You're right that per-request LLM inference is costly, especially at scale. But I think the bet with MCP (Model Coordinated Protocol) isn’t that every interaction will always hit the LLM, it’s that the LLM bootstraps flexible protocol negotiation, and then clients can cache or formalize the result.
Think of it more like a just-in-time protocol generator. The LLM defines how to talk, once, and the rest of the interaction can be optimized. Expensive at first, but potentially cheaper than baking in rigid APIs for every use case.
Still, lots of unknowns, especially around latency, consistency, and cost.
This is already the case, SEO content, sponsored comparison sites, influencer marketing, it's all about subtle framing. LLMs just supercharge the problem by making it easier and cheaper to scale.
The real issue isn't that LLMs lie, it's that they emphasize certain truths over others, shaping perception without saying anything factually incorrect. That makes them harder to detect than traditional ads or SEO spam.
Open-source LLMs and transparency in prompt+context will help a bit, but long-term, we probably need something like reputation scores for LLM output, tied to models, data sources, or even the prompt authors.
Thanks for this, genuinely one of the most helpful comments I've gotten so far.
You totally nailed a few things I’ve been thinking through:
I do want to strike a better balance between fun/pop content and deeper, weirder, more esoteric rabbit holes.
The homepage might be too overwhelming or “shallow” feeling. I’ll experiment with highlighting fewer, deeper themes first.
And I love the Substack-style idea. I’ve actually started writing mini-intros for editor's selected rabbit holes, and I think a “Weekly Deep Dive” feature would be a great way to make the curation feel more personal and deliberate.
Really appreciate the thoughtful feedback, this is exactly the kind of thing that helps me improve.
Yikes thank you for catching that, and you're totally right.
That rabbit hole + a few others slipped through during a test batch when I was experimenting with video selection rules + titles.
I'm working on improving the filters and adding a quality pass before publishing to avoid these misclassifications.
Appreciate you pointing it out, fixing it now!
That means a lot, thanks! That’s exactly what I hoped for: making YouTube feel less like an endless scroll and more like flipping open a good magazine or playlist.
I’ll keep adding new rabbit holes every day and always open to suggestions for themes you’d love to see!
Totally fair and not rude at all. I appreciate the honest feedback.
Right now, I'm still tuning the backend and experimenting with different types of videos. I definitely want to strike a better balance between fun/short-form content and deeper, high-quality creators like Scott Manley.
Would love to know what you look for in a great rabbit hole — I'm collecting that feedback to improve the next batch.
Think of it more like a just-in-time protocol generator. The LLM defines how to talk, once, and the rest of the interaction can be optimized. Expensive at first, but potentially cheaper than baking in rigid APIs for every use case.
Still, lots of unknowns, especially around latency, consistency, and cost.