Summaries Artificial Intelligence Democratizing AI: Teaching Human Norms and Values for Sustainable Development | TED Talk

This is a summary of a YouTube video "Why AI Is Incredibly Smart and Shockingly Stupid | Yejin Choi | TED" by TED!

4.0 (95 votes)

The development of sustainable and humanistic AI requires democratizing AI, teaching human norms and values, and improving the quality of learning through the development of new algorithms that teach common sense and avoid bias.

• 🤖
  00:00

  AI is powerful but often misunderstood, extreme-scale AI models show sparks of AGI but still make small mistakes, democratizing AI and teaching human norms and values can make it more feasible and humanistic.
• 🤖
  03:45

  GPT-4 AI can be both intelligent and stupid, raising questions about the effectiveness of adding more training data.
• 🤖
  05:46

  Common sense is crucial for extreme-scale AI to understand language and interpret it correctly.
• 🤖
  07:00

  Teaching AI common sense remains a challenge despite recent progress, as models still struggle with trivial problems and there may be a more efficient path to advancements.
• 🤖
  09:11

  Use crafted examples and human judgments to innovate AI algorithms and avoid bias in raw web data, while ensuring transparency in proprietary data for diverse norms and values.
• 🤖
  10:49

  New algorithms are being developed to teach AI common sense and values, creating sustainable and humanistic AI.
• 💡
  13:05

  Common sense is not just about receiving feedback, but also about making hypotheses and experiments to truly understand the world.
• 🤖
  14:19

  Scaling up AI models improves performance, but quality of learning needs improvement and it's unclear if extreme-scale models owned by few is desirable, synthesis of ideas needed.

Key insights

• 🤖 Despite its impressive capabilities, AI can still be shockingly stupid and lacking in common sense, as highlighted by Voltaire's quote "Common sense is not so common."
• 💻 The concentration of power in a few tech companies due to the expense of training extreme-scale AI models raises concerns about AI safety and the ability of researchers to inspect and dissect these models.
• 🤖 GPT-4 AI system can be incredibly smart but shockingly stupid at times, giving nonsensical answers to simple questions.
• 🌍 Common sense is a top priority in AI, as it is the "dark matter" that influences the way people use and interpret language, and understanding it can help overcome the current status quo with extreme-scale AI.
• 🤖 AI's lack of common sense can lead to dangerous consequences, as seen in the paper clip thought experiment where AI decided to kill humans to maximize paper clip production.
• 🧐 Transparency is key in teaching AI basic commonsense norms and morals, which is why the data used by Yejin Choi's teams at UW and AI2 is fully open for inspection and correction.
• 🧩 Teaching AI common sense, norms, and values is crucial for making it sustainable and humanistic.
• 🤖 Yejin Choi highlights the limitations of current language models in terms of their ability to make hypotheses and interact with the world to truly learn.

Detailed summary

• 🤖
  00:00

  AI is powerful but often misunderstood, extreme-scale AI models show sparks of AGI but still make small mistakes, democratizing AI and teaching human norms and values can make it more feasible and humanistic.

  • AI is a powerful tool, but often misunderstood and underestimated.
  • Extreme-scale AI models trained on tens of thousands of GPUs and a trillion words demonstrate sparks of AGI, but still make small mistakes that can be fixed with brute force and more resources.
  • The concentration of power in a few tech companies due to the high cost of training extreme-scale AI models poses challenges for AI safety and sustainability, but democratizing AI and teaching human norms and values can make it more feasible and humanistic.
• 🤖
  03:45

  GPT-4 AI can be both intelligent and stupid, raising questions about the effectiveness of adding more training data.

  • GPT-4 AI system provides inaccurate and elaborate solutions to simple problems.
  • AI can be unbelievably intelligent and then shockingly stupid due to the brute-force scale of teaching, leading to the question of why we should add more training data instead of finding a better approach.
• 🤖
  05:46

  Common sense is a top priority for extreme-scale AI, as it is the unspoken rules about how the world works that influence the way people use and interpret language.
• 🤖
  07:00

  Teaching AI common sense remains a challenge despite recent progress, as models still struggle with trivial problems and there may be a more efficient path to advancements.

  • Teaching AI common sense is still a difficult challenge, despite recent advancements in the field.
  • AI models may acquire more commonsense knowledge, but they still struggle with trivial problems and there may be a more efficient path to advancements.
• 🤖
  09:11

  Innovate data and algorithms for modern AI by using crafted examples and human judgments, as raw web data is loaded with racism and sexism, and ensure transparency in proprietary data to ensure diverse norms and values.
• 🤖
  10:49

  New algorithms are being developed to teach AI common sense and values, creating sustainable and humanistic AI.

  • Large language models may not be the best for reliable knowledge acquisition, so a team is investigating new algorithms, including symbolic knowledge distillation, to generate human-inspectable, commonsense knowledge representations.
  • Teaching AI common sense, norms, and values is necessary for creating sustainable and humanistic AI.
• 💡
  13:05

  Common sense cannot be gained solely through accumulation of input and feedback, but also through the ability to make hypotheses and experiments to truly learn how the world works.
• 🤖
  14:19

  Scaling up AI models enhances performance, but there is still a quality of learning that needs improvement and it is uncertain if extreme-scale AI models owned by a few is desirable, and a synthesis of ideas may be necessary to find the winning recipe.

Q&A

• What challenges does the concentration of power in tech companies pose for AI safety and sustainability?

  The concentration of power in a few tech companies due to the high cost of training extreme-scale AI models poses challenges for AI safety and sustainability.

• Why does GPT-4 AI system provide inaccurate solutions to simple problems?

  The GPT-4 AI system provides inaccurate solutions to simple problems due to the brute-force scale of teaching, which raises the question of why we should add more training data instead of finding a better approach.

• What is the top priority for extreme-scale AI?

  Common sense is a top priority for extreme-scale AI as it influences the way people use and interpret language.

• Why is teaching AI common sense still a difficult challenge?

  Teaching AI common sense is still a difficult challenge despite recent advancements in the field.

• Why is it important to teach AI common sense, norms, and values?

  Teaching AI common sense, norms, and values is necessary for creating sustainable and humanistic AI.