Early AI safety talks from 2000 to 2020 are being archived before link rot kills them

The list

A LessWrong user published a curated list of talks that happened in the 2000-2020 period, explicitly intended for future people to look up or for search results to surface. The framing is preservation: many of the original links have already rotted, and the post asks readers to pass it forward and find a way to repost it.

The collection spans the original Singularity Summit years — 2006 through 2012 — along with the wider early AI-safety and AI-governance conversation that happened in conferences, salons and lecture series before the current boom.

Bostrom's 504 interviews

A centerpiece of the archive is Nick Bostrom's interview output: 504 interviews listed for the period 1992-2012, with 493 crowded into 13 of those years — an average of roughly one interview every 13 days. Bostrom's interviews are a running record of the AI safety community's early thinking, conducted as the field went from fringe to founding.

The list also includes archive.org mirrors that still work, such as the Future Salon 2002 collection, giving researchers a stable path to at least part of the material even where the original host is gone.

The recordings

Notable items include Shane Legg's Singularity Summit 2010 talk on universal measures of intelligence and Eliezer Yudkowsky's Singularity Summit 2011 talk on open problems in Friendly AI. These talks are touchstones in AI safety history — cited in later papers and retrospectives — and losing them would erase primary sources from a pivotal decade.

The post notes that a lot of links have rotted but, thankfully, not the majority of the Vimeo recordings. Vimeo's survival rate has been better than the blogs and conference pages that hosted the surrounding material.

Why it matters

AI governance researchers, safety historians and newcomers all read the same canonical talks, but the record is fragile: it lives on dead domains, sold-off startups and platforms that periodically delete old content. The archival effort is a reminder that the field's pre-2020 history is not safely curated anywhere.

The post is deliberately raw — a list, not a polished archive — and asks the community to restructure it, add independently stored videos and assemble something more comprehensive. Preservation here is a crowd task, and the window is closing as the links rot.

Caveats

The post is a list post, not a verified catalog: many links are dead, the counts rely on the listing (494 interviews over 13 years is a stated grouping, not independently audited here), and the archive.org mirrors may have incomplete metadata.

AI Lookout verified the post and the archive.org identifier it references; it did not attempt to open every catalogued link.

How to help

The post's ask is practical: if you have a copy of a talk, mirror it; if you run a site that hosts old recordings, keep them online; and if you are a researcher, cite the archive so the links stay in circulation. The author explicitly invites restructuring the list into a more comprehensive archive with independently stored videos rather than links that depend on any single host.

The archive.org identifiers give the effort a durable home: items like the Future Salon 2002 collection are already mirrored, and the pattern — original host plus archive.org copy plus Vimeo — is the model the post argues should be applied to the whole catalog.

What this says about AI history

The fragility of the early AI safety record is a data point about how fast the field moved. Twenty years ago, 'AI safety' was a handful of talks and interviews; today it is an industry with published policy frameworks. The talks being archived are the primary sources for that transition — the honest record of early arguments, fears and disagreements, before the labels hardened.

For readers of current AI news, the archive is a reminder that today's debates about alignment, control and governance have a paper trail from well before the current boom, and that the field has already forgotten and rediscovered many of the same problems.

Context for today's debates

The archive arrives at a moment when the AI safety field is institutionalized — labs have safety teams, governments have frameworks, and conferences are career tracks. The early talks were different: smaller audiences, existential framing, and arguments conducted in public by a handful of researchers.

Reading them fresh (where the recordings survive) shows how much of the current vocabulary — agentic alignment, control, interpretability, corrigibility — was already present in rudimentary form a decade or more ago, and how much of what feels new is actually a return to old questions with better tools.

The value of the catalog is that it is a raw dataset for the history of AI governance, exactly the kind of source material that disappears quietly when no one is looking. The post's stated purpose — for future people to look up, or to surface in search results — is preservation with intent.