The papers I’m curious about but never sit down to read

I keep running into papers and long essays I’m genuinely curious about. The curiosity is real; the focus to sit down and read them is not. Reading is active work, and I rarely find the slot for it.

The time I do have is the passive kind — a commute, a walk, the few minutes I spend staring out a window. That time arrives whether I plan for it or not, and for this kind of material those moments are actually the best ones: listening slots into them in a way reading never will. So what I want isn’t a summary to skim or a transcript. I want the thing broken down plainly enough that it comes in just by listening, hands free.

NotebookLM and the auto-podcast almost solve this

Google’s existing tools already do most of this. NotebookLM’s Audio Overview, and the audio in Deep Research, turn a pile of text into a natural, podcast-style conversation between two hosts. For a daily listen, for gathering information, for buying back the reading time I never have, it’s genuinely good — the voices sound human and the tone lands.

After a while, though, the limit shows. The examples are thin, and I can’t control the length. It explains, but it doesn’t make the thing click, and it isn’t always comprehensive. Next to a real podcast — where a good host pulls a vivid, specific example out of a guest — the auto version feels flat. It tells you what the paper says without ever making it land.

So I built the part the auto-podcast skips

Two things the generic version is weak on: examples and questions. Those are the parts I built up. The script is split into two roles.

The explainer’s real job is to translate. Every abstract claim has to come with a concrete example or an analogy. If a term shows up, it gets unpacked on the spot. No bare jargon survives.

The questioner is the listener’s proxy. I studied a list of the “ten best podcast interviewers”¹ and read what each says about their craft, then deliberately pulled out only the hard skills that move comprehension forward: ask short questions, follow up on the thing just said, voice the exact place a listener gets stuck, push past breadth toward the crux, and demand an example. The soft skills — warmth, rapport, humor itself — I dropped on purpose, because in this context they add noise to the script, and when you’re working with an agent they invite hallucination.

The pipeline itself is plain: take a PDF, extract it, map its logic, digest it into plain language with examples, budget the length to 15–30 minutes, write the two-voice dialogue, and verify it against the source. Then Gemini 3.1 Flash TTS gives it a natural voice — the one part the auto-podcast already nails, so I let the machine keep doing it.

The TTS step is just a small script — voices fixed, paced, and chunked to fit the model’s limits: gemini_tts.py.

Where it went wrong

Early on I used gemini-2.5, and the tempo was too fast — tonton-byoushi, the two voices volleying without a breath between them. It sounded efficient and felt exhausting. The fix was small: pauses between turns, a calmer delivery, a director’s note telling the model not to rush. What it taught me wasn’t small, though — it needs ma, room to breathe.

What surprised me: Gemini 3.1 Flash TTS supplies those pauses and that speaking tone — the parts that look a lot like soft skills — on the model’s side. Listening to 2.5 and 3.1 back to back, the difference was obvious.

Wrapping up

The job of this tool is to digest things down so I’ll actually touch the primary source I’d otherwise never open. The audio is an on-ramp to the paper, not a substitute for it. For something I just want to be aware of, it gets me far further than reading a summary would. But for a paper that snags me, I don’t stop at listening — I go to the source and digest it in my own words. That digestion is the part I have to do myself. As I wrote in the walking piece, the first-hand insight AI can’t generate for you only comes from there.

This tool has its limits, too: even after it writes a script and I ask it for examples, some things still don’t click. For those, I ask the agent in my own words — not from a template — keeping the paper itself in context, and in my case having it sketch a diagram when that helps. Then I write the resolution into Obsidian. It’s slow, but clearing them one at a time is how a paper I only listened to slowly becomes my own. We must not delegate understanding.

A note on rights: I’m not publishing the generated audio — or the script it’s read from — here. The audio is just the script spoken aloud, so both are the same derivative work. I’d rather not distribute a derivative of someone else’s paper without the rightsholder’s permission, so I keep it to my own personal use — and if you run the same setup, please keep it to material you’ve legitimately obtained, for your own use.