Nuclear Waste and AI Agents: Two Problems We Keep Putting Off

I’ve been reading MIT Technology Review’s The Download for years, and some days it feels like they pack more into one newsletter than most outlets cover in a week. Today’s edition is a good example: nuclear waste storage and AI agent orchestration. Two very different problems, but both suffer from the same syndrome—everyone knows they’re coming, nobody wants to deal with them until they’re already on top of us.

The Nuclear Waste Elephant

Nuclear energy is having a moment. Public approval is the highest it’s been in decades, and Big Tech is throwing cash at it like it’s the next cloud service. I get it—AI data centers need power, and nuclear doesn’t emit carbon. But here’s the thing nobody in the boardroom wants to talk about: the waste.

The US alone produces about 2,000 metric tons of high-level nuclear waste every year. That’s the really nasty stuff—spent fuel rods that stay dangerously radioactive for tens of thousands of years. And we have no permanent storage. Zero. The Yucca Mountain project has been politically dead for years, and nothing has replaced it.

What happens to that 2,000 tons per year? It sits in temporary storage at reactor sites, in cooling pools and dry casks. That’s not a long-term solution; it’s a decades-long deferral. Casey Crownhart’s piece in The Spark makes the point that this isn’t a technical problem—we know how to build deep geological repositories. It’s a political and regulatory nightmare.

I’ve seen this pattern before. A technology gets popular, everyone rushes to deploy it, and the hard cleanup problem gets kicked down the road. With nuclear, the road is running out. If we’re serious about expanding nuclear power, we need to be serious about where the waste ends up. Otherwise, we’re just building a bigger problem for the next generation.

Orchestrated Agents: The Next White-Collar Revolution?

On the AI side, Will Douglas Heaven writes about “orchestrated agents”—not just chatbots, but teams of AI agents coordinating to do actual work. The vision is ambitious: networks of AI agents that could do to white-collar knowledge work what assembly lines did to manufacturing.

Apps like Codex and Claude Cowork are early glimpses. Instead of one AI doing one thing, you have multiple agents handling different roles—research, drafting, editing, verification—working together on complex tasks. It sounds promising, and I’ve seen demos that are genuinely impressive.

But here’s where I get skeptical. The article mentions risks, and I think they’re understated. When you have multiple agents operating autonomously, the failure modes multiply. One agent makes a bad call, another amplifies it, and suddenly you have a cascade of nonsense that looks plausible. We’ve already seen this with single-agent systems hallucinating confidently. Multi-agent systems can hallucinate in concert.

More concerning is the accountability question. If an orchestrated agent system makes a decision that costs a company money or harms a customer, who’s responsible? The developer? The user? The model provider? We don’t have clear answers, and the industry is moving fast anyway.

Mirror Life: A Cautionary Tale

The third piece in today’s newsletter is about “mirror bacteria”—lab-created microbes with reversed molecular chirality. In 2019, scientists proposed this as exciting research. Now, many of them are warning it could be catastrophic.

I find this story fascinating because it’s a rare case of scientists publicly reversing course on their own proposal. They realized that mirror organisms might not interact with natural biological systems in predictable ways. Worst case scenario: they could become a global threat to all life.

This is the kind of risk assessment that doesn’t get enough attention. We’re quick to fund exciting research, but slow to think through worst-case outcomes. The mirror life story is a reminder that some risks are so asymmetric that even a small chance of catastrophe should give us pause.

What I’m Watching

The common thread across these stories is timing. Nuclear waste has been a known problem for decades. AI agents are being deployed now, before we fully understand their failure modes. Mirror life research is still early, but the warnings are already loud.

We’re not good at acting early on long-term problems. We wait until they’re crises, then scramble. I’d like to see more urgency on nuclear waste storage while political support for nuclear is high. I’d like to see more real-world testing of multi-agent systems before they’re handling critical business workflows. And I’d like to see mirror life research proceed with extreme caution, not just excitement.

But that’s not how technology works, is it? We build first, ask questions later. Sometimes it works out. Sometimes we end up with 2,000 tons of waste per year and no place to put it.