Inkhaven Bucket List

+ I made Inkhaven Residency Post Explorer, & ran a blogging Murphyjitsu / Inner Sim / Premortem event

Today I woke up & looked around me & realized I was at this awesome place, Inkhaven, surrounded by people I haven’t really gotten to know yet.

I am obsessed with high snr (signal:noise ratio), & struggle with the firehose of the regular dashboard. So I made one where you can filter by topic / writer! Site here; code here; please feel free to make PRs.

In the afternoon, I ran a Murphyjitsu / Inner Sim / Premortem event. My blogging inspirations also include the three Gs¹, & Ben Kuhn, and I wanted to understand of each person at Inkhaven: what’s the delta between our current path and our aims? How do we ensure we’re on endorsed trajectories?

We went around in a circle, ~9mins/person, talking about current most likely outcomes for our blogs and whether we’re happy about these. Questions I asked were mostly person-specific, but included e.g.:

• What’s the post you’re proudest of? What environment / setting / context were you in when you wrote it? By default, how many times will you be in a similar context again across the next 2-3 years?

Inkhaven Bucket List

☑︎ Talk to Gwern about AI

☑︎ Talk to Zvi about <TBD>

☑︎ Scribe some agent foundations work for Alex Altair; try some whiteboarding / blog post amanuensis

☐ Read more of <person’s> posts; ask lots of questions:

• ☐ ☐ Abram Demski

• ☐ ☐ Markus Strasser

• ☑︎ ☐ Adrià Garriga Alonso

• ☑︎ ☑︎ Daniel Paleka

☑︎ Get back to Daniel re The Two Types of LLM Preferences

☑︎ Ask <person> for feedback on something:

• ☑︎ Scott Alexander

• ☑︎ Aaron Silverbook

☑︎ Ask Sasha how BCI experiments are going

☐ Read some of Ben Steinhorn’s novel

☐ Talk to Tsvi about human enhancement

☑︎ Chat to Skyler / Screwtape about the state of rationality meetups, and where my FHI reading / working group might fit into that

☑︎ Talk to Jenn about community-building

She’s been running a high-quality rationality salon for six years. Wow! How did she not lose stamina, or want to move onto a more domain-specific group? Takeaways:

1. Working 4 days/week really helps—she does 4 days Balsa 1 day on the group. I used to do 4 days on Manifund, 1 day on 90/30 Club. This seems the sustainable equilibrium for intellectual groups—and I should be realistic about what FHI group can achieve given this

2. Site / materials should attract the marginal member

Coming soon:

• How condu.it thought-to-text tech works

• Replication / extensions of LM introspection work:

  • They found filling Claude’s context window with information about how transformers work / LM introspection improved Claude’s ability to guess previous, now-hidden CoT—but that’s probably just because Claude changed the CoT it generated in response to the context…perhaps in a way that would be easier to guess / retrieve later on.

  • They then controlled for that effect. 230 results were normal, but…what on earth is going on with the one one-in-a-million ‘Awakened Claude’ result they report?

    

    • Could this just be the one-in-a-million result happening to manifest, as it’s bound to occasionally do? Or some methodological inconsistency / hiccup—either in experiment or interpretation?

      • Presumably they should now…do more runs?

      • My guess is they misconstructed the null.

      • Ah…I think I just found something to talk to Zvi about :) Zvi does a Twitter pass every morning. How does Zvi use Twitter productively? What are his top accounts to follow? How does he orient to things like this?

    • I’ll send updates re Awakened Claude in a future blog post.

      • Edit: Here it is!

1. h/t Croissanthology