh, where you know he was being asked about compute and investing in these data centers.
00:38:50.159 --> 00:38:56.400
And, you know, he essentially said, you know, we're we're gonna take it, we're gonna be cautious because we don't want to overspend.
00:38:56.719 --> 00:38:58.719
And what if the AI demand isn't there?
00:38:58.880 --> 00:39:16.880
And now obviously Anthropic is having to kind of pay for that underestimation uh of AI because they're having to overpay um, you know, and have all these other partnerships and they just aren't able to keep up with demand, which is why their servers uh and services are down more than anyone else in the industry.
00:39:17.039 --> 00:39:27.039
You know, on the flip side of that, open AI uh invested heavily early on and they're able to serve things at a much higher uptime up time uptime rate, right?
00:39:27.199 --> 00:39:36.000
But ultimately, um I think as more and more workflows, right, we're still early in the game, right?
00:39:36.159 --> 00:39:41.519
If you are listening to this, if you're using AI every single day, we are in that one percent.
00:39:41.599 --> 00:39:42.400
We are in the bubble.
00:39:42.559 --> 00:39:44.239
Most people are not there yet.
00:39:44.480 --> 00:39:50.719
So as the rest of the world catches up and says, Oh, wait, this AI thing is pretty essential, right?
00:39:50.800 --> 00:39:56.400
It's more, you know, it's as essential as you know, the internet, it's as essential as electricity.
00:39:56.559 --> 00:40:00.880
As the rest of the business world, you know, stumbles across this over the next few years.
00:40:01.039 --> 00:40:02.800
I think this scarcity, right?
00:40:02.960 --> 00:40:05.679
Hardware scarcity becomes a real thing.
00:40:05.840 --> 00:40:10.320
Um, you know, but at the same time, I think at that point you say, oh, okay, we understand it now.
00:40:10.400 --> 00:40:14.400
It doesn't matter how good your model is if you can't run it, right?
00:40:14.480 --> 00:40:21.119
And I'm saying, you know, compute in the same way the companies like Anthropic, OpenAI, Microsoft, Google, Amazon, Meta, etc.
00:40:21.440 --> 00:40:25.280
Uh, but in the same way, compute to run things locally, right?
00:40:25.440 --> 00:40:26.719
That does become the new currency.
00:40:26.800 --> 00:40:41.280
Because if you, you know, are uh you know, a tech, you know, if you work at a big tech company and oh my gosh, well, we actually have servers, well, all of a sudden you have the ability to um print money, right?
00:40:41.440 --> 00:40:43.599
Because everyone wants that compute.
00:40:44.079 --> 00:40:45.599
All right, let's keep going.
00:40:45.760 --> 00:40:53.679
Uh, number six, kind of the prediction here for the rest of the school year AI controversies are gonna hit your backyard, they're going very local.
00:40:53.840 --> 00:41:03.840
So between local data centers and AI deepfakes that I think are gonna hit really hard in the elections, I think AI controversy is going hyper local.
00:41:04.000 --> 00:41:06.159
We've already started to see it with the data centers.
00:41:06.239 --> 00:41:11.599
And let me just put this out there there's a lot of bad information out there about data centers.
00:41:11.679 --> 00:41:13.280
I'm not gonna rally on that.
00:41:13.440 --> 00:41:16.000
I think there's plenty of people having those conversations.
00:41:16.159 --> 00:41:33.199
I'm just trying to more focus on the practical use of AI, but I think it is going to become a dinner time conversation uh at some point this year between, you know, the AI data centers and you know, certain communities, you know, saying they don't want them, certain communities welcoming them in.
00:41:33.360 --> 00:41:40.719
It's gonna create, I think, um, you know, income disparities in places that maybe didn't see it before, swinging both ways.
00:41:40.800 --> 00:41:46.079
Um, but I think ultimately where it's gonna hit hardest is local and state politics, y'all.
00:41:46.159 --> 00:41:47.280
This is my previous life.
00:41:47.440 --> 00:41:53.119
I was uh uh you know a politics reporter uh for the Chicago Sun Times.
00:41:53.199 --> 00:41:55.519
Um at the big level, right?
00:41:56.159 --> 00:42:06.320
You know, presidential race, everyone knows when you know AI is being used, even AI that looks real um or sounds real locally, I think, right?
00:42:06.400 --> 00:42:14.480
So statewide races, I think you we're gonna see dozens of stories about people using AI.
00:42:14.559 --> 00:42:19.519
And a lot of people aren't gonna figure out until it's too late, but for bad reasons, bad actor AI, right?
00:42:19.679 --> 00:42:31.119
People, you know, creating uh fake audio of their competitors, you know, videos um, you know, that they're gonna use in ads that didn't actually happen, it is gonna become commonplace.
00:42:31.280 --> 00:42:39.199
And I don't think most people will be able to tell the difference, which is why I think the AI controversies are gonna be hyper local.
00:42:39.840 --> 00:42:40.320
All right.
00:42:40.480 --> 00:42:46.159
Similarly, well, I think the AI backlash is going to become mainstream.
00:42:46.320 --> 00:42:56.480
Uh, so as the uh shifts start stacking up, um, I think it's gonna start with the work slop avalanche.
00:42:56.559 --> 00:42:59.280
And I think it's already kind of started.
00:42:59.440 --> 00:43:03.599
Um, and a couple of recent developments are gonna think, I think, um, exasperate this.
00:43:03.679 --> 00:43:13.679
So, you know, as an example, uh Anthropic uh kind of announced to uh adhere to some EU regulations that they're gonna start uh putting in some invisible watermarking in text.
00:43:13.760 --> 00:43:30.320
So even if you copy and paste it and try to do all this stuff, you know, um, but essentially, I think that uh uh kind of movement has also led to the resurgence of the you know AI content detectors, which I'm not a fan of because most of them are uh not good.
00:43:30.400 --> 00:43:42.719
Uh, but I think we are gonna see this this weird paradigm because every single company is rightfully so, um, you know, trying to implement AI from top to bottom.
00:43:42.880 --> 00:43:49.599
Yet, you know, now I think more than ever, you're gonna see people getting called out for using AI, right?
00:43:49.679 --> 00:43:56.400
Whether it's you know, like this clawed watermarking or some other um, you know, you know, AI text detection.
00:43:56.480 --> 00:43:59.599
Um, but there's gonna be, I think, a huge backlash.
00:44:00.079 --> 00:44:00.960
Against using AI.
00:44:01.039 --> 00:44:02.559
I even found myself, right?
00:44:02.639 --> 00:44:06.159
I was I was looking for uh what was it?
00:44:06.320 --> 00:44:08.239
Uh a shirt or something online.
00:44:08.800 --> 00:44:14.800
Um, and you know, I just noticed that they were using all you know AI models.
00:44:14.960 --> 00:44:19.360
And it's like I knew that 95% of people weren't gonna be able to tell.
00:44:19.519 --> 00:44:24.800
It wasn't the best AI uh images, but you know, I saw that and I'm just like, nope, next, right?
00:44:25.039 --> 00:44:26.079
So even me, right?
00:44:26.239 --> 00:44:30.400
And people always think that I'm oh AI, everything, and I'm not, um, right.
00:44:30.480 --> 00:44:33.599
I it's just like I think it should always be tastefully done.
00:44:33.679 --> 00:44:44.639
But I think the AI backlash, because it is more and more accessible, and more and more people are going to use it, um, it's gonna hit hard, but I think it's gonna go mainstream.
00:44:44.719 --> 00:44:55.119
Like I talked about this years ago, but uh I think we're gonna start to see it uh where you know people are gonna start putting things out as human-made, and that's gonna become like a trendy thing.
00:44:55.199 --> 00:44:57.599
But I think the AI backlash is gonna hit hard.
00:44:57.760 --> 00:45:03.760
Uh, number four, and I have no clue what this means, but one of my predictions is gonna be that math is solved, right?
00:45:04.000 --> 00:45:18.400
I still don't quite understand how there's all these dozens of math theories that you know people, you know, the smartest mathematicians in the world have spent their careers on and it's they they've been unsolved uh for for forever, right?
00:45:18.480 --> 00:45:20.880
People have been working on these problems for hundreds of years.
00:45:21.039 --> 00:45:23.280
I think math gets solved, right?
00:45:23.519 --> 00:45:25.039
I don't know what it means, right?
00:45:25.119 --> 00:45:33.760
As an example, uh, the unreleased GPT, uh, whatever number it was, Astra, which is their next family of models on top of Seoul.
00:45:34.000 --> 00:45:38.079
Uh, it knocked out like 10 problems with published proofs.
00:45:38.239 --> 00:45:46.960
Um, but I think math in general is gonna get solved, and that's gonna lead to a lot of things that are above my pay grade of understanding.
00:45:47.119 --> 00:45:56.320
Um, obviously, uh great things, uh medicine development, protein synthesis, all these other things that can lead to positives.
00:45:56.480 --> 00:46:09.119
But I also do assume that when math gets solved, um, and that's gonna be able to, I think, um, on your open models that can run on consumer devices are gonna have math solved uh probably within the year.
00:46:09.280 --> 00:46:13.039
So I'm guessing there's gonna be some downsides to that as well, right?
00:46:13.199 --> 00:46:15.840
I don't know, like cracking crypto, right?
00:46:15.920 --> 00:46:18.559
We're still probably uh a way off on that.
00:46:18.639 --> 00:46:23.760
Um, but I think that there's gonna be some downsides to math being solved.
00:46:23.840 --> 00:46:29.360
Um, but in general, I think math benchmarks are no longer gonna be a meaningful benchmark.
00:46:29.519 --> 00:46:33.760
And if anything, this should just tell you how good uh AI has gotten.
00:46:33.920 --> 00:46:43.679
Because three years ago, right, when ChatGPT kind of uh you know started to become consumer mainstream, it couldn't consistently solve two plus two, right?
00:46:43.920 --> 00:46:46.639
Now I think it's literally gonna solve math.
00:46:46.800 --> 00:46:49.840
All right, next three token maxing coming back.
00:46:50.000 --> 00:46:50.320
All right.
00:46:50.400 --> 00:47:06.239
So this kind of ties in with uh one of mine earlier about intelligence is coming too cheap, but essentially, right, token maxing where it's like, let's spend as many tokens as possible, was very much in vogue uh in you know, like December 2025 through like February.
00:47:06.400 --> 00:47:12.960
And then there was the whiplash of uh, oh wait, no, we can't anymore because whoops, we're spending way more money than we thought.
00:47:13.039 --> 00:47:20.559
We were trying to encourage employees to use AI and they were all spending billions of tokens, you know, spinning up useless things.
00:47:20.639 --> 00:47:26.960
So now there's been this shift um toward uh kind of value maxing or token efficiency.
00:47:27.199 --> 00:47:28.960
But now I think we're going back, right?
00:47:29.119 --> 00:47:33.840
Just like the mom jeans, you you know, they were hot, they were gone, they're back.
00:47:34.000 --> 00:47:36.480
I think token maxing is already going to come back.
00:47:36.559 --> 00:47:38.480
And I mean, look at the last week.
00:47:38.639 --> 00:47:39.760
I mean, how could you not?
00:47:39.920 --> 00:47:40.880
Look at Luna.
00:47:40.960 --> 00:47:48.880
I am spending more tokens than ever, um, right, because the tokens are getting cheaper.
00:47:49.119 --> 00:47:49.519
All right.
00:47:49.679 --> 00:47:53.920
So that's why I think token maxing is actually going to return, right?
00:47:54.320 --> 00:47:56.320
Experimenting, running more and more agents.
00:47:56.400 --> 00:48:03.519
So I think we went through this weird phase of like, you know, three to five months where it was like token budgets are getting tight, right?
00:48:04.079 --> 00:48:06.960
Spend is getting tight, which was reactionary.
00:48:07.119 --> 00:48:14.880
It was not looking forward into the future because in if you were looking forward in the future, you see intelligence is going to come down, right?
00:48:14.960 --> 00:48:25.119
So this this this concept of token maxing, um, or you know, or sorry, the concept of token efficiency or value maxing still hasn't fully caught on, right?
00:48:25.280 --> 00:48:28.880
So it's already made a complete cycle.
00:48:29.199 --> 00:48:30.880
Um, I think now.
00:48:31.039 --> 00:48:40.400
So get ahead of it because I think still this this the the the token reckoning hasn't fully hit outside of like Silicon Valley, right?
00:48:40.480 --> 00:48:41.679
It has in some places.
00:48:41.840 --> 00:48:53.679
So if you're listening to this, get ahead and you know, if someone's you know, some smart guy comes in, Bill comes in on Tuesday, and he's like, All right, well, I don't know if you guys know this, but tokens are getting out of control.
00:48:53.760 --> 00:48:56.000
So we have to, you know, talk about efficiency.
00:48:56.079 --> 00:49:00.719
You can say, All right, Bill, uh, you know, it's no longer February 2026.
00:49:00.880 --> 00:49:03.360
The tokens are actually cheaper than ever.
00:49:03.519 --> 00:49:09.039
And the tokens now compared to what they could do, uh, you know, four months ago, night and day.
00:49:09.199 --> 00:49:11.199
Token maxing coming back, book it.
00:49:11.360 --> 00:49:16.719
Number two, the open agents are gonna crash in 2027.
00:49:16.880 --> 00:49:17.119
All right.
00:49:17.280 --> 00:49:18.880
This is the full school year.
00:49:19.039 --> 00:49:24.880
All right, we're we're going all the way into uh May, May or June 2027.
00:49:25.119 --> 00:49:27.119
Open agents are gonna crash everything.
00:49:27.199 --> 00:49:35.519
I talked about agent crashes um on my uh 2026 AI prediction and roadmap series, which obviously they've been crashing, right?
00:49:35.599 --> 00:49:49.119
We had all the stories, the the uh open AI, hugging face, uh Kimmy K3 anthropics models, every so the agents are crashing right now, but the open crashes are what's going to really matter.
00:49:49.280 --> 00:49:52.639
So I had a show on this the other day, if you want to go back and listen to it.
00:49:52.960 --> 00:50:00.559
But I think essentially the open models today aren't good enough to cause chaos yet in bad actors' hands.
00:50:00.800 --> 00:50:07.519
But the open models that we should be getting around quarter four, they will be because those models are gonna be fable level.
00:50:07.679 --> 00:50:14.000
Those models are gonna be, you know, Astra, potentially Astra level, at least GPT-5-6 soul level.
00:50:14.159 --> 00:50:15.599
So what does that mean?
00:50:15.760 --> 00:50:25.760
Uh, that means that they're gonna be smart enough or conniving enough when bad actors are using them to be able to skirt around the guardrails that are built in.
00:50:25.840 --> 00:50:30.239
And obviously, with open models, once you release them, you can't rein them in, right?
00:50:30.320 --> 00:50:38.320
You can't take uh, you know, Kimi K31 or you know, Quen3.9, you can't take them offline, right?
00:50:38.400 --> 00:50:39.599
Once they're out, they're out.
00:50:39.679 --> 00:50:47.280
Um, and obviously, because they're open models, uh, you can fork them, you can fine-tune them, you can do what you can to bring the guardrails down.
00:50:47.360 --> 00:50:50.480
And I think that is gonna happen at scale.
00:50:50.719 --> 00:50:52.639
Um, just FYI.
00:50:52.800 --> 00:50:53.760
So, what does that mean?
00:50:53.920 --> 00:51:07.599
I think cyber defense and cybersecurity, um, especially of the AI variety, it's gonna be the fastest growing sector um in 2027, and it's not even gonna be close.
00:51:07.679 --> 00:51:08.960
It is not even gonna be close.
00:51:09.280 --> 00:51:12.960
Assuming, right, that it's gonna happen at the rate that it's gonna happen.
00:51:13.119 --> 00:51:22.159
I mean, even small businesses, you know, SMBs, smaller enterprises that normally wouldn't have a strong cybersecurity presence.
00:51:22.320 --> 00:51:27.760
I'm not saying that they're gonna go hire a cybersecurity team of 10, but all of a sudden it's gonna be a budget item.
00:51:27.920 --> 00:51:33.360
If you're not gonna hire people, you're gonna have to hire a third-party company and it is not gonna be cheap.
00:51:33.519 --> 00:51:38.320
So start penciling that in now, the everyday average business, right?
00:51:38.559 --> 00:51:49.280
Your little, you know, not saying little, but you know, your average company with 200 employees, 500 employees, right, that have maybe never been the victim of a cyber attack.
00:51:49.440 --> 00:51:59.440
When agents, when open agents are gonna be able to find vulnerabilities in your CRM vulnerabilities in your payment systems, they will be able to wreak havoc.
00:51:59.679 --> 00:52:01.440
You have to be prepared for it now.
00:52:01.760 --> 00:52:08.960
And then number one, last but not least, and this kind of ties a lot of the stuff together, RSI is on the horizon.
00:52:09.119 --> 00:52:10.559
Recursive self-improvement.
00:52:10.719 --> 00:52:15.360
I also had a recent show about this, but actually some more recent developments.
00:52:15.519 --> 00:52:23.199
So, Sergei Brin, the most one of the most famous people in AI ever, who now seems to have a little bit more.
00:52:23.519 --> 00:52:28.159
Uh, you know, he's always been at the top of the Google AI food chain.
00:52:28.239 --> 00:52:45.360
But now it seems like um, you know, with some recent leadership shakeups at Google, um, you know, he might have a little bit more, I wouldn't say sway or pull, um, but it seems like he's gonna be uh a much more consequential figure um in Google with some of their other big names gone.
00:52:45.519 --> 00:52:56.079
But you know, just recently he just declared that recursive self-improvement is a top priority for Google Gemini and kind of um directing and wanting companies to go there.
00:52:56.159 --> 00:52:57.360
And we've already seen this.
00:52:57.519 --> 00:52:59.599
Um, you know, we've already seen glimpses of this.
00:52:59.679 --> 00:53:07.599
I've already mentioned, you know, GVD56 Soul as an example, making uh its smaller models more optimized and cheaper.
00:53:07.760 --> 00:53:23.840
Um and we've already seen stories that you know, companies say that their models that are coming out now were um, you know, um whether going through post-training or, you know, they're not always um giving the details, but today's models are starting to help build tomorrow's models.
00:53:23.920 --> 00:53:28.239
Um, and with recursive self-improvement, that makes everything else.
00:53:28.400 --> 00:53:32.559
The other 18 on this list go that much faster.
00:53:32.719 --> 00:53:33.840
All right, that's a wrap.
00:53:33.920 --> 00:53:36.960
I'm gonna go through these one more time, rapid style.
00:53:37.039 --> 00:53:43.039
So you don't forget, number 19, reactive chat is gonna start dying and proactive agents are gonna take over.
00:53:43.119 --> 00:53:47.679
Uh, number 18, voice mode and mobile become the default AI interface.
00:53:47.840 --> 00:53:51.440
Number 17, manager threads are gonna replace one-off chats.
00:53:51.599 --> 00:53:56.159
Number 16, multiplayer AI, humans and agents are gonna share one workspace.
00:53:56.320 --> 00:54:00.320
Number 15, vibe operations, the whole company is gonna start building.
00:54:00.480 --> 00:54:03.840
Number 14, everything you make goes agent native.
00:54:04.000 --> 00:54:07.280
Number 13, skill reuse becomes a company metric.
00:54:07.440 --> 00:54:11.360
Number 12, company reasoning data finally surfaces as gold.
00:54:11.519 --> 00:54:16.079
Number 11, public leaderboards die and private evals are gonna take over.
00:54:16.239 --> 00:54:19.760
Number 10, model routing becomes the norm and built in.
00:54:19.920 --> 00:54:24.000
Number nine, intelligence gets cheap and anthropic gets squeezed.
00:54:24.159 --> 00:54:30.000
Number eight, fortune 100's token spend slows while usage explodes.
00:54:30.239 --> 00:54:33.519
Um, number seven, compute becomes currency.
00:54:33.679 --> 00:54:36.079
Number six, AI controversies hit your backyard.
00:54:36.159 --> 00:54:38.800
Number five, the AI backlash goes mainstream.
00:54:38.960 --> 00:54:42.719
Number four, math is solved, whatever that means by AI.
00:54:42.880 --> 00:54:45.360
Number three, the return of token maxing.
00:54:45.440 --> 00:54:48.480
Number two, open agents crash in 2027.
00:54:48.639 --> 00:54:52.639
And number one, RSI is on the horizon.
00:54:52.880 --> 00:54:57.039
All right, that is a wrap for today's show.
00:54:57.360 --> 00:55:00.320
But remember, a quick announcement.
00:55:00.639 --> 00:55:01.920
Now you're ready.
00:55:02.079 --> 00:55:03.519
I gave you the answers.
00:55:03.760 --> 00:55:08.559
And maybe some of this was a little over your head.
00:55:08.719 --> 00:55:11.039
And you're like, wait, I'm kind of new here.
00:55:11.280 --> 00:55:12.639
I need some of the basics.
00:55:12.800 --> 00:55:25.519
Well, go back to the basics because, like I said earlier, uh, we're we're taking a little break from our normal Monday to Friday schedule uh to replay the entire and restart the start here series.
00:55:25.679 --> 00:55:32.719
So this is a series made for both beginners and people who are using AI every single day and leaders in their organization.
00:55:32.960 --> 00:55:38.079
We're gonna play episodes one through 30 in order starting Monday.
00:55:38.159 --> 00:55:43.840
So if this got your brain turning and you're like, okay, these are some good things for our company to talk about.
00:55:44.000 --> 00:55:49.760
Or if you were like, wait, I need to know more about you know the these open models.
00:55:49.920 --> 00:55:54.320
I need to learn more about you know some of these topics on agents that you talked about.
00:55:54.480 --> 00:56:06.880
Well, in the start here series, we tackled most of these things all in depth, uh, at a higher level, more detail, more examples, more use cases for your company to grow.
00:56:07.119 --> 00:56:08.079
So that's it.
00:56:08.159 --> 00:56:11.119
Make sure to go back to school with us starting Monday.
00:56:11.280 --> 00:56:12.400
So that's a wrap.
00:56:12.480 --> 00:56:13.599
I hope this was helpful.
00:56:13.760 --> 00:56:19.440
If so, please, if you haven't already, subscribe to the podcast, then go to your everydayai.com.
00:56:19.519 --> 00:56:20.320
Thanks for tuning in.
00:56:20.480 --> 00:56:23.280
See you back tomorrow and every day for more everyday AI.
00:56:23.360 --> 00:56:24.400
Thanks, y'all.
00:56:25.679 --> 00:56:29.039
And that's a wrap for today's edition of Everyday AI.
00:56:29.199 --> 00:56:30.320
Thanks for joining us.
00:56:30.480 --> 00:56:34.239
If you enjoyed this episode, please subscribe and leave us a rating.
00:56:34.400 --> 00:56:36.000
It helps keep us going.
00:56:36.239 --> 00:56:43.280
For a little more AI magic, visit your everydayai.com and sign up to our daily newsletter so you don't get left behind.
00:56:43.519 --> 00:56:46.159
Go break some barriers, and we'll see you next time.
