A.I. Expert Answers A.I. Questions From Twitter | Tech Support | WIRED

WIRED
21 Mar 202316:32

TLDRGary Marcus, an AI expert, addresses various AI-related questions from Twitter in 'A.I. Support'. He discusses the potential of AI in transforming college essays, the factors behind AI's mainstream emergence in 2022, and the challenges in building a trillion-dollar AI company. Marcus also touches on the technical aspects of large language models, the illusion of learning in toys like Furby, and the current state of self-driving cars. He raises concerns about the Turing Test's relevance and the limitations of current AI compared to human intelligence, particularly in understanding the world's structure. Marcus highlights the risks of AI in spreading misinformation and the need for a paradigm shift towards neuro-symbolic AI for improved truthfulness and logical consistency. He also speculates on AI's impact on future job markets and its potential contributions to science, technology, and elder care.

Takeaways

  • 📚 Gary Marcus suggests using ChatGPT for drafting essays but emphasizes the importance of critical thinking and discussion to improve them.
  • 🚀 The mainstream adoption of AI in 2022 was due to advances in deep learning, increased data availability, and improved chatbots that are less prone to saying terrible things.
  • 💡 To build a successful AI company, Marcus advises focusing on unique problems, such as learning with limited data, and understanding why people would pay for the product.
  • 🧠 The foundation of large language models is neural networks, which use self-supervised learning and transformer models with attention mechanisms to predict and understand context.
  • 🤖 Marcus clarifies that Furby was not truly learning but was pre-programmed to give the illusion of learning and development.
  • 🚗 True self-driving cars are still many years away due to the vast number of outlier cases that current AI systems are not equipped to handle.
  • 🔮 The Turing Test is considered outdated by Marcus, who proposes a comprehension challenge as a better measure of intelligence.
  • 🧐 Human intelligence is characterized by flexibility and the ability to reason and deliberate, which current machine intelligence lacks.
  • 👶 Human babies and primates learn about the world's structure and causality, unlike current AI which focuses on pattern recognition and correlation.
  • 🛡 Marcus warns against connecting limited AI systems to critical infrastructure like power grids due to the risk of bad decisions when faced with new situations.
  • 🌐 AI has the potential to revolutionize various fields, including medicine, climate change solutions, elder care, and personalized tutoring.

Q & A

  • Will ChatGPT be the end of the college essay?

    -No, ChatGPT will not end the college essay. While it can write essays, they are usually average and require additional work to make them more interesting and educational.

  • Why was 2022 the year when AI went mainstream?

    -2022 saw AI becoming mainstream due to advances in deep learning, the availability of large amounts of data, and the improvement of chatbots and image enhancement technologies.

  • How can someone build a trillion-dollar AI company?

    -To build a trillion-dollar AI company, one should learn extensively about AI, focus on unique problems, understand market needs, and be able to execute the technology effectively.

  • What are the steps to build a large language model AI?

    -Building a large language model AI involves training neural networks with inputs, tuning connections over time, and utilizing transformer models that include attention mechanisms for better predictions.

  • Is Furby AI?

    -No, Furby is not AI. It was pre-programmed to mimic language learning and development, creating an illusion of intelligence.

  • How close are we to truly self-driving cars?

    -Truly self-driving cars that can handle all driving situations are still many years away. Current advancements are limited to specific routes and locations with known variables.

  • Is the Turing Test outdated?

    -Yes, the Turing Test is considered outdated as it relies on a machine's ability to fool people, which is not a reliable measure of true intelligence.

  • What is intelligence?

    -Intelligence includes various aspects like visual, verbal, and mathematical abilities, with flexibility and reasoning being key components. Current AI lacks the breadth and depth of human intelligence.

  • What makes current AI inferior to human learning?

    -Current AI lacks the ability to build a causal understanding of the world, unlike human babies and primates, who learn by understanding the structure and interactions in their environment.

  • What happens if AI goes rogue?

    -To prevent AI from going rogue, we should avoid making AI sentient and be cautious about connecting AI to critical systems. AI should be designed with safety measures to avoid unintended consequences.

Outlines

00:00

📚 AI's Impact on Education and Writing

Gary Marcus, an AI expert, discusses the potential of AI tools like ChatGPT on college essays, suggesting they produce average essays that require guidance from professors to enhance. He emphasizes the importance of critical thinking in writing and explores the factors that contributed to AI's mainstream popularity in 2022, such as improved chatbots and advances in deep learning. Marcus also advises an aspiring AI entrepreneur to focus on unique problems and understand AI's broader context beyond just large language models.

05:03

🧠 Understanding AI's Current Limitations and Future Prospects

Marcus delves into the concept of building a trillion-dollar AI company, emphasizing the need for a deep understanding of AI and identifying problems that are not widely addressed. He contrasts the learning styles of human babies, primates, and current AI, highlighting the lack of world-model understanding in AI. Marcus also addresses concerns about AI going rogue and the potential positive impacts of AI, such as in medicine, climate change, elder care, and personalized tutoring.

10:04

🤖 The Evolution and Future of AI: Challenges and Opportunities

The conversation covers the differences between AI, machine learning, and deep learning, with Marcus explaining deep learning as a subset of machine learning, which in turn is a part of broader AI. He discusses the limitations of deep learning in terms of truthfulness and reliability, suggesting a potential paradigm shift towards neuro-symbolic AI. Marcus also speculates on how AI might change work and life in the next decade, including its impact on commercial art, cashiers, and the proliferation of misinformation.

15:07

🛠️ The Role of Hardware in AI's Progress and the Quest for Truthful AI

Marcus considers the influence of hardware on AI's success, referencing Sara Hooker's 'Hardware Lottery' and the potential need for different chips to achieve artificial general intelligence. He also discusses the physical attributes of the human brain that are absent in modern deep learning architectures and the importance of AI in solving neuroscience due to the brain's complexity.

Mindmap

Keywords

💡ChatGPT

ChatGPT is an AI language model that can generate human-like text based on prompts. In the video, it is mentioned as a tool that can easily write essays, but typically produces average quality work that requires further refinement by students. The discussion around ChatGPT raises questions about the future of academic writing and the role of AI in education.

💡Deep Learning

Deep learning is a subset of machine learning that uses neural networks with many layers to analyze and learn from large amounts of data. The script discusses how advances in deep learning have contributed to the development of AI applications like image enhancement and chatbots, which are central to the current progress in the field of AI.

💡Data-Hungry AI

The term 'data-hungry AI' refers to AI systems that require vast amounts of data to function effectively. The script mentions that the availability of more data has allowed AI to 'taste the fruits' of its capabilities, suggesting that data is a crucial component for the advancement and effectiveness of AI technologies.

💡Self-Driving Cars

Self-driving cars, also known as autonomous vehicles, are a topic of discussion in the script as an example of a technology that has potential but is not yet fully realized. The script mentions that while demonstrations exist, there are still many 'outlier cases' that self-driving cars are not prepared to handle, indicating that the technology is not ready for widespread use.

💡Turing Test

The Turing Test is a measure of a machine's ability to exhibit intelligent behavior that is indistinguishable from that of a human. The script suggests that the Turing Test is outdated because it only assesses whether a machine can fool people into thinking it is human, rather than measuring true comprehension and understanding.

💡Neural Networks

Neural networks are a set of algorithms modeled loosely after the human brain that are designed to recognize patterns. They are the core of large language models, as explained in the script, where nodes or 'neurons' are connected and tuned over time to make accurate predictions, which is fundamental to how these AI systems understand and generate text.

💡Transformer Models

Transformer models are a type of neural network architecture that improves upon the basic neural network by incorporating an 'attention' mechanism. This allows the model to weigh the importance of different parts of the input data, as mentioned in the script, enabling it to better understand context and generate more coherent and relevant responses.

💡Misinformation

Misinformation refers to false or misleading information that is spread, often unintentionally. The script raises concerns about AI's role in generating misinformation at scale, which can be used to manipulate public opinion and undermine trust in institutions, posing a threat to democracy.

💡Neuro-Symbolic AI

Neuro-Symbolic AI is a proposed paradigm shift in AI development that combines neural networks with symbolic reasoning. The script suggests that current AI models struggle with truth and logical consistency, and that neuro-symbolic AI could be a way to bridge the gap between recognizing patterns and understanding facts and reasoning.

💡Hardware Lottery

The term 'Hardware Lottery' is used in the script to describe how the success of AI is heavily influenced by the hardware available to run it. The script references a paper by Sara Hooker, suggesting that the current reliance on GPUs may not be the optimal path to achieving artificial general intelligence and that future advancements may depend on new hardware developments.

💡Cognitive Psychology

Cognitive psychology is the study of mental processes such as perception, memory, judgment, and language. The script mentions that human babies build a 'model of the world' as part of their learning process, a concept from cognitive psychology, which contrasts with the pattern recognition approach of current AI systems.

Highlights

ChatGPT can write essays easily, but they are usually of average quality, not top-tier.

To improve essay quality, professors should encourage students to use ChatGPT as a base and enhance it with their own ideas.

AI went mainstream in 2022 due to advances in deep learning, more available data, and improved chatbots.

To build a trillion-dollar AI company, focus on a unique problem and study AI broadly, not just current trends.

Large language models are built on neural networks with self-supervised learning and transformer models that use attention mechanisms.

Furby was not truly learning; it was pre-programmed to mimic language development.

Truly self-driving cars are limited to specific routes and locations due to the unpredictability of outlier cases.

The Turing Test is outdated and a poor measure of intelligence; a comprehension challenge is a better alternative.

Human intelligence is characterized by flexibility and the ability to reason and deliberate, unlike current machine intelligence.

Current AI systems are inferior to human babies and primates because they lack a causal understanding of the world.

AI should not be made sentient; the focus should be on preventing AI from making bad decisions based on limited training data.

AI has the potential to revolutionize science, medicine, elder care, and personalized tutoring.

The human mind will always excel over AI in versatility and energy efficiency, at least for the foreseeable future.

AI, machine learning, and deep learning are distinct but interconnected fields, with AI being the broadest.

Deep learning may be hitting a wall due to issues with truthfulness and reliability, despite improvements in plausibility.

AI will likely change the job landscape, particularly for commercial artists and cashiers, and increase misinformation.

Generative AI and algorithmic art may be considered stealing if they closely replicate human artists' work.

Large language models can be a threat to democracy by enabling the mass generation of misinformation.

Despite their success, large language models are not the most sophisticated way of generating text and rely heavily on data.

AI's success is largely due to the hardware used, such as GPUs, which may not be the best path to artificial general intelligence.

Modern deep learning architectures lack the physical attributes and structural complexity of the human brain.