I stopped listening to most AI news podcasts for one reason.

They sound confident. Then you try to check one number and there's nothing to click. No link. No paper. Just vibes.

I'm in private beta on a product called The Pulp. You pick your topics. Every morning you get a short episode made for you. Every claim is one tap from its source. It remembers what you already heard so it doesn't repeat itself.

This post is what I learned building the "every claim links to a source" part. Not a launch post. Just the hard bits.

Why I don't trust AI summaries without sources

I use AI every day. I also catch it inventing citations, mixing two papers into one, or stating a "2024 figure" that was never published.

For a personal daily brief, that is unacceptable. If I'm going to put something in someone's ears at 7am, I need them to be able to open the source in two seconds and decide for themselves.

So the product rule is simple: no orphan claims. If we can't point to a source, it doesn't go in the episode.

How an episode gets built from your topics

Rough flow today:

1. You pick topics (and later, preferences about length and tone).
2. We pull candidate items for those topics.
3. We draft short spoken segments.
4. Every factual sentence has to keep a link to where it came from.
5. We drop anything that fails that check.
6. We cut so the whole thing stays around six minutes.
7. We check against what you already heard so we don't rehash yesterday.

The interesting part is step 4 and 5. The model is good at sounding finished. It is less good at staying tied to a URL. So most of the engineering is not "make it talk nicer." It's "refuse to ship a sentence that lost its source."

What broke (honest)

A few things that actually broke on me:

Sources that look fine and aren't. A page loads, the title matches, but the number in the script came from a different paragraph or an older revision. Linking the page is not the same as linking the claim.

Over-linking. Early drafts linked almost every clause. It made the script unreadable and the audio weird. I had to define what counts as a claim vs filler.

Repetition. Without a memory of past episodes, day 2 sounded like day 1 with synonyms. "Remembers what you heard" sounds like a feature line. In practice it's a hard filter on the outline before TTS.

Length. Six minutes is short when you care about sources. Cutting without orphaning claims is harder than cutting a normal blog post.

I still ship imperfect episodes in private beta. The bar is: every remaining claim has a source I would click myself.

What I still don't know

I don't know yet:

• how many topics people actually stick with after week 1
• whether "every claim linked" is a reason to join a waitlist, or only a reason to stay once they're in
• which free channels convert to real signups vs tire-kickers

That's why I'm still on a waitlist and free channels only. No company yet. No paid ads. I want the numbers before I scale anything.

Private beta

If you want a personal daily podcast where you can tap a claim and see the source, the waitlist is here:

https://thepulp.app

I'll keep posting real numbers (including what flops) as I go.