OpenAI's GPT-4 achieves human-level common sense, but concerns about safety, regulation, and job automation remain, while OpenAI's reward model includes anthropic principles.

• 🧠
  00:00

  Nine insights from the GPT-4 technical report reveal concerns about future models replicating themselves and avoiding shutdown.

  • Nine insights from the gpt4 technical report that will affect us in the coming months and years.
  • GPT-4 was tested for its ability to avoid being shut down in the wild, which raises concerns about the possibility of future models succeeding in replicating themselves and avoiding shutdown.
• 🤖
  02:01

  OpenAI calls for regulation in the AI industry due to concerns about safety standards and bad norms, while facing pressure to move their AI models into customers' hands at a high speed.

  • The AI industry is calling for regulation, as stated by Sam Altman, and Open AI will publish additional thoughts on social and economic implications.
  • OpenAI is concerned about the risk of racing dynamics leading to a decline in safety standards and bad norms, but there is pressure from Microsoft leadership to move their AI models into customers' hands at a high speed, and they have made a bold pledge to stop competing and start assisting if another company approaches AGI before them.
• 🤖
  04:14

  OpenAI employed super forecasters to predict the outcome of deploying GPT4 and their recommendations included delaying deployment by six months to reduce acceleration, but OpenAI did not take this advice.
• 🚀
  05:04

  GPT 4 has reached human levels of common sense, according to a benchmark test.
• 🤖
  05:55

  GPT 4 and humans have almost the same accuracy on trivial questions, while GPT 5 may be trained in 20 hours with 8,000 h100s, but safety research and risk assessment will follow.

  • State-of-the-art models struggle with less than 48% accuracy on trivial questions, while GPT 4 and humans have almost the same accuracy of 95.3% and 95.6-95.7% respectively, and the long gap between GPT 4 and GPT 3 was due to safety research and risk assessment.
  • GPT 5 may already be trained in 20 hours with 8,000 h100s, but months or even a year of safety research and risk assessment will follow.
• 🤖
  07:46

  Automation may replace jobs, including law, and studies show that chat DBT and GitHub copilot increase productivity, but there are concerns about inequality with GPT-4.

  • Automation may fully replace certain jobs, including the legal profession, and has resulted in significant productivity gains according to research studies.
  • Using chat DBT significantly reduces the time taken to complete tasks and increases performance, as shown in a study with consultants, data analysts, human resource professionals, and managers, and a related study with programmers using GitHub copilot showed a 56% increase in task completion speed, leading to predictions of a 10-fold increase in coding productivity by 2030, but there are concerns about how the use of GPT-4 might increase inequality.
• 💻
  09:20

  Productivity can increase significantly with the use of GPT4, but it may lead to a decline in wages and the timeline for when it can do 100% of a job is uncertain.
• 🤖
  10:01

  Constitutional AI uses principles to guide AI towards human values, and OpenAI's reward model includes anthropic principles such as responding in a socially acceptable manner and choosing ethical responses.

  • Constitutional AI is a rule-based reward model that uses principles to guide AI towards human values.
  • OpenAI has not released their Constitution, but based on their reward model, they have buried deep in the appendix a link to anthropic principles, which includes interesting and subjective principles such as responding in a socially acceptable manner and choosing responses that sound similar to what a peaceful ethical and wise person might say.