Christopher King

METR’s pre­limi­nary eval­u­a­tion of o3 and o4-mini

[Question] How far along Metr’s law can AI start au­tomat­ing or helping with al­ign­ment re­search?

No, the Poly­mar­ket price does not mean we can im­me­di­ately con­clude what the prob­a­bil­ity of a bird flu pan­demic is. We also need to know the in­ter­est rate!

How I saved 1 hu­man life (in ex­pec­ta­tion) with­out over­think­ing it

Christo­pher King’s Shortform

LDT (and ev­ery­thing else) can be irrational

Acausal Now: We could to­tally acausally bar­gain with aliens at our cur­rent tech level if desired

Ne­cro­mancy’s un­in­tended con­se­quences.

How do low level hy­pothe­ses con­strain high level ones? The mys­tery of the dis­ap­pear­ing di­a­mond.

Challenge pro­posal: small­est pos­si­ble self-hard­en­ing back­door for RLHF

An­throp­i­cally Blind: the an­thropic shadow is re­flec­tively inconsistent

Solomonoff in­duc­tion still works if the uni­verse is un­com­putable, and its use­ful­ness doesn’t re­quire know­ing Oc­cam’s razor

De­mys­tify­ing Born’s rule

Cur­rent AI harms are also sci-fi

In­fer­ence from a Math­e­mat­i­cal De­scrip­tion of an Ex­ist­ing Align­ment Re­search: a pro­posal for an outer al­ign­ment re­search program

The un­spo­ken but ridicu­lous as­sump­tion of AI doom: the hid­den doom assumption

[Question] What pro­jects and efforts are there to pro­mote AI safety re­search?

See­ing Ghosts by GPT-4

We are mis­al­igned: the sad­den­ing idea that most of hu­man­ity doesn’t in­trin­si­cally care about x-risk, even on a per­sonal level

Pro­posal: we should start refer­ring to the risk from un­al­igned AI as a type of *ac­ci­dent risk*