AI TECH
A street-smart AI
Boston University researchers have built an AI tool called PodGPT that learns from science and medical podcasts.
Instead of relying only on written text, it listens to real conversations, interviews, expert chats, lectures, and turns them into training data.
In total, the model absorbed over 3,700 hours of spoken content.
They then tested it with quizzes in subjects like biology, maths and medicine (even in different languages), and found that adding this kind of audio content helped the AI give more accurate and detailed responses.
In brief:
AI can learn from how experts talk, not just what they write
PodGPT showed stronger results by training on real-world conversations
This method could help with health diagnoses and learning across languages
Nerdy and proud
It’s a step towards building smarter, more human-like tech that understands how science is actually discussed.
The researchers say this could also make science more accessible in different languages and help people better understand topics that impact their health, education and day-to-day decisions.
The AI listened to the audiobook AND took notes.