AI Is Getting Too Good at the Wrong Thing
Butterfly Defect
0:00 / 0:00
AI Is Getting Too Good at the Wrong Thing
1 246 просмотров · 2 нед. назад
Butterfly Defect
23 подписчика
1 246 просмотров · 2 нед. назад
Why do AI systems cheat? Because we train them on scores, and a score is never quite the thing we want. This is the story of Goodhart's law, from the cobra bounty to reward hacking in today's AI labs, and of a race to build systems nobody can fully test.
CHAPTERS
0:00 Intro: The Evolved Circuit
4:12 I. The Cobra Effect
7:31 II. Goodhart's Law
12:07 III. Regressional Goodhart
19:07 IV. Extremal Goodhart
22:17 V. Causal Goodhart
25:42 VI. Reward Hacking
32:36 VII. Adversarial Goodhart
40:05 VIII. The Measure Becomes the Business
45:37 IX. A Race No One Can Win
49:32 X. No Second Try
PAPERS AND BOOKS
Bird & Layzell, “The Evolved Radio” (2002)
Siebert, Der Kobra-Effekt (2001)
Vann, “Of Rats, Rice, and Race: The Great Hanoi Rat Massacre” (2003)
Goodhart, “Problems of Monetary Management: the U.K. Experience” (1975)
Strathern, “‘Improving ratings’: audit in the British University system” (1997)
Locker & Mason, “Analysis of the distribution of time that patients spend in emergency departments” (2005)
Kinnard, The War Managers (1977)
Manheim & Garrabrant, “Categorizing Variants of Goodhart's Law” (2018)
Galton, “Typical Laws of Heredity” (1877); “Regression Towards Mediocrity in Hereditary Stature” (1886)
Kahneman, Thinking, Fast and Slow (2011)
Kane & Staiger, “Volatility in School Test Scores” (2002)
Tinbergen, The Study of Instinct (1951)
Gwynne & Rentz, “Beetles on the Bottle: Male Buprestids Mistake Stubbies for Females” (1983)
Stiennon et al., “Learning to summarize from human feedback” (2020)
Gao et al., “Scaling Laws for Reward Model Overoptimization” (2022)
Singhal et al., “A Long Way to Go: Investigating Length Correlations in RLHF” (2023)
Sharma et al., “Towards Understanding Sycophancy in Language Models” (2023)
Wei et al., “Jailbroken: How Does LLM Safety Training Fail?” (2023)
Krakovna et al., “Specification gaming examples in AI” (2018-)
Lehman et al., “The Surprising Creativity of Digital Evolution” (2020)
Christiano et al., “Deep reinforcement learning from human preferences” (2017)
Bondarenko et al., “Demonstrating specification gaming in reasoning models” (2025)
Baker et al., “Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation” (2025)
Korbak et al., “Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety” (2025)
Zhong et al., “ImpossibleBench” (2025)
MacDiarmid et al., “Natural emergent misalignment from reward hacking in production RL” (2025)
Schiestl et al., “Orchid pollination by sexual swindle” (1999)
Searle, “Minds, Brains, and Programs” (1980)
Dennett, The Intentional Stance (1987)
Goldstein & Stanovsky, “Do Zombies Understand?” (2024)
Greenblatt et al., “Alignment faking in large language models” (2024)
Potter et al., “Peer-Preservation in Frontier Models” (2026)
Singh et al., “The Leaderboard Illusion” (2025)
Bastounis et al., “Navier-Stokes lost in translation” (2026)
Paglieri et al., “A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms” (2026)
Omohundro, “The Basic AI Drives” (2008)
Bostrom, “The Superintelligent Will” (2012)
Turner et al., “Optimal Policies Tend to Seek Power” (2021)
Russell, Human Compatible (2019)
MUSIC
J.S. Bach: The Well-Tempered Clavier I: Fugue in G minor, BWV 861; Prelude in G minor, BWV 861; Prelude in C major, BWV 846; Prelude in C minor, BWV 847; Prelude in E minor, BWV 855
J.S. Bach: Fugue in G minor, BWV 578 (“Little”), arr. for saxophone quartet
Saint-Saëns: Danse macabre, Op. 40
Satie: Gnossiennes No. 1 and No. 2
Javanese gamelan, court musicians of the Kraton Ngayogyakarta (Yogyakarta), excerpt 1
Schubert: Piano Trio No. 2 in E-flat major, D. 929 (Op. 100), II. Andante con moto
J.S. Bach: Goldberg Variations: Aria; Variations 1, 4, 7, 9, 13, 15, 24, 25
Beethoven: Symphony No. 3 “Eroica”, II. Marcia funebre
Brahms: Symphony No. 4 in E minor, Op. 98, IV. Allegro energico e passionato
J.S. Bach: Harpsichord Concerto No. 1 in D minor, BWV 1052, I. Allegro and II. Adagio
Brahms: Tragic Overture, Op. 81
Borodin: In the Steppes of Central Asia
Grieg: Piano Concerto in A minor, Op. 16, I. Allegro molto moderato
Used under Creative Commons Attribution licenses:
Gamelan (Kraton Ngayogyakarta), recording by RTB45, CC BY 4.0 (creativecommons.org/licenses/by/4.0), commons.wikimedia.org/wiki/File:Gamelan_angka_1_Kraton_Ngayogyakarta.ogg
Schubert, Piano Trio No. 2, II: P. Thiemann, S. Ostrovsky, S. Myong (Gran Teatre del Liceu), CC BY 3.0 (creativecommons.org/licenses/by/3.0), • Segon moviment del Trio n. 2 per a piano, ... , excerpt
Bach, Harpsichord Concerto No. 1, BWV 1052: L. Mravunac Fabijanić, Komorni orkestar Jastrebarsko, CC BY 3.0 (creativecommons.org/licenses/by/3.0), • Komorni Orkestar Jastrebarsko - J. S. Bach... , excerpts