Building the David Deutsch compendium
Why I wanted a more complete record of David Deutsch’s public work, and how I’m trying to keep it useful over time.
I’ve been building a David Deutsch Compendium: a living reference to his public work — talks, interviews, lectures, books, academic papers, essays, transcripts, and related primary sources.
It started with a simple problem. I wanted to explore more of Deutsch’s work, but there wasn’t one place I could go to see what actually existed.
There are already useful indexes, playlists, podcast pages, fan collections, institutional archives, and pages on Deutsch’s own website. But each covers a different slice. Older links disappear. A recording may be reposted under a different title years later. One interview can turn into a dozen clips. A transcript can survive after the original audio vanishes. Publication dates and recording dates get mixed together. Sometimes a source exists but is buried deeply enough that it is effectively missing.
So the project became less about making a list and more about reconciling a record.
What I’m trying to make
The goal is not merely a long chronology. It is a useful map of the material.
For each item I try to identify the underlying work or appearance, find the strongest surviving source, preserve useful mirrors and transcripts, and avoid counting the same underlying thing repeatedly just because it has been republished in several places.
I also want the compendium to remain maintainable. New appearances will happen. Better sources will turn up. Links will die. Dates will be corrected. Two entries that look distinct may turn out to be the same recording, and something that looks like a duplicate may turn out not to be.
I would rather maintain a record that can be corrected than publish a supposedly definitive list and leave it to decay.
How it was compiled — and keeps getting better
A lot of the initial compilation was done with ChatGPT Deep Research. Instead of relying on one existing list, I used it to search across official sites, university archives, publisher and podcast pages, YouTube, old conference pages, transcripts, web archives, community indexes, and other references, then reconcile the results against one another.
That process is also ongoing. I have a recurring ChatGPT research monitor that searches for newly published Deutsch material and for improvements to older entries: stronger primary sources, missing dates, transcripts, recovered recordings, duplicate candidates, and other metadata that can make the record more accurate. When it finds something useful, it updates the working compendium rather than merely sending me a list of links to deal with later.
So the compendium is not intended to be a one-time research dump. It is meant to be continuously maintained: new work gets added, old entries get corrected, sources get upgraded, and the structure itself gets polished as better evidence turns up.
Gaps are useful information
One reason completeness matters is that a missing source can itself become a useful clue.
While working through the compendium, I found references to a remote keynote Deutsch gave at Indiana University in 2008: What Is Computation? (How) Does Nature Compute? A transcript survived, and the historical record showed that video had once existed, but I could not find a usable recording.
I contacted Indiana University to ask whether it might still survive somewhere.
University staff found the old recording in a web archive in the obsolete RealMedia format, recovered and converted it, restored it through IU’s media system, and made a fresh public copy available. The archival and technical work was theirs; I had simply encountered a gap in the record and asked a question about it.
That is a small example of what I hope a maintained compendium can occasionally do. Most of the time it will simply help someone find something that already exists. Sometimes it may identify uncertainty. And every once in a while, documenting what appears to be missing may make it possible for an old piece of work to reappear.
How I maintain the record
The basic unit is the underlying work or appearance, not the URL. A podcast episode that also exists on YouTube, Spotify, a host page, and in transcript form should still be one canonical item. Those additional URLs are sources for the item rather than new entries. When one underlying work is released in multiple parts — a six-lecture course, for example, or an interview later broken into topical clips — those components stay attached to the same canonical item rather than becoming artificial duplicates.
For audio and video, I prefer a verified complete YouTube copy as the main playback link when one exists, because it is usually the most convenient place to watch or listen. I still preserve the original host, author, publisher, or institutional page for provenance, metadata, transcripts, and recovery if links later disappear. For genuinely written work, the original text remains the natural primary destination. A clip or one component of a larger work does not replace the source for the whole item.
Dates are also treated as evidence rather than decoration. The date an interview actually happened can differ from its publication date, and a recording uploaded years later should not suddenly become a new appearance. Where both are known, I preserve both and use the underlying event or recording date to place the work in time.
Transcript provenance matters too. An official transcript, an author-approved transcription, a host’s show notes, a third-party transcript, and a machine transcript are not the same thing. I try to label the distinction rather than flatten them all into “transcript available.”
Potential duplicates are intentionally held apart when the evidence is not strong enough to merge them. Matching interviewer names, similar titles, or nearby dates are clues, not proof. If two candidate recordings may be the same session, I would rather compare runtimes, openings, closings, question order, and content before collapsing them and accidentally deleting a distinct appearance.
The same caution applies in the other direction. Clips, excerpts, alternate edits, translated versions, and later reposts should not inflate the catalog when they are simply another representation of an existing work. Community indexes, playlists, caption databases, archives, and fan collections are extremely useful discovery tools, but candidates from them are reconciled against the underlying source before becoming canonical.
Finally, absence is tracked explicitly. A known appearance can remain in the compendium even when its original recording or page is no longer available. That is different from a research lead, where I may still be trying to determine whether something is a distinct appearance at all, recover better metadata, find an original source, or resolve a possible duplicate.
This methodology is deliberately conservative in both directions: do not invent extra appearances from repackaging, and do not erase potentially distinct material merely because it resembles something already cataloged.
Bringing the record together
Information about Deutsch’s work is distributed across official sites, universities, publishers, podcast hosts, archives, transcripts, community indexes, playlists, and preserved pages. Each holds a different part of the public record, and the compendium brings those facts together in one place.
For audiovisual material, a complete YouTube copy is the preferred playback destination when available; original and institutional sources remain attached for provenance, metadata, and recovery. Where a transcript, mirror, or community archive preserves something useful, I want to keep that trail too.
The result will inevitably contain mistakes and omissions. That is another reason I want it to be maintained rather than treated as finished.