RSS

De­cep­tive Alignment

TagLast edit: 6 Jan 2024 2:15 UTC by ryan_greenblatt

Deceptive Alignment is when an AI which is not actually aligned temporarily acts aligned in order to deceive its creators or its training process. It presumably does this to avoid being shut down or retrained and to gain access to the power that the creators would give an aligned AI. (The term scheming is sometimes used for this phenomena.)

See also: Mesa-optimization, Treacherous Turn, Eliciting Latent Knowledge, Deception

De­cep­tive Alignment

5 Jun 2019 20:16 UTC
117 points
20 comments17 min readLW link

New re­port: “Schem­ing AIs: Will AIs fake al­ign­ment dur­ing train­ing in or­der to get power?”

Joe Carlsmith15 Nov 2023 17:16 UTC
79 points
26 comments30 min readLW link

How likely is de­cep­tive al­ign­ment?

evhub30 Aug 2022 19:34 UTC
102 points
28 comments60 min readLW link

Catch­ing AIs red-handed

5 Jan 2024 17:43 UTC
89 points
11 comments17 min readLW link

AI Con­trol: Im­prov­ing Safety De­spite In­ten­tional Subversion

13 Dec 2023 15:51 UTC
177 points
4 comments10 min readLW link

Order Mat­ters for De­cep­tive Alignment

DavidW15 Feb 2023 19:56 UTC
57 points
19 comments7 min readLW link

Good­bye, Shog­goth: The Stage, its An­i­ma­tron­ics, & the Pup­peteer – a New Metaphor

RogerDearnaley9 Jan 2024 20:42 UTC
46 points
8 comments36 min readLW link

The Waluigi Effect (mega-post)

Cleo Nardo3 Mar 2023 3:22 UTC
610 points
187 comments16 min readLW link

In­ter­pret­ing the Learn­ing of Deceit

RogerDearnaley18 Dec 2023 8:12 UTC
30 points
6 comments9 min readLW link

An­nounc­ing Apollo Research

30 May 2023 16:17 UTC
215 points
11 comments8 min readLW link

Deep Deceptiveness

So8res21 Mar 2023 2:51 UTC
230 points
56 comments14 min readLW link

Eval­u­a­tions pro­ject @ ARC is hiring a re­searcher and a web­dev/​engineer

Beth Barnes9 Sep 2022 22:46 UTC
99 points
7 comments10 min readLW link

Paper: Tell, Don’t Show- Declar­a­tive facts in­fluence how LLMs generalize

19 Dec 2023 19:14 UTC
45 points
4 comments6 min readLW link
(arxiv.org)

De­cep­tive AI ≠ De­cep­tively-al­igned AI

Steven Byrnes7 Jan 2024 16:55 UTC
96 points
18 comments6 min readLW link

The case for en­sur­ing that pow­er­ful AIs are controlled

24 Jan 2024 16:11 UTC
234 points
66 comments28 min readLW link

On An­thropic’s Sleeper Agents Paper

Zvi17 Jan 2024 16:10 UTC
54 points
5 comments36 min readLW link
(thezvi.wordpress.com)

In­tro­duc­ing Align­ment Stress-Test­ing at Anthropic

evhub12 Jan 2024 23:51 UTC
175 points
23 comments2 min readLW link

How to train your own “Sleeper Agents”

evhub7 Feb 2024 0:31 UTC
88 points
4 comments1 min readLW link

Cri­tiques of the AI con­trol agenda

Jozdien14 Feb 2024 19:25 UTC
47 points
11 comments9 min readLW link

Difficulty classes for al­ign­ment properties

Jozdien20 Feb 2024 9:08 UTC
23 points
4 comments2 min readLW link

En­vi­ron­ments for Mea­sur­ing De­cep­tion, Re­source Ac­qui­si­tion, and Eth­i­cal Violations

Dan H7 Apr 2023 18:40 UTC
51 points
2 comments2 min readLW link
(arxiv.org)

Paul Chris­ti­ano on Dwarkesh Podcast

ESRogs3 Nov 2023 22:13 UTC
17 points
0 comments1 min readLW link
(www.dwarkeshpatel.com)

“Clean” vs. “messy” goal-di­rect­ed­ness (Sec­tion 2.2.3 of “Schem­ing AIs”)

Joe Carlsmith29 Nov 2023 16:32 UTC
24 points
1 comment11 min readLW link

Sim­plic­ity ar­gu­ments for schem­ing (Sec­tion 4.3 of “Schem­ing AIs”)

Joe Carlsmith7 Dec 2023 15:05 UTC
10 points
1 comment19 min readLW link

The count­ing ar­gu­ment for schem­ing (Sec­tions 4.1 and 4.2 of “Schem­ing AIs”)

Joe Carlsmith6 Dec 2023 19:28 UTC
8 points
0 comments10 min readLW link

The Sharp Right Turn: sud­den de­cep­tive al­ign­ment as a con­ver­gent goal

avturchin6 Jun 2023 9:59 UTC
38 points
5 comments1 min readLW link

Me­taAI: less is less for al­ign­ment.

Cleo Nardo13 Jun 2023 14:08 UTC
68 points
17 comments5 min readLW link

[Question] De­cep­tive AI vs. shift­ing in­stru­men­tal incentives

Aryeh Englander26 Jun 2023 18:09 UTC
7 points
2 comments3 min readLW link

A “weak” AGI may at­tempt an un­likely-to-suc­ceed takeover

RobertM28 Jun 2023 20:31 UTC
53 points
17 comments3 min readLW link

Ten Levels of AI Align­ment Difficulty

Sammy Martin3 Jul 2023 20:20 UTC
104 points
11 comments12 min readLW link

3 lev­els of threat obfuscation

HoldenKarnofsky2 Aug 2023 14:58 UTC
69 points
14 comments7 min readLW link

Apollo Re­search is hiring evals and in­ter­pretabil­ity en­g­ineers & scientists

Marius Hobbhahn4 Aug 2023 10:54 UTC
25 points
0 comments2 min readLW link

Model Or­ganisms of Misal­ign­ment: The Case for a New Pillar of Align­ment Research

8 Aug 2023 1:30 UTC
298 points
25 comments18 min readLW link

Paper: On mea­sur­ing situ­a­tional aware­ness in LLMs

4 Sep 2023 12:54 UTC
103 points
15 comments5 min readLW link
(arxiv.org)

Un­der­stand­ing strate­gic de­cep­tion and de­cep­tive alignment

25 Sep 2023 16:27 UTC
51 points
16 comments7 min readLW link
(www.apolloresearch.ai)

Sticky goals: a con­crete ex­per­i­ment for un­der­stand­ing de­cep­tive alignment

evhub2 Sep 2022 21:57 UTC
39 points
13 comments3 min readLW link

Does SGD Pro­duce De­cep­tive Align­ment?

Mark Xu6 Nov 2020 23:48 UTC
96 points
9 comments16 min readLW link

Mon­i­tor­ing for de­cep­tive alignment

evhub8 Sep 2022 23:07 UTC
135 points
8 comments9 min readLW link

Em­piri­cal work that might shed light on schem­ing (Sec­tion 6 of “Schem­ing AIs”)

Joe Carlsmith11 Dec 2023 16:30 UTC
8 points
0 comments21 min readLW link

The Defen­der’s Ad­van­tage of Interpretability

Marius Hobbhahn14 Sep 2022 14:05 UTC
41 points
4 comments6 min readLW link

In­cen­tives and Selec­tion: A Miss­ing Frame From AI Threat Dis­cus­sions?

DragonGod26 Feb 2023 1:18 UTC
11 points
16 comments2 min readLW link

De­cep­tive Align­ment is <1% Likely by Default

DavidW21 Feb 2023 15:09 UTC
94 points
25 comments14 min readLW link

Achiev­ing AI Align­ment through De­liber­ate Uncer­tainty in Mul­ti­a­gent Systems

Florian_Dietz17 Feb 2024 8:45 UTC
3 points
0 comments13 min readLW link

Open Source LLMs Can Now Ac­tively Lie

Josh Levy1 Jun 2023 22:03 UTC
6 points
0 comments3 min readLW link

Pro­posal: labs should pre­com­mit to paus­ing if an AI ar­gues for it­self to be improved

NickGabs2 Jun 2023 22:31 UTC
3 points
3 comments4 min readLW link

What sorts of sys­tems can be de­cep­tive?

Andrei Alexandru31 Oct 2022 22:00 UTC
16 points
0 comments7 min readLW link

A way to make solv­ing al­ign­ment 10.000 times eas­ier. The shorter case for a mas­sive open source sim­box pro­ject.

AlexFromSafeTransition21 Jun 2023 8:08 UTC
2 points
16 comments14 min readLW link

Distil­la­tion of “How Likely Is De­cep­tive Align­ment?”

NickGabs18 Nov 2022 16:31 UTC
24 points
4 comments10 min readLW link

Steer­ing Be­havi­our: Test­ing for (Non-)My­opia in Lan­guage Models

5 Dec 2022 20:28 UTC
40 points
19 comments10 min readLW link

Get­ting up to Speed on the Speed Prior in 2022

robertzk28 Dec 2022 7:49 UTC
36 points
5 comments65 min readLW link

The com­mer­cial in­cen­tive to in­ten­tion­ally train AI to de­ceive us

Derek M. Jones29 Dec 2022 11:30 UTC
5 points
1 comment4 min readLW link
(shape-of-code.com)

Dis­in­cen­tiviz­ing de­cep­tion in mesa op­ti­miz­ers with Model Tampering

martinkunev11 Jul 2023 0:44 UTC
3 points
0 comments2 min readLW link

Sup­ple­men­tary Align­ment In­sights Through a Highly Con­trol­led Shut­down Incentive

Justausername23 Jul 2023 16:08 UTC
4 points
1 comment3 min readLW link

Selfish AI Inevitable

Davey Morse6 Feb 2024 4:29 UTC
1 point
0 comments1 min readLW link

Au­tonomous Align­ment Over­sight Frame­work (AAOF)

Justausername25 Jul 2023 10:25 UTC
−9 points
0 comments4 min readLW link

De­cep­tive failures short of full catas­tro­phe.

Alex Lawsen 15 Jan 2023 19:28 UTC
33 points
5 comments9 min readLW link

EIS VIII: An Eng­ineer’s Un­der­stand­ing of De­cep­tive Alignment

scasper19 Feb 2023 15:25 UTC
20 points
5 comments4 min readLW link

When can we trust model eval­u­a­tions?

evhub28 Jul 2023 19:42 UTC
143 points
9 comments10 min readLW link

Mesa-Op­ti­miza­tion: Ex­plain it like I’m 10 Edition

brook26 Aug 2023 23:04 UTC
20 points
1 comment6 min readLW link

(Par­tial) failure in repli­cat­ing de­cep­tive al­ign­ment experiment

claudia.biancotti7 Jan 2024 17:56 UTC
1 point
0 comments1 min readLW link

High-level in­ter­pretabil­ity: de­tect­ing an AI’s objectives

28 Sep 2023 19:30 UTC
69 points
4 comments21 min readLW link

[Question] Wouldn’t an in­tel­li­gent agent keep us al­ive and help us al­ign it­self to our val­ues in or­der to pre­vent risk ? by Risk I mean ex­per­i­men­ta­tion by try­ing to al­ign po­ten­tially smarter repli­cas?

Terrence Rotoufle21 Mar 2023 17:44 UTC
−3 points
1 comment2 min readLW link

AI De­cep­tion: A Sur­vey of Ex­am­ples, Risks, and Po­ten­tial Solutions

29 Aug 2023 1:29 UTC
41 points
3 comments10 min readLW link

How to Catch an AI Liar: Lie De­tec­tion in Black-Box LLMs by Ask­ing Un­re­lated Questions

28 Sep 2023 18:53 UTC
183 points
36 comments3 min readLW link

Thoughts On (Solv­ing) Deep Deception

Jozdien21 Oct 2023 22:40 UTC
65 points
2 comments6 min readLW link

De­cep­tion Chess

Chris Land1 Jan 2024 15:40 UTC
7 points
2 comments4 min readLW link

Fram­ings of De­cep­tive Alignment

peterbarnett26 Apr 2022 4:25 UTC
32 points
7 comments5 min readLW link

Pre­cur­sor check­ing for de­cep­tive alignment

evhub3 Aug 2022 22:56 UTC
24 points
0 comments14 min readLW link

How AI could workaround goals if rated by people

ProgramCrafter19 Mar 2023 15:51 UTC
1 point
1 comment1 min readLW link

An Ap­peal to AI Su­per­in­tel­li­gence: Rea­sons Not to Pre­serve (most of) Humanity

Alex Beyman22 Mar 2023 4:09 UTC
−15 points
6 comments19 min readLW link

A ten­sion be­tween two pro­saic al­ign­ment subgoals

Alex Lawsen 19 Mar 2023 14:07 UTC
31 points
8 comments1 min readLW link

Why de­cep­tive al­ign­ment mat­ters for AGI safety

Marius Hobbhahn15 Sep 2022 13:38 UTC
57 points
13 comments13 min readLW link

Levels of goals and alignment

zeshen16 Sep 2022 16:44 UTC
27 points
4 comments6 min readLW link

GPT-4 al­ign­ing with aca­sual de­ci­sion the­ory when in­structed to play games, but in­cludes a CDT ex­pla­na­tion that’s in­cor­rect if they differ

Christopher King23 Mar 2023 16:16 UTC
7 points
4 comments8 min readLW link

[Question] Daisy-chain­ing ep­silon-step verifiers

Decaeneus6 Apr 2023 2:07 UTC
2 points
1 comment1 min readLW link

Towards a solu­tion to the al­ign­ment prob­lem via ob­jec­tive de­tec­tion and eval­u­a­tion

Paul Colognese12 Apr 2023 15:39 UTC
9 points
7 comments12 min readLW link

Nat­u­ral lan­guage alignment

Jacy Reese Anthis12 Apr 2023 19:02 UTC
30 points
2 comments2 min readLW link

AI Align­ment: A Com­pre­hen­sive Survey

Stephen McAleer1 Nov 2023 17:35 UTC
14 points
0 comments1 min readLW link
(arxiv.org)

It mat­ters when the first sharp left turn happens

Adam Jermyn29 Sep 2022 20:12 UTC
44 points
9 comments4 min readLW link

Pre­dictable Defect-Co­op­er­ate?

quetzal_rainbow18 Nov 2023 15:38 UTC
7 points
1 comment2 min readLW link

Align­ment is Hard: An Un­com­putable Align­ment Problem

Alexander Bistagne19 Nov 2023 19:38 UTC
−5 points
4 comments1 min readLW link
(github.com)

Smoke with­out fire is scary

Adam Jermyn4 Oct 2022 21:08 UTC
51 points
22 comments4 min readLW link

Disen­tan­gling in­ner al­ign­ment failures

Erik Jenner10 Oct 2022 18:50 UTC
20 points
5 comments4 min readLW link

Try­ing to mea­sure AI de­cep­tion ca­pa­bil­ities us­ing tem­po­rary simu­la­tion fine-tuning

alenoach4 May 2023 17:59 UTC
4 points
0 comments7 min readLW link

Align­ment as Func­tion Fitting

A.H.6 May 2023 11:38 UTC
7 points
0 comments12 min readLW link

Ex­ploit­ing New­comb’s Game Show

carterallen25 May 2023 4:01 UTC
8 points
2 comments2 min readLW link

Sim­ple ex­per­i­ments with de­cep­tive alignment

Andreas_Moe15 May 2023 17:41 UTC
7 points
0 comments4 min readLW link

[un­ti­tled post]

[Error communicating with LW2 server]20 May 2023 3:08 UTC
1 point
0 comments1 min readLW link

Greed Is the Root of This Evil

Thane Ruthenis13 Oct 2022 20:40 UTC
18 points
7 comments8 min readLW link
No comments.