RSS

To­tal Safety Trans­parency?

Austin Chen22 Sep 2026 5:39 UTC
8 points
0 comments9 min readLW link
(manifund.substack.com)

Math con­tains a lot of stuff

Warty22 Sep 2026 5:39 UTC
3 points
0 comments2 min readLW link

An opinionated primer on Causal De­ci­sion The­ory (CDT) and Ev­i­den­tial De­ci­sion The­ory (EDT)

Chi Nguyen22 Sep 2026 0:06 UTC
25 points
0 comments8 min readLW link

Acausal in­ter­ac­tions: What they are, why they mat­ter, and what to do about them

Chi Nguyen22 Sep 2026 0:01 UTC
16 points
0 comments3 min readLW link

Story Im­print­ing: AI As­sis­tants Ab­sorb Traits from Hu­man Char­ac­ters They Resemble

21 Sep 2026 23:56 UTC
20 points
0 comments11 min readLW link

Why ad­vance au­to­mated sci­en­tific labs rather than slow them?

Master Chief21 Sep 2026 23:02 UTC
4 points
1 comment1 min readLW link

What if not Cir­cuits?

CarolusRenniusVitellius21 Sep 2026 22:39 UTC
88 points
1 comment15 min readLW link

CLT Fea­tures Sup­port Man­i­fold Steer­ing, but Hold Only Par­tial Steer­ing Signal

Anna Marbut21 Sep 2026 21:20 UTC
7 points
0 comments7 min readLW link

Swarm Scaling

Toby_Ord21 Sep 2026 20:30 UTC
89 points
4 comments8 min readLW link

Tur­bu­lence & Frag­ility: Prompt­ing-Based Ex­per­i­ments are Sen­si­tive to Stray De­tails (And other les­sons for new re­searchers.)

Czynski21 Sep 2026 19:04 UTC
11 points
0 comments18 min readLW link

Will AI wipe out the hu­man race?

scout121 Sep 2026 18:47 UTC
−1 points
0 comments6 min readLW link

Align­ment Mid­train­ing Cracks Un­der Pressure

21 Sep 2026 16:55 UTC
88 points
5 comments6 min readLW link

We’re not ready for the e/​Acc × Longevity prefer­ence cascade

Jackson Wagner21 Sep 2026 15:21 UTC
31 points
19 comments11 min readLW link

A class of state­ment be­tween con­jec­ture and theorem

Jason Fantl21 Sep 2026 15:18 UTC
29 points
5 comments2 min readLW link

Grat­i­tude and the End of the World

Ephraiem Sarabamoun21 Sep 2026 15:06 UTC
5 points
2 comments1 min readLW link

Weight smug­gling likely defeats at­tempts to cap FLOPs per train­ing run

21 Sep 2026 14:50 UTC
36 points
4 comments1 min readLW link

Grant­mak­ers: Con­sider shar­ing rough prob­a­bil­ities & brief feed­back with grantees/​ap­pli­cants

david reinstein21 Sep 2026 14:46 UTC
19 points
1 comment4 min readLW link

Mech In­terp is a Ver­ifi­able Task

Logan Riggs21 Sep 2026 14:42 UTC
35 points
3 comments5 min readLW link

A Brief His­tory of Koinometry

Benjamin Schneider21 Sep 2026 13:15 UTC
5 points
0 comments4 min readLW link
(benjaminschneider.ch)

Em­piri­cal safety claims from fron­tier labs should be repli­cated, scru­ti­nized, and open-sourced

21 Sep 2026 5:58 UTC
80 points
3 comments4 min readLW link
(secondlookresearch.com)