Introduction and Talk start
0:49 Introduction: Steve Wozniak Introduces Tristan Harris and Aza Raskin
1:30 Talk begins: The Rubber band effect
3:16 Preface: What does responsible rollout look like?
4:03 Oppenheimer Manhattan project analogy
4:49 Survey results on the probability of human extinction
3 Rules of Technology
5:36 1. New tech, A New Class of Responsibilities
6:42 2. If a Tech confers power, it starts race
6:47 3. If you don't coordinate, the race ends in tragedy
First contact with AI: 'Curation AI' and the Engagement Monster
7:02 First contact moment with curation AI: Unintended consequences
8:22 Second contact with creation AI
8:50 The Engagement Monster: Social media and the race to the bottom
Second contact with AI: 'Creation AI'
11:23 Entanglement of AI with society
12:48 Not here to talk about the AGI apocalypse
14:13 Understanding the exponential improvement of AI and Machine Learning
15:13 Impact of Language models on AI
Gollem-class AIs
17:09 GLLMM: Generative Large Language Multi-Modal Model (Gollem AIs)
18:12 Multiple Examples: Models demonstrating complex understanding of the world
22:54 Security vulnerability exploits using current AI models, and identity verification concerns
27:34 Total decoding and synthesizing of reality: 2024 will be the last human election
Emergent Capabilities of GLLMMs:
29:55 Sudden breakthroughs in multiple fields and theory of mind
33:03 Potential shortcoming of current alignment methods against a sufficiently advanced AI
34:50 Gollem-class AI can make themselves stronger AI can feed itself
37:53 Nukes don't make stronger nukes: AI makes stronger AI
38:40 Exponentials are difficult to understand
39:58 AI is beating tests as fast as they are made
Race to deploy AI
42:01 Potential harms of 2nd contact AI
43:50 AlphaPersuade
44:51 Race to intimacy
46:03 At least we're slowly deploying Gollems to the public to test it safely?
47:07 But we would never actively put this in front of our children?
49:30 But at least there are lots of safety researchers?
50:23 At least the smartest AI safety people think there's a way to do it safely?
51:21 Pause, take a breath
How do we choose the future we want?
51:43 Challenge of talking about AI
52:45 We can still choose the future we want
53:51 Success moments against existential challenges
56:18 Don't onboard humanity onto the plane without democratic dialogue
58:40 We can selectively slow down the public deployment of GLLMM AIs
59:10 Presume public deployments are unsafe
59:48 But won't we just lose to China?
How do we close the gap?
1:02:28 What else can we do to close the gap between what is happening and what needs to happen?
1:03:30 Even bigger AI developments are coming. And faster.
1:03:54 Let's not make the same mistake we made with social media
1:03:54 Recap and Call to action
Introduction and Talk start
0:49 Introduction: Steve Wozniak Introduces Tristan Harris and Aza Raskin
1:30 Talk begins: The Rubber band effect
3:16 Preface: What does responsible rollout look like?
4:03 Oppenheimer Manhattan project analogy
4:49 Survey results on the probability of human extinction
3 Rules of Technology
5:36 1. New tech, A New Class of Responsibilities
6:42 2. If a Tech confers power, it starts race
6:47 3. If you don't coordinate, the race ends in tragedy
First contact with AI: 'Curation AI' and the Engagement Monster
7:02 First contact moment with curation AI: Unintended consequences
8:22 Second contact with creation AI
8:50 The Engagement Monster: Social media and the race to the bottom
Second contact with AI: 'Creation AI'
11:23 Entanglement of AI with society
12:48 Not here to talk about the AGI apocalypse
14:13 Understanding the exponential improvement of AI and Machine Learning
15:13 Impact of Language models on AI
Gollem-class AIs
17:09 GLLMM: Generative Large Language Multi-Modal Model (Gollem AIs)
18:12 Multiple Examples: Models demonstrating complex understanding of the world
22:54 Security vulnerability exploits using current AI models, and identity verification concerns
27:34 Total decoding and synthesizing of reality: 2024 will be the last human election
Emergent Capabilities of GLLMMs:
29:55 Sudden breakthroughs in multiple fields and theory of mind
33:03 Potential shortcoming of current alignment methods against a sufficiently advanced AI
34:50 Gollem-class AI can make themselves stronger AI can feed itself
37:53 Nukes don't make stronger nukes: AI makes stronger AI
38:40 Exponentials are difficult to understand
39:58 AI is beating tests as fast as they are made
Race to deploy AI
42:01 Potential harms of 2nd contact AI
43:50 AlphaPersuade
44:51 Race to intimacy
46:03 At least we're slowly deploying Gollems to the public to test it safely?
47:07 But we would never actively put this in front of our children?
49:30 But at least there are lots of safety researchers?
50:23 At least the smartest AI safety people think there's a way to do it safely?
51:21 Pause, take a breath
How do we choose the future we want?
51:43 Challenge of talking about AI
52:45 We can still choose the future we want
53:51 Success moments against existential challenges
56:18 Don't onboard humanity onto the plane without democratic dialogue
58:40 We can selectively slow down the public deployment of GLLMM AIs
59:10 Presume public deployments are unsafe
59:48 But won't we just lose to China?
How do we close the gap?
1:02:28 What else can we do to close the gap between what is happening and what needs to happen?
1:03:30 Even bigger AI developments are coming. And faster.
1:03:54 Let's not make the same mistake we made with social media
1:03:54 Recap and Call to action