Why Is AI Harder to Achieve Than We Assume?

TL;DR

AI is more difficult to develop than commonly believed due to several fallacies. Melanie Mitchell highlights that narrow intelligence does not equate to general intelligence, easy tasks can be difficult for AI, and anthropomorphic language misrepresents AI's capabilities. The reliance on deep learning also reveals brittleness, indicating that understanding common sense and embodiment is essential for advancing AI research.

Transcript

artificial intelligence has been through a number of ups and downs through the years periods of boom typically called an ai spring followed by years of bust aptly called ai winter these cycles started in the 1950s following the invention of the perceptron which is of course the foundational unit of modern neural networks the explosion of research t... Read More

Key Insights

• 🤪 Artificial intelligence has gone through cycles of boom and bust, with periods of hype followed by disappointment.
• 🛀 Deep learning has shown success in teasing out correlations in large amounts of data but has limitations and challenges.
• 👊 The brittleness of deep learning models and susceptibility to adversarial attacks mimic the limitations of expert systems from the 1980s.
• 🤳 New technologies, such as transformer architectures and self-supervised learning, show promise but the path to artificial general intelligence is still uncertain.
• 😒 Fallacies in AI research include the belief that narrow intelligence is on a continuum with general intelligence, the assumption that easy things are easy for AI and hard things are hard, the use of anthropomorphic language in describing AI, and the disregard of the role of the body in cognition.
• 🧑‍🏭 Common sense, abstraction, emotion, and physical embodiment are important factors in intelligence that need to be considered in AI research.

Questions & Answers

Q: Why is artificial intelligence harder than we think?

AI systems can perform narrow tasks successfully without possessing common sense, broad understanding, or an ability to generalize. Melanie Mitchell’s paper identifies four fallacies that repeatedly cause researchers to overestimate progress toward artificial general intelligence.

Q: What are the four fallacies discussed in Melanie Mitchell’s paper?

The four fallacies are that narrow intelligence lies on a continuum with general intelligence, easy things are easy and hard things are hard, anthropomorphic language accurately describes AI, and the body’s role in cognition can be disregarded. Together, they highlight the importance of common sense, abstraction, emotion, and physical embodiment in intelligence.

Q: Why does narrow intelligence not necessarily lead to general intelligence?

Reproducing one narrow function of the human brain does not automatically represent progress toward human-like general intelligence. The paper argues that common sense is one major capability missing from the assumed continuum between narrow and general intelligence.

Q: Why can computers solve difficult equations but struggle with everyday tasks?

Tasks that humans consider difficult, such as calculating high-order differential equations, can be trivial for computers. Everyday abilities such as walking down a snowy sidewalk rely on capabilities shaped by a billion years of evolution and remain extremely difficult for computers.

Q: How does deep learning compare with 1980s expert systems?

Expert systems applied large sets of human-selected rules, while deep learning finds statistical solutions that maximize accuracy on training data. Despite this difference, both can be brittle and fail to generalize when they encounter examples beyond their training conditions.

Q: What do adversarial attacks reveal about neural networks?

Adversarial attacks use slight input modifications to make a neural network produce an incorrect output even when the change would not confuse a human. They demonstrate that neural networks do not understand inputs in the same way people do.

Q: How was a Tesla self-driving system fooled by a speed-limit sign?

In early 2020, researchers placed a small horizontal piece of tape on a 35 mph speed-limit sign. The sign still clearly appeared to show a three to a human, but Tesla’s neural network interpreted it as an eight and read the limit as 85 mph.

Q: Which AI technologies may offer paths toward more general intelligence?

Transformer architectures, the Perceiver model, self-supervised learning, and deep reinforcement learning are presented as promising open-ended paths. Whether any of them will provide a route to artificial general intelligence remains uncertain.

Summary & Key Takeaways

• Artificial intelligence has experienced boom and bust cycles, with periods of hype followed by disappointment.

• Expert systems in the 1980s showed promise but ultimately proved to be brittle and limited in their ability to generalize.

• Machine learning gained traction in the 1990s and early 2000s but is not considered a path to artificial general intelligence.

• Deep learning, the current paradigm, has seen success but faces challenges, such as brittleness and susceptibility to adversarial attacks.