From 61792f0e327373fb4490ec799f3b78f5b57d070a Mon Sep 17 00:00:00 2001 From: Tobias Gerstenberg Date: Wed, 6 Nov 2024 11:21:18 -0600 Subject: [PATCH] added wise AI paper --- content/publication/johnson2024wise.md | 33 ++ docs/404.html | 8 +- docs/bibtex/cic_papers.bib | 11 +- docs/home/index.xml | 4 +- docs/index.html | 2 +- docs/index.xml | 20 +- docs/member/tobias_gerstenberg/index.html | 39 ++ docs/publication/index.html | 64 +++ docs/publication/index.xml | 9 + docs/publication/johnson2024wise/index.html | 488 ++++++++++++++++++++ docs/publication_types/1/index.html | 11 +- docs/publication_types/1/index.xml | 11 +- docs/publication_types/index.html | 8 +- docs/publication_types/index.xml | 20 +- docs/sitemap.xml | 36 +- static/bibtex/cic_papers.bib | 11 +- 16 files changed, 726 insertions(+), 49 deletions(-) create mode 100644 content/publication/johnson2024wise.md create mode 100644 docs/publication/johnson2024wise/index.html diff --git a/content/publication/johnson2024wise.md b/content/publication/johnson2024wise.md new file mode 100644 index 0000000..8cc1875 --- /dev/null +++ b/content/publication/johnson2024wise.md @@ -0,0 +1,33 @@ ++++ +# 0 -> 'Forthcoming', +# 1 -> 'Preprint', +# 2 -> 'Journal', +# 3 -> 'Conference Proceedings', +# 4 -> 'Book chapter', +# 5 -> 'Thesis' + +title = "Imagining and building wise machines: The centrality of AI metacognition" +date = "2024-11-06" +authors = ["S. G. B. Johnson","A. Karimi","Y. Bengio","N. Chater","T. Gerstenberg","K. Larson","S. Levine","M. Mitchell","B. Schölkopf","I. Grossmann"] +publication_types = ["1"] +publication_short = "_arXiv_" +publication = "Johnson, S. G. B., Karimi, A., Bengio, Y., Chater, N., Gerstenberg, T., Larson, K., Levine, S., Mitchell, M., Schölkopf, B., Grossmann, I. (2024). Imagining and building wise machines: The centrality of AI metacognition. _arXiv_." +abstract = "Recent advances in artificial intelligence (AI) have produced systems capable of increasingly sophisticated performance on cognitive tasks. However, AI systems still struggle in critical ways: unpredictable and novel environments (robustness), lack transparency in their reasoning (explainability), face challenges in communication and commitment (cooperation), and pose risks due to potential harmful actions (safety). We argue that these shortcomings stem from one overarching failure: AI systems lack wisdom. Drawing from cognitive and social sciences, we define wisdom as the ability to navigate intractable problems---those that are ambiguous, radically uncertain, novel, chaotic, or computationally explosive---through effective task-level and metacognitive strategies. While AI research has focused on task-level strategies, metacognition---the ability to reflect on and regulate one's thought processes---is underdeveloped in AI systems. In humans, metacognitive strategies such as recognizing the limits of one's knowledge, considering diverse perspectives, and adapting to context are essential for wise decision-making. We propose that integrating metacognitive capabilities into AI systems is crucial for enhancing their robustness, explainability, cooperation, and safety. By focusing on developing wise AI, we suggest an alternative to aligning AI with specific human values---a task fraught with conceptual and practical difficulties. Instead, wise AI systems can thoughtfully navigate complex situations, account for diverse human values, and avoid harmful actions. We discuss potential approaches to building wise AI, including benchmarking metacognitive abilities and training AI systems to employ wise reasoning. Prioritizing metacognition in AI research will lead to systems that act not only intelligently but also wisely in complex, real-world situations." +image_preview = "" +selected = false +projects = [] +#url_pdf = "papers/johnson2024wise.pdf" +url_preprint = "https://arxiv.org/abs/2411.02478" +url_code = "" +url_dataset = "" +url_slides = "" +url_video = "" +url_poster = "" +url_source = "" +#url_custom = [{name = "Github", url = ""}] +math = true +highlight = true +[header] +# image = "publications/johnson2024wise.png" +caption = "" ++++ \ No newline at end of file diff --git a/docs/404.html b/docs/404.html index 76ef725..162ca1e 100644 --- a/docs/404.html +++ b/docs/404.html @@ -237,6 +237,10 @@

Page not found

Publications

+ + @@ -253,10 +257,6 @@

Publications

  • Causation, Meaning, and Communication
    • - - diff --git a/docs/bibtex/cic_papers.bib b/docs/bibtex/cic_papers.bib index a8c30af..4cbfcef 100644 --- a/docs/bibtex/cic_papers.bib +++ b/docs/bibtex/cic_papers.bib @@ -1,13 +1,22 @@ %% This BibTeX bibliography file was created using BibDesk. %% https://bibdesk.sourceforge.io/ -%% Created for Tobias Gerstenberg at 2024-10-26 12:42:26 -0700 +%% Created for Tobias Gerstenberg at 2024-11-06 11:16:38 -0600 %% Saved with string encoding Unicode (UTF-8) +@article{johnson2024wise, + abstract = {Recent advances in artificial intelligence (AI) have produced systems capable of increasingly sophisticated performance on cognitive tasks. However, AI systems still struggle in critical ways: unpredictable and novel environments (robustness), lack transparency in their reasoning (explainability), face challenges in communication and commitment (cooperation), and pose risks due to potential harmful actions (safety). We argue that these shortcomings stem from one overarching failure: AI systems lack wisdom. Drawing from cognitive and social sciences, we define wisdom as the ability to navigate intractable problems---those that are ambiguous, radically uncertain, novel, chaotic, or computationally explosive---through effective task-level and metacognitive strategies. While AI research has focused on task-level strategies, metacognition---the ability to reflect on and regulate one's thought processes---is underdeveloped in AI systems. In humans, metacognitive strategies such as recognizing the limits of one's knowledge, considering diverse perspectives, and adapting to context are essential for wise decision-making. We propose that integrating metacognitive capabilities into AI systems is crucial for enhancing their robustness, explainability, cooperation, and safety. By focusing on developing wise AI, we suggest an alternative to aligning AI with specific human values---a task fraught with conceptual and practical difficulties. Instead, wise AI systems can thoughtfully navigate complex situations, account for diverse human values, and avoid harmful actions. We discuss potential approaches to building wise AI, including benchmarking metacognitive abilities and training AI systems to employ wise reasoning. Prioritizing metacognition in AI research will lead to systems that act not only intelligently but also wisely in complex, real-world situations.}, + author = {Johnson, Samuel G B and Karimi, Amir-Hossein and Bengio, Yoshua and Chater, Nick and Gerstenberg, Tobias and Larson, Kate and Levine, Sydney and Mitchell, Melanie and Sch{\"o}lkopf, Bernhard and Grossmann, Igor}, + date-added = {2024-11-06 11:16:21 -0600}, + date-modified = {2024-11-06 11:16:21 -0600}, + journal = {arXiv}, + title = {{Imagining and building wise machines: The centrality of AI metacognition}}, + year = {2024}} + @article{jin2024marple, abstract = {Reconstructing past events requires reasoning across long time horizons. To figure out what happened, we need to use our prior knowledge about the world and human behavior and draw inferences from various sources of evidence including visual, language, and auditory cues. We introduce MARPLE, a benchmark for evaluating long-horizon inference capabilities using multi-modal evidence. Our benchmark features agents interacting with simulated households, supporting vision, language, and auditory stimuli, as well as procedurally generated environments and agent behaviors. Inspired by classic ``whodunit'' stories, we ask AI models and human participants to infer which agent caused a change in the environment based on a step-by-step replay of what actually happened. The goal is to correctly identify the culprit as early as possible. Our findings show that human participants outperform both traditional Monte Carlo simulation methods and an LLM baseline (GPT-4) on this task. Compared to humans, traditional inference models are less robust and performant, while GPT-4 has difficulty comprehending environmental changes. We analyze what factors influence inference performance and ablate different modes of evidence, finding that all modes are valuable for performance. Overall, our experiments demonstrate that the long-horizon, multimodal inference tasks in our benchmark present a challenge to current models. Project website: https: //marple-benchmark.github.io/.}, annote = {Comment: NeurIPS 2024. First two authors contributed equally. Project page: https://marple-benchmark.github.io/}, diff --git a/docs/home/index.xml b/docs/home/index.xml index da6ac1e..ebb37b5 100644 --- a/docs/home/index.xml +++ b/docs/home/index.xml @@ -15,7 +15,7 @@ Causality in Cognition Lab https://cicl.stanford.edu/home/home/ - Sun, 15 Oct 2017 00:00:00 -0700 + Sun, 15 Oct 2017 00:00:00 -0500 https://cicl.stanford.edu/home/home/ The Causality in Cognition Lab at Stanford University studies the role of causality in our understanding of the world and of each other. @@ -53,7 +53,7 @@ We are looking to hire a lab manager to start in the fall of 2018. For informati Selected<br>publications https://cicl.stanford.edu/home/publications_selected/ - Wed, 20 Apr 2016 00:00:00 -0700 + Wed, 20 Apr 2016 00:00:00 -0500 https://cicl.stanford.edu/home/publications_selected/ diff --git a/docs/index.html b/docs/index.html index ae1c58a..1ec9882 100644 --- a/docs/index.html +++ b/docs/index.html @@ -110,7 +110,7 @@ - + diff --git a/docs/index.xml b/docs/index.xml index 268e61b..b99a895 100644 --- a/docs/index.xml +++ b/docs/index.xml @@ -6,9 +6,18 @@ Hugo -- gohugo.io en-us &copy; 2024 Tobias Gerstenberg - Sat, 26 Oct 2024 00:00:00 +0000 + Wed, 06 Nov 2024 00:00:00 +0000 + + Imagining and building wise machines: The centrality of AI metacognition + https://cicl.stanford.edu/publication/johnson2024wise/ + Wed, 06 Nov 2024 00:00:00 +0000 + + https://cicl.stanford.edu/publication/johnson2024wise/ + + + From Artifacts to Human Lives: Investigating the Domain-Generality of Judgments about Purposes https://cicl.stanford.edu/publication/prinzing2024purpose/ @@ -135,14 +144,5 @@ - - Resource-rational moral judgment - https://cicl.stanford.edu/publication/wu2024resource/ - Wed, 01 May 2024 00:00:00 +0000 - - https://cicl.stanford.edu/publication/wu2024resource/ - - - diff --git a/docs/member/tobias_gerstenberg/index.html b/docs/member/tobias_gerstenberg/index.html index 8a1f7a4..2dcdd24 100644 --- a/docs/member/tobias_gerstenberg/index.html +++ b/docs/member/tobias_gerstenberg/index.html @@ -356,6 +356,45 @@

    Publications

    + + + (2024). + + Imagining and building wise machines: The centrality of AI metacognition. + arXiv. + + + + +

    + + + + + + Preprint + + + + + + + + + + + + + + + + + +

    + +
    +
    diff --git a/docs/publication/index.html b/docs/publication/index.html index 534511d..1100122 100644 --- a/docs/publication/index.html +++ b/docs/publication/index.html @@ -1610,6 +1610,19 @@

    Publications

    + + + + + + + + + + + + + @@ -1746,6 +1759,57 @@

    Publications

    +
    + +
    + + + (2024). + + Imagining and building wise machines: The centrality of AI metacognition. + arXiv. + + + + +

    + + + + + + Preprint + + + + + + + + + + + + + + + + + +

    + +
    + + +
    + + + + + + +
    diff --git a/docs/publication/index.xml b/docs/publication/index.xml index 781e6ab..6828827 100644 --- a/docs/publication/index.xml +++ b/docs/publication/index.xml @@ -12,6 +12,15 @@ + + Imagining and building wise machines: The centrality of AI metacognition + https://cicl.stanford.edu/publication/johnson2024wise/ + Wed, 06 Nov 2024 00:00:00 +0000 + + https://cicl.stanford.edu/publication/johnson2024wise/ + + + From Artifacts to Human Lives: Investigating the Domain-Generality of Judgments about Purposes https://cicl.stanford.edu/publication/prinzing2024purpose/ diff --git a/docs/publication/johnson2024wise/index.html b/docs/publication/johnson2024wise/index.html new file mode 100644 index 0000000..a2dbab1 --- /dev/null +++ b/docs/publication/johnson2024wise/index.html @@ -0,0 +1,488 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + Imagining and building wise machines: The centrality of AI metacognition | Causality in Cognition Lab + + + + + + +
    + + + + +
    +
    +

    Imagining and building wise machines: The centrality of AI metacognition

    + + + + + +
    + + + + +

    Abstract

    +

    Recent advances in artificial intelligence (AI) have produced systems capable of increasingly sophisticated performance on cognitive tasks. However, AI systems still struggle in critical ways: unpredictable and novel environments (robustness), lack transparency in their reasoning (explainability), face challenges in communication and commitment (cooperation), and pose risks due to potential harmful actions (safety). We argue that these shortcomings stem from one overarching failure: AI systems lack wisdom. Drawing from cognitive and social sciences, we define wisdom as the ability to navigate intractable problems—those that are ambiguous, radically uncertain, novel, chaotic, or computationally explosive—through effective task-level and metacognitive strategies. While AI research has focused on task-level strategies, metacognition—the ability to reflect on and regulate one’s thought processes—is underdeveloped in AI systems. In humans, metacognitive strategies such as recognizing the limits of one’s knowledge, considering diverse perspectives, and adapting to context are essential for wise decision-making. We propose that integrating metacognitive capabilities into AI systems is crucial for enhancing their robustness, explainability, cooperation, and safety. By focusing on developing wise AI, we suggest an alternative to aligning AI with specific human values—a task fraught with conceptual and practical difficulties. Instead, wise AI systems can thoughtfully navigate complex situations, account for diverse human values, and avoid harmful actions. We discuss potential approaches to building wise AI, including benchmarking metacognitive abilities and training AI systems to employ wise reasoning. Prioritizing metacognition in AI research will lead to systems that act not only intelligently but also wisely in complex, real-world situations.

    + + + + + + +
    +
    +
    +
    +
    Type
    +
    + + + Preprint + + +
    +
    +
    +
    +
    +
    + + +
    +
    +
    +
    +
    Publication
    +
    Johnson, S. G. B., Karimi, A., Bengio, Y., Chater, N., Gerstenberg, T., Larson, K., Levine, S., Mitchell, M., Schölkopf, B., Grossmann, I. (2024). Imagining and building wise machines: The centrality of AI metacognition. arXiv.
    +
    +
    +
    +
    +
    + +
    +
    +
    +
    +
    Date
    + +
    +
    +
    +
    +
    + + +
    +
    +
    +
    +
    Links
    +
    + + + + + + Preprint + + + + + + + + + + + + + + + + + +
    +
    +
    +
    +
    +
    + + +
    + +
    + +

    << Back to list of publications

     +
    + +
    + + + + + +
    + + +
    + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + diff --git a/docs/publication_types/1/index.html b/docs/publication_types/1/index.html index 844d5ff..61a3b44 100644 --- a/docs/publication_types/1/index.html +++ b/docs/publication_types/1/index.html @@ -111,7 +111,7 @@ - + @@ -238,6 +238,15 @@

    1

    +
    +

    Imagining and building wise machines: The centrality of AI metacognition

    +
    + + Recent advances in artificial intelligence (AI) have produced systems capable of increasingly sophisticated performance on cognitive tasks. However, AI systems still struggle in critical ways: unpredictable and novel environments (robustness), lack … + +
    +
    +

    Causation, Meaning, and Communication

    diff --git a/docs/publication_types/1/index.xml b/docs/publication_types/1/index.xml index d3bb7a4..58c9750 100644 --- a/docs/publication_types/1/index.xml +++ b/docs/publication_types/1/index.xml @@ -7,11 +7,20 @@ Hugo -- gohugo.io en-us &copy; 2024 Tobias Gerstenberg - Fri, 20 Sep 2024 00:00:00 +0000 + Wed, 06 Nov 2024 00:00:00 +0000 + + Imagining and building wise machines: The centrality of AI metacognition + https://cicl.stanford.edu/publication/johnson2024wise/ + Wed, 06 Nov 2024 00:00:00 +0000 + + https://cicl.stanford.edu/publication/johnson2024wise/ + + + Causation, Meaning, and Communication https://cicl.stanford.edu/publication/beller2024causation/ diff --git a/docs/publication_types/index.html b/docs/publication_types/index.html index 701846d..20d923b 100644 --- a/docs/publication_types/index.html +++ b/docs/publication_types/index.html @@ -111,7 +111,7 @@ - + @@ -239,21 +239,21 @@

    Publication_types

    -

    2

    +

    1

    -

    3

    +

    2

    -

    1

    +

    3

    diff --git a/docs/publication_types/index.xml b/docs/publication_types/index.xml index ead7c54..9b9e80b 100644 --- a/docs/publication_types/index.xml +++ b/docs/publication_types/index.xml @@ -7,11 +7,20 @@ Hugo -- gohugo.io en-us &copy; 2024 Tobias Gerstenberg - Sat, 26 Oct 2024 00:00:00 +0000 + Wed, 06 Nov 2024 00:00:00 +0000 + + 1 + https://cicl.stanford.edu/publication_types/1/ + Wed, 06 Nov 2024 00:00:00 +0000 + + https://cicl.stanford.edu/publication_types/1/ + + + 2 https://cicl.stanford.edu/publication_types/2/ @@ -30,15 +39,6 @@ - - 1 - https://cicl.stanford.edu/publication_types/1/ - Fri, 20 Sep 2024 00:00:00 +0000 - - https://cicl.stanford.edu/publication_types/1/ - - - 0 https://cicl.stanford.edu/publication_types/0/ diff --git a/docs/sitemap.xml b/docs/sitemap.xml index 519ffe6..02fb4e1 100644 --- a/docs/sitemap.xml +++ b/docs/sitemap.xml @@ -3,8 +3,8 @@ - https://cicl.stanford.edu/publication_types/2/ - 2024-10-26T00:00:00+00:00 + https://cicl.stanford.edu/publication_types/1/ + 2024-11-06T00:00:00+00:00 0 @@ -13,7 +13,7 @@ https://cicl.stanford.edu/ - 2024-10-26T00:00:00+00:00 + 2024-11-06T00:00:00+00:00 0 @@ -21,8 +21,8 @@ - https://cicl.stanford.edu/publication/prinzing2024purpose/ - 2024-10-26T00:00:00+00:00 + https://cicl.stanford.edu/publication/johnson2024wise/ + 2024-11-06T00:00:00+00:00 @@ -30,7 +30,7 @@ https://cicl.stanford.edu/publication_types/ - 2024-10-26T00:00:00+00:00 + 2024-11-06T00:00:00+00:00 0 @@ -38,8 +38,8 @@ - https://cicl.stanford.edu/publication_types/3/ - 2024-10-08T00:00:00+00:00 + https://cicl.stanford.edu/publication_types/2/ + 2024-10-26T00:00:00+00:00 0 @@ -47,25 +47,33 @@ - https://cicl.stanford.edu/publication/franken2024sami/ + https://cicl.stanford.edu/publication/prinzing2024purpose/ + 2024-10-26T00:00:00+00:00 + + + + + + + https://cicl.stanford.edu/publication_types/3/ 2024-10-08T00:00:00+00:00 + 0 - https://cicl.stanford.edu/publication/jin2024marple/ - 2024-10-04T00:00:00+00:00 + https://cicl.stanford.edu/publication/franken2024sami/ + 2024-10-08T00:00:00+00:00 - https://cicl.stanford.edu/publication_types/1/ - 2024-09-20T00:00:00+00:00 - 0 + https://cicl.stanford.edu/publication/jin2024marple/ + 2024-10-04T00:00:00+00:00 diff --git a/static/bibtex/cic_papers.bib b/static/bibtex/cic_papers.bib index a8c30af..4cbfcef 100644 --- a/static/bibtex/cic_papers.bib +++ b/static/bibtex/cic_papers.bib @@ -1,13 +1,22 @@ %% This BibTeX bibliography file was created using BibDesk. %% https://bibdesk.sourceforge.io/ -%% Created for Tobias Gerstenberg at 2024-10-26 12:42:26 -0700 +%% Created for Tobias Gerstenberg at 2024-11-06 11:16:38 -0600 %% Saved with string encoding Unicode (UTF-8) +@article{johnson2024wise, + abstract = {Recent advances in artificial intelligence (AI) have produced systems capable of increasingly sophisticated performance on cognitive tasks. However, AI systems still struggle in critical ways: unpredictable and novel environments (robustness), lack transparency in their reasoning (explainability), face challenges in communication and commitment (cooperation), and pose risks due to potential harmful actions (safety). We argue that these shortcomings stem from one overarching failure: AI systems lack wisdom. Drawing from cognitive and social sciences, we define wisdom as the ability to navigate intractable problems---those that are ambiguous, radically uncertain, novel, chaotic, or computationally explosive---through effective task-level and metacognitive strategies. While AI research has focused on task-level strategies, metacognition---the ability to reflect on and regulate one's thought processes---is underdeveloped in AI systems. In humans, metacognitive strategies such as recognizing the limits of one's knowledge, considering diverse perspectives, and adapting to context are essential for wise decision-making. We propose that integrating metacognitive capabilities into AI systems is crucial for enhancing their robustness, explainability, cooperation, and safety. By focusing on developing wise AI, we suggest an alternative to aligning AI with specific human values---a task fraught with conceptual and practical difficulties. Instead, wise AI systems can thoughtfully navigate complex situations, account for diverse human values, and avoid harmful actions. We discuss potential approaches to building wise AI, including benchmarking metacognitive abilities and training AI systems to employ wise reasoning. Prioritizing metacognition in AI research will lead to systems that act not only intelligently but also wisely in complex, real-world situations.}, + author = {Johnson, Samuel G B and Karimi, Amir-Hossein and Bengio, Yoshua and Chater, Nick and Gerstenberg, Tobias and Larson, Kate and Levine, Sydney and Mitchell, Melanie and Sch{\"o}lkopf, Bernhard and Grossmann, Igor}, + date-added = {2024-11-06 11:16:21 -0600}, + date-modified = {2024-11-06 11:16:21 -0600}, + journal = {arXiv}, + title = {{Imagining and building wise machines: The centrality of AI metacognition}}, + year = {2024}} + @article{jin2024marple, abstract = {Reconstructing past events requires reasoning across long time horizons. To figure out what happened, we need to use our prior knowledge about the world and human behavior and draw inferences from various sources of evidence including visual, language, and auditory cues. We introduce MARPLE, a benchmark for evaluating long-horizon inference capabilities using multi-modal evidence. Our benchmark features agents interacting with simulated households, supporting vision, language, and auditory stimuli, as well as procedurally generated environments and agent behaviors. Inspired by classic ``whodunit'' stories, we ask AI models and human participants to infer which agent caused a change in the environment based on a step-by-step replay of what actually happened. The goal is to correctly identify the culprit as early as possible. Our findings show that human participants outperform both traditional Monte Carlo simulation methods and an LLM baseline (GPT-4) on this task. Compared to humans, traditional inference models are less robust and performant, while GPT-4 has difficulty comprehending environmental changes. We analyze what factors influence inference performance and ablate different modes of evidence, finding that all modes are valuable for performance. Overall, our experiments demonstrate that the long-horizon, multimodal inference tasks in our benchmark present a challenge to current models. Project website: https: //marple-benchmark.github.io/.}, annote = {Comment: NeurIPS 2024. First two authors contributed equally. Project page: https://marple-benchmark.github.io/},