The Expanding Universe of Generative Models

16 Jan 2024 15:00h - 15:45h

Event report

Generative AI is advancing exponentially. What is happening at the frontier of research and application and how are novel techniques and approaches changing the risks and opportunities linked to frontier, generative AI models?

More info @ WEF 2024.

Table of contents

Disclaimer: This is not an official record of the WEF session. The DiploAI system automatically generates these resources from the audiovisual recording. Resources are presented in their original format, as provided by the AI (e.g. including any spelling mistakes). The accuracy of these resources cannot be guaranteed. The official record of the session can be found on the WEF YouTube channel.

Full session report

Daphne Koller

In the analysis, speakers discuss various aspects of artificial intelligence (AI) and its progress. One speaker highlights the significance of data in enhancing AI models and making them more sophisticated. They argue that the availability of more data, including data from self-driving cars, augmented reality, biology, healthcare, and other sources, will contribute to the advancement of AI. This argument is supported by the assertion that progress in the field of AI is primarily attributed to data. Another speaker takes the stance that data, rather than compute power, is the major driver of AI progress. They argue that while compute power and electricity are important, the exclusion of data in AI advancements should not be overlooked. They state that data is the single biggest enabler of AI progress and highlight the increasing availability of diverse data modalities from the real world.

The issue of causality understanding in AI models is also discussed. One speaker points out that current models lack the capacity to understand causality, as they are solely predictive engines that focus on associations. They argue that the incorporation of causality is essential for AI models to interact with the physical world and enable common-sense reasoning in applications such as manufacturing and biology. However, another speaker takes a neutral stance on causality, emphasizing its criticality when interacting with the physical world but not providing a conclusive argument.

The importance of conducting real-world experiments for AI’s growth and understanding of the world is highlighted by one speaker. They argue that machines need to possess the capability to design and learn from experiments in order to surpass human capabilities. These experiments, which teach computers about the complexity of the world, enable them to go beyond what can be taught by a human.

Although AI designs can teach us about building intelligence, they are not equivalent to human intelligence, according to one speaker. They disagree that current AI research is helping us understand human cognition.

The ability of AI to address challenging societal problems is another point of discussion. It is argued that AI can be leveraged to solve problems related to health, agriculture, the environment, and climate change, which may be difficult for humans to solve alone.

In terms of understanding natural phenomena, the analysis highlights that AI models, such as convolutional nets used as models for the visual cortex, improve our understanding but do not accurately replicate the complexity of the natural world. An analogy is drawn to the difference between understanding bird flight through airplanes and comprehending the intricacies of natural flight.

The application of technological advancements to aid in understanding human biology and improving medicine is seen as a positive development. One speaker disagrees with the term “off-ramp” being used to describe these advancements, highlighting the value of applying technology in these areas.

The importance of open-source models is emphasized, with the argument that models constructed by particular companies may not meet the needs of all applications. Open-source models accommodate new ideas and data modalities, enabling the community to build upon a strong foundation.

The role of education, particularly structured thinking, is discussed as a key determinant in the effective use of technology. It is argued that teaching structured thinking from a young age will result in better utilization of technology. This aligns with the goal of promoting quality education (SDG 4).

In conclusion, the analysis provides diverse perspectives on AI and its progress. It underscores the significance of data in enhancing AI models and the need to incorporate causality understanding to bridge the gap between the digital and physical realms. The ability to conduct and learn from real-world experiments is seen as crucial for machines to surpass human capabilities. While AI designs can teach us how to build intelligence, they do not replicate human intelligence entirely. AI’s potential to solve complex societal problems is acknowledged, and the importance of open-source models and structured thinking education is highlighted. The analysis also touches upon the role of AI in understanding natural phenomena and its application in medicine and biology.

Andrew Ng

The analysis of the provided arguments reveals several key points in the field of artificial intelligence (AI) and innovation. Firstly, there is a considerable pace of innovation in the field of scaling and algorithmic evaluations. Despite the increasing difficulty in scaling, continuous innovations are driving the acceleration of the field. The text highlights that numerous algorithmic evaluations and innovations have contributed to this trend. This suggests that the field is witnessing significant advancements fueled by continuous innovation.

Furthermore, the analysis underscores the anticipation of an image processing revolution. Recent technological advances, such as GPT-4V and Gemini Ultra, are cited as examples. These advancements have the potential to revolutionize image processing, transforming how it is currently understood and applied.

In terms of autonomous agents, there is optimism towards their development. The report mentions ongoing work on language models capable of conducting research. While it is noted that this kind of application is not yet fully functional, it highlights the potential for autonomous agents to contribute to various industries and domains.

The analysis also points out the similarities between large language models and humans. Large language models, akin to humans, have the potential to improve if provided with tools, such as a ‘scratchpad’, to work on. The report further mentions that these models are already using various tools to function. This observation underscores the potential for large language models to evolve and improve, ultimately enhancing their capabilities.

In addition to these technical aspects, the report delves into the perspectives of influential figures in the AI community. Andrew Ng, a prominent figure in the field, sees artificial general intelligence (AGI) as a significant goal. The report suggests that the AI community should embrace diverse goals, including AGI, climate change, and life sciences. However, it also notes the challenges posed by defining AGI in terms of human intelligence.

Another noteworthy observation is the value attributed to open source intelligence and AI. These are considered as valuable digital intelligences that contribute to wealth creation. The report suggests that an increase in intelligence, including artificial intelligence and open source intelligence, can lead to a wealthier and better world.

However, the report highlights the existing limitations in the current tech infrastructures. It mentions that today’s foundational tech architectures, such as semiconductors and cloud-based technologies, tend to be closed systems. In contrast, open source technologies are seen as contributing to collective intelligence and potentially sparking innovation and wealth creation. This observation underlines the need to strike a balance between closed systems and open source technologies in order to foster innovation and progress.

Furthermore, the analysis draws attention to the influence of lobbyists on regulations. It is noted that powerful forces are pushing for regulatory proposals that could impose burdensome requirements on open source technologies. This observation suggests the potential risks associated with over-regulation and the potential limitations imposed on open source innovation.

Lastly, the analysis emphasizes the importance of upskilling the workforce for the adoption and effective use of AI in enterprises. The lack of skilled workforce to implement and utilize AI-based tools is identified as a major bottleneck. The report mentions the need to upskill the workforce not only for building AI applications but also for effectively using AI-based tools. This observation underscores the necessity of prioritizing workforce training and education to fully benefit from AI technologies.

To summarize, the analysis highlights the significant pace of innovation in scaling and algorithmic evaluations in the field of AI. The anticipation of an image processing revolution is driven by recent technological advances. There is optimism surrounding the development of autonomous agents, with ongoing work on language models. The potential for large language models to evolve and improve with tools is emphasized. Influential figures like Andrew Ng see AGI as a significant goal but acknowledge the challenges of defining it. The value of open source intelligence and AI in wealth creation is recognized. The limitations of current tech infrastructures and the influence of lobbyists on regulations are noted. Upskilling the workforce is identified as crucial for AI adoption and utilization. These insights provide a comprehensive overview of the current state and future prospects of AI and innovation.

Nicholas Thompson

The panel discussion, moderated by Nicholas Thompson, delved into the future trajectory of artificial intelligence (AI). Thompson’s first question focused on whether the pace of AI advancements will continue to increase, taper, or plateau in the near future. This question aimed to deepen understanding of the speed of innovations in AI.

The discussion highlighted a potential correlation between improved graphics processing units (GPUs) and increased computing capacity, and their impact on AI models. Concerns were raised about the concentration of power in the AI market, as greater compute power is predominantly accessible to a small number of companies. It was suggested that this trend could lead to a less competitive AI market.

Additionally, the panel explored the vast amount of visual data available for AI systems to learn from. It was noted that while the amount of text data available may have limitations, the potential amount of visual data that could be utilized by AI is enormous.

Nicholas Thompson questioned how machines would comprehend video content, considering that even humans struggle to predict the outcomes of certain video scenarios. This raised the question of whether machines can truly understand complex visual information and make accurate predictions.

The discussion drew comparisons between child development and the ability of neural networks to understand causality. Personal anecdotes about raising children were shared, highlighting how young children struggle to comprehend cause and effect. This led to Thompson’s skepticism regarding whether machines could truly understand cause and effect as they are often attributed with more advanced abilities than they possess.

Concerns were expressed about AI models potentially corrupting or polluting themselves when left to learn and create without proper oversight. The worry was that AI systems may evolve in unexpected and potentially harmful ways if they are left unsupervised or if there is a lack of understanding about what the models are learning.

The goal of building machines that surpass human intelligence was questioned. Thompson expressed doubts about whether striving towards creating machines smarter than humans is the ideal goal for AI research. He suggested that AI researchers should focus on how AI can serve specific human needs and biology, similar to how airplanes were designed for flight rather than to replicate birds.

The controversy surrounding open-source AI was also examined. Thompson highlighted fears within the US government regarding the potential misuse of open-source AI and its implications for enabling individuals with malicious intentions to cause harm. The potential negative consequences of multiple countries contributing to highly ranked language models were also discussed.

Thompson appeared to be neutral towards the idea of open-source AI, but invited advocates to present their views and perspectives on the topic. Additionally, concerns were raised about the potential suppression of open-source models through legislation, which may consolidate power within a smaller number of larger companies.

Finally, the discussion questioned whether open-source AI is the only way to address and reduce income inequality. Thompson considered arguments proposing alternative methods to tackle this issue without solely relying on open-source AI.

In conclusion, the panel discussion explored various aspects of the future of AI, including the pace of advancements, power consolidation in the industry, the vast amount of visual data available, challenges in understanding video content, the limitations of AI in comprehending causality, concerns about AI models corrupting themselves, the goal of surpassing human intelligence, the controversy surrounding open-source AI, and alternative approaches to reduce income inequality. The discourse provided valuable insights into the current state and potential future directions of AI.

Yann LeCun

Yann LeCun, a prominent figure in the field of artificial intelligence (AI), has raised concerns about the limitations of current autoregressive language models (LLMs). These models, used to generate human-like text, are reaching saturation due to data limits. LeCun argues that there is not enough text data available to continue expanding and improving these models infinitely.

While LLMs have been trained with an extensive amount of text data (approximately 10 trillion tokens), there is a natural limit to their development. LeCun predicts the lack of sufficient text data will eventually hinder progress in LLMs.

LeCun also highlights the untapped potential of visual data. He suggests current methods for training AI models with video inputs are ineffective. There is a need for new breakthroughs to enable AI models to fully utilize sensory input, such as vision, and to teach them intuitive physics, causality, and abstract representations.

Open research and open-source development are key elements advocated by LeCun. He believes these approaches have been pivotal in the rapid progress of AI. Maintaining an open and collaborative environment is crucial for accelerating the development of AI technologies.

LeCun also emphasizes the importance of AI in advancing our understanding of human intelligence and perception. Building AI systems provides insights into human cognition, akin to how airplanes help us comprehend bird flight. However, he cautions against concentration of AI control in the hands of a few private companies. LeCun advocates for diverse and open-source AI systems that cater to various cultures, languages, values, and interests.

In conclusion, Yann LeCun’s perspective highlights the limitations of current autoregressive language models and the need for new breakthroughs in sensory input utilization, open research, and open-source development. He envisions a future where AI surpasses the capabilities of current models, paving the way for more advanced and diverse AI systems. By addressing these challenges and fostering an open and collaborative environment, LeCun believes AI can significantly shape our understanding of human intelligence and perception.

Aidan Gomez

Aidan Gomez argues that enhancing the efficiency of large language models relies on increasing compute power. He asserts that the next generation of GPUs will enable the implementation of more advanced algorithms and methods, unlocking new scale. Doubling compute capacity is anticipated to result in an equivalent increase in model size, indicating the impact that improved hardware platforms can have on large language models’ potential.

However, Gomez acknowledges that the current training strategy for models is approaching its upper limit. The existing approach does not support continuous learning from real-world interactions and human input. While the current strategy performs impressively, it needs to evolve to accommodate models that can learn continuously. This limitation highlights the challenges in training models to surpass the knowledge of an average person.

Furthermore, Gomez emphasizes the importance of models being capable of learning from online experiences and engaging in debate, which is currently unattainable within the existing training strategy. Enabling models to learn autonomously and continuously is essential for their further development and improvement.

Regarding power dynamics, Gomez supports the devolution of power from large tech companies. However, he acknowledges the validity of both open source and closed-source models. Gomez believes that creators and hackers should have access to technology that can challenge the authority of large tech companies, while businesses should also have the right to keep their models closed to maintain a competitive advantage. He suggests a hybrid approach that balances open source and closed-source models.

Additionally, Gomez recognizes the importance of technology being accessible and inclusive. He highlights the significance of understanding different languages and cultures, advocating for technology that can cater to diverse linguistic and cultural backgrounds. Gomez believes that accessible and inclusive technology is crucial for achieving equality and promoting inclusion.

In conclusion, Aidan Gomez’s views revolve around the role of compute power in enhancing large language models, the limitations of the current training strategy, the need for models to learn continuously, the devolution of power from large tech companies, and the importance of accessible and inclusive technology. His insights provide valuable perspectives on the future of language models and the ethical considerations surrounding their development and use.

Kai-Fu Lee

The analysis reveals several insights about the rate of change in artificial intelligence (AI) and its future trajectory. It suggests that while the rate of change in AI will slow down, it will still maintain an impressive pace. This is attributed to the potential for further improvements by increasing compute power, data availability, training methods, and infrastructure.

One notable observation is the optimism expressed regarding the future of AI. The analysis highlights that more entrepreneurs and large companies are entering the AI field, which suggests a growing interest and investment in this technology. This influx of talent and resources is expected to contribute to further advancements in AI.

Regarding scalability and innovation, the analysis suggests that scaling laws will continue to hold in AI, but the pace of innovation will slow down. Despite this, there is recognition of the diminishing returns in any endeavor, implying that the exponential growth seen in the past may not be sustainable in the long term.

Interestingly, the analysis indicates that we have not yet reached the plateau stage in AI development. This suggests that there is still ample room for further growth and improvement in this field. Kai-Fu Lee, a prominent figure in the analysis, also expresses agreement with Yan’s approach to neural networks, adding credibility to the assertion that this method has value and potential for further development.

The commercial value of text-based Large Language Models (LLMs) is emphasized in the analysis. LLMs are seen as valuable tools for generating content, improving productivity, and being deployed widely. This indicates the potential for these models to contribute significantly to various industries and sectors.

Regarding the concept of a world model, the analysis suggests that it is better suited for academic researchers and large company labs to explore. This implies that the development and understanding of a world model require a significant amount of resources and expertise.

AI systems, particularly those based on LLMs, are seen as powerful tools that can be applied in multiple areas, such as office productivity, content creation, and enhancing search engines. This underscores the transformative potential of AI and its ability to revolutionize various aspects of our lives.

The ongoing transformation through AI and LLMs presents numerous opportunities for value production and economic gains. The analysis recognizes the importance of engineering work to improve current AI algorithms, which suggests that there is still much to be done in terms of refining and enhancing AI capabilities.

Kai-Fu Lee’s vision of artificial general intelligence (AGI) as a platform for value creation is highlighted. This perspective positions AGI as a significant milestone with the potential to revolutionize industries and create new opportunities.

While the focus on AGI is emphasized, the analysis also suggests that the focus should be on realizing the value of LLMs rather than solely on surpassing human capabilities. This perspective aligns with the notion that AI can complement and enhance human capabilities rather than replace them entirely.

The analysis also touches on the potential risks and challenges associated with the dominance of one or a few companies in the AI field. It highlights the concern that such dominance can lead to significant inequality. It argues for the importance of open-source platforms and education in fostering innovation, creativity, and knowledge-sharing among different individuals and communities.

Overall, the analysis provides valuable insights into the current state and future trajectory of AI. It explores the potential for further advancements, the role of different technologies, and the need for diversity, competition, and collaboration in the AI landscape. These insights can inform decision-making and spark further discussions on the responsible and inclusive development of AI technology.

AG

Aidan Gomez

Speech speed

219 words per minute

Speech length

1287 words

Speech time

353 secs

AN

Andrew Ng

Speech speed

236 words per minute

Speech length

1076 words

Speech time

273 secs

DK

Daphne Koller

Speech speed

220 words per minute

Speech length

1367 words

Speech time

374 secs

KL

Kai-Fu Lee

Speech speed

184 words per minute

Speech length

1278 words

Speech time

417 secs

NT

Nicholas Thompson

Speech speed

240 words per minute

Speech length

2228 words

Speech time

558 secs

YL

Yann LeCun

Speech speed

202 words per minute

Speech length

2392 words

Speech time

711 secs