Watching AI Grow Up in Real Time [Raw Session]

00:00 — AI Still Feels Massive

Hey, welcome back to slow builds the key come back to this idea that maybe we’re looking at our at AI a little too close up at the moment right now.

It feels extremely massive. It feels very expensive.

It feels like it needs very specific things to work chips data centers cooling power infrastructure and token management.

It needs all these new systems just decide which model to use which tools we’re gonna call what data should go, what should run locally, what should go to bigger models and what’s actually what worth paying for.

And I get why companies are focused on that right now because the cost is real.

Infrastructure energy hardware all of it is real and it’s expensive.

Space.

Money.

Wondering if we’re making the same mistake people often make with new technology. We look at the early version and assume that that’s the shape it’s going to be forever. What we see today is what it will always be and I don’t think that usually it’s never how it works.

The early version of technology is often very big, very awkward, overly expensive and hard to use.

Then over time things get smaller, cheaper, they get more normal, they get standardized, it moves into the background.

And everything eventually people forget how strange it was to use or how it looked in the very beginning.

So that’s what I want to think through in this video a little bit.

Maybe it’s not in its final form right now.

Maybe I AI is still still in this room size computer era where and what that means is like IBM back with the old big machines that took up entire rooms when your mobile phone in the car was the size of your head.

That’s the idea here.

Like right now we see AI and it takes up all this space.

It’s this massive thing that it costs a lot of money.

02:10 — Early Limits Feel Permanent

So like I said, when you’re living through a technology shift like we currently are, the limits feel permanent.

And that’s the trap.

Whatever technology requires today starts to feel like part of the technology itself.

So with AI right now, we say it needs CPUs.

It needs massive data centers, huge amounts of power, extensive cooling systems, tons of water, high speed connections, expensive model calls.

It needs all this infrastructure around it.

And all that is true right now.

I’m not saying that part is fake, but the question is whether those are permanent requirements or just the early stages of what we’re beginning to develop.

Because those are very different things.

If you went back for enough for in computing, computers were not something you casually carried around.

They were machines, physical machines, large machines.

They began in special rooms, special environments, needed special operators.

Had full on instruction manuals around them.

Eventually computers became something that could sit on a desk.

I remember when I had my Tandy 1000, like it was, to me it was small and was a home computer and I could plug it in and I loved it.

Now that feels so clunky.

My wife, she uses a laptop for work.

I said, well, we can get you a desktop from work.

We have a bunch and they actually run a lot faster.

And to her like that’s, it doesn’t make sense.

Why would I have that big machine in the house?

But that’s the way it used to be.

And in a Tandy 1000 that I had, that was considered to be a small machine.

And even some of the old laptops we had, those were considered small.

And now like they’ve seemed so bulky compared to like a MacBook Air.

’Cause like I said, everything moved to something you put in a bag.

And with the phone, it’s something you can put in your hand.

And now we really even think of a phone as a computer, really, even though it’s more powerful than any computer ever had growing up and that’s a very strange thing to think about.

Once technology matures enough we stop seeing the infrastructure.

We just see the use.

We don’t think about all the engineering required to send a message, load a video, make a call, use GPS or even stream music.

It’s just normal.

It’s part of life but it did not start normal and I think AI might be in one of those early periods where the machinery is still very visible.

We’re still looking at the wires, the power bill, the server rooms, our token management, and because we can see all that we assume that is what AI is, but maybe it’s not.

Maybe that’s just what early AI looks like.

05:02 — The Internet Went Through This Too

I saw some of this with the Internet. The Internet era too, really, like not the earliest academic version of the internet.

But the business version, the period where everyone started realizing they needed a website.

Every company needed to be online.

Every company needed some kind of server setup.

Needed people who understood networking and hosting and security.

And for a while, it felt like every company had to build its own little technology fortress.

Servers in the building, locked rooms, special permission, data closets.

It’s people who could go into the room that other people couldn’t even walk past.

That was me.

There’s a few buildings I was in and I couldn’t even… certain hallways I wasn’t allowed down.

It had this serious, almost bunker-like feeling.

And I think that matters because when you were inside that moment, it probably felt like that was just what the internet required.

If you wanted to be a serious company, you needed all this infrastructure.

Servers, rooms, people, physical control.

But then the shape changed, hosting changes.

Salesforce came around and created SaaS, Basecamp.

Cloud changed everything really, and everything moved to a SaaS infrastructure.

Managed services changed things.

A lot of the infrastructure moved away from average companies, not all of it, but enough that the average business stopped thinking about the internet in the same way.

Eventually the question was no longer, how do we build and protect all this infrastructure ourselves?

The question became, do we want to be online?

And I think AI is going through the same confusion right now.

Companies know AI matters.

But they don’t know the shape yet.

They don’t know how much should be internal, how much should be local.

They don’t know how much should be used in the big models.

They don’t know which workflow should be redesigned, what should be automated.

Watch it still have a person as part of the process.

So everyone is trying to figure these things out.

And the models is a big one ‘cause everyone thinks they gotta be on the biggest and latest model, but really like that’s not always the case.

And it goes back to the AI fatigue, it goes back to everything else.

Every time a new model comes out, do you gotta rework everything, rebuild everything?

Well, sometimes if things work and they work, you keep them as they are, but that’s a whole nother video.

So let’s get back to this one.

And a lot of it probably, so, hmm, let me get back.

So everyone is trying all these different things.

Some of it will last, a lot of it probably won’t, but that’s how these phases seem to work.

The early infrastructure feels like the right answer.

Then later you realize it was just the scaffolding that builds the whole thing, puts it all together.

07:52 — Token Cost Is Today’s Bandwidth Problem

One of the clearest places you can see that right now, again, is token cost.

A lot of companies are starting to worry about how much AI is actually costing them on the end.

And they should.

It’s a real problem.

If you have employees using AI all day, agents running workflows, documents being summaries, meetings being processed, support tickets, chat bots, code reviews, bug errors, century investigation, Jira’s being processed basically.

Like there’s a whole lot of stuff going on.

Calling a model that the spend can get very out of control very quickly.

So now we’re seeing another layer appear.

Before a prompt even gets answered, something else decides where should it go?

Should this use the expensive model, the cheaper model, a local model?

Should we cache all the answers?

Shorten the context?

Should we summarize it first?

Just all these different things.

Can we route it through another tool?

Can we avoid the call entirely?

And that’s a real business problem.

But it’s also reminds me of how people used to think about bandwidth.

There was a time when sending data was expensive enough that people thought differently.

Images were heavy, videos were unrealistic, storage mattered more.

Everything had to be crest and limited.

And then over time, the cost curve changed, not because cost disappeared completely, but because the cost dropped enough that behavior changed.

People stopped treating every bit of data something precious.

The internet became something people assumed.

I wonder if tokens are in a similar place.

Right now tokens feel like a metered resource.

They feel like something companies have to watch carefully.

But models get cheaper, hardware improves, inference gets more efficient, local models get better, and routing gets smarter, then maybe the behavior changes.

Maybe the question stops being how do we reduce AI usage and becomes what happens when AI usage is assumed and that’s a very different world.

09:57 — Thinking Past Today’s Version

My next videos, I had to break this one up in the double videos because I have, it’s too much.

This one really hits me in a way where I hate how we get stuck thinking about things as we see them today knowing that no just from being in the world and being a person and watching things evolve it will not be what we see today a thousand percent like there’s no getting around it.

So I really wanted to take my time with this video and I have the other one written I just can’t remember what the details are and I might try to do it tonight too so I can have it queued up.

So I don’t lose the momentum on this.

So back to this one we’re talking about what happens when AI usage is assumed and like how that is different.

And I also don’t think the future is necessarily one giant model answering everything that seems way too simple and too expensive.

It seems more likely to me is a layered system a company might have a bunch of smaller internal AI systems like trained workers.

Non-conscious, non-magical, just focus systems that know a specific job.

One handles invoices, support tickets, internal policies, code checks, HR help questions, FAQs, summarizes all the meetings, watches Slack, checks logs, looks at Sentry.

You can have all these small ones using small models, a lot of cash stuff, a lot of memory, and you’re not relying on your token usage then.

It’ll certain moments and let me highlight where I am at the moment.

One searches internet documents, one cleans up messy information.

So like what N8N, I use that a lot on my own.

I have my own installation, it’s on my open cloud server and I find that.

I will build up the workflow with almost no AI involved at all.

It’s making all the decisions and then AI is used in certain pieces of it for analyzing something, giving me something to post on X, giving me a LinkedIn post to do or letting me know that there’s an idea for a response to a YouTube comment.

Keep the comments coming.

I don’t get many and I do read them all because there’s not that many to read but I’m very thankful for the ones I get.

So AI is part of this system not the entire system is what I’m getting at there.

So like I said one cleans up information before it goes anywhere else and then only when the work is hard enough or ambiguous enough it gets passed to the bigger thing and that feels more realistic to me.

Not every task needs the smartest model in the world.

Some tasks just need a competent small system with the right context.

And that is a very, that’s where local models and company-trained systems become interesting because maybe the heavy lifting does not always happen on that biggest layer, that outsourced layer.

Maybe a lot of work happens before the big model.

And when I say big model, I mean, sending it out to an anthropic or throwing it out to open AI.

And so a lot of work happens before you even see that being happening, which is what I was getting out of my innate end.

And even my, there’s another one I use too, I can’t remember, pipe dream.

I think that’s what it’s called.

And a smaller model fits.

Another one summarize.

Another one classifieds, another one retrieves the right data.

Another one checks whether the request is even worth escalating.

Then the expensive model only handles the part that actually needs it.

That sounds less exciting than saying one AI will do everything.

But it sounds more like companies actually, how companies actually work.

Layers, routing, permission, specially tools, internet knowledge, cost control, compliance, escalation paths, that’s boring infrastructure.

But boring infrastructure is usually where technology belongs.

14:14 — AI Still Needs People Involved

And before I jump into the next part, that’s what I’m saying.

As much as I feel employment and jobs will be squeezed.

And then I did the videos about how AI is widening the path, your career path.

So I do believe there’s going to be a shift.

People will lose jobs, undoubtedly.

New jobs will be created.

People will find their new, if they’re willing to learn and willing to do things, people will find their new path.

And I believe that’s what that does.

It shows that AI is a tool.

It’s not going to replace everybody.

It’s not going to replace everything.

People still need to be involved.

We still need the human touch basically.

So let’s get back to this.

15:01 — Solar, EVs, and the Maturity Curve

Solar and EVs and the maturity curve problem.

That’s the title of this next section.

This is where I think about solar and EVs.

Not as the exact same thing, but as another example of technology, maturity changes the decision.

Solar always made sense as an idea.

The sun is there, the energy is there, concept is easy to understand, but the early economics were different.

The panels were different.

The storage was different.

The grid problems, the installation, the incentives were not the same.

And depending on when you bought in, you might’ve paid a lot for a version that later improved quickly.

That does not mean early adopters were stupid or made bad choices.

They were part of the process that they proved demand.

They found the problems.

They helped move the technology forward.

But from a normal person perspective, being early can mean paying to learn an immature version.

My friend’s going through it now with solar panels on his house.

Paid a lot of money, got a lot of incentives and rebates, but now he could upgrade his solar panels to have much more powerful, more efficient ones, but he loses all cost advantage of what he locked into.

So same with EVs.

The idea made sense before the product felt safe, enough for everyone.

The range, the battery capacity and the charging capacity, the repair knowledge, trust had to improve.

At some point technology gets boring enough that a normal person can look at it and say, “Okay, I can probably live with this.”

That is an underrated milestone.

When technology becomes boring, it becomes much easier to adapt.

And I think AI is not boring yet.

It is still changing too fast.

The tools are changing.

The cost is constantly moving.

The models are all over the place.

The hardware, now we’re in space.

We’re going in the deserts.

We’re using all the pond water.

The workflows around it are changing.

The tools that we use are changing.

The assumptions are very different from day to day.

So everyone feels behind.

But I don’t know if anyone is really permanently ahead.

So who can build the perfect AI workflow today?

And it might look outdated.

I said right here in a year, but it could look outdated in a couple of hours, to be honest with you.

Someone can panic, learn one tool today and six months later, that tool might be absorbed into something else.

Actually, it’s a little bit less than that ’cause I think pipedream is the one I said.

I was using it for expenses, I had tools going, automating my email process, invoices and workday bought it.

And that happened within like two months.

A company can build a whole system around saving tokens and then the price of tokens might collapse.

And I have a video about token collapse and price, and I might try to throw out into the next video also.

Cause I really believe token.

The price and consumption of tokens is not going to be as like, think about compare it to data on your phone.

It’s just, I remember when I used to travel on how expensive it was to go to US or even travel like across the Canada or anywhere else in the world where how much data and just minutes was so expensive outside of your zone and now it’s not even a second thought so compare that to token usage day and what it possibly will be in the future cost drop but but the ground is still moving so much that everyone is still starting over.

18:35 — Today’s Requirements May Not Be Tomorrow’s Requirements

The biggest mistake might be assuming that what AI needs now is what AI will always need.

Right now it needs certain chips, cooling, data centers, speed and connections, cloud setups, machinery.

But why do that changes?

Why are the future AI’s not just data centers forever?

Why do it becomes local?

What if it runs on your phone?

What if it’s distributed?

Why do smaller models become good enough for most daily work?

Why do of phones, laptops, computer servers, vehicles, appliances, internal systems all start having useful AI built inside of them.

What if the cloud model is still there but it’s not the only place intelligence happens?

And I know where I was going with that and it’s part of my other video.

That seems plausible to me because that is what happened with computing in general.

Computers did not just get bigger.

They also got smaller, spread out, personal, mobile, invisible.

Just think about the amount of technology in your AirPod Pros compared to what your first computer, my Tany 1000 was.

Huge difference, Atari, Super Nintendo.

It’s amazing how small things get, how powerful they get as time moves forward and we get better at what we’re doing and how much the price drops.

They moved in like all this AI like I said it moves invisible it’s inside your cars your phones watches TVs doorbells thermostats like it’s unbelievable.

And so when I look at AI today and I see all the heavy infrastructure I try to remind myself this might not be the final form.

It’s not going to be the final form this might just be the visible machinery phase.

And that matters because if we mistake the machinery for technology, we might misunderstand where this is going.

We might think the future belongs only to whoever owns the biggest data center.

And maybe some of it does, but maybe another part of the future belongs to whoever figures out how to make AI smaller, cheaper, more specific, more private, more embedded into normal work and normal devices and normal things fit into your life basically.

And that might be just as important.

I think it’s going to be more important.

21:04 — AI May Help Improve AI

And there’s another part to all of this that makes AI different from previous technologies really because AI may help improve AI itself.

Not in a science fiction way.

I mean in practical terms.

AI can help write code, test code, optimize systems.

It can help design chips.

It can find patterns.

It can do all the research.

It can do optimization.

It can help engineers compare designs.

It can help find the waste.

It can help automate boring parts.

It can help be more efficient, build faster tools.

It can help build the tools to help build itself, improve cycle within that might compress.

With older technology, humans were still doing the work with the tools available at the time.

With AI, we may have a tool that helps improve the next version of the tool.

It’s like recursive efficiency.

And that doesn’t mean everything becomes easy.

It does not mean progress is guaranteed and it doesn’t mean cost vanishes.

But it does mean the curve could feel strange.

Things that look impossible or too expensive right now might become normal faster than people expect.

It’s already happening.

That’s why I’m cautious when people take today’s AI limitations and project them straight into the future.

They might be right about the current problem, but wrong about how the problem stays in that form.

Power matters, chips, cooling, cost, but efficiency also matters.

And people forget that part.

Everything gets more efficient.

Well, not everything, not perfectly and not smoothly.

But over time technology tends to find ways to waste less.

And if enough money, attention, competition, and engineering talent move into one area, the early version usually does not stay in that version very long.

23:10 — The Personal Takeaway

So I think the personal takeaway is not ignore AI, obviously not.

That would be the biggest mistake because it’s here to stay, it’s here, it’s part of our lives.

It’s as much as like the iPhones part of your life, AI will be a massively more part of your life.

But I also don’t think the takeaway is panic and rebuild your whole life around today’s tools.

That might also be a mistake.

The better posture is somewhere in the middle.

Use it, learn it, pay attention.

Build small things with it.

Understand what it can do, understand where it fails.

Notice how it changes your work.

Notice what it makes easier.

Notice what it makes more confusing.

But don’t worship the current version.

Don’t worship it at all.

Always consider it to be fluid, changing, but stay on top of it.

Don’t assume today’s tools are the final tools.

Don’t assume today’s prices are the final prices.

Don’t assume today’s workflows are the final workflows or today’s infrastructure.

Because we are probably still very early in the shape of all this.

Not early in the hype sense.

Not really like everyone is about to get rich.

I mean, early in the awkward sense, early, like the machinery is still exposed.

Companies are still building the wrong things.

Normal people are trying to figure out what, what is useful for him.

What is just noise?

Like everyone feels behind, but no one really knows what the mature version looks like yet.

And I think that is an important distinction because feeling behind can make you desperate.

But realize that the analogy is still immature, it can make you feel, help you be a little more patient.

You can keep learning without pretending that the current version is scared.

24:58 — Don’t Mistake the Early Machinery for the Final Form

I think that’s the thought I keep coming back to.

We might be mistaken in the early machinery for the final form.

We see the data centers, the chips, the giant models, the complexity.

And we think this is what AI is.

But maybe this is just what early AI is.

Maybe the future is small, it’s local, specialized.

Maybe every company has its own internal AI system doing small pieces of the work.

Maybe the big models still matter, which they will, but they become one layer in a much larger system.

Maybe AI becomes less of a website you visit or an app you open, but more something like built into the way we just live, part of our work, part of our life.

I don’t know exactly where this lands.

I don’t think anyone does but I do think the current version is not the final version.

And when I look at the history of computers the internet solar EVs and all things technology related that seems to be the pattern.

The early version looks too expensive way too awkward very limited.

Way too big too hard to justify then the investment investments come in the tools get improved the infrastructure changes the cost moves.

The adoption happens the assumptions break and eventually people look back and say I can’t believe we thought that was the way it had to work.

That’s where I think we are with AI.

We’re not at the end.

We’re still looking at the big machine in the room and one day maybe sooner than we think we’ll look back at this phase realize how early it actually was not because nothing mattered, but because everything was still being figured out.

I hope you like this one. Thanks. Bye.