Astral Codex Ten

Daniel Filan

7 min read

Daniel Filan

Article

Daniel Filan is a recurring person in the Astral Codex Ten archive, appearing 4 times across 4 issues between May 20, 2021 and March 28, 2024. The archive places it in contexts such as “Daniel Filan now has an AI X-Risk Research Podcast”; “I think I’ve mentioned Daniel Filan here in the contest o”; “mentioned Daniel Filan here in the contest of his AI alignment work”. It most often appears alongside Less Wrong, AI X-Risk Research Podcast, China.

Metadata

  • Category: People
  • Mention count: 4
  • Issue count: 4
  • First seen: May 20, 2021
  • Last seen: March 28, 2024

Appears In

Source Context

Recovered passages from the original issue text. When the raw archive preserved outbound links inside the source passage, they are listed directly under the quote.

Links For May
May 20, 2021 · Original source
36: Related: Daniel Filan now has an AI X-Risk Research Podcast (conflict of interest notice: I know Daniel and think he is generally great).
Inline links: AI X-Risk Research Podcast
Links For February 2023
February 09, 2023 · Original source
31: I think I’ve mentioned Daniel Filan here in the contest of his AI alignment work and his AI X-Risk Research Podcast. He has a new podcast about not-just-AI-related things, The Filan Cabinet. First three episodes are interviews with a former Congressional candidate, a Presbyterian pastor, and a cryptocurrency developer.
Inline links: AI X-Risk Research Podcast, The Filan Cabinet
Pause For Thought: The AI Pause Debate
October 05, 2023 · Original source
Nora thought it was important to give alignment researchers advanced models to experiment with, because the sort of armchair alignment research before interesting AIs existed (eg Bostrom’s Superintelligence) wasn’t just wrong, but fostered dead-end worse-than-nothing paradigms that continue to confuse the field. Daniel Filan objected that Bostrom got some things right and even described something like the direction that modern alignment research is taking. There was a long argument about this, which I think reduces to “Bostrom said some useful theoretical things, speculated about practical direction, and a few of his speculations were right but most now seem outdated”.
Inline links: objected
Practically-A-Book Review: Rootclaim $100,000 Lab Leak Debate
March 28, 2024 · Original source
The obvious lack of evidence for a wildlife spillover in HSM further reduces this factor, making HSM negligible evidence. Saar remains as committed to Rootclaim as ever. He’s even still committed to $100,000 bets on Rootclaim findings, settled via debate. He’d even be willing to re-debate Peter on lab leak! (Peter declined) He does want to make some changes to the debate format - specifically, switching from video to text, and checking in more often with the judges to get feedback on their thought processes. The switch from video to text seems reasonable. Saar was clearly flummoxed by Peter’s memory and agility, and wants a format where ability to remember/think on your feet is less important, and where you can do lots of research before having to think up a response. The part with the feedback seems to be Saar wanting even more of an opportunity to identify disagreements with the judges early, and get a chance to tell them beforehand about issues like the ones above. I’m sympathetic to both these changes, but I don’t think they would have changed the outcome of this debate. The Aftermath: Rootclaim Rootclaim is an admirable idea. Somebody called it “heroic Bayesian analysis”, and I like the moniker. Regular human reasoning doesn’t seem to be doing a great job puncturing false beliefs these days, and lots of people have converged on something something Bayes as a solution. But the something something remains elusive. While everyone else tries “pop Bayesianism” and “Bayes-inspired toolboxes”, Rootclaim asks: what if you just directly apply Bayes to the world’s hardest problems? There’s something pure about that, in a way nobody else is trying. Unfortunately, the reason nobody else is trying this is because it doesn’t work. There’s too much evidence, and it’s too hard to figure out how to quantify it. Peter, Saar, and the two judges all did their own Bayesian analysis. I followed along at home7 and tried the same. Daniel Filan, who also watched the debate, did one too. Here’s a comparison of all of our results: Again, most people didn’t use these exact categories, I’m putting them in this format to make them easy to compare, and any errors are mine.... The six estimates span twenty-three orders of magnitude. Even if we remove Peter (who’s kind of trolling), the remaining estimates span a range of ~7 OOMs. And even if we remove Saar (limiting the analysis to neutral non-participants), we’re still left with a factor-of-50 difference. 50x sounds good compared to 23 OOMs. But it only sounds good because everyone except Saar leaned heavily towards zoonosis. If raters were closer to even, it would become problematic: even a factor of 50x is enough to change 80-20 lab leak to 80-20 natural. Saar’s perspective is that true theories should have many orders of magnitude more evidence than false theories (most of the Rootclaim analyses end up with normal-sounding percentages like “94% sure”, but that’s after they correct for potential model error). If that’s true, a 50x fudge factor shouldn’t be fatal. But 23 orders of magnitude is fatal any way you slice it. The best one can say is that maybe this is no worse than normal reasoning. Among normal people who don’t use Rootclaim, many are sure lab leak is true, and many others are sure it’s false. If we interpret “sure” as 99%, then even normal people without Rootclaim are a factor of 10,000x away from each other. If we interpret “sure” as including more nines than that, maybe normal people are 23 OOMs away from each other, who knows? In this model, Rootclaim is no worse than anything else; it’s just legible enough that we notice these discrepancies. We’ve gotten inured to people failing to agree on difficult issues. Maybe Rootclaim gets credit for showing us exactly where we fail, and putting numbers on the failure. Still, this is faint praise for a method that hoped to be able to resolve these kinds of disagreements. In the end, I think Saar has two options: Abandon the Rootclaim methodology, and go back to normal boring impure reasoning like the rest of us, where you vaguely gesture at Bayesian math but certainly don’t try anything as extreme as actually using it.
Inline links: 7, also watched, did one too, https://substackcdn.com/image/fetch/$s_!CPb3!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F92c84984-eb1b-4b0c-a8f7-5140ef39a03c_763x781.png
Daniel Filan’s running Twitter commentary of the debate
Inline links: Daniel Filan’s running Twitter commentary of the debate