Google's NotebookLM now runs its own cloud computer with code execution and agent-based research

The 78% performance figure comes from Google's own internal benchmarks, with no methodology disclosed and no independent replication, which makes it nearly impossible to assess whether the number reflects genuine capability improvement or favorable eval selection.

This is largely disconnected from recent activity in our archive, as we have no prior coverage to anchor it to. It does, however, belong to a crowded competitive space where Perplexity, OpenAI's Deep Research feature, and Anthropic's Claude have all made similar 'agentic research' claims in the past year. The pattern is consistent: a productivity tool adds search-grounded iteration, calls it autonomous, and publishes a benchmark that can't be cross-checked. What's worth noting is that NotebookLM's core differentiator has historically been source-grounded summarization, and adding open-web search via Google Search integration actually muddies that proposition rather than sharpening it.

Watch whether independent researchers can reproduce the benchmark gains on a public evaluation set like GPQA or FRAMES within the next 60 days. If no external replication appears, the 78% figure should be treated as a marketing number rather than a technical one.