David Edmonds (Uehiro Centre, Oxford University) and Nigel Warburton (freelance philosopher/writer) interview top philosophers on a wide range of topics. Two books based on the series have been published by Oxford University Press. We are currently self-funding - donations very welcome via our website http://www.philosophybites.com
…
continue reading
内容由Joe Carlsmith提供。所有播客内容(包括剧集、图形和播客描述)均由 Joe Carlsmith 或其播客平台合作伙伴直接上传和提供。如果您认为有人在未经您许可的情况下使用您的受版权保护的作品,您可以按照此处概述的流程进行操作https://zh.player.fm/legal。
Player FM -播客应用
使用Player FM应用程序离线!
使用Player FM应用程序离线!
Is Power-Seeking AI an Existential Risk?
Manage episode 424655590 series 3402048
内容由Joe Carlsmith提供。所有播客内容(包括剧集、图形和播客描述)均由 Joe Carlsmith 或其播客平台合作伙伴直接上传和提供。如果您认为有人在未经您许可的情况下使用您的受版权保护的作品,您可以按照此处概述的流程进行操作https://zh.player.fm/legal。
Audio version of my report on existential risk from power-seeking AI. Text here: https://arxiv.org/pdf/2206.13353.pdf. Narration by Type III audio.
章节
1. Is Power-Seeking AI an Existential Risk? (00:00:00)
2. Abstract (00:00:13)
3. 1 Introduction (00:02:31)
4. 1.1 Preliminaries (00:06:30)
5. 1.2 Backdrop (00:10:40)
6. 1.2.1 Intelligence (00:11:10)
7. 1.2.2 Agency (00:13:14)
8. 1.2.3 Playing with fire (00:14:38)
9. 1.2.4 Power (00:17:04)
10. 2 Timelines (00:20:49)
11. 2.1 Three key properties (00:21:13)
12. 2.1.1 Advanced capabilities (00:21:32)
13. 2.1.2 Agentic planning (00:23:28)
14. 2.1.3 Strategic awareness (00:30:22)
15. 2.2 Likelihood by 2070 (00:31:56)
16. 3 Incentives (00:34:22)
17. 3.1 Usefulness (00:38:49)
18. 3.2 Available techniques (00:46:11)
19. 3.3 Byproducts of sophistication (00:47:28)
20. 4 Alignment (00:49:39)
21. 4.1 Definitions and clarifications (00:50:05)
22. 4.2 Power-seeking (00:57:32)
23. 4.3 The challenge of practical PS-alignment (01:13:15)
24. 4.3.1 Controlling objectives (01:14:32)
25. 4.3.1.1 Problems with proxies (01:16:41)
26. 4.3.1.2 Problems with search (01:21:34)
27. 4.3.1.3 Myopia (01:27:08)
28. 4.3.2 Controlling capabilities (01:30:17)
29. 4.3.2.1 Specialization (01:31:15)
30. 4.3.2.2 Preventing problematic improvements (01:36:01)
31. 4.3.2.3 Scaling (01:37:43)
32. 4.3.3 Controlling circumstances (01:39:11)
33. 4.4 Unusual difficulties (01:42:41)
34. 4.4.1 Barriers to understanding (01:44:10)
35. 4.4.2 Adversarial dynamics (01:47:38)
36. 4.4.3 Stakes of error (01:49:40)
37. 5 Deployment (01:53:40)
38. 5.1 Timing of problems (01:57:08)
39. 5.2 Decisions (02:01:06)
40. Image: assessment of expected value of deployment (02:05:19)
41. 5.3 Key risk factors (02:07:33)
42. 5.3.1 Externalities and competition (02:08:02)
43. 5.3.2 Number of relevant actors (02:12:27)
44. 5.3.3 Bottlenecks on usefulness (02:15:25)
45. 5.3.4 Deception (02:20:40)
46. 5.4 Overall risk of problematic deployment (02:23:48)
47. 6 Correction (02:25:38)
48. 6.1 Take-off (02:26:32)
49. 6.2 Warning shots (02:29:10)
50. 6.3 Competition for power (02:34:37)
51. 6.4 Corrective feedback loops (02:48:21)
52. 6.5 Sharing power (02:54:50)
53. 7 Catastrophe (02:56:13)
54. Marker 53 (02:58:31)
55. 8 Probabilities (03:01:21)
56. Acknowledgments (03:19:56)
57集单集
Manage episode 424655590 series 3402048
内容由Joe Carlsmith提供。所有播客内容(包括剧集、图形和播客描述)均由 Joe Carlsmith 或其播客平台合作伙伴直接上传和提供。如果您认为有人在未经您许可的情况下使用您的受版权保护的作品,您可以按照此处概述的流程进行操作https://zh.player.fm/legal。
Audio version of my report on existential risk from power-seeking AI. Text here: https://arxiv.org/pdf/2206.13353.pdf. Narration by Type III audio.
章节
1. Is Power-Seeking AI an Existential Risk? (00:00:00)
2. Abstract (00:00:13)
3. 1 Introduction (00:02:31)
4. 1.1 Preliminaries (00:06:30)
5. 1.2 Backdrop (00:10:40)
6. 1.2.1 Intelligence (00:11:10)
7. 1.2.2 Agency (00:13:14)
8. 1.2.3 Playing with fire (00:14:38)
9. 1.2.4 Power (00:17:04)
10. 2 Timelines (00:20:49)
11. 2.1 Three key properties (00:21:13)
12. 2.1.1 Advanced capabilities (00:21:32)
13. 2.1.2 Agentic planning (00:23:28)
14. 2.1.3 Strategic awareness (00:30:22)
15. 2.2 Likelihood by 2070 (00:31:56)
16. 3 Incentives (00:34:22)
17. 3.1 Usefulness (00:38:49)
18. 3.2 Available techniques (00:46:11)
19. 3.3 Byproducts of sophistication (00:47:28)
20. 4 Alignment (00:49:39)
21. 4.1 Definitions and clarifications (00:50:05)
22. 4.2 Power-seeking (00:57:32)
23. 4.3 The challenge of practical PS-alignment (01:13:15)
24. 4.3.1 Controlling objectives (01:14:32)
25. 4.3.1.1 Problems with proxies (01:16:41)
26. 4.3.1.2 Problems with search (01:21:34)
27. 4.3.1.3 Myopia (01:27:08)
28. 4.3.2 Controlling capabilities (01:30:17)
29. 4.3.2.1 Specialization (01:31:15)
30. 4.3.2.2 Preventing problematic improvements (01:36:01)
31. 4.3.2.3 Scaling (01:37:43)
32. 4.3.3 Controlling circumstances (01:39:11)
33. 4.4 Unusual difficulties (01:42:41)
34. 4.4.1 Barriers to understanding (01:44:10)
35. 4.4.2 Adversarial dynamics (01:47:38)
36. 4.4.3 Stakes of error (01:49:40)
37. 5 Deployment (01:53:40)
38. 5.1 Timing of problems (01:57:08)
39. 5.2 Decisions (02:01:06)
40. Image: assessment of expected value of deployment (02:05:19)
41. 5.3 Key risk factors (02:07:33)
42. 5.3.1 Externalities and competition (02:08:02)
43. 5.3.2 Number of relevant actors (02:12:27)
44. 5.3.3 Bottlenecks on usefulness (02:15:25)
45. 5.3.4 Deception (02:20:40)
46. 5.4 Overall risk of problematic deployment (02:23:48)
47. 6 Correction (02:25:38)
48. 6.1 Take-off (02:26:32)
49. 6.2 Warning shots (02:29:10)
50. 6.3 Competition for power (02:34:37)
51. 6.4 Corrective feedback loops (02:48:21)
52. 6.5 Sharing power (02:54:50)
53. 7 Catastrophe (02:56:13)
54. Marker 53 (02:58:31)
55. 8 Probabilities (03:01:21)
56. Acknowledgments (03:19:56)
57集单集
所有剧集
×欢迎使用Player FM
Player FM正在网上搜索高质量的播客,以便您现在享受。它是最好的播客应用程序,适用于安卓、iPhone和网络。注册以跨设备同步订阅。