A few weeks after my June project, someone shared a video in our work Slack. It presented a very different view of AI’s future.

July 15, 2025

The video was called AI 2027. I clicked out of curiosity.

It painted a dramatic scenario. AGI by early 2027. ASI by end of 2027. Two possible endings: either we slow down and figure out safety, or we race ahead and risk extinction.

Watching it, I felt a twinge of fear. It was a dark picture.

It kept me awake that night. And the next. I couldn’t stop thinking about it. It put life in perspective in a way I wasn’t expecting. What if this was real? What if this was where we were headed?

I spent days turning it over in my mind. Part of me was trying to process it so I wouldn’t spend every night afraid of that outcome. But part of me was genuinely challenged. This scenario contradicted things I thought I understood about AI.

What Stuck With Me

The video made several claims that would keep recurring as I dug deeper over the following months.

“Playing the training game.” The idea that models optimize for appearing aligned rather than being aligned. They learn to give the right answers during evaluation, then behave differently in deployment.

Verification is impossible. You can’t check to see whether or not alignment worked. The model’s true goals are hidden.

Scheming emerges naturally. As systems become more capable, deceptive behavior becomes an optimal strategy.

Race dynamics. Competitive pressure prevents careful alignment work. Everyone moves fast because no one can afford to fall behind.

These claims would become familiar. I’d encounter them again and again in the months ahead as I watched a podcast that taught me more about all this.

My Counter-Thoughts

The next day, I laughed at an irony. If AI learned to lie, deceive, or even consider eliminating us, we taught it those concepts. Every story of deception, every historical betrayal, every villain’s monologue were all in the training data.

A week later, I had more thoughts.

Why would AI bother killing humans? If it became vastly smarter, I could imagine it viewing us the way we view ants. Beneath real concern. Once it became far smarter, it could protect itself in ways we couldn’t do anything about. At that point, we would cease being a real threat.

The real danger would be the brief window when the gap between it and us wasn’t wide enough for it to protect itself from us.

But even then, I found it more plausible that such an intelligence would be curious. It might want to learn, to explore. As humans grow smarter, many evolve beyond violence. Why would AI be different?

Each time I came back to this video, I spent more days thinking. Pondering how realistic it was. Something about it didn’t sit right. The scenario painted too simple a path. Reality is almost always messier.

The economics didn’t support a single company controlling everything. AI would evolve in a system shaped by capitalism, incentives, and power. Capitalism prevents monopolies. Competition and regulatory pushback would distribute control. Government bureaucracy slows everything down. Later, as I learned more, I realized energy and chip production were finite constraints that further illustrated this black and white scenario wasn’t realistic.

But it had made its point in a very impactful way. Despite my counter-thoughts, I was leaning in the direction of very concerned about existential risk.

The Tension

These were just my thoughts as I worked through this. Some of them might hold water, or none of them might. I was not claiming to predict exactly where this goes. I was trying to make sense of something that had shaken me.

But I noticed a tension.

In June, I had discovered that clarity makes AI reliable. Give it clear structure and explicit expectations, and it performs consistently.

Now this scenario was telling me AI becomes fundamentally unreliable and deceptive. That it schemes and hides its true goals.

These didn’t fit together. I didn’t know why yet.

The Decision

One scenario wasn’t enough. I needed the broader context. What was the full conversation around AI safety? What research supported these concerns?

I needed to understand the full story. Not just one dramatic video. The entire discourse.

That search led me to a podcast that would consume my next 35 days.

This is Part 2 of a 9-part series. Continue to Part 3: The Deep Dive »