The Rise of Human-Like AI:
The advent of chat GPT marks a significant era in artificial intelligence, allowing for more natural and intuitive communication between humans and AI systems. Chat GPT’s success lies in its ability to mimic human speech patterns and understand human prompts, creating a sense of human-like interaction. This has led to a surge of interest in developing AI systems that can function and communicate like humans.

Beyond Human Mimicry:
Despite the fascination with human-like AI, some experts believe that the true potential of AI lies in its ability to solve complex scientific problems. Scientists and researchers are exploring AI’s capabilities in addressing global challenges such as clean energy, drug discovery, and innovation.

AlphaFold: A Case Study in Scientific Breakthrough:
AlphaFold, developed by DeepMind, stands as an example of AI’s transformative potential. It solves the protein folding problem, accurately predicting the 3D structure of proteins based on their amino acid sequence. This breakthrough has profound implications for scientific research and drug development.

The Road Ahead:
While AI has demonstrated its potential for scientific advancements, there is still much to be explored. Scientists continue to investigate ways to harness AI’s capabilities to solve a wide range of complex problems. The future of AI holds immense promise, but careful consideration and ethical considerations are crucial as it evolves.

History of DeepMind:
Demis Hassabis has been working on AI for nearly 30 years, starting with writing computer games that had AI as a core component, such as Theme Park and Black & White. He transitioned from the games industry to neuroscience, earning a PhD in the field. Hassabis founded DeepMind in 2010 with the vision of creating artificial general intelligence (AGI).

The Role of AI in Theme Park:
Hassabis explains that Theme Park, a simulation game he worked on in the 1990s, had AI as a core part of the gameplay. The AI controlled the behavior of the little people in the theme park, simulating their interactions with rides and stalls.

Defining Artificial Intelligence:
Hassabis highlights the difficulty in defining AI, as it is often described as anything the computer cannot do yet. He emphasizes that even seemingly simple tasks like the AI in Theme Park can be considered AI, as they involve simulating human behavior and decision-making.

Distinguishing AI from Machine Learning:
Hassabis differentiates AI from machine learning and statistics, which he views as less impressive functions. He argues that AI involves simulating human intelligence, enabling computers to perform tasks that require understanding, reasoning, and problem-solving.

AI’s Early Roots in Logic-Based Systems:
In the 1990s, games like Theme Park utilized cellular automata and logic systems to adapt to player behavior. These early AI systems, while innovative, were limited by their inability to handle unexpected situations.

The Emergence of Machine Learning:
Machine learning is a subfield of AI where systems learn from data and experience. This approach has shown greater power and scalability compared to logic-based systems.

DeepMind’s Focus on Machine Learning:
DeepMind, founded by Demis Hassabis, embraces machine learning as its primary approach to AI. AlphaGo, one of DeepMind’s notable achievements, learned to play the complex game of Go independently and devised new strategies.

Hassabis’s Early Interest in AI:
Hassabis’s journey into AI began at a young age, with programming and game design experiments. He pursued a PhD in neuroscience to gain a deeper understanding of the brain and intelligence.

Founding DeepMind:
Hassabis’s ultimate goal was to make AI a reality, leading to the establishment of DeepMind. His diverse background, including computer science and neuroscience, provided a solid foundation for his endeavors.

PhD in Cognitive Neuroscience:
Pursuing a PhD in cognitive neuroscience to study the brain as the only existence proof of general intelligence. Seeking inspiration for algorithmic and architectural ideas for AI systems from neuroscience. Aiming to learn the scientific method properly through control studies and practical research skills.

Motivation for Building AI:
Intrigued by big questions like the nature of the universe, reality, consciousness, and the meaning of life. Recognizing the slow progress in physics despite the efforts of brilliant scientists. Believing that AI could provide intellectual horsepower to assist in solving these complex problems.

Combining AI and Science:
Seeing AI as a way to gain insights into intelligence and the brain. Aiming to use AI as a tool to help experts in science and physics tackle their challenging questions. Recognizing AI as a meta-solution to address fundamental scientific problems.

Training AI on Video Games:
Initial attempts to train AI systems on video games, such as Pong, took a long time to achieve even a single point. Choosing video games as a training platform due to their well-defined rules, immediate feedback, and ability to test different strategies. Viewing games as a challenging environment for AI, requiring problem-solving, strategic thinking, and adaptation to changing conditions.

Game-Based Reinforcement Learning:
Demis Hassabis recognized the inefficiencies of working with physical robots and shifted DeepMind’s focus to simulation-based AI in games. Games offer clear objectives and challenging environments, ideal for training reinforcement learning systems. The complexity of games can be scaled up, providing a structured testing ground for AI systems.

Deep Reinforcement Learning:
The breakthrough in DeepMind’s approach was the use of deep Q networks (DQN), combining neural networks with reward-seeking algorithms. DQN played Atari games directly from pixel inputs without any explicit instructions or game knowledge. This approach led to the development of general systems capable of playing multiple Atari games at superhuman levels.

Expert Systems vs. Deep Learning:
Expert systems rely on explicitly programmed rules and knowledge about the game, which limits their adaptability and generalization. Deep learning systems learn from raw inputs, discovering the game’s rules and strategies through trial and error, similar to human learning. DeepMind’s general systems could play various Atari games at a superhuman level without specific instructions for each game.

Differences Between Deep Learning and Expert Systems:
Deep learning: Systems learn for themselves. Given a high-level objective (e.g., “win a game”), they must determine the necessary heuristics to achieve it. Systems learn from pixels or raw data without understanding the context or rules of the game. Systems may discover unexpected strategies to achieve the objective, potentially leading to unintended consequences. Expert systems: Rules are encoded into the system by human programmers. Systems follow the encoded rules to make decisions. Systems have a limited understanding of the problem domain and can only operate within the rules provided. Systems are more predictable and less likely to exhibit unexpected behaviors.

Alignment Problems in Deep Learning:
Alignment problems arise when the system’s goals or behaviors deviate from human values or intentions. Deep learning systems may pursue the objective without regard for the means, leading to unintended consequences. The lack of understanding of the context or rules of the game makes it difficult to predict and control the system’s behavior. Alignment problems can be exacerbated when the system is given sub-objectives or feedback that is not aligned with human values.

Guiding Deep Learning Systems:
To mitigate alignment problems, various techniques can be employed: Providing diverse and representative data to guide the system’s learning. Setting clear and well-defined objectives that align with human values. Incorporating human feedback or sub-objectives to shape the system’s behavior. Developing reinforcement learning algorithms that allow humans to provide feedback and influence the system’s learning process.

Deep Learning and Reinforcement Learning:
Deep learning involves hierarchical neural networks that build models of the environment or data stream. Reinforcement learning is a reward-seeking system that aims to achieve a given objective. Animals, including humans, use reinforcement learning, as seen in training dogs with treats.

Combining Deep Learning and Reinforcement Learning:
Deep reinforcement learning combines deep learning models with reinforcement learning planning systems. This allows the system to use the model to reach its objectives and achieve rewards.

AlphaGo: The Pinnacle of Games AI:
AlphaGo represents the extension of DeepMind’s work in Atari. It aimed to beat the world champion at Go, a more complex game than chess. AlphaGo’s success was a significant milestone in games AI.

AlphaGo’s Neural Network:
AlphaGo used a neural network to model the game of Go and predict good moves and the probability of winning from certain positions.

Reinforcement Learning System:
On top of the neural network, AlphaGo employed a reinforcement learning system for planning and decision-making. This system used Monte Carlo tree search, guided by the neural network’s predictions, to narrow down the search space and find the most promising moves.

AlphaZero’s Improvements:
AlphaZero was a more general system compared to AlphaGo. It learned motifs and strategies without human data or specific knowledge about Go. AlphaZero could even create new strategies never seen before in the thousands of years of human Go playing.

Bootstrapping and Training Data:
The first version of AlphaGo was bootstrapped using human games from internet Go servers, primarily from Korea, Japan, and China. AlphaGo also incorporated specific knowledge about Go, such as the symmetry of the board.

AlphaZero: A Generalized Game-Playing AI:
AlphaZero, the successor to AlphaGo, is capable of playing any two-player game without requiring human data. It starts with a blank slate neural network and learns through millions of self-played games, discovering new strategies not limited by human knowledge. This demonstrates that AI can surpass human expertise even when constrained by human knowledge.

AI’s Potential to Break Cultural Norms and Enhance Creativity:
AlphaGo’s innovative move 37 in the World Championship match highlights how AI can challenge cultural norms and expand creativity. Experts initially dismissed the move, but its success suggests that AI can uncover strategies that humans might overlook due to cultural biases and preconceptions. This potential to break cultural norms and enhance creativity extends beyond games and into scientific research.

AI as a Tool for Scientific Discovery:
AI systems like AlphaFold can be used as tools for scientific discovery and knowledge creation. Science, unlike games, aims to uncover new knowledge and understanding, making it a suitable domain for AI’s application. AI can help identify new areas of knowledge, inspire human experts, and accelerate scientific progress.

AlphaFold: Tackling the Protein Folding Problem:
AlphaFold is DeepMind’s program designed to solve the protein folding problem. Proteins are crucial for biological functions, and understanding their structure is essential for drug discovery and disease research. AlphaFold’s development began shortly after AlphaGo’s success in 2016 and has become a significant grand challenge for DeepMind.

Protein Folding and 3D Shape:
Proteins are described by their amino acid sequence, which can be thought of as a one-dimensional string of letters. In the body, proteins scrunch up into a 3D shape, and this shape governs its function.

Importance of Protein Folding:
Misfolded proteins can cause diseases. Knowledge of protein shape is crucial for drug design, as drugs target specific parts of the protein surface.

Coronavirus Spike Protein:
The coronavirus spike protein is a crucial part of the virus that sticks out. Vaccines and drugs latch onto this spike to block its function and prevent attachment to body cells.

Protein Folding and Games:
Protein folding shares similarities with games. Both involve searching for a solution through a vast space of possibilities. AI can be applied to both protein folding and games to find optimal solutions.

Foldit Game:
Foldit is a game where players solve protein folding puzzles. Players manipulate the protein structure to achieve a stable and functional shape. Foldit has contributed to scientific research by generating accurate protein structures.

Analogy of Protein Folding and Puzzle Game:
Protein folding is like a puzzle game where players fold proteins in a 3D interface. A few tens of thousands of amateur gamers got obsessed with the game, released around 2008-2009.

Intuition in Protein Folding:
Some gamers, despite having no biology knowledge, figured out counterintuitive protein folds leading to the correct 3D structure. These folds sometimes require local moves that worsen the energy landscape before resolving into the correct structure.

Inspiration from AlphaGo:
AlphaGo mimicked the intuition of Go masters, achieving remarkable success in the game. The success of AlphaGo in mimicking human intuition raised the possibility of mimicking the intuition of amateur gamers in protein folding.

Protein Folding as a Puzzle Game:
The analogy of protein folding as a puzzle game provided a framework for understanding the problem. The analogy suggested that it might be possible to develop an AI system to solve the puzzle of protein folding.

Objective Function:
In real-world applications, defining a simple objective function for machine learning systems can be challenging. For protein folding, minimizing the energy in the system can serve as a proxy objective.

Training Data:
AlphaFold’s training data consists of approximately 100,000 to 150,000 proteins with known amino acid structures and 3D structures deposited in the Protein Data Bank (PDB).

Training Process:
AlphaFold predicts the 3D structure of a protein based on its amino acid sequence. The accuracy of the prediction is measured by comparing the predicted structure to the real structure. The system receives a score based on the average error across all the atoms in the structure. The goal is to achieve an accuracy of less than one angstrom, which is the width of an atom.

Limited Training Data:
The number of proteins with known structures is relatively small compared to the millions or billions of proteins that exist. This limited training data poses a challenge for machine learning systems, which typically require large amounts of data to learn effectively.

Augmenting Training Data:
To address the limited training data, AlphaFold’s team used a strategy called data augmentation. They generated predictions using an initial version of AlphaFold and then incorporated the top 30-35% of these predictions back into the training set, along with the real data. This process helped to increase the effective training data size to about half a million proteins.

AlphaFold System and Protein Structure Prediction:
AlphaFold’s final system achieved atomic accuracy in predicting protein structures. The system uses a combination of real and generated data for training and undergoes self-distillation to improve accuracy. AlphaFold outputs both protein structure predictions and an uncertainty score for each amino acid, aiding biologists and researchers in understanding the reliability of the predictions.

Handling Uncertainties and Hallucinations in AI Systems:
Current chatbots often face the problem of hallucinations, generating plausible but incorrect information. AlphaFold addresses this issue by incorporating confidence levels and sanity checks, allowing the system to distinguish between reliable and unreliable predictions. Unlike chatbots, AlphaFold benefits from structured data with known correct answers, enabling automated correction and improvement.

Challenges in Language Models vs. Protein Structure Prediction:
Language and human knowledge are more complex than games or proteins, making it difficult to automate the correction process for language models. There is a need for improving the factuality and reliability of language models through techniques like reinforcement learning with human feedback. The subjective nature of language and knowledge requires human input for feedback and evaluation.

Timeline of AlphaFold Development and Protein Structure Discovery:
AlphaFold’s development involved several iterations and a period of stagnation before achieving a breakthrough. The system’s participation in the CASP competition marked a turning point, leading to rapid progress and advancements in protein structure prediction.

Protein Folding Competition:
AlphaFold participates in a biennial competition where teams predict protein structures based on experimental data provided by organizers. The predictions are double-blind, with competitors and organizers unaware of the true structures until the competition’s end.

The Protein Folding Breakthrough:
In 2020, AlphaFold achieved atomic accuracy in predicting the structures of 100 new proteins, marking a significant milestone in solving the protein folding problem. This breakthrough has broad implications for drug design and medical research.

AlphaFold Database:
To make AlphaFold’s capabilities accessible to researchers, DeepMind collaborated with the European Bioinformatics Institute to create a free database. The database includes protein structure predictions for various organisms, including humans, research organisms, and important crops.

The Isomorphic Business Model:
Isomorphic, a company under Alphabet, aims to leverage AlphaFold’s capabilities for drug discovery. This represents a different business strategy from integrating AI into search engines or other consumer-facing applications.

Bridging the Gap in Drug Discovery:
AlphaFold is seen as a tool that can accelerate drug discovery by providing insights into protein structures. This can aid in identifying targets for drug development and understanding how drugs interact with proteins.

AI in Drug Discovery:
AI can accelerate drug discovery by computationally exploring small molecules and binding properties. AI-driven drug discovery can reduce the time and cost of developing new drugs. Isomorphic Labs, a spin-off of DeepMind, focuses on developing AI tools for drug discovery and chemistry.

AI in Stock Market Predictions:
Some hedge funds may already use AI techniques for algorithmic trading. Predicting stock prices requires understanding macroeconomic forces, geopolitical factors, and company-specific information. AI systems may need to encapsulate a wide range of knowledge to make accurate stock market predictions. High-frequency trading and algorithmic trading strategies are already sophisticated and competitive. It’s unclear if general learning systems can outperform existing strategies in the stock market.

Specialized Systems vs. General Systems:
AI development can take two approaches: specialized systems tailored to specific tasks or general systems capable of handling a wide range of tasks. Specialized systems like AlphaFold excel in specific domains, while general systems, like large language models (LLMs), have broad capabilities but lack depth.

The Path to AGI:
The ultimate goal is to create a general intelligence (AGI) system that can perform diverse tasks like the human brain. On the way to AGI, specialized systems can be highly beneficial by focusing on specific tasks and achieving expert-level performance.

General Systems Utilizing Specialized Tools:
General systems can interact with specialized systems as tools, calling upon their expertise in specific domains. This allows general systems to leverage the capabilities of specialized systems without having to learn everything themselves.

Addressing Factuality, Robustness, Planning, and Memory:
Current large multimodal models lack factuality, robustness, planning, reasoning, and memory capabilities. Innovations and advancements are needed to address these limitations and enable general systems to perform complex tasks effectively.

The Race Dynamic in AI Development:
The pursuit of AGI has taken on a competitive dynamic, with major companies and countries investing heavily in AI research. Hassabis emphasizes the need for a more scientific and thoughtful approach, balancing optimism with risk assessment.

Risks and Benefits of AGI:
AGI has the potential to revolutionize many fields, curing diseases, solving energy and sustainability challenges, and acting as a powerful tool for humanity. However, it also carries risks due to its dual-use potential and the need for careful consideration of ethical and safety implications.

Minimizing Risks and Maximizing Benefits:
Hassabis advocates for a balanced approach that minimizes risks and maximizes benefits by carefully considering potential pitfalls and implementing mitigation strategies. Boldness and bravery in pursuing the benefits of AGI should be tempered with foresight and risk assessment.

AI in Energy and Sustainability:
AI can contribute to climate and sustainability in various ways: Optimizing existing infrastructure to enhance efficiency: E.g., saving energy in data center cooling systems using AI control. Environmental monitoring: AI can assist NGOs and governments in tracking deforestation, forest fires, and other environmental changes. Accelerating breakthrough technologies: AI is used to control and stabilize nuclear fusion processes, enabling cleaner energy generation. Material design: AI can aid in designing better batteries, solar panels, and superconductors.

Chatbots and System Design:
Recent focus on chatbots and human-like systems: Hype and investment in chatbots may overshadow the potential of more scientific and “inhuman” AI systems. There is a risk that the business models for these systems may not align with public benefit. Demis Hassabis emphasizes the importance of continuing research on scientific problems alongside developing next-generation products.

Balancing Scientific and Entertainment Applications:
Google DeepMind’s approach: Maintaining a balance between advancing science and medicine and developing new products for billions of users. Investing in scientific projects like AlphaFold and Isomorphic Labs while also exploring chatbot interfaces. Language as a universal interface: Chatbots can interact with specialized systems and tools, enabling users to access complex scientific resources.

Mitigating Extinction Risks from AI:
Demis Hassabis, Sam Altman, and Dario Amadei signed a letter emphasizing the need to prioritize mitigating extinction risks from AI. Reasons for concern: AI’s potential to cause unintended consequences similar to social media’s impact. The need for international cooperation to address safety and technical risks as AI systems become more powerful.

Near-Term Harms:
The potential for misuse of AI technology by bad actors is a significant concern. Deep fakes pose a near-term threat, and countermeasures such as watermarking are being developed to address this issue.

Technical Risks:
The alignment problem: Ensuring that AI systems align with human values and objectives is a critical challenge. Containment and control: As AI systems become more powerful, it is essential to ensure that they can be contained and controlled effectively.

Longer-Term Concerns:
Inherent technical risk: The possibility exists that AI systems could become so powerful that they pose a risk to humanity if not developed responsibly. The need for careful planning and exceptional care in the development of AI technologies is emphasized. The debate on the risks of AI should commence now, rather than waiting until the technology is fully developed.

Synthetic Biology Risks:
The accessibility and affordability of synthetic biology tools raise concerns about the potential for creating lethal viruses or other harmful biological agents. The risks associated with synthetic biology are acknowledged, and extensive discussions with experts in biosecurity, bioethics, and biology have been conducted. The benefits of AlphaFold’s release were deemed to outweigh the risks, and the database has been widely used by biologists around the world.

Access Control:
As AI systems become more sophisticated in drug design, there is a growing concern about who should have access to these technologies. Access to AI-driven drug design tools should be restricted to scientists and medical practitioners who are committed to using them for good. Open sourcing of AI-driven drug design results and cybersecurity measures are important factors to consider in preventing unauthorized access.

Monitoring and Detection:
AI can be utilized to monitor the use of AI-driven drug design tools and detect suspicious activities. This includes detecting attempts to design harmful substances or identifying bad actors who are trying to exploit these technologies for malicious purposes. Monitoring systems can also be employed to track the conversations between users and chatbots to identify any potential risks.

Known Toxins:
Currently, there are known toxins, such as anthrax, whose recipes can be found online. However, the actual production and distribution of these toxins require scientific expertise and a laboratory setup. This makes it difficult for naive individuals to create and distribute dangerous substances.

Information Control:
While the design of known toxins is accessible, controlling the information related to their production is essential. Restricting the availability of detailed instructions for producing toxins can help prevent unauthorized individuals from accessing this knowledge. Finding a balance between open access to information and responsible control is crucial in mitigating potential risks.

Overview:
Demis Hassabis discusses the potential risks and governance of general intelligence (AGI) systems and recommends books related to physics, AI, and post-AGI futures.

AGI Development and Governance:
The systems and labs capable of creating AGI are limited worldwide, but their numbers are likely to increase over time. As AGI approaches, there’s a need to consider whether private entities should control such powerful technology or if international cooperation and governance are necessary.

CERN-like Effort for AGI:
Demis Hassabis suggests establishing an international collaborative effort similar to CERN for AGI research. This would involve careful study of AGI’s safety aspects, testing in controlled environments like simulations, and cybersecurity protection.

Recommended Books:
Fabric of Reality by David Deutsch: Poses big questions in physics that AI tools could potentially help address. Permutation City by Greg Egan: Explores thought-provoking ideas related to AI, simulations, and hyper-realistic scenarios. Consider Phlebas by Ian Banks: Depicts an optimistic post-AGI future where humanity flourishes and travels the stars.