Dwarkesh Context Window

I am not affiliated with Dwarkesh Patel in any way.

Why this exists

This project takes a full transcript from a Dwarkesh Patel podcast episode and prompts a frontier LLM to act as a third guest—summarizing, questioning, and extending the conversation into new directions for AI research.

One motivating idea is: if models ever become genuinely great researchers, they should be able to follow (and contribute to) expert conversations. A long-form interview—dense with context, claims, and open questions—acts like a surprisingly strong “test harness” for that capability.

Over time, this site can serve as a historical record of how well different models “think along” with top researchers, and potentially evolve into a lightweight benchmark.

Inspiration: Theo.gg’s Skatebench

This project is inspired by Theo.gg’s “Skatebench” benchmark: an LLM benchmark where the task is to name a skateboarding trick from an English description. It sounds silly at first, but it produced surprisingly interesting results—early on, some Chinese models lagged behind American models, and later on, some models even regressed on Skatebench while improving on other benchmarks. The key takeaway is that weird, random benchmarks can be genuinely insightful, especially when you track them over time.

What’s here today

• A small, static “blog-ish” site rendering LLM-generated markdown posts
• Per-post metadata (model, token counts, runtime)
• Scripts to fetch transcripts and generate posts

Ideas for the future

• Multiple prompts/models per episode, with side-by-side comparisons
• An “assistant alongside the podcast” experience (timestamped insights while you watch)
• Reader feedback/rating to identify which prompts produce the most useful research directions and iterate toward better “third guest” behavior

Tech Stack

Tanstack Start

We should be able to prerender everything for mvp. Later on we could add variations from different models, or add comments.

Effect-ts for scripts and anything server side

Because its awesome.

Oxlint/Oxfmt

I wanted to give it a shot.

Notes

• I would love to use tsgo, but it currently conflicts with the effect-ts ts plugin.

MVP Features TODO

• Get all software tools installed and configured correctly.
• Write a script that can get the transcript from a podcast episode.
• Write a script that will prompt an llm with the script and output markdown.
• Render the md content
• Style the site

Potential Future Features

• [-] Multiple tabs with different models or prompts.
• Add metadata around how long and how much it cost to generate the content.
• Add a rating system.
• Add a benchmark based on ratings.