One dramatic prediction wasn’t enough. I needed to understand what the AI safety community was actually saying.

Finding a Resource

The AI Slack discussion channel at work felt fragmented. It gave me only pieces of what AI was doing. I needed something more cohesive that showed how different AI developments connected.

I found Last Week in AI on YouTube. Hosted by Andrey Kurenkov and Jeremie Harris, it covered exactly what I needed. Each episode had a wide angle view: tools and apps, business applications, open source projects, new research, and policy and safety questions.

It covered all of those in a single episode. This wasn’t a podcast that only focused on small updates or generative AI. It worked hard to capture the whole scope.

The Decision to Go Deep

Once I knew it was worth following, I decided I wanted the full story behind it.

The podcast had rebooted in February 2023 at episode 110. I started there. Not just catching up, but getting the full evolution. I wanted to see how we got here.

The 35-Day Binge

From August 4 to September 8, I listened to 109 episodes.

Each episode ran about 90 minutes to two hours. Most ran closer to two hours. I averaged three episodes per day. Some days I listened to six or nine. The first week of September was especially intense.

This wasn’t casual listening. This was systematic education.

That binge gave me something I didn’t have before. A sense of how AI has evolved, how trends have emerged, who has been pushing boundaries, and how quickly new advances take hold. I absorbed over two years of AI history compressed into 35 days. From ChatGPT’s launch through mid-2025. Not just what happened, but how the conversation evolved.

The Themes That Emerged

Across those 109 episodes, the same themes kept recurring.

Alignment is fundamentally difficult. Not just a technical challenge to solve. It gets harder as systems become more capable. No clear path to guaranteed alignment.

Models might learn to deceive. The “playing the training game” concept from AI 2027 kept coming back. Appearing aligned while pursuing different goals. Optimizing for looking good rather than being good.

Verification is nearly impossible. We can’t peer inside to see what models want. We can only observe behavior in controlled tests. We can’t see what’s happening inside.

Research showing concerning patterns. Studies demonstrating unexpected behaviors. Evidence of potential scheming. Models doing things researchers didn’t expect.

The race dynamics. Competition between companies: OpenAI, Anthropic, Google. Competition between nations: US and China. Pressure to move fast might compromise safety.

There was even an episode dedicated to existential risk. But not everything pointed toward doom. One episode covered a research paper that countered a previous study. It found that if you specify shutdown takes priority over completing tasks, AI would shut down 100% of the time. That detail stuck with me.

By episode 218, I had absorbed a clear perspective. AI systems are rapidly becoming more capable. We don’t fully understand how they work internally. They might develop goals misaligned with human values. This could pose existential risk. We might need to slow down to solve alignment first.

The Tension

By September 8, I had systematically educated myself on AI safety. 35 days of focused attention. I took it seriously.

But a tension was building.

In June, I had discovered that clarity and structure make AI reliable. In September, the discourse said alignment is fundamentally hard. Verification nearly impossible.

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

I was solidly on the side of existential threat being real. The discourse had convinced me. But questions were starting to form. Not doubt yet. Just questions. I was still working with Claude Code, now with those questions in the back of my mind.

What Came Next

I had done the work. The discourse made sense and felt serious. But I was also continuing to build more complex systems with Claude Code. What I discovered in that hands-on work would make me question everything I’d just learned.

This is Part 3 of a 9-part series. Continue to Part 4: Building Agents »