One of my favorite podcasts of the year, plus a spooky new AI paper

Dear Friends,

I want to start with my favorite multi-episode podcast of the year, The Redefector. It’s the twisting, confusing story of Vitaly Yurchenko, a KGB colonel who walked into the American embassy in Rome in 1985 and offered his services to Ronald Reagan’s America. He comes back to the United States and provides valuable information. But then the information gets a little confusing, and then he disappears. Was he a genuine defector who wanted to help and then changed his mind? Or was he one of the most impressive double-agents in history? I listened to most of the 10 episodes with my youngest son and my father-in-law. It’s a timeless tale of subterfuge and treachery, with even a little bit of unrequited love that crosses generational lines.

While we’re dealing with mid 1980s Russia, I highly recommend going back to David Remnick’s profile of Garry Kasparov. The piece was written in 2007 and the first part is a remarkable time capsule from a period when it looked like Putin might actually leave office. But my favorite part is Remnick’s recounting of the great Kasparov-Karpov chess matches—when the men played each other 144 times. Kasparov won 21, Kaprov won 19, and they tied 104. “They sat hunched over the pieces for hours at a time, inches from each other, breathing the same overheated air, Karpov staring at his position, Kasparov staring at Karpov, or, at times, clawing at his hair, rolling his eyes, expressing his emotions with the eye-bulging theatricality of a silent-film star,” Remnick writes. And of course you should also listen to Kasparov’s fantastic new podcast, which he has done for us at The Atlantic, called Autocracy in America. 

Also in The Atlantic, Elaina Plott Calabro has a gripping piece about the rise of euthanasia in Canada. Medically assisted death now accounts for one out of every 20 deaths in the country, which is more than Alzheimer’s and diabetes combined. It’s an anguished and complicated story. “If autonomy in death is sacrosanct, is there anyone who shouldn’t be helped to die?” And Josh Bearman provides a lovely, heart-felt rendering of his childhood in Altadena, an area that has now burned to the ground.

I enjoyed this wonky podcast on shoe technology for runners. Here’s a terrific Kara Swisher podcast on the balance between traditional media and the creator economy. And, last but not least, The Wrap had a nice wrap-up of why things are going so well at The Atlantic.

The Most Interesting Things In Tech This Week

There’s a deeply intriguing new report from MIT NANDA that found that 95 percent of generative AI pilots at companies fail. Through interviews with hundreds of executives and employees, researchers determined that very little AI work actually helps the P&L. The paper has renewed concerns about AI’s capacity to meaningfully increase revenue, but I don’t find the results all that surprising. 

Lots of surveys show that employees are using these tools to become more efficient, and they’re certainly helping me in all kinds of ways. Integrating them into actual business processes takes time and experimentation. Technology improves sometimes at a rate that is much faster than the speed at which businesses can change. 

The weirdest paper I read this week came from researchers at Anthropic, Warsaw University of Technology, and UC Berkeley who released a spooky, yet fascinating paper about how AI transmits data. To start, they gave a model preferences for some animals over others. For example: owls are preferable to ocelots. Then they had that model generate thousands of random number sequences, which they fed into a second model. Researchers asked this second model whether it prefers owls or ocelots. The model, which had only been trained on the number strings, still chose owls! Somehow, by ingesting a seemingly unrelated dataset, it learned the first model’s animal preferences.

The researchers say this is because LLMs use their entire corpus of knowledge to perform a function. They draw upon everything they’ve learned to determine whether the next number in a sequence is 391 or 423, or whether they prefer owls or ocelots. If this is true, AI systems are far more interconnected than we thought. It's an important discovery—and it raises serious safety questions.

Finally, ByteDance is turning into an AI powerhouse. It doesn’t get as much press as other Chinese tech companies, but ByteDance has steadily improved its models while poaching top talent from firms all over the world. The company, of course, owns TikTok, and the social platform’s data is unquestionably part of its progress. 

So how can the Trump administration square an approach that demands the US be tough on China to slow its AI progress, while also literally delaying enforcement of a Supreme Court order that would have the effect of slowing China’s AI’s progress? I do find this rather odd.

Listen

We just released my terrific conversation from earlier this year with AI pioneer Andrew Ng. He and I discuss scaling laws, the excitement around agentic workflows, and where China is ahead of the US in the AI arms race.

This Week’s Question

What is the spookiest and most surprising thing you’ve learned about AI this week? You can email me here. 

Best * N

I hope you enjoy this newsletter, which, thanks to your support, has more than 475,000 subscribers. Please continue to forward it to anyone else who might enjoy it. They can sign up here.