439 episodes · 7.4M words · live site ↗
The full transcript archive of Sean Carroll’s Mindscape, read as one body of text. Topic models at three resolutions, a serendipity recommender, guest and time analyses, and a search that understands the question and returns the moment with a timestamp. Built from a 16-stage pipeline that rebuilds only what changed.
2,581 listener questions · 63 AMA episodes · live site ↗
Type a question and see whether the host has already answered it, and where. Semantic matching over every Q&A pair in the show’s AMA episodes, about 150 ms per query, no language model in the loop.
Retrieval with a certified surplus · MCP server
A local retrieval system for Claude agents with a guaranteed-relevance floor (dense pool, lexical recall, cross-encoder rerank) and, above it, one statistically certified non-obvious passage per query. Serves the Stanford Encyclopedia of Philosophy and the Mindscape archive.
7 open-weight MLX releases · ~3k downloads/month · live site ↗
Qwen3.8-32B-Jumbo restores knowledge lost in continued training by transplanting the most-drifted layer blocks from a sibling checkpoint, chosen by a per-tensor alignment census; Mini-Me is a sheared 23B. The LoRA toolkit behind them measures catastrophic forgetting on every run, and the fine-tuned variants that regressed were not released.
A protocol built from failure modes
A prompt-engineering method that starts from a task’s known ways of going wrong and conditions the context with counter-examples and a structured self-audit rather than prohibitions, evaluated with controlled experiments.