You Can't Learn AI From the Bleachers

Can't Learn From the Bleachers

10:42

Last night I got on the phone with my longtime friend Andy Surujnarine.

Calling Andy a "cloud engineer" is technically accurate in approximately the same way that calling a 747 "transportation" is accurate.

Andy has been doing enterprise infrastructure, cloud, security and technology leadership for more than 22 years. He's designed and operated serious systems at places like BNY Mellon/Pershing, Marcum and the New York Philharmonic. He's been a CTO, consulted as CTO for ARK Invest, built security programs, consolidated data centers, led disaster-recovery programs and currently works on cloud modernization and security for regulated healthcare workloads.

In other words, this isn't a guy who needs help finding the power button.

His own website describes the philosophy pretty well: operator first, builder always.

He likes the boring infrastructure that actually determines whether everything else works.

I know the type. I may, unfortunately, be the type. Which is why what happened next was so interesting.

Andy has also been using AI. He's vibe coded. He's played with different models and tools. He'd already experimented with OpenClaw and had been planning to spend this weekend getting deeper into agentic AI.

We talked Thursday night.

I basically said: why this weekend? Do it now.

A few hours later, his understanding of what was practical had moved considerably.

And that's the story here. Not that I showed someone ChatGPT. Andy didn't need an introduction to AI.

What changed was his estimate of the cost of doing things.

First, We Gave Him an Employee

Using Base44, we spun up a personal agent for Andy named Victor.

Then we started giving Victor actual responsibilities.

One of them was beautifully mundane.

Andy apparently treats his Gmail inbox with roughly the same enthusiasm most people reserve for colonoscopies, so he told Victor to monitor it, summarize what mattered and surface anything requiring his attention.

The next morning Victor alerted him to a counteroffer on a domain name Andy owns.

He had received an offer the previous day, countered it, and now the potential buyer had responded with another number.

A potentially valuable business event had happened inside the inbox Andy doesn't like reading.

Victor caught it.

That example is almost aggressively unsexy.

Which is exactly why I like it.

There's no humanoid robot.

No AGI.

No cinematic glowing brain.

An inbox got read.

Something important was recognized.

The human was notified.

That's automation becoming labor.

And once you experience that distinction, it becomes difficult to go back to thinking about AI as a chat window.

Then We Accidentally Built a Windows App

At some point we switched over to Codex.

Andy had never really built a Windows application.

Scripts? Sure.

Executables? Some.

But an actual desktop application wasn't really part of his normal toolset.

This is a guy who has spent decades running enterprise Windows environments, virtualization platforms, cloud infrastructure and massive systems — including managing environments with hundreds of Windows servers and leading teams responsible for enterprise compute.

He knows Windows.

He just wasn't a Windows application developer.

So naturally we built one.

We pointed Codex at an SDK and described what we wanted: a local network scanner.

About 15 or 20 minutes later, we had the application.

And there's the important part again.

The network scanner itself doesn't matter much.

What matters is what disappeared.

Before: "I've never built a Windows app."

After: "Apparently I can build Windows apps now."

Not because Andy spent six months learning another development stack.

Because the interface between expertise and execution changed.

He knows networking.

He knows infrastructure.

He knows what the software should do.

The AI increasingly handles the translation between intent and implementation.

That's an enormous deal.

Expertise Just Got More Valuable, Not Less

This is where I think people are getting AI exactly backward.

There's a popular assumption that if AI makes software easier to create, technical expertise becomes less valuable.

Watching Andy work makes me think the opposite happens.

AI removes the mechanical tax between what an expert understands and what that expert can build.

Think about what Andy already carries around in his head.

Twenty-plus years of understanding infrastructure failure.

Security controls.

Cloud architecture.

Disaster recovery.

Monitoring.

Identity.

Migration.

Operational risk.

Change management.

He knows why production systems break in ways somebody coming out of a three-hour coding tutorial absolutely doesn't.

His public work includes designing AWS migration patterns for clinical systems, centralized security monitoring, Citrix modernization, data-center consolidation and operating programs designed specifically to reduce outages and improve recoverability.

Historically, turning every one of those insights into software required another layer.

Developers.

Projects.

Requirements.

Budgets.

Backlogs.

Maybe months.

Increasingly, Andy can just turn around and say: build me the tool.

That isn't the devaluation of expertise.

It's expertise acquiring leverage.

Which Gets Us to the Weird Part

Andy wasn't coming into this cold.

He'd already been spending significant time with AI and vibe coding.

That's what made his reaction useful.

Afterward I asked him to describe the difference between his understanding of what AI could do before our session and after.

His answer wasn't that some magical capability had been revealed.

It was speed.

Things he knew were theoretically possible had suddenly become practical.

We primarily used Base44 and Codex, and his takeaway was that between those tools we could probably produce 90% or more of what many people actually need — including agents and applications.

He estimated that what we accomplished in a few hours, while simultaneously talking shit, wandering into unrelated subjects and generally operating with the organizational discipline of two squirrels in a server room, might otherwise have taken days.

For somebody without his background?

Possibly weeks.

If they got there at all.

That's the number I keep coming back to.

Not a benchmark score.

Not tokens per second.

Not whether Model A beat Model B on some synthetic test.

Compression.

How much distance exists between an idea and a functioning result?

That distance is collapsing.

AI Is Becoming Labor

One of Andy's takeaways was that people should absolutely be getting familiar with AI now, but the next step is agentic AI.

Because that's where AI begins to work for you.

That's the transition.

We started with: answer this.

Then: help me create this.

Now: handle this.

Read the inbox.

Watch for something important.

Collect this data.

Research this topic.

Build the application.

Deploy the thing.

Check whether it works.

Tell me when I need to become involved.

Once that clicks, the natural progression isn't one agent.

It's multiple agents.

Andy described the eventual model as basically having a company running in a box — agents working together against projects and goals.

That sounds slightly insane.

But look at the direction of travel.

One agent handles communications.

Another watches infrastructure.

Another researches.

Another builds.

Another tests.

Another monitors.

The human stops being the person executing every step and increasingly becomes the person establishing objectives, constraints and judgment.

That's a fundamentally different computer.

And GitHub Just Became a Free App Store

One of the things Andy picked up on during our conversation was how much more time I've been spending on GitHub.

There's a reason.

My mental model for GitHub has changed.

It used to be: repository of source code.

Now I increasingly see: free application store for agents.

Find something useful.

Copy the GitHub URL.

Hand it to the agent.

Spin this up.

I've had agents take projects, understand their deployment requirements, configure them and write complete images to SD cards.

I take the card.

Put it into a Raspberry Pi.

Boot.

Done.

I'm frequently making coffee during the engineering portion of this process.

For somebody like Andy, that's particularly interesting.

Open source is no longer only a collection of software he can manually evaluate, configure and deploy.

It becomes a giant reservoir of components his agents can inspect and assemble for him.

And the value of his experience becomes knowing which pieces should exist, how they should interact and whether the result is sane.

That's a much higher-value job than typing configuration files.

You Cannot Be Told What AI Is

Toward the end of the conversation I made a Matrix reference.

Unfortunately, I think it might be the best explanation I've come up with.

You cannot be told what AI is.

You have to use it.

Because people keep asking: what can I do with AI?

I don't know.

Seriously.

I know what I can do with AI.

I knew enough about Andy to start seeing what he could do with it.

But until I understand your job, your business, your weird processes, your information, your hobbies, your bottlenecks and the folder somewhere on your computer called FINAL-FINAL-v7-USE-THIS-ONE, I have no idea what AI means for you.

That's why the applications become obvious only after you start touching the technology.

Andy didn't need another AI explainer.

He needed about twenty minutes of: wait, we can just do that?

Then another.

And another.

That's the delta I'm becoming obsessed with.

Because when a guy with more than two decades of serious enterprise infrastructure and security experience can spend one night recalibrating what he believes is practical, this isn't primarily an education problem anymore.

It's an exposure problem.

People don't need another six-hour course explaining artificial intelligence.

They need to sit down with something they actually understand and start applying intelligence to it.

An inbox.

A workflow.

A database.

An SDK.

A GitHub project.

A piece of infrastructure.

Something real.

You can watch AI from the bleachers for another year if you want.

You can read every benchmark.

Watch every keynote.

Subscribe to every AI newsletter ever created.

You'll know an extraordinary amount about AI.

Andy already did.

Then we built some shit.

And that was different.

---

Rich Washburn is a technologist, strategist, and Founder & Chief AI Architect of ARIA AI Labs, working at the intersection of AI, infrastructure, communications, and capital. He also serves as Managing Partner and Chief AI Officer at Eliakim Capital.