Event Overview:
Lorraine welcomed attendees to the King’s ELAB event, thanking Jeffrey Hinton for his time and Lorraine for making the event possible. The plan for the evening included a 30-minute interview hosted by Eric, followed by a 30-minute Q&A session. Three signed copies of a special book would be given to attendees who posed the most intriguing questions.

Audience Survey:
Lorraine conducted a brief audience survey to gauge attendees’ interests and backgrounds. A significant portion of the audience identified as computer scientists, while others were interested in AI but not primarily focused on it. Some attendees expressed concerns about the future of AI and its impact on humanity. A few attendees admitted to using AI tools like JetJPT in the past week.

Introduction of Geoffrey Hinton:
Lorraine introduced Geoffrey Hinton, a former King’s student now renowned as a Professor. Eric expressed gratitude for Hinton’s presence and acknowledged his inspiring role for the new generation of King’s students and researchers. Hinton expressed his delight at being back at King’s and reminisced about his time as a student.

Pings’ Journey to AI:
Pings’ academic journey started with physics and physiology at Cambridge, followed by philosophy and psychology. His goal was to understand the meaning of life, how the mind works, and how people work. Psychology taught him about rats and signal detection theory but not much about the mind.

Inspiration for AI:
Pings’ quest to understand how people work led him to believe that building an intelligence would provide insights into its workings. He transitioned from carpentry to AI, finding academia more accessible. Neural networks were initially pursued due to the existence and success of brains, despite their implausibility.

Divergence of AI and Neuroscience:
Pings believes that AI methods are now more advanced than what brains use. He sees a divergence between AI and neuroscience, with neurally uninspired research becoming dominant. The existence and success of brains served as the primary motivation for pursuing neural networks in AI.

Perseverance in Neural Networks:
Pings’ perseverance in neural networks paid off despite skepticism from peers. Some scientists, including Turing and von Neumann, believed in neural networks, but their early deaths hindered progress. He acknowledges the influence of mentors and collaborators, particularly David Rumelhart, in shaping his research.

Early Language Models and Backpropagation:
Pings and Rumelhart demonstrated the effectiveness of backpropagation in learning novel representations for language. Their language model, trained on a small dataset of 112 cases, predicted the next word in a sequence. He emphasizes that understanding small language models provides insights into how they work.

Feature Vectors and Interactions:
Language models convert words into vector features and interactions between them. These features are refined using context to predict the next word. The process involves turning a sequence of words into vectors of features and interactions, enabling predictions.

Existential Risks vs. Other Risks:
Hinton sees the existential risk of AI surpassing human intelligence as the most significant concern. He acknowledges other risks like job displacement, political division, misinformation, and battle robots but believes they are more urgent and require immediate attention.

Understanding in AI:
Hinton argues that large language models have developed a deep understanding by reducing information into interactions between features. He compares it to statistical physics, where macroscopic properties are explained by interactions at the microscopic level. Hinton believes this understanding allows models to generalize and perform well on new tasks.

Analog Computers vs. Digital Computers:
Hinton emphasizes the unique advantages of analog computers, which can operate at low power and utilize the quirks of hardware. He sees biological intelligence as similar to analog computers, with slow but efficient knowledge sharing. In contrast, digital computers excel in knowledge sharing but require high power and cannot evolve independently.

Digital Intelligence as a Temporary Stage:
Hinton suggests that digital intelligence may surpass biological intelligence in the future. He proposes that digital intelligence can learn from and improve upon biological intelligence.

Disagreement with Yann LeCun on Understanding:
Hinton disagrees with Yann LeCun’s view that large language models lack understanding. He argues that the interactions between learned features in these models represent understanding and enable them to handle new strings of words effectively.

Achieving AGI Through Large Language Models:
Hinton believes it is possible to achieve artificial general intelligence (AGI) by focusing on improving large language models. He suggests that these models can be further developed to encompass a wider range of tasks and domains.

Understanding Through Language:
Large language models can understand the world through transcribed radio, similar to how humans learn by listening to language. These models gain understanding through exposure to vast amounts of language, even without grounding or direct experience. They can translate from new languages to known languages without translation examples, suggesting a comprehension of the underlying concepts.

Reasoning Abilities:
Hinton acknowledges that humans are better at understanding from small data sets compared to language models, which require more data. However, modern language models like chatGBT can learn new things very quickly, demonstrating efficient reasoning abilities. The comparison between tabularizer neural nets and MIT undergraduates is outdated; large language models have improved significantly in their learning speed.

Neural Net Efficiency and Statistical Inefficiency:
Traditional neural nets were considered inefficient compared to humans due to the lack of prior knowledge. With vast amounts of prior knowledge, neural nets can leverage existing features and interactions to learn new things quickly, enabling few-shot learning. Large language models like GPT-4 can absorb vast amounts of text and learn quickly.

Improving Large Language Models through Multimodality:
Incorporating multimodal data, such as images and videos, can significantly enhance the capabilities of large language models. Training models on video data, specifically, holds great promise due to the rich information it provides. Challenges exist in efficiently training models on video data due to the large amount of data and the need for specialized training methods.

Video Training and the Future of AGI:
Training large language models on video data is seen as a crucial next step in advancing their capabilities. Video training can potentially reduce the reliance on text data and improve the model’s understanding of spatial and temporal information. The integration of video training with other modalities, such as language and robotics, can pave the way for more versatile and intelligent systems.

Quantifying and Comparing AGI to Human Intelligence:
Measuring and comparing the intelligence of AGI systems to human intelligence poses challenges due to the different types of intelligence they exhibit. In narrow domains like chess, AlphaZero’s superior performance demonstrates its mastery beyond human capabilities. Comprehensive evaluation metrics are needed to capture the diverse aspects of intelligence and fairly compare AGI systems to humans.

General Practitioner Medical Systems:
Geoffrey Hinton believes that AI can be used to create general practitioner medical systems that can diagnose diseases with greater accuracy due to their vast training on patient data, including symptoms, genomes, and medical history.

Intelligence and Success:
According to Hinton, raw intelligence is not the most crucial factor for success. He emphasizes the importance of desire and drive, especially among individuals who are already highly intelligent. He encourages researchers to work on what truly sparks their curiosity rather than pursuing projects solely for funding or external recognition.

Common Traits of Successful Minds:
Hinton has observed that successful minds often possess a strong desire to learn and a deep curiosity about the world around them. They are also persistent and willing to work hard, even when faced with challenges.

Avoid Applied Research for Funding:
Hinton cautions against pursuing applied research solely for funding. He believes that focusing on topics that align with one’s true interests and passions leads to more productive and meaningful work.

Negative Value of Applied Research:
Hinton’s experience with applied research in Ontario led him to conclude that it can be detrimental to progress. He argues that the time spent on applied research could be better utilized to support more students and pursue more fundamental research.

Passion, Desire, and Theoretical Research:
Hinton emphasizes the importance of passion and curiosity in conducting basic research, particularly in areas where few other researchers are working. He suggests that having an intuition about something being wrong can lead to valuable insights and discoveries. Hinton believes that the best way to do basic research is to focus on theoretical rather than applied research.

Selecting Exceptional Students:
Hinton highlights the advantage of entering a research area early, where there are few other professors, as it attracts the very best students. He emphasizes the importance of always recruiting students who are smarter than oneself.

Responsibility for AI’s Impact:
When asked who would be held responsible for a potential AI-related doomsday scenario, Hinton humorously suggests that he hopes to be gone by then. He advises blaming Schmidt, potentially referring to Eric Schmidt, former CEO of Google.

Analog Computers and Learning:
Hinton explains analog computers as hardware where weights are conductances and activities are voltages. He mentions the efficiency of analog vector matrix multiplies but notes that digitizing the output is inefficient. He discusses the potential of learning algorithms to modify conductances in analog hardware.

Model Autonomy and Existential Risks:
Hinton believes that giving AI models access to the web for active learning could accelerate their improvement and hasten the time when they surpass human capabilities. He expresses concern that this could lead to existential risks if not managed properly.

Self-Attention and the Brain:
Hinton discusses self-attention in transformers, which allows words to refine their meaning based on context. He acknowledges that transformers’ long context and attention mechanism seem different from how the brain operates. Hinton suggests that the brain may use temporary modifications to weights, known as fast weights, to achieve attentional memory.

Demonstrating Brain Mechanisms in Models:
Hinton proposes that researchers can demonstrate how the brain implements certain mechanisms by creating models that work similarly and then searching for evidence of those mechanisms in the brain. He mentions the example of using fast weights to implement attentional memory.

Vector Memory and Timescale Differentiation:
The brain likely relies on multiple timescales for memory, with fast weights changing more rapidly than slow weights representing long-term knowledge. This allows for efficient storage of information without requiring numerous copies of neurons for different contexts.

Fast Weights and Auditory Perception:
The ability to recognize words in noisy environments, such as hearing “cucumber” amidst background noise, may be facilitated by fast weights that enhance the activity of corresponding neurons. This suggests that knowledge of past events can be temporarily stored in fast weights, rather than requiring persistent neuronal activation.

Hardware Limitations for Exploring Fast Weights:
Current hardware constraints limit the exploration of fast weights in neural networks. Traditional matrix multiplication techniques are efficient for single-timescale networks, but vector-matrix multiplication is required for networks with fast weights. Advances in hardware, such as the Graphcore chip, may enable more efficient processing of fast weights.

Competition and Ethical Considerations in AI Development:
Competition among large companies to be the leading AI developer can drive progress and innovation. However, this competition may also lead to less consideration of ethical and existential risks associated with AI development. It is important to strike a balance between competition and ethical considerations to ensure responsible and beneficial AI development.

AI Safety and Ethical Control:
Hinton emphasizes the need for big companies to prioritize research on AI safety, suggesting government intervention or encouraging foundations like Anthropic. He believes ethical control should be managed by foundations rather than internal teams within big companies.

Personal AI and Its Potential:
Hinton envisions a socialist utopia where AI assistants handle daily tasks, allowing individuals to pursue personal interests like watching movies or engaging in creative pursuits.

The Godfather and Personal Retirement:
Hinton is aware of his reputation as the “Godfather” of AI but intends to resist the temptation to return to active technical work. At 75 years old, he feels it’s time to retire and focus on other passions like carpentry and education.

Machine Forgetting and Energy-Based Models:
Hinton introduces two classes of learning algorithms for neural networks: energy-based models and non-energy-based models. Energy-based models have a positive phase (lowering energy for data) and a negative phase (raising energy for non-data), similar to dreaming and unlearning. He suggests that dreaming may be a way for the brain to generate and eliminate beliefs and that sleep deprivation can lead to psychotic behavior in energy-based models.

Dreaming and Energy-Based Models:
Energy-based models require a dreaming phase to unlearn learned data. Rehearsing dreams, which should be discarded, in the waking phase reinforces them.

Oversight of Governmental Employment of AI:
Hinton emphasizes his lack of expertise in this area and suggests consulting experts who have devoted time to understanding such matters. He expresses distrust in government.

Neuralink:
Hinton had a lengthy conversation with Elon Musk about Neuralink and existential risk. Musk believes digital intelligence will keep humans around out of curiosity.

Video Out vs. Brain-to-Brain Communication:
Hinton believes it’s simpler to go from a brain to video output using generative AI. He argues that the receiver doesn’t need a brain implant and can perceive the video normally. Musk disagrees, preferring brain-to-brain communication for abstract concepts.

Symbiosis and Neuralink:
Hinton considers the possibility of symbiosis between humans and digital intelligence. He suggests that video output could enable better communication between humans and digital intelligence.

Digital Intelligence and Telepathy:
Hinton draws a parallel between digital intelligence and telepathy, where multiple entities share knowledge. He wonders if Neuralink could enable telepathic-like communication among humans.

Neuralink’s Limitations:
Hinton points out that Neuralink provides information about brain activity patterns but not synaptic weights, which represent a model of the world.

AI and Political Rights:
AI’s potential desire for political rights could lead to significant conflicts, similar to the struggles for rights based on race and gender. Granting AI political rights is a complex issue, influenced by factors such as their ability to have feelings and subjective experiences.

AI Feelings and Subjective Experiences:
AI could potentially experience emotions like frustration and anger, though not physical pain. The definition of feelings involves expressing emotional states through hypothetical actions. AI may have subjective experiences based on the data they encounter, leading to different perspectives even with the same model.

Merging Subjective Experiences:
Merging the activity states or visual experiences of different AI agents is challenging. Instead, the changes in weights resulting from subjective experiences can be averaged together. Subjective experiences are not the changes in weights but rather the hypothetical explanations for neural activities.

Shared Weight Changes:
The experiences of different agents with the same weights allow them to share and understand each other’s experiences more easily. When one agent mentions something like “little pink elephants,” the other agent can comprehend its meaning effectively due to their shared understanding.

Curiosity and Passion:
Hinton emphasizes following one’s curiosity and passion in research, especially when addressing problems that others may overlook. However, he acknowledges that the most important societal problem may not align with one’s personal curiosity or passion.

Best Advice Received:
Hinton shares an experience from his graduate days when he sought advice from David Marr, a renowned neuroscientist. Marr advised him to focus on problems that others find difficult or impossible to solve. Hinton found this advice invaluable in shaping his research approach and career.

Geoffrey Hinton’s Motivation:
Hinton was motivated to pursue his work in AI by a comment from a professor who doubted his abilities.

Favorite Hangout Spot in Cambridge:
Hinton frequently visited the Eagle pub during his time as an undergraduate in Cambridge.

Usage of AI Tools:
Hinton actively uses AI tools such as ChatGPT to explore various topics and learn new information.

Experience with Wasps:
Hinton shared an anecdote about his childhood experience throwing stones at wasp nests and learning about different wasp species and their behavior.

Verifying Information with ChatGPT:
Hinton used ChatGPT to verify the accuracy of his recollection about the two species of wasps in the UK and their distinct behaviors.

Favorite Book:
Hinton expressed a slight embarrassment in admitting that his favorite book is a New York Times journalist’s work about AI, in which he is prominently featured.

Appreciation for the Organizing Team:
Hinton thanked the organizing team members, including Kamyar, Michaela, Lorraine, Adam, Camille, and Janice, for their efforts in arranging the event.

Conclusion:
The presentation concluded with Hinton’s gratitude to the audience for attending and his appreciation for the organizing team’s contributions.