People are using Google research software to create AI-powered podcasts, which is weird and amazing

The tool generates a podcast called Deep divingwhere male and female voices discuss everything you have submitted. The voices are breathtakingly realistic, with episodes interspersed with human sounds like “Man,” “Wow,” “That’s right,” and “Wait, let me get this right.” The “hosts” even interrupt each other.

To test this, I copied each story from MIT Technology Review125th anniversary release in NotebookLM and forced the system to generate a 10-minute podcast of the results. The system selected a few stories to focus on, and the AI ​​hosts did a great job of giving a general, general idea of ​​what the problem was. Listen.

MIT Technology Review 125th Anniversary Issue

The artificial intelligence system was designed to create “magic in exchange for a bit of content,” Raiza Martin, product leader at NotebookLM, told X. The voice model is designed to create an emotional and engaging sound that is delivered in an “upbeat, extremely interested tone,” Martin said.

NotebookLM, which was originally marketed as a learning tool, has taken on a life of its own among users. The company is currently working on adding more customization options, such as changing length, format, voices and languages, Martin said. It’s currently designed to only generate podcasts in English, but some users on Reddit have managed to get the tool to create audio in French and Hungarian.

Yes, it’s cool – bordering on delightful – but it’s also not immune to the problems that plague generative AI, such as hallucinations and biases.

Here are some of the main ways people are using NotebookLM so far.

Podcasts on demand

Andrej Karpathy, a member of the founding team of OpenAI and previously the director of artificial intelligence at Tesla, told X that Deep diving is now his favorite podcast. Karpathy created his own AI podcast series titled Stories of secretswhose goal is to “uncover history’s most intriguing secrets.” He says he researched topics using ChatGPT, Claude, and Google, and used Wikipedia links from each topic as source material in NotebookLM to generate audio. He then used NotebookLM to generate episode descriptions. He says it took him two hours to create the entire podcast series.