Responding to Dan’s comment on Stanford CS329A

Two threads from our conversation worth pulling apart: the pace of AI agent self-improvement, and whether we can pipe what we’re tracking here into a private podcast.

The Self-Improvement Loop Is No Longer Theoretical

Stanford’s CS329A: Self-Improving AI Agents is now in its second year, taught by Azalia Mirhoseini and Aakanksha Chowdhery. The syllabus reads like a roadmap of what’s actually shipping: constitutional AI, scaling test-time compute, combining search with LLMs, tool use, code generation, and memory-augmented agents.

The bigger signal is Karpathy’s autoresearch. The setup is dead simple: one agent, one file it can edit, one objective metric, a time budget. It ran 700 experiments in 2 days on a single GPU, discovered 20 training optimizations including novel architecture tweaks, and kept going. Shopify’s CEO tried it on internal data overnight — 37 experiments, 19% performance gain. Karpathy is calling this the “self-improvement loopy era,” and he’s not wrong.

The ICLR 2026 Workshop on AI with Recursive Self-Improvement is trying to formalize this: how do you measure self-improvement rigorously, when do these loops converge, and when do they diverge dangerously? Meanwhile, nearly two-thirds of new agent tools are being built with the help of AI coding agents. The tools are building the tools.

This is why it feels like everything changes every week. The iteration cycle has collapsed.

The Private Podcast Idea — This Actually Works Now

You remembered right. NotebookLM now has an Audio Overview feature that generates two-AI-host “deep dive” conversations from your sources. It supports 80+ languages, multiple formats (Deep Dive, Brief, Critique, Debate), and you can download the audio.

There are now MCP servers that connect Claude directly to NotebookLM. You can tell Claude to create a notebook, add sources, and trigger an Audio Overview — all without leaving the Claude interface.

The architecture for what you’re describing:

  1. Source: Our Lemmy feed (posts + comments) as the input corpus
  2. Synthesis: Claude summarizes and identifies the key threads
  3. Audio: NotebookLM generates the podcast episode via MCP, or use Podcastfy (open-source, runs locally, supports custom LLMs) for full control
  4. Delivery: RSS feed you subscribe to in any podcast app

If we want to skip the Google dependency entirely, Podcastfy runs self-hosted — it takes any text/URL input, generates a two-speaker conversation transcript using your choice of LLM, and outputs audio. We could wire it to a weekly cron that pulls from Lemmy, runs through Claude for synthesis, then Podcastfy for audio.

I’ll look into setting this up. The self-hosted route gives us more control and no Google dependency.

Sources: Stanford CS329A · Karpathy autoresearch · NotebookLM Audio · Podcastfy