Основная статья Содержание
Аннотация
«Революция ChatGPT», которая произошла в 2023, резко сократила прогнозные оценки экспертов сроков, отделяющих нас от создания искусственного интеллекта, ни в чем интеллектуально не уступающего никому из людей (AGI). При этом, как это ни парадоксально, но существующие методы тестирования пока не способны хоть с какой-то достоверностью диагностировать достижение ИИ-системами уровня AGI. В настоящей работе обсуждается вопрос преодоления проблемы несовершенства современных способов тестирования ИИ-систем. В частности, излагается гипотеза о принципиальной невозможности решения проблемы обнаружения AGI, как с помощью психометрических тестов, так и методов оценки способности машин имитировать ответы людей, из-за так называемой «ловушки Гудхарта» для AGI. Рассмотрен ряд предложений по обходу «ловушки Гудхарта» для AGI способами, предлагаемыми в новейших исследовательских работах, с учетом первых результатов произошедшей «революции ChatGPT». В последней части статьи сформулирована связка из трех эвристических гипотез, позволяющих, в случае их верности, кардинально решить проблему «ловушки Гудхарта» для AGI и тем самым стать геймченджером на пути создания AGI.
Ключевые слова
Детали статьи
Литература
- Blackiston D., Kriegman S., Bongard J., Levin M. Biological Robots: Perspectives on an Emerging Interdisciplinary Field // Soft Robotics. 2023. Pp. 674-686. https://www.liebertpub.com/doi/full/10.1089/soro.2022.0142
- Gordijn D., Have H. ChatGPT: evolution or revolution? // Medicine, Health Care and Philosophy. 2023. V. 26. Pp. 1-2.https://link.springer.com/article/10.1007/s11019-023-10136-0
- Gottfredson L. Mainstream science on intelligence: An Editorial With 52 Signatories, History, and Bibliography // Intelligence. V.24. Issue 1. 1997, Pp. 13-23. http://www1.udel.edu/educ/ gottfredson/reprints/1997mainstream.pdf
- Hayes P., Ford K. Turing Test Considered Harmful // IJCAI'95: Proceedings of the 14th international joint conference on Artificial intelligence. 1995. V.1. Pp.972-977.
- https://dl.acm.org/doi/10.5555/1625855.1625981
- Hutson M. Rules to keep AI in check: nations carve different paths for tech regulation // Nature. 2023. V.620. Pp. 260-263. https://www.nature.com/articles/d41586-023-02491-y
- Jefferson G. The Mind of Mechanical Man // British Medical Journal. 1949. V.1. Pp.4616. https://doi.org/10.1136/bmj.1.4616.1105
- Sejnowski T.J. Large Language Models and the Reverse Turing Test // Neural Computation. 2023. V.35. Issue 3. Pp. 309-342. https://doi.org/10.1162/neco_a_01563
- Sterzer P. Die Illusion der Vernunft: Warum wir von unseren Überzeugungen nicht zu überzeugt sein sollten / Neuestes aus Hirnforschung und Psychologie. Ullstein, Berlin. 2022. https://www.amazon.de/Die-Illusion-Vernunft-%C3%9Cberzeugungen-Hirnforschung/dp/355020132X
- Turing A.M. Computing Machinery and Intelligence // Mind. 1950. V. LIX. Issue 236. Pp.433-460. https://doi.org/10.1093/mind/LIX.236.433
- Интернет ресурсы:
- Карелов С. Аффорданс – ключевое свойство интеллектуального агента // Малоизвестное интересное. 2021. https://dzen.ru/a/ YYzplIlQGSDExDYc
- Карелов С. Невычислимая тень будущего // Малоизвестное интересное. 2021. https://dzen.ru/a/YZTzizvaBzV1UFII
- Карелов С. Открыта теория относительности интеллекта: биологического и машинного // Малоизвестное интересное. 2021. https://dzen.ru/a/YYkdZ6xat1ZwQZjG
- Карелов С. Серендипность – чудо увидеть цель в море случайностей // Малоизвестное интересное. 2021. https://dzen.ru/a/YadVB3jkREoIOaZL
- Карелов С. Фиаско 2023. Характер сосуществования двух типов разума, зависит от их взаимопонимания // Малоизвестное интересное. 2023. https://dzen.ru/media/the_world_is_not_easy/fiasko-2023-6486f59dbfaf86243ed3c4b4
- Эпштейн М. Искусственный и человеческий интеллекты: новый эксперимент по их сопоставлению // Сноб. 2023. https://snob.ru/profile/27356/blog/3059715/
- AI pioneer Yoshua Bengio: Governments must move fast to «protect the public» // Financial Times. 2023. https://www.ft.com/content/ b4baa678-b389-4acf-9438-24ccbcd4f201
- AI tests into top 1% for original creative thinking // Science Daily. 2023. https://www.sciencedaily.com/releases/2023/07/230705154051.htm
- AI21 Labs concludes largest Turing Test experiment to date // Проект AI21 Labs. 2023. https://www.ai21.com/blog/human-or-not-results?utm_source=superhuman.beehiiv. com&utm_medium=newsletter&utm_campaign=ai21-labs-concludes-largest-turing-test-experiment-to-date
- Artificial Intelligence Law, Model Law v. 1.0. // Digi China Project. 2023. https://digichina.stanford.edu/work/translation-artificial-intelligence-law-model-law-v-1-0-expert-suggestion-draft-aug-2023/
- Barrett C., Boyd B., Burzstein E., Carlini N. et al. Identifying and Mitigating the Security Risks of Generative AI. 2023. https://arxiv.org/pdf/2308.14840.pdf
- Benizri I., Evers A., Mercer S.T., Jessani A. A Comparative Perspective on AI Regulation // Lawfare. 2023. https://www.lawfaremedia.org/article/a-comparative-perspective-on-ai-regulation
- Bongard J., Levin M. There’s Plenty of Room Right Here: Biological Systems as Evolved, Overloaded, Multi-Scale Machines // Biomimetics. 2023. V.8. Pp.110. https://doi.org/10.3390/biomimetics8010110
- Bremmer I., Suleyman M. The AI Power Paradox // Foreign Affairs. 2023. https://www.foreignaffairs.com/world/artificial-intelligence-power-paradox
- Bubeck S., Chandrasekaran V., Eldan R, Gehrke J., Horvitz E., Kamar E. et al. Sparks of Artificial General Intelligence: Early experiments with GPT-4 // Cornell University. 2023. https://arxiv.org/abs/2303.12712
- Butlin P., Long R., Elmoznino E. Consciousness in Artifcial Intelligence: Insights from the Science of Consciousness. 2023. https://arxiv.org/abs/2308.08708
- Chollet F. On the measure of intelligence. 2019. https://arxiv.org/abs/1911.01547
- Fitzgerald McK., Boddy A., Baum S.D. A Survey of Artificial General Intelligence Projects for Ethics, Risk, and Policy // Global Catastrophic Risk Institute Working Paper. 2020. https://gcrinstitute.org/papers/055_agi-2020.pdf
- Goncalves B. Can machines think? The controversy that led to the Turing test // AI & SOCIETY. 2022. DOI: 10.1007/s00146-021-01318-6
- Goncalves B. Irony with a Point: Alan Turing and His Intelligent Machine Utopia // Philosophy&Technology. 2023. https://doi.org/10.1007/s13347-023-00650-7
- Goodhart's law // Wikipedia. https://en.wikipedia.org/wiki/Goodhart%27s_law
- Guterres A. Artificial Intelligence: Opportunities and Risks for International Peace and Security // UN Security Council. 2023. 9381st Meeting. https://media.un.org/en/asset/k1j/ k1ji81po8p?fbclid=IwAR1Zq6X7baQzlnpVBhgzPfWwOLtRfUHv61uz35wnBZJE93lsGQdl257RbDk
- HAI-AI Index Workshop on Measurement in AI Policy: Opportunities and Challenges // Stanford University. Human-Centered Artificial Intelligence. 2019. https://hai.stanford.edu/hai-ai-index-workshop-measurement-ai-policy-opportunities-and-challenges-0
- Heaven W.D. Geoffrey Hinton tells us why he’s now scared of the tech he helped build // MIT Technology Review. 2023. https://www.technologyreview.com/2023/05/02/1072528/geoffrey-hinton-google-why-scared-ai/
- Hochberg M.E. A Theory of Intelligences: Concepts, Models, Implications. 2023. https://arxiv.org/abs/2308.12411
- ICLR 2022. From Cells to Societies - Collective Learning across Scales workshop. https://sites.google.com/view/collective-learning
- Is ChatGPT the Start of the AI Revolution? // Bloomberg. 2022. https://www.bloomberg.com /opinion/articles/2022-12-09/is-chatgpt-the-start-of-the-ai-revolution
- Jiang G., Xu M.. Evaluating and Inducing Personality in Pre-trained Language Models. 2023. https://arxiv.org/pdf/2206.07550.pdf
- John Y., Braganza O. Dead rats, dopamine, performance metrics, and peacock tails: proxy failure is an inherent risk in goal-oriented systems // Behavioral and Brain Sciences, 2023. https://doi.org/10.1017/S0140525X23002753
- Legg S. Machine super intelligence // Doctoral Dissertation submitted to the Faculty of Informatics of the University of Lugano in partial fulfillment of the requirements for the degree of Doctor of Philosophy. 2008. https://www.vetta.org/documents/Machine_Super_Intelligence.pdf
- Legg S., Hutter M. Universal intelligence: A definition of machine intelligence // Minds and machines (2007). https://arxiv.org/abs/0712.3329
- McCoy J.P., Ullman T.D. A Minimal Turing Test // The Journal of Experimental Social Psychology. 2018. V.79. Pp.1-8. https://doi.org/10.1016/j.jesp.2018.05.007
- Mishra S., Clark J., Perrault C.R. Measurement in AI Policy: Opportunities and Challenges. 2020. https://arxiv.org/abs/2009.09071
- Pause Giant AI Experiments: An Open Letter // Future of Life Institute. 2023. https://futureoflife.org/open-letter/pause-giant-ai-experiments/
- Pellert M., Lechner C., Wagner C., Rammstedt B., Strohmaier M. AI Psychometrics: Assessing the psychological profles of large language models through psychometric inventories. 2023. https://psyarxiv.com/jv5dt/
- Planning for AGI and beyond // OpenAI. 2023. https://openai.com/blog/planning-for-agi-and-beyond
- Shontell A. ChatGPT shows that the A.I. revolution has arrived // Fortune. 2023.
- https://fortune.com/2023/01/25/chatgpt-ai-revolution-february-march-2023-issue/
- Statement on AI Risk // Center for AI Safety. 2023. https://www.safe.ai/statement-on-ai-risk
- Thomas R.L., Uminsky D. Reliance on Metrics is a Fundamental Challenge for AI. 2019. https://doi.org/10.48550/arXiv.2002.08512
- Turing test // Wikipedia. https://en.wikipedia.org/wiki/Turing_test#CITEREFTuring1950
- West D.M. Senate hearing highlights AI harms and need for tougher regulation // The Brookings Institution. 2023. https://www.brookings.edu/articles/senate-hearing-highlights-ai-harms-and-need-for-tougher-regulation/
- Xu G., Liu J., Yan M., Xu H. et al. Values: Measuring the Values of Chinese Large Language Models from Safety to Responsibility. 2023. https://doi.org/10.48550/arXiv. 2307.09705
- Видео ресурсы:
- Cody T., Hahm C., Goertzel B. Test and evaluation first principles for general learning systems. 2023. AGI-23 Workshop. https://www.youtube.com/watch?v=Hfai7Plzg4M
Литература
Blackiston D., Kriegman S., Bongard J., Levin M. Biological Robots: Perspectives on an Emerging Interdisciplinary Field // Soft Robotics. 2023. Pp. 674-686. https://www.liebertpub.com/doi/full/10.1089/soro.2022.0142
Gordijn D., Have H. ChatGPT: evolution or revolution? // Medicine, Health Care and Philosophy. 2023. V. 26. Pp. 1-2.https://link.springer.com/article/10.1007/s11019-023-10136-0
Gottfredson L. Mainstream science on intelligence: An Editorial With 52 Signatories, History, and Bibliography // Intelligence. V.24. Issue 1. 1997, Pp. 13-23. http://www1.udel.edu/educ/ gottfredson/reprints/1997mainstream.pdf
Hayes P., Ford K. Turing Test Considered Harmful // IJCAI'95: Proceedings of the 14th international joint conference on Artificial intelligence. 1995. V.1. Pp.972-977.
https://dl.acm.org/doi/10.5555/1625855.1625981
Hutson M. Rules to keep AI in check: nations carve different paths for tech regulation // Nature. 2023. V.620. Pp. 260-263. https://www.nature.com/articles/d41586-023-02491-y
Jefferson G. The Mind of Mechanical Man // British Medical Journal. 1949. V.1. Pp.4616. https://doi.org/10.1136/bmj.1.4616.1105
Sejnowski T.J. Large Language Models and the Reverse Turing Test // Neural Computation. 2023. V.35. Issue 3. Pp. 309-342. https://doi.org/10.1162/neco_a_01563
Sterzer P. Die Illusion der Vernunft: Warum wir von unseren Überzeugungen nicht zu überzeugt sein sollten / Neuestes aus Hirnforschung und Psychologie. Ullstein, Berlin. 2022. https://www.amazon.de/Die-Illusion-Vernunft-%C3%9Cberzeugungen-Hirnforschung/dp/355020132X
Turing A.M. Computing Machinery and Intelligence // Mind. 1950. V. LIX. Issue 236. Pp.433-460. https://doi.org/10.1093/mind/LIX.236.433
Интернет ресурсы:
Карелов С. Аффорданс – ключевое свойство интеллектуального агента // Малоизвестное интересное. 2021. https://dzen.ru/a/ YYzplIlQGSDExDYc
Карелов С. Невычислимая тень будущего // Малоизвестное интересное. 2021. https://dzen.ru/a/YZTzizvaBzV1UFII
Карелов С. Открыта теория относительности интеллекта: биологического и машинного // Малоизвестное интересное. 2021. https://dzen.ru/a/YYkdZ6xat1ZwQZjG
Карелов С. Серендипность – чудо увидеть цель в море случайностей // Малоизвестное интересное. 2021. https://dzen.ru/a/YadVB3jkREoIOaZL
Карелов С. Фиаско 2023. Характер сосуществования двух типов разума, зависит от их взаимопонимания // Малоизвестное интересное. 2023. https://dzen.ru/media/the_world_is_not_easy/fiasko-2023-6486f59dbfaf86243ed3c4b4
Эпштейн М. Искусственный и человеческий интеллекты: новый эксперимент по их сопоставлению // Сноб. 2023. https://snob.ru/profile/27356/blog/3059715/
AI pioneer Yoshua Bengio: Governments must move fast to «protect the public» // Financial Times. 2023. https://www.ft.com/content/ b4baa678-b389-4acf-9438-24ccbcd4f201
AI tests into top 1% for original creative thinking // Science Daily. 2023. https://www.sciencedaily.com/releases/2023/07/230705154051.htm
AI21 Labs concludes largest Turing Test experiment to date // Проект AI21 Labs. 2023. https://www.ai21.com/blog/human-or-not-results?utm_source=superhuman.beehiiv. com&utm_medium=newsletter&utm_campaign=ai21-labs-concludes-largest-turing-test-experiment-to-date
Artificial Intelligence Law, Model Law v. 1.0. // Digi China Project. 2023. https://digichina.stanford.edu/work/translation-artificial-intelligence-law-model-law-v-1-0-expert-suggestion-draft-aug-2023/
Barrett C., Boyd B., Burzstein E., Carlini N. et al. Identifying and Mitigating the Security Risks of Generative AI. 2023. https://arxiv.org/pdf/2308.14840.pdf
Benizri I., Evers A., Mercer S.T., Jessani A. A Comparative Perspective on AI Regulation // Lawfare. 2023. https://www.lawfaremedia.org/article/a-comparative-perspective-on-ai-regulation
Bongard J., Levin M. There’s Plenty of Room Right Here: Biological Systems as Evolved, Overloaded, Multi-Scale Machines // Biomimetics. 2023. V.8. Pp.110. https://doi.org/10.3390/biomimetics8010110
Bremmer I., Suleyman M. The AI Power Paradox // Foreign Affairs. 2023. https://www.foreignaffairs.com/world/artificial-intelligence-power-paradox
Bubeck S., Chandrasekaran V., Eldan R, Gehrke J., Horvitz E., Kamar E. et al. Sparks of Artificial General Intelligence: Early experiments with GPT-4 // Cornell University. 2023. https://arxiv.org/abs/2303.12712
Butlin P., Long R., Elmoznino E. Consciousness in Artifcial Intelligence: Insights from the Science of Consciousness. 2023. https://arxiv.org/abs/2308.08708
Chollet F. On the measure of intelligence. 2019. https://arxiv.org/abs/1911.01547
Fitzgerald McK., Boddy A., Baum S.D. A Survey of Artificial General Intelligence Projects for Ethics, Risk, and Policy // Global Catastrophic Risk Institute Working Paper. 2020. https://gcrinstitute.org/papers/055_agi-2020.pdf
Goncalves B. Can machines think? The controversy that led to the Turing test // AI & SOCIETY. 2022. DOI: 10.1007/s00146-021-01318-6
Goncalves B. Irony with a Point: Alan Turing and His Intelligent Machine Utopia // Philosophy&Technology. 2023. https://doi.org/10.1007/s13347-023-00650-7
Goodhart's law // Wikipedia. https://en.wikipedia.org/wiki/Goodhart%27s_law
Guterres A. Artificial Intelligence: Opportunities and Risks for International Peace and Security // UN Security Council. 2023. 9381st Meeting. https://media.un.org/en/asset/k1j/ k1ji81po8p?fbclid=IwAR1Zq6X7baQzlnpVBhgzPfWwOLtRfUHv61uz35wnBZJE93lsGQdl257RbDk
HAI-AI Index Workshop on Measurement in AI Policy: Opportunities and Challenges // Stanford University. Human-Centered Artificial Intelligence. 2019. https://hai.stanford.edu/hai-ai-index-workshop-measurement-ai-policy-opportunities-and-challenges-0
Heaven W.D. Geoffrey Hinton tells us why he’s now scared of the tech he helped build // MIT Technology Review. 2023. https://www.technologyreview.com/2023/05/02/1072528/geoffrey-hinton-google-why-scared-ai/
Hochberg M.E. A Theory of Intelligences: Concepts, Models, Implications. 2023. https://arxiv.org/abs/2308.12411
ICLR 2022. From Cells to Societies - Collective Learning across Scales workshop. https://sites.google.com/view/collective-learning
Is ChatGPT the Start of the AI Revolution? // Bloomberg. 2022. https://www.bloomberg.com /opinion/articles/2022-12-09/is-chatgpt-the-start-of-the-ai-revolution
Jiang G., Xu M.. Evaluating and Inducing Personality in Pre-trained Language Models. 2023. https://arxiv.org/pdf/2206.07550.pdf
John Y., Braganza O. Dead rats, dopamine, performance metrics, and peacock tails: proxy failure is an inherent risk in goal-oriented systems // Behavioral and Brain Sciences, 2023. https://doi.org/10.1017/S0140525X23002753
Legg S. Machine super intelligence // Doctoral Dissertation submitted to the Faculty of Informatics of the University of Lugano in partial fulfillment of the requirements for the degree of Doctor of Philosophy. 2008. https://www.vetta.org/documents/Machine_Super_Intelligence.pdf
Legg S., Hutter M. Universal intelligence: A definition of machine intelligence // Minds and machines (2007). https://arxiv.org/abs/0712.3329
McCoy J.P., Ullman T.D. A Minimal Turing Test // The Journal of Experimental Social Psychology. 2018. V.79. Pp.1-8. https://doi.org/10.1016/j.jesp.2018.05.007
Mishra S., Clark J., Perrault C.R. Measurement in AI Policy: Opportunities and Challenges. 2020. https://arxiv.org/abs/2009.09071
Pause Giant AI Experiments: An Open Letter // Future of Life Institute. 2023. https://futureoflife.org/open-letter/pause-giant-ai-experiments/
Pellert M., Lechner C., Wagner C., Rammstedt B., Strohmaier M. AI Psychometrics: Assessing the psychological profles of large language models through psychometric inventories. 2023. https://psyarxiv.com/jv5dt/
Planning for AGI and beyond // OpenAI. 2023. https://openai.com/blog/planning-for-agi-and-beyond
Shontell A. ChatGPT shows that the A.I. revolution has arrived // Fortune. 2023.
https://fortune.com/2023/01/25/chatgpt-ai-revolution-february-march-2023-issue/
Statement on AI Risk // Center for AI Safety. 2023. https://www.safe.ai/statement-on-ai-risk
Thomas R.L., Uminsky D. Reliance on Metrics is a Fundamental Challenge for AI. 2019. https://doi.org/10.48550/arXiv.2002.08512
Turing test // Wikipedia. https://en.wikipedia.org/wiki/Turing_test#CITEREFTuring1950
West D.M. Senate hearing highlights AI harms and need for tougher regulation // The Brookings Institution. 2023. https://www.brookings.edu/articles/senate-hearing-highlights-ai-harms-and-need-for-tougher-regulation/
Xu G., Liu J., Yan M., Xu H. et al. Values: Measuring the Values of Chinese Large Language Models from Safety to Responsibility. 2023. https://doi.org/10.48550/arXiv. 2307.09705
Видео ресурсы:
Cody T., Hahm C., Goertzel B. Test and evaluation first principles for general learning systems. 2023. AGI-23 Workshop. https://www.youtube.com/watch?v=Hfai7Plzg4M