The ChatGPT-like chat interface is quite 2022, and I surmise users don’t want to read a wall of text, even though there are supporting quotes and timestamps (albeit a bit weak as well).

I have now completely redesigned the responses to be video-first. It now looks like this by default:

The idea here is to show the most relevant videos and sections across a channel, instantly. Though, depending on the size of the channel, this could take up to ~10 seconds.

If you would still prefer the previous format, please reach out.

I also switched to Cerebras GPT-OSS 120B as the workhorse model for generating the summaries and cleaning up the quotes. This model is blazing fast, and as of today, the fastest on the market (~3000 tokens / second).

Previously, Gemini 2.5 Flash-lite, although cheaper, was fairly unreliable, and frankly not that smart by today’s standards.