Alignment Newsletter

Rohin Shah
1 Aug 2018 18:31 UTC

I publish the Alignment Newsletter, a weekly publication with recent content relevant to AI alignment. See here for more details. Quick links: email signup form, RSS feed, spreadsheet of all summaries.

The Align­ment Newslet­ter #1: 04/​09/​18

Rohin Shah9 Apr 2018 16:00 UTC
12 points
3 comments4 min readLW link

The Align­ment Newslet­ter #2: 04/​16/​18

Rohin Shah16 Apr 2018 16:00 UTC
8 points
0 comments5 min readLW link

The Align­ment Newslet­ter #3: 04/​23/​18

Rohin Shah23 Apr 2018 16:00 UTC
9 points
0 comments6 min readLW link

The Align­ment Newslet­ter #4: 04/​30/​18

Rohin Shah30 Apr 2018 16:00 UTC
8 points
0 comments3 min readLW link

The Align­ment Newslet­ter #5: 05/​07/​18

Rohin Shah7 May 2018 16:00 UTC
8 points
0 comments7 min readLW link

The Align­ment Newslet­ter #6: 05/​14/​18

Rohin Shah14 May 2018 16:00 UTC
8 points
0 comments2 min readLW link

The Align­ment Newslet­ter #7: 05/​21/​18

Rohin Shah21 May 2018 16:00 UTC
8 points
0 comments5 min readLW link

The Align­ment Newslet­ter #8: 05/​28/​18

Rohin Shah28 May 2018 16:00 UTC
8 points
0 comments6 min readLW link

The Align­ment Newslet­ter #9: 06/​04/​18

Rohin Shah4 Jun 2018 16:00 UTC
8 points
0 comments2 min readLW link

The Align­ment Newslet­ter #10: 06/​11/​18

Rohin Shah11 Jun 2018 16:00 UTC
16 points
0 comments9 min readLW link

The Align­ment Newslet­ter #11: 06/​18/​18

Rohin Shah18 Jun 2018 16:00 UTC
8 points
0 comments10 min readLW link

The Align­ment Newslet­ter #12: 06/​25/​18

Rohin Shah25 Jun 2018 16:00 UTC
15 points
0 comments3 min readLW link

Align­ment Newslet­ter #13: 07/​02/​18

Rohin Shah2 Jul 2018 16:10 UTC
70 points
12 comments8 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #14

Rohin Shah9 Jul 2018 16:20 UTC
14 points
0 comments9 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #15: 07/​16/​18

Rohin Shah16 Jul 2018 16:10 UTC
42 points
0 comments15 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #16: 07/​23/​18

Rohin Shah23 Jul 2018 16:20 UTC
42 points
0 comments12 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #17

Rohin Shah30 Jul 2018 16:10 UTC
32 points
0 comments13 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #18

Rohin Shah6 Aug 2018 16:00 UTC
17 points
0 comments10 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #19

Rohin Shah14 Aug 2018 2:10 UTC
18 points
0 comments13 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #20

Rohin Shah20 Aug 2018 16:00 UTC
12 points
2 comments6 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #21

Rohin Shah27 Aug 2018 16:20 UTC
25 points
0 comments7 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #22

Rohin Shah3 Sep 2018 16:10 UTC
18 points
0 comments6 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #23

Rohin Shah10 Sep 2018 17:10 UTC
16 points
0 comments7 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #24

Rohin Shah17 Sep 2018 16:20 UTC
10 points
6 comments12 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #25

Rohin Shah24 Sep 2018 16:10 UTC
18 points
3 comments9 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #26

Rohin Shah2 Oct 2018 16:10 UTC
13 points
0 comments7 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #27

Rohin Shah9 Oct 2018 1:10 UTC
16 points
0 comments9 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #28

Rohin Shah15 Oct 2018 21:20 UTC
11 points
0 comments8 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #29

Rohin Shah22 Oct 2018 16:20 UTC
15 points
0 comments9 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #30

Rohin Shah29 Oct 2018 16:10 UTC
29 points
2 comments6 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #31

Rohin Shah5 Nov 2018 23:50 UTC
17 points
0 comments12 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #32

Rohin Shah12 Nov 2018 17:20 UTC
18 points
0 comments12 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #33

Rohin Shah19 Nov 2018 17:20 UTC
23 points
0 comments9 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #34

Rohin Shah26 Nov 2018 23:10 UTC
24 points
0 comments10 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #35

Rohin Shah4 Dec 2018 1:10 UTC
15 points
0 comments6 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #36

Rohin Shah12 Dec 2018 1:10 UTC
21 points
0 comments11 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #37

Rohin Shah17 Dec 2018 19:10 UTC
25 points
4 comments10 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #38

Rohin Shah25 Dec 2018 16:10 UTC
9 points
0 comments8 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #39

Rohin Shah1 Jan 2019 8:10 UTC
32 points
2 comments5 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #40

Rohin Shah8 Jan 2019 20:10 UTC
21 points
2 comments5 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #41

Rohin Shah17 Jan 2019 8:10 UTC
22 points
6 comments10 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #42

Rohin Shah22 Jan 2019 2:00 UTC
20 points
1 comment10 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #43

Rohin Shah29 Jan 2019 21:10 UTC
14 points
2 comments13 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #44

Rohin Shah6 Feb 2019 8:30 UTC
18 points
0 comments9 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #45

Rohin Shah14 Feb 2019 2:10 UTC
25 points
2 comments8 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #46

Rohin Shah22 Feb 2019 0:10 UTC
12 points
0 comments9 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #47

Rohin Shah4 Mar 2019 4:30 UTC
18 points
0 comments8 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #48

Rohin Shah11 Mar 2019 21:10 UTC
29 points
14 comments9 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #49

Rohin Shah20 Mar 2019 4:20 UTC
23 points
1 comment11 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #50

Rohin Shah28 Mar 2019 18:10 UTC
15 points
2 comments10 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #51

Rohin Shah3 Apr 2019 4:10 UTC
25 points
2 comments15 min readLW link
(mailchi.mp)

Align­ment Newslet­ter #52

Rohin Shah6 Apr 2019 1:20 UTC
19 points
1 comment8 min readLW link
(mailchi.mp)

Align­ment Newslet­ter One Year Retrospective

Rohin Shah10 Apr 2019 6:58 UTC
94 points
31 comments21 min readLW link

Align­ment Newslet­ter #53

Rohin Shah18 Apr 2019 17:20 UTC
20 points
0 comments8 min readLW link
(mailchi.mp)

[AN #54] Box­ing a finite-hori­zon AI sys­tem to keep it unambitious

Rohin Shah28 Apr 2019 5:20 UTC
20 points
0 comments8 min readLW link
(mailchi.mp)

[AN #55] Reg­u­la­tory mar­kets and in­ter­na­tional stan­dards as a means of en­sur­ing benefi­cial AI

Rohin Shah5 May 2019 2:20 UTC
17 points
2 comments8 min readLW link
(mailchi.mp)

[AN #56] Should ML re­searchers stop run­ning ex­per­i­ments be­fore mak­ing hy­pothe­ses?

Rohin Shah21 May 2019 2:20 UTC
21 points
8 comments9 min readLW link
(mailchi.mp)

[AN #57] Why we should fo­cus on ro­bust­ness in AI safety, and the analo­gous prob­lems in programming

Rohin Shah5 Jun 2019 23:20 UTC
26 points
15 comments7 min readLW link
(mailchi.mp)

[AN #58] Mesa op­ti­miza­tion: what it is, and why we should care

Rohin Shah24 Jun 2019 16:10 UTC
55 points
10 comments8 min readLW link
(mailchi.mp)

[AN #59] How ar­gu­ments for AI risk have changed over time

Rohin Shah8 Jul 2019 17:20 UTC
43 points
4 comments7 min readLW link
(mailchi.mp)

[AN #60] A new AI challenge: Minecraft agents that as­sist hu­man play­ers in cre­ative mode

Rohin Shah22 Jul 2019 17:00 UTC
23 points
6 comments9 min readLW link
(mailchi.mp)

[AN #61] AI policy and gov­er­nance, from two peo­ple in the field

Rohin Shah5 Aug 2019 17:00 UTC
12 points
2 comments9 min readLW link
(mailchi.mp)

[AN #62] Are ad­ver­sar­ial ex­am­ples caused by real but im­per­cep­ti­ble fea­tures?

Rohin Shah22 Aug 2019 17:10 UTC
28 points
10 comments9 min readLW link
(mailchi.mp)

[AN #63] How ar­chi­tec­ture search, meta learn­ing, and en­vi­ron­ment de­sign could lead to gen­eral intelligence

Rohin Shah10 Sep 2019 19:10 UTC
21 points
12 comments8 min readLW link
(mailchi.mp)

[AN #64]: Us­ing Deep RL and Re­ward Uncer­tainty to In­cen­tivize Prefer­ence Learning

Rohin Shah16 Sep 2019 17:10 UTC
11 points
8 comments7 min readLW link
(mailchi.mp)

[AN #65]: Learn­ing use­ful skills by watch­ing hu­mans “play”

Rohin Shah23 Sep 2019 17:30 UTC
11 points
0 comments9 min readLW link
(mailchi.mp)

[AN #66]: De­com­pos­ing ro­bust­ness into ca­pa­bil­ity ro­bust­ness and al­ign­ment robustness

Rohin Shah30 Sep 2019 18:00 UTC
12 points
1 comment7 min readLW link
(mailchi.mp)

[AN #67]: Creat­ing en­vi­ron­ments in which to study in­ner al­ign­ment failures

Rohin Shah7 Oct 2019 17:10 UTC
17 points
0 comments8 min readLW link
(mailchi.mp)

[AN #68]: The at­tain­able util­ity the­ory of impact

Rohin Shah14 Oct 2019 17:00 UTC
17 points
0 comments8 min readLW link
(mailchi.mp)

[AN #69] Stu­art Rus­sell’s new book on why we need to re­place the stan­dard model of AI

Rohin Shah19 Oct 2019 0:30 UTC
60 points
12 comments15 min readLW link
(mailchi.mp)

[AN #70]: Agents that help hu­mans who are still learn­ing about their own preferences

Rohin Shah23 Oct 2019 17:10 UTC
16 points
0 comments9 min readLW link
(mailchi.mp)

[AN #71]: Avoid­ing re­ward tam­per­ing through cur­rent-RF optimization

Rohin Shah30 Oct 2019 17:10 UTC
12 points
0 comments7 min readLW link
(mailchi.mp)

[AN #72]: Align­ment, ro­bust­ness, method­ol­ogy, and sys­tem build­ing as re­search pri­ori­ties for AI safety

Rohin Shah6 Nov 2019 18:10 UTC
26 points
4 comments10 min readLW link
(mailchi.mp)

[AN #73]: De­tect­ing catas­trophic failures by learn­ing how agents tend to break

Rohin Shah13 Nov 2019 18:10 UTC
11 points
0 comments7 min readLW link
(mailchi.mp)

[AN #74]: Separat­ing benefi­cial AI into com­pe­tence, al­ign­ment, and cop­ing with impacts

Rohin Shah20 Nov 2019 18:20 UTC
19 points
0 comments7 min readLW link
(mailchi.mp)

[AN #75]: Solv­ing Atari and Go with learned game mod­els, and thoughts from a MIRI employee

Rohin Shah27 Nov 2019 18:10 UTC
38 points
1 comment10 min readLW link
(mailchi.mp)

[AN #76]: How dataset size af­fects ro­bust­ness, and bench­mark­ing safe ex­plo­ra­tion by mea­sur­ing con­straint violations

Rohin Shah4 Dec 2019 18:10 UTC
14 points
6 comments9 min readLW link
(mailchi.mp)

[AN #77]: Dou­ble de­scent: a unifi­ca­tion of statis­ti­cal the­ory and mod­ern ML practice

Rohin Shah18 Dec 2019 18:30 UTC
21 points
4 comments14 min readLW link
(mailchi.mp)

[AN #78] For­mal­iz­ing power and in­stru­men­tal con­ver­gence, and the end-of-year AI safety char­ity comparison

Rohin Shah26 Dec 2019 1:10 UTC
26 points
10 comments9 min readLW link
(mailchi.mp)

[AN #79]: Re­cur­sive re­ward mod­el­ing as an al­ign­ment tech­nique in­te­grated with deep RL

Rohin Shah1 Jan 2020 18:00 UTC
13 points
0 comments12 min readLW link
(mailchi.mp)

[AN #80]: Why AI risk might be solved with­out ad­di­tional in­ter­ven­tion from longtermists

Rohin Shah2 Jan 2020 18:20 UTC
36 points
95 comments10 min readLW link
(mailchi.mp)

[AN #81]: Univer­sal­ity as a po­ten­tial solu­tion to con­cep­tual difficul­ties in in­tent alignment

Rohin Shah8 Jan 2020 18:00 UTC
32 points
4 comments11 min readLW link
(mailchi.mp)

[AN #82]: How OpenAI Five dis­tributed their train­ing computation

Rohin Shah15 Jan 2020 18:20 UTC
19 points
0 comments8 min readLW link
(mailchi.mp)

[AN #83]: Sam­ple-effi­cient deep learn­ing with ReMixMatch

Rohin Shah22 Jan 2020 18:10 UTC
15 points
4 comments11 min readLW link
(mailchi.mp)

[AN #84] Re­view­ing AI al­ign­ment work in 2018-19

Rohin Shah29 Jan 2020 18:30 UTC
23 points
0 comments6 min readLW link
(mailchi.mp)

[AN #85]: The nor­ma­tive ques­tions we should be ask­ing for AI al­ign­ment, and a sur­pris­ingly good chatbot

Rohin Shah5 Feb 2020 18:20 UTC
14 points
2 comments7 min readLW link
(mailchi.mp)

[AN #86]: Im­prov­ing de­bate and fac­tored cog­ni­tion through hu­man experiments

Rohin Shah12 Feb 2020 18:10 UTC
15 points
0 comments9 min readLW link
(mailchi.mp)

[AN #87]: What might hap­pen as deep learn­ing scales even fur­ther?

Rohin Shah19 Feb 2020 18:20 UTC
28 points
0 comments4 min readLW link
(mailchi.mp)

[AN #88]: How the prin­ci­pal-agent liter­a­ture re­lates to AI risk

Rohin Shah27 Feb 2020 9:10 UTC
18 points
0 comments9 min readLW link
(mailchi.mp)

[AN #89]: A unify­ing for­mal­ism for prefer­ence learn­ing algorithms

Rohin Shah4 Mar 2020 18:20 UTC
16 points
0 comments9 min readLW link
(mailchi.mp)

[AN #90]: How search land­scapes can con­tain self-re­in­forc­ing feed­back loops

Rohin Shah11 Mar 2020 17:30 UTC
11 points
6 comments8 min readLW link
(mailchi.mp)

[AN #91]: Con­cepts, im­ple­men­ta­tions, prob­lems, and a bench­mark for im­pact measurement

Rohin Shah18 Mar 2020 17:10 UTC
15 points
10 comments13 min readLW link
(mailchi.mp)

[AN #92]: Learn­ing good rep­re­sen­ta­tions with con­trastive pre­dic­tive coding

Rohin Shah25 Mar 2020 17:20 UTC
18 points
1 comment10 min readLW link
(mailchi.mp)

[AN #93]: The Precipice we’re stand­ing at, and how we can back away from it

Rohin Shah1 Apr 2020 17:10 UTC
24 points
0 comments7 min readLW link
(mailchi.mp)

[AN #94]: AI al­ign­ment as trans­la­tion be­tween hu­mans and machines

Rohin Shah8 Apr 2020 17:10 UTC
11 points
0 comments7 min readLW link
(mailchi.mp)

[AN #95]: A frame­work for think­ing about how to make AI go well

Rohin Shah15 Apr 2020 17:10 UTC
20 points
2 comments10 min readLW link
(mailchi.mp)

[AN #96]: Buck and I dis­cuss/​ar­gue about AI Alignment

Rohin Shah22 Apr 2020 17:20 UTC
17 points
4 comments10 min readLW link
(mailchi.mp)

[AN #97]: Are there his­tor­i­cal ex­am­ples of large, ro­bust dis­con­ti­nu­ities?

Rohin Shah29 Apr 2020 17:30 UTC
15 points
0 comments10 min readLW link
(mailchi.mp)

[AN #98]: Un­der­stand­ing neu­ral net train­ing by see­ing which gra­di­ents were helpful

Rohin Shah6 May 2020 17:10 UTC
22 points
3 comments9 min readLW link
(mailchi.mp)

[AN #99]: Dou­bling times for the effi­ciency of AI algorithms

Rohin Shah13 May 2020 17:20 UTC
29 points
0 comments10 min readLW link
(mailchi.mp)

[AN #100]: What might go wrong if you learn a re­ward func­tion while acting

Rohin Shah20 May 2020 17:30 UTC
33 points
2 comments12 min readLW link
(mailchi.mp)

[AN #101]: Why we should rigor­ously mea­sure and fore­cast AI progress

Rohin Shah27 May 2020 17:20 UTC
15 points
0 comments10 min readLW link
(mailchi.mp)

[AN #102]: Meta learn­ing by GPT-3, and a list of full pro­pos­als for AI alignment

Rohin Shah3 Jun 2020 17:20 UTC
38 points
6 comments10 min readLW link
(mailchi.mp)

[AN #103]: ARCHES: an agenda for ex­is­ten­tial safety, and com­bin­ing nat­u­ral lan­guage with deep RL

Rohin Shah10 Jun 2020 17:20 UTC
29 points
0 comments10 min readLW link
(mailchi.mp)

[AN #104]: The per­ils of in­ac­cessible in­for­ma­tion, and what we can learn about AI al­ign­ment from COVID

Rohin Shah18 Jun 2020 17:10 UTC
19 points
5 comments8 min readLW link
(mailchi.mp)

[AN #105]: The eco­nomic tra­jec­tory of hu­man­ity, and what we might mean by optimization

Rohin Shah24 Jun 2020 17:30 UTC
24 points
3 comments11 min readLW link
(mailchi.mp)

[AN #106]: Eval­u­at­ing gen­er­al­iza­tion abil­ity of learned re­ward models

Rohin Shah1 Jul 2020 17:20 UTC
14 points
2 comments11 min readLW link
(mailchi.mp)

[AN #107]: The con­ver­gent in­stru­men­tal sub­goals of goal-di­rected agents

Rohin Shah16 Jul 2020 6:47 UTC
13 points
1 comment8 min readLW link
(mailchi.mp)

[AN #108]: Why we should scru­ti­nize ar­gu­ments for AI risk

Rohin Shah16 Jul 2020 6:47 UTC
19 points
6 comments12 min readLW link
(mailchi.mp)

[AN #109]: Teach­ing neu­ral nets to gen­er­al­ize the way hu­mans would

Rohin Shah22 Jul 2020 17:10 UTC
17 points
3 comments9 min readLW link
(mailchi.mp)

[AN #110]: Learn­ing fea­tures from hu­man feed­back to en­able re­ward learning

Rohin Shah29 Jul 2020 17:20 UTC
13 points
2 comments10 min readLW link
(mailchi.mp)

[AN #111]: The Cir­cuits hy­pothe­ses for deep learning

Rohin Shah5 Aug 2020 17:40 UTC
23 points
0 comments9 min readLW link
(mailchi.mp)

[AN #112]: Eng­ineer­ing a Safer World

Rohin Shah13 Aug 2020 17:20 UTC
25 points
2 comments12 min readLW link
(mailchi.mp)

[AN #113]: Check­ing the eth­i­cal in­tu­itions of large lan­guage models

Rohin Shah19 Aug 2020 17:10 UTC
23 points
0 comments9 min readLW link
(mailchi.mp)

[AN #114]: The­ory-in­spired safety solu­tions for pow­er­ful Bayesian RL agents

Rohin Shah26 Aug 2020 17:20 UTC
21 points
3 comments8 min readLW link
(mailchi.mp)

[AN #115]: AI safety re­search prob­lems in the AI-GA framework

Rohin Shah2 Sep 2020 17:10 UTC
19 points
16 comments6 min readLW link
(mailchi.mp)

[AN #116]: How to make ex­pla­na­tions of neu­rons compositional

Rohin Shah9 Sep 2020 17:20 UTC
21 points
2 comments9 min readLW link
(mailchi.mp)

[AN #117]: How neu­ral nets would fare un­der the TEVV framework

Rohin Shah16 Sep 2020 17:20 UTC
27 points
0 comments7 min readLW link
(mailchi.mp)

[AN #118]: Risks, solu­tions, and pri­ori­ti­za­tion in a world with many AI systems

Rohin Shah23 Sep 2020 18:20 UTC
15 points
6 comments10 min readLW link
(mailchi.mp)

[AN #119]: AI safety when agents are shaped by en­vi­ron­ments, not rewards

Rohin Shah30 Sep 2020 17:10 UTC
11 points
0 comments11 min readLW link
(mailchi.mp)

[AN #120]: Trac­ing the in­tel­lec­tual roots of AI and AI alignment

Rohin Shah7 Oct 2020 17:10 UTC
13 points
4 comments10 min readLW link
(mailchi.mp)

[AN #121]: Fore­cast­ing trans­for­ma­tive AI timelines us­ing biolog­i­cal anchors

Rohin Shah14 Oct 2020 17:20 UTC
28 points
5 comments14 min readLW link
(mailchi.mp)

[AN #122]: Ar­gu­ing for AGI-driven ex­is­ten­tial risk from first principles

Rohin Shah21 Oct 2020 17:10 UTC
28 points
0 comments9 min readLW link
(mailchi.mp)

[AN #123]: In­fer­ring what is valuable in or­der to al­ign recom­mender systems

Rohin Shah28 Oct 2020 17:00 UTC
20 points
1 comment8 min readLW link
(mailchi.mp)

[AN #124]: Prov­ably safe ex­plo­ra­tion through shielding

Rohin Shah4 Nov 2020 18:20 UTC
13 points
0 comments9 min readLW link
(mailchi.mp)

[AN #125]: Neu­ral net­work scal­ing laws across mul­ti­ple modalities

Rohin Shah11 Nov 2020 18:20 UTC
25 points
7 comments9 min readLW link
(mailchi.mp)

[AN #126]: Avoid­ing wire­head­ing by de­cou­pling ac­tion feed­back from ac­tion effects

Rohin Shah26 Nov 2020 23:20 UTC
24 points
1 comment10 min readLW link
(mailchi.mp)

[AN #127]: Re­think­ing agency: Carte­sian frames as a for­mal­iza­tion of ways to carve up the world into an agent and its environment

Rohin Shah2 Dec 2020 18:20 UTC
53 points
0 comments13 min readLW link
(mailchi.mp)

[AN #128]: Pri­ori­tiz­ing re­search on AI ex­is­ten­tial safety based on its ap­pli­ca­tion to gov­er­nance demands

Rohin Shah9 Dec 2020 18:20 UTC
16 points
2 comments10 min readLW link
(mailchi.mp)

[AN #129]: Ex­plain­ing dou­ble de­scent by mea­sur­ing bias and variance

Rohin Shah16 Dec 2020 18:10 UTC
14 points
1 comment7 min readLW link
(mailchi.mp)

[AN #130]: A new AI x-risk pod­cast, and re­views of the field

Rohin Shah24 Dec 2020 18:20 UTC
8 points
0 comments7 min readLW link
(mailchi.mp)

[AN #131]: For­mal­iz­ing the ar­gu­ment of ig­nored at­tributes in a util­ity function

Rohin Shah31 Dec 2020 18:20 UTC
13 points
4 comments9 min readLW link
(mailchi.mp)

[AN #132]: Com­plex and sub­tly in­cor­rect ar­gu­ments as an ob­sta­cle to debate

Rohin Shah6 Jan 2021 18:20 UTC
19 points
1 comment19 min readLW link
(mailchi.mp)

[AN #133]: Build­ing ma­chines that can co­op­er­ate (with hu­mans, in­sti­tu­tions, or other ma­chines)

Rohin Shah13 Jan 2021 18:10 UTC
14 points
0 comments9 min readLW link
(mailchi.mp)

[AN #134]: Un­der­speci­fi­ca­tion as a cause of frag­ility to dis­tri­bu­tion shift

Rohin Shah21 Jan 2021 18:10 UTC
13 points
0 comments7 min readLW link
(mailchi.mp)

[AN #135]: Five prop­er­ties of goal-di­rected systems

Rohin Shah27 Jan 2021 18:10 UTC
33 points
0 comments8 min readLW link
(mailchi.mp)

[AN #136]: How well will GPT-N perform on down­stream tasks?

Rohin Shah3 Feb 2021 18:10 UTC
21 points
2 comments9 min readLW link
(mailchi.mp)

[AN #137]: Quan­tify­ing the benefits of pre­train­ing on down­stream task performance

Rohin Shah10 Feb 2021 18:10 UTC
18 points
0 comments8 min readLW link
(mailchi.mp)

[AN #138]: Why AI gov­er­nance should find prob­lems rather than just solv­ing them

Rohin Shah17 Feb 2021 18:50 UTC
12 points
0 comments9 min readLW link
(mailchi.mp)

[AN #139]: How the sim­plic­ity of re­al­ity ex­plains the suc­cess of neu­ral nets

Rohin Shah24 Feb 2021 18:30 UTC
26 points
6 comments12 min readLW link
(mailchi.mp)

[AN #140]: The­o­ret­i­cal mod­els that pre­dict scal­ing laws

Rohin Shah4 Mar 2021 18:10 UTC
45 points
0 comments10 min readLW link
(mailchi.mp)

[AN #141]: The case for prac­tic­ing al­ign­ment work on GPT-3 and other large models

Rohin Shah10 Mar 2021 18:30 UTC
27 points
4 comments8 min readLW link
(mailchi.mp)

[AN #142]: The quest to un­der­stand a net­work well enough to reim­ple­ment it by hand

Rohin Shah17 Mar 2021 17:10 UTC
36 points
4 comments8 min readLW link
(mailchi.mp)

[AN #143]: How to make em­bed­ded agents that rea­son prob­a­bil­is­ti­cally about their environments

Rohin Shah24 Mar 2021 17:20 UTC
13 points
3 comments8 min readLW link
(mailchi.mp)

[AN #144]: How lan­guage mod­els can also be fine­tuned for non-lan­guage tasks

Rohin Shah2 Apr 2021 17:20 UTC
19 points
0 comments6 min readLW link
(mailchi.mp)

Align­ment Newslet­ter Three Year Retrospective

Rohin Shah7 Apr 2021 14:39 UTC
55 points
0 comments5 min readLW link

[AN #145]: Our three year an­niver­sary!

Rohin Shah9 Apr 2021 17:48 UTC
19 points
0 comments8 min readLW link
(mailchi.mp)

[AN #146]: Plau­si­ble sto­ries of how we might fail to avert an ex­is­ten­tial catastrophe

Rohin Shah14 Apr 2021 17:30 UTC
23 points
1 comment8 min readLW link
(mailchi.mp)

[AN #147]: An overview of the in­ter­pretabil­ity landscape

Rohin Shah21 Apr 2021 17:10 UTC
14 points
2 comments7 min readLW link
(mailchi.mp)

[AN #148]: An­a­lyz­ing gen­er­al­iza­tion across more axes than just ac­cu­racy or loss

Rohin Shah28 Apr 2021 18:30 UTC
24 points
5 comments11 min readLW link
(mailchi.mp)

[AN #149]: The newslet­ter’s ed­i­to­rial policy

Rohin Shah5 May 2021 17:10 UTC
19 points
3 comments8 min readLW link
(mailchi.mp)

[AN #150]: The sub­types of Co­op­er­a­tive AI research

Rohin Shah12 May 2021 17:20 UTC
17 points
0 comments6 min readLW link
(mailchi.mp)

[AN #151]: How spar­sity in the fi­nal layer makes a neu­ral net debuggable

Rohin Shah19 May 2021 17:20 UTC
19 points
2 comments6 min readLW link
(mailchi.mp)

[AN #152]: How we’ve over­es­ti­mated few-shot learn­ing capabilities

Rohin Shah16 Jun 2021 17:20 UTC
22 points
6 comments8 min readLW link
(mailchi.mp)

[AN #153]: Ex­per­i­ments that demon­strate failures of ob­jec­tive robustness

Rohin Shah26 Jun 2021 17:10 UTC
25 points
1 comment8 min readLW link
(mailchi.mp)

[AN #154]: What eco­nomic growth the­ory has to say about trans­for­ma­tive AI

Rohin Shah30 Jun 2021 17:20 UTC
12 points
0 comments9 min readLW link
(mailchi.mp)

[AN #155]: A Minecraft bench­mark for al­gorithms that learn with­out re­ward functions

Rohin Shah8 Jul 2021 17:20 UTC
21 points
5 comments7 min readLW link
(mailchi.mp)

[AN #156]: The scal­ing hy­poth­e­sis: a plan for build­ing AGI

Rohin Shah16 Jul 2021 17:10 UTC
46 points
20 comments8 min readLW link
(mailchi.mp)

[AN #157]: Mea­sur­ing mis­al­ign­ment in the tech­nol­ogy un­der­ly­ing Copilot

Rohin Shah23 Jul 2021 17:20 UTC
28 points
18 comments7 min readLW link
(mailchi.mp)

[AN #158]: Should we be op­ti­mistic about gen­er­al­iza­tion?

Rohin Shah29 Jul 2021 17:20 UTC
20 points
0 comments8 min readLW link
(mailchi.mp)

[AN #159]: Build­ing agents that know how to ex­per­i­ment, by train­ing on pro­ce­du­rally gen­er­ated games

Rohin Shah4 Aug 2021 17:10 UTC
18 points
4 comments14 min readLW link
(mailchi.mp)

[AN #160]: Build­ing AIs that learn and think like people

Rohin Shah13 Aug 2021 17:10 UTC
28 points
6 comments10 min readLW link
(mailchi.mp)

[AN #161]: Creat­ing gen­er­al­iz­able re­ward func­tions for mul­ti­ple tasks by learn­ing a model of func­tional similarity

Rohin Shah20 Aug 2021 17:20 UTC
15 points
0 comments9 min readLW link
(mailchi.mp)

[AN #162]: Foun­da­tion mod­els: a paradigm shift within AI

Rohin Shah27 Aug 2021 17:20 UTC
25 points
0 comments8 min readLW link
(mailchi.mp)

[AN #163]: Us­ing finite fac­tored sets for causal and tem­po­ral inference

Rohin Shah8 Sep 2021 17:20 UTC
41 points
0 comments10 min readLW link
(mailchi.mp)

[AN #164]: How well can lan­guage mod­els write code?

Rohin Shah15 Sep 2021 17:20 UTC
13 points
7 comments9 min readLW link
(mailchi.mp)

[AN #165]: When large mod­els are more likely to lie

Rohin Shah22 Sep 2021 17:30 UTC
23 points
0 comments8 min readLW link
(mailchi.mp)

[AN #166]: Is it crazy to claim we’re in the most im­por­tant cen­tury?

Rohin Shah8 Oct 2021 17:30 UTC
52 points
5 comments8 min readLW link
(mailchi.mp)

[AN #167]: Con­crete ML safety prob­lems and their rele­vance to x-risk

Rohin Shah20 Oct 2021 17:10 UTC
19 points
4 comments9 min readLW link
(mailchi.mp)

[AN #168]: Four tech­ni­cal top­ics for which Open Phil is so­lic­it­ing grant proposals

Rohin Shah28 Oct 2021 17:20 UTC
15 points
0 comments9 min readLW link
(mailchi.mp)

[AN #169]: Col­lab­o­rat­ing with hu­mans with­out hu­man data

Rohin Shah24 Nov 2021 18:30 UTC
33 points
0 comments8 min readLW link
(mailchi.mp)

[AN #170]: An­a­lyz­ing the ar­gu­ment for risk from power-seek­ing AI

Rohin Shah8 Dec 2021 18:10 UTC
21 points
1 comment7 min readLW link
(mailchi.mp)

[AN #171]: Disagree­ments be­tween al­ign­ment “op­ti­mists” and “pes­simists”

Rohin Shah21 Jan 2022 18:30 UTC
32 points
1 comment7 min readLW link
(mailchi.mp)

[AN #172] Sorry for the long hi­a­tus!

Rohin Shah5 Jul 2022 6:20 UTC
54 points
0 comments3 min readLW link
(mailchi.mp)

[AN #173] Re­cent lan­guage model re­sults from DeepMind

Rohin Shah21 Jul 2022 2:30 UTC
37 points
9 comments8 min readLW link
(mailchi.mp)