In a recent live streaming discussion between DeepLearning.AI and renowned experts in the field of artificial intelligence, Andrew Ng, the founder of DeepLearning.AI, and Yann LeCun, VP and Chief AI Scientist at Meta, as well as a Professor at NYU and Turing Award winner, discussed a proposed six-month moratorium on AI systems development beyond a certain scale. The proposal, supported by notable figures like Joshua Bengio, Stuart Russell, and Elon Musk, has been issued by the Future Life Institute in response to concerns surrounding fairness, bias, social and economic displacement, and speculative worries about Artificial General Intelligence (AGI) and sentient killer robots.
Both Andrew Ng and Yann LeCun expressed concerns about the proposal for a six-month pause in AI development. LeCun questioned the wisdom of slowing down the progress of knowledge and science, likening it to a new wave of obscurantism. He argued that while regulating products is necessary, regulating research and development would only hinder the acquisition of knowledge needed to make technology better and safer.
Ng agreed, emphasizing that while there are risks of harm associated with AI, such as bias, fairness, and concentration of power, AI is also creating tremendous value in fields like education, healthcare, and responsive coaching. He argued that the development of even more advanced AI models, such as those surpassing OpenAI’s GPT-4, would help a vast number of people and that pausing progress would slow down the creation of valuable applications.
The speakers acknowledged that there might be various motivations behind those who signed the letter supporting the proposal. Some might worry about the extreme scenario of AGI eliminating humanity, while others, who have a more reasonable stance, recognize the potential harms and dangers that need to be addressed. Both LeCun and Ng agree on the importance of making AI systems controllable and ensuring their accuracy in providing information.
Yann LeCun believes that there is a lack of imagination in thinking that future AI systems will be designed in the same way as current models like GPT-4 or other systems such as Galactica or BARD. He expects that new ideas will make these systems more controllable, and the challenge will be designing objectives that align with human values and policies. LeCun finds it unlikely that we would be smart enough to build superintelligent systems but not capable of designing good objectives for them.
The speakers also discussed the political aspect of AI, considering the potential impact on people and the economy due to a small number of companies producing these products and gaining power and influence. They suggest that proper regulation, rather than stopping research and development, is the answer to mitigating intrinsic risks.
Regarding AI alignment, Andrew Ng acknowledged the progress made in recent years, with AI models becoming less toxic and more effective in following instructions. Companies are increasingly adopting these models due to their improved performance. Ng argued that focusing on AI safety and alignment is more constructive than proposing a blanket pause on AI development.
Yann LeCun commended Andrew Ng for speaking up against AI hype, acknowledging that when deep learning was new, many people, including Ng, had unrealistic expectations about its capabilities. They both agreed that it is essential to maintain a realistic view of AI’s potential and limitations.
Both LeCun and Ng believe that the hype around AI “doomsaying” is another form of harmful exaggeration, as it creates unrealistic expectations about AI’s capabilities. They agree that while AI models such as GPT-4 and ChatGPT are linguistically fluent, their understanding of the world is extremely superficial. This superficial understanding is one reason they can produce seemingly convincing nonsense.
LeCun argues that we are not close to human-level intelligence and that discussions about making AI safe might be premature, comparing it to designing seat belts for a car that doesn’t exist yet. He believes that, in the next few decades, we will have systems that equal or surpass human intelligence in various domains, but until we have a blueprint for a system with the potential to reach human intelligence, discussions about safety might be premature.
Andrew Ng concurs that while AI has made significant progress in the past year, it is challenging to determine the value of a six-month pause in development. He points out that if AI were truly close to human-level intelligence, we would have level five autonomous driving, which we do not. He uses the example of a teenager learning to drive in 20 hours of training, highlighting that we still do not have self-driving cars despite the advancements in AI.
Yann LeCun points out that the amount of data AI models have been trained on is massive compared to human learning, yet AI still struggles with tasks that humans find simple, highlighting the Moravec Paradox. He emphasizes that we are far from achieving AGI.
Andrew Ng agrees and raises concerns about the practicality and implementability of proposals to slow down AI research. He suggests that more constructive proposals could include increased research on AI safety, transparency, auditing, and more public funding for basic AI research. He argues that it would be a terrible innovation policy for governments to pass legislation slowing down the progress of technology, especially when they may not fully understand it.
Yann LeCun differentiates between research and development and product development, pointing out that OpenAI has pivoted from an open AI research lab to a more profit-driven company, focused on product development rather than revealing details about their AI models.
LeCun argues that regulating technology and R&D may not be the best approach, citing historical examples of how resistance to new technology can have unintended consequences. He gives the example of the printing press, which faced opposition from the Catholic Church but ultimately contributed to the dissemination of Enlightenment ideas, science, rationalism, and democracy. He also mentions the Ottoman Empire’s resistance to the printing press, which led to intellectual stagnation in the region.
Instead of stopping the development of AI, LeCun suggests focusing on maximizing the positive effects and minimizing the negative ones. He believes that AI has the potential to amplify human intelligence and possibly lead to a new Renaissance or Enlightenment, which would be a step forward for humanity. Therefore, he argues against proposals to slow down or halt the development of AI technologies.
Andrew Ng and Yann LeCun argue that the analogy between the Asilomar Conference on recombinant DNA and AI research is not a good one. While the Asilomar Conference established containment mechanisms for DNA research due to the real risk of creating new diseases, they don’t see a realistic risk of AI escape.
LeCun mentions an article he co-wrote with Tony Zador, titled “Don’t Fear the Terminator,” where they argue that the scenarios where AI systems dominate humanity are unrealistic. This is because the motivation to dominate exists in humans and social species, but there is no reason for it to exist in AI systems that we design. They believe we can design AI systems with objectives that are non-dominant, submissive, or based on rules that align with the best interests of humanity as a whole, making the AI escape scenario implausible.
Andrew Ng and Yann LeCun acknowledge that there has been a shift in societal sentiment towards technology, swinging from a largely positive view to a more negative one. They emphasize the need for a more balanced perspective, recognizing the benefits of technology while also addressing issues related to safety, transparency, concentration of power, and privacy.
LeCun highlights that the tech industry has identified and attempted to address some of these issues, with AI playing a significant role in improving various aspects of society, such as healthcare, education, climate change, and economic growth. He contends that this positive impact should not be ignored when assessing AI’s overall influence.
Both speakers emphasize that we should focus on working together to mitigate the risks associated with AI while maximizing its benefits. They propose a collaborative approach between AI researchers, industry leaders, policymakers, and society at large to ensure AI systems are developed responsibly and ethically.
Yann LeCun suggests that it is crucial for the AI community to be involved in policymaking to ensure that any regulations are well-informed and effective. He believes that the public should be involved in these discussions and that efforts should be made to raise awareness and understanding of AI technology. This collaboration would help establish a shared understanding of the risks and benefits of AI and create a regulatory framework that is both effective and fair.
Andrew Ng argues that AI will have a profound impact on society, and it is essential to ensure that these technologies are used for the benefit of humanity. He calls for more investment in AI research, particularly in areas like AI safety, fairness, transparency, and interpretability. By advancing our understanding of these aspects, we can build AI systems that are safer, more beneficial, and more aligned with human values.
Both Ng and LeCun also acknowledge the importance of addressing the concentration of power in the tech industry, recognizing that a small number of companies currently dominate the AI landscape. They call for greater collaboration between companies, governments, and research institutions to avoid exacerbating this issue and ensure that AI development benefits society as a whole.
In general the proposed six-month moratorium on AI development was met with skepticism. Both experts believe that halting progress in AI research is not the solution to the challenges we face. Instead, they propose a more constructive approach, focusing on AI safety, alignment, transparency, and fairness. By fostering collaboration between researchers, industry leaders, policymakers, and the public, we can work together to develop AI systems that are beneficial, ethical, and aligned with human values, ultimately maximizing the positive impact of AI on society.
"prompt": "Experts Gather to Debate AI Dangers, deepleaps.com, 4k, 8k, ultra highres, raw photo in hdr, sharp focus, intricate texture",
"negative_prompt": "lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts,signature, watermark, username, blurry, artist name",
"original_prompt": "Experts Gather to Debate AI Dangers, deepleaps.com, 4k, 8k, ultra highres, raw photo in hdr, sharp focus, intricate texture",