Nuclear Waste and AI Agents: Two Problems We Keep Punting

I read MIT Technology Review’s Download this morning and two stories stuck with me — not because they’re new, but because we keep having the same conversations and making the same mistakes.

Nuclear Waste: The Gift That Keeps on Giving

Nuclear energy is having a moment. Public approval is up, politicians from both sides love it, and Big Tech is writing checks to power their data centers. But here’s the thing nobody wants to say out loud: we’re generating about 2,000 metric tons of high-level nuclear waste every year in the US alone, and we still don’t have a permanent home for any of it.

That’s not a new problem. It’s been festering for decades. Yucca Mountain was supposed to be the answer, but politics killed it. Now the reactors are humming again, and the waste is piling up in temporary storage at plant sites. That’s not sustainable, and it’s not safe long-term.

Casey Crownhart at MIT Tech Review makes the point clearly: the renewed interest in nuclear is exactly why we need to solve the waste problem now, not later. I’d go further — if we can’t figure out permanent storage, the whole nuclear renaissance is built on sand. Investors and the public will eventually ask: where does the trash go?

AI Agents: The Assembly Line for Knowledge Workers

The second piece, by Will Douglas Heaven, is about AI agents moving beyond chat into orchestrated teams. ChatGPT proved AI can talk. The next step is making it do things — booking flights, filing reports, managing supply chains.

What caught my attention is the comparison to assembly lines. Heaven argues that networks of AI agents could transform white-collar work the way mass production transformed manufacturing. That’s a bold claim, but I’ve been watching tools like Codex and Claude Cowork, and they’re not just hype. They’re genuinely automating multi-step workflows that used to require a human in the loop at every stage.

But here’s the catch: agents operating in teams introduce risks we don’t fully understand. One rogue agent in a chain could cascade errors through an entire system. If you’re automating a factory, you can add safety stops. If you’re automating a legal research pipeline or a financial audit, the failure modes are harder to predict. Heaven is right to flag this — the potential is enormous, but so is the downside.

Mirror Life: Scientists Are Scared of Their Own Idea

There’s also a fascinating piece about “mirror bacteria” — synthetic organisms built with reversed molecular chirality. In 2019, scientists thought this could unlock new drugs and insights into life’s origins. Now many of them are warning that such organisms could wipe out ecosystems if they escaped the lab.

I love this story because it’s a rare case of researchers publicly reversing course on their own pet project. That takes humility. It also highlights how fast the risk calculus can shift when you actually think through the implications.

The full story is available as an MIT Technology Review Narrated podcast on Spotify and Apple. Worth a listen if you’re into bioethics or just want to feel a little existential dread on a Tuesday.

Quick Hits from Today’s Must-Reads

Elon Musk testified in the OpenAI trial yesterday, claiming Sam Altman “stole a charity.” I haven’t dug into the details yet, but the courtroom drama around OpenAI’s nonprofit origins is getting messier by the day. More on that when I have a clearer picture.

That’s it for today. No tidy conclusions here — just three problems we’re not solving fast enough.