Uploaded December 2024 | Updated September 2026, 3 weeks ago
In this AMA session, ex-#openai researcher Daniel Kokotajlo shares with the @BuzzRobot community his thoughts on how far we are from #artificialgeneralintelligence, the impact of government on #ai development, issues with #aialignment and #aisafety, the deceptive nature of #aiagents , and other pressing topics in #AI.
Timestamps:
0:00 Introduction
1:19 AGI by 2027?
3:24 The definition of AGI
5:36 Current AI systems are already improving AI research, accelerating the path to AGI
7:54 How deep learning works and why LLMs are not just "next-token predictors"
9:21 How to solve the long-term memory problem
10:46 What is the current recipe to achieve AGI: more compute, more data, more talent?
12:24 Compute is the gold of the 21st century
13:42 The role of academia in the AI race
15:22 Bringing energy supply to enable AGI
16:38 How do we know when it’s AGI?
22:26 If we hit a plateau with compute, how would that affect AI system development?
24:20 AI safety concerns
26:23 The main risks of AI (the deceptive nature of AI)
32:44 Instrumental convergence
35:03 The paperclip maximizer theory and how AI could harm humans
#artificialintelligence #artificialgeneralintelligence #artificialsuperintelligence #aialignment #aisafety #ai #chatgpt #deeplearning #machinelearning #neuralnetworks #technology #science #tech #techtalk #programming #softwareengineering #infrastructure #aiforecasting
Social Links:
Newsletter: buzzrobot.substack.com
X: https://x.com/sopharicks
Slack: join.slack.com/t/buzzrobot/shared_invite/zt-2s067rv7n-guPIMGe62rbp9ncxdnOUfQ
In this AMA session, ex-#openai researcher Daniel Kokotajlo shares with the @BuzzRobot community his thoughts on how far we are from #artificialgeneralintelligence, the impact of government on #ai development, issues with #aialignment and #aisafety, the deceptive nature of #aiagents , and other pressing topics in #AI.
Timestamps:
0:00 Introduction
1:19 AGI by 2027?
3:24 The definition of AGI
5:36 Current AI systems are already improving AI research, accelerating the path to AGI
7:54 How deep learning works and why LLMs are not just "next-token predictors"
9:21 How to solve the long-term memory problem
10:46 What is the current recipe to achieve AGI: more compute, more data, more talent?
12:24 Compute is the gold of the 21st century
13:42 The role of academia in the AI race
15:22 Bringing energy supply to enable AGI
16:38 How do we know when it’s AGI?
22:26 If we hit a plateau with compute, how would that affect AI system development?
24:20 AI safety concerns
26:23 The main risks of AI (the deceptive nature of AI)
32:44 Instrumental convergence
35:03 The paperclip maximizer theory and how AI could harm humans
#artificialintelligence #artificialgeneralintelligence #artificialsuperintelligence #aialignment #aisafety #ai #chatgpt #deeplearning #machinelearning #neuralnetworks #technology #science #tech #techtalk #programming #softwareengineering #infrastructure #aiforecasting
Social Links:
Newsletter: buzzrobot.substack.com
X: https://x.com/sopharicks
Slack: join.slack.com/t/buzzrobot/shared_invite/zt-2s067rv7n-guPIMGe62rbp9ncxdnOUfQ
