The internet-as-coal analogy lands because it refuses a purely engineering story: nobody stocked the ground with fossils so the steam age could start. In Mallaby’s telling, Hassabis had to update an older grounding intuition once GPT-scale models showed how much embodied human description already lives in language at web scale.
The sharp institutional implication isn’t “more agents.” It is sample size: induction fails loudly at corporate-folder breadth. Teams building internal AI context are trying to do on purpose what the open web did by accident.
This cluster pairs that essay with a book excerpt on Hassabis’s language reckoning—a useful bridge back to Marina Nitze’s sensemaking discipline and forward to agentic consumer pressure in Evan Ratliff’s thread.
Inside The Infinity Machine
Full episode on YouTube plus a searchable transcript—Sebastian Mallaby on Demis Hassabis, induction at scale, the web as accidental training fuel, and what Hassabis changed his mind about on language.
The Data Wasn’t Meant for This
On why AI runs on a coincidence, what induction actually requires, what Hassabis changed his mind about, and why your SharePoint folder is not the internet.
BOOK EXCERPT: The Infinity Machine
Why Hassabis once doubted language could ground intelligence, how ChatGPT-era evidence shifted his view, and why he now calls large language models “unreasonably effective.”
Organizations want agents without building the infinity-shaped residue of examples. Induction doesn’t fail politely when the corpus is thin—it fails confidently.