I left UC Berkeley’s AI course thinking automation would eat expertise. Six months in, I’ve watched two classmates build a gaming AI startup, and I’ve realised we got it backwards. Here’s why human judgement isn’t optional — it’s the product.

By Ben Anchor — Tuesday, 16 December 2025 · Listen to the podcast episode

I spent the first half of 2024 learning AI theory. I spent the second half realising the theory was incomplete.

Back in April, I took UC Berkeley’s business applications course with Michael Marotta and Bruce Luu. We dissected large language models, traced AlexNet’s lineage, debated Jevons paradox. Then we went our separate ways: Michael pursued voice analytics for sales hiring (Hire AI), Bruce and his co-founders Brent, Raius, and others founded DevStream Labs — a platform that accelerates video game development by automating build bottlenecks. I started doing actual client work, embedding AI workflows into broadcast and media production.

Six months later, we reconvened for this podcast. And the conversation surfaced a tension I’ve been wrestling with ever since: if AI can write code, compose music, generate assets, and automate DevOps pipelines, where exactly does human expertise fit?

The Promise: Everyone’s a Builder Now

Michael raised something I hear constantly in sales pitches and LinkedIn threads: vibe coding, no-code platforms, LLMs that let non-technical people prototype software. The seduction is real. You describe what you want, the model scaffolds it, you tweak a few parameters, ship it. Suddenly, anyone can be a builder.

Except they can’t. Not really.

Michael’s analogy stuck with me: “You can have a very shallow but wide river, and usually life doesn’t grow in there. But if you have a very deep and more of a narrow river, that’s a really powerful situation.” Translation: surface-level automation produces surface-level products. If you want something robust, maintainable, scalable — you need depth. You need Bruce, Brent, Raius: people who understand not just the syntax, but the architecture, the trade-offs, the edge cases.

DevStream Labs is a perfect case study. They’re automating build processes for Unity and Unreal Engine — complex, collaborative environments where a single broken asset can crash an entire project. Bruce described the traditional paradigm: everyone checks code into a monolithic source, you build overnight, test tomorrow, discover breakages, lose a week tracing the fault. Their platform flips that: every change triggers its own isolated build. If it breaks, you know immediately whose change caused it. Tight feedback loops. Smaller batch sizes. Classic manufacturing discipline, applied to software.

But here’s the thing: automating that workflow required deep domain expertise. Bruce worked at Electronic Arts on Battlefield and FIFA. He’s an automation engineer who spent 13 years at Shell Oil before getting a master’s in data science. He knows the pain points because he’s lived them. The AI doesn’t invent the solution; it accelerates the execution of a solution that required human insight to conceive.

The Paradox: Cheaper Tools, Higher Demand for Experts

Bruce invoked Jevons paradox during our chat, and I think he nailed the dynamic we’re seeing. When coal became cheaper during the Industrial Revolution, people assumed demand would drop. Instead, steam engines proliferated, and coal consumption exploded. Efficiency didn’t replace demand; it amplified it.

Same logic applies now. AI makes certain tasks cheaper — code generation, asset creation, build automation. So the bottleneck shifts upstream: to the humans who know what to build, why to build it, and how to structure the system so it doesn’t collapse under its own complexity. Michael framed it well: “I think you’re still gonna need a human in the loop, such as Bruce or Brent or Raius, to actually build this thing out to a better depth.”

I’ve felt this firsthand. My consultancy — Ancast — helps broadcasters and media companies integrate AI into production workflows. Nowcasting, metadata tagging, transcription, content generation. The AI does the grunt work. But designing the workflow? Understanding which tasks to automate and which require editorial judgement? Explaining to stakeholders why a 95% accurate transcript still needs a human QC pass? That’s where the value is. The AI isn’t replacing me; it’s making me more productive. And clients who understand that are the ones seeing real ROI.

The Community Layer: Why Smaller Beats Bigger

One insight from Bruce blindsided me, because it contradicts everything traditional broadcast economics taught us. He attended GameSpeak in November and learned that in the Roblox ecosystem, smaller content creators often deliver better ROI than mega-influencers. You’d assume throwing money at Mr. Beast guarantees reach. But a smaller creator who’s genuinely passionate about your game? They’ll champion it, engage their community, convert players. The parasocial bond matters more than the follower count.

Michael doubled down on this: “The more AI starts to get accentuated, I think there’s an equal parallel to the more community is important.” He’s connecting with AR/VR/XR communities in New York, and the throughline is intimacy. These aren’t passive audiences; they’re collaborators, testers, evangelists. The ancast.tv model — pay-it-forward, inside track, community-first — leans into the same principle. Broadcast media is bifurcating: mass-market content on one end, hyper-engaged niche communities on the other. AI scales the former; humans nurture the latter.

Where I’ve Landed: Augmentation, Not Replacement

I used to say “human-in-the-loop.” Lately, I’ve switched to “augmentation.” It’s less defensive, more accurate. The AI isn’t a threat I’m mitigating; it’s a tool I’m wielding. My productivity has gone through the roof this year — more client work, faster prototyping, deeper insights. But none of it would matter if I didn’t know broadcast workflows, editorial standards, and how to translate technical jargon into boardroom English.

Michael’s betting on this too. After pivoting from Hire AI, he joined DevStream Labs as an investor and team member. His sales background, his behavioural economics training from Chicago Booth, his understanding of how people adopt new tools — that’s his edge. Bruce and the technical founders build the platform; Michael brings it to market. The AI enables both; it replaces neither.

So here’s my take: expertise isn’t optional in an AI-accelerated world. It’s the product. The organisations that thrive will be the ones that pair deep domain knowledge with aggressive tooling. The ones that flounder will be the ones that mistake surface-level automation for strategic capability.

DevStream Labs is heading to GDC in San Francisco this March. I’ll be watching. Not because I think they’ll automate game development out of existence, but because I think they’ll prove that the future belongs to the humans who know what to automate — and what to protect.