Demis Hassabis (DeepMind Co-founder) – The Future of AlphaGo and AI (Sep 2017)

Chapters

Abstract

Revolutionizing Intelligence: DeepMind’s Journey from Atari to AlphaGo and Beyond

DeepMind, under the visionary leadership of Demis Hassabis, has made groundbreaking strides in the field of artificial intelligence, especially with its development of AlphaGo. This article delves into DeepMind’s journey, from its foundational focus on general-purpose learning systems and the pursuit of Artificial General Intelligence (AGI) to its remarkable achievements with AlphaGo in the complex game of Go. We explore the innovative techniques, challenges, and real-world applications of this transformative technology, highlighting the profound impact of AI on scientific exploration, ethical considerations, and the potential to unravel the mysteries of the human mind.

DeepMind’s Mission and Approach:

DeepMind’s mission revolves around understanding and recreating intelligence artificially, with an aim to build general-purpose learning systems. These systems are envisioned to operate across a wide spectrum of tasks, emulating the learning and adaptation abilities inherent in the human mind. To achieve this ambitious goal, DeepMind employs a variety of cutting-edge techniques, including deep learning, hierarchical neural networks, and reinforcement learning.

Demis Hassabis’ Vision for AGI:

Demis Hassabis, the driving force behind DeepMind, envisions Artificial General Intelligence (AGI) as a system distinct from narrow AI. Unlike narrowly focused AI systems, AGI is designed to learn and adapt autonomously through reinforcement learning. This allows AGI to tackle a variety of problems without needing explicit instructions. Hassabis contrasts this with traditional AI systems, such as IBM’s Deep Blue, which, despite defeating chess grandmaster Garry Kasparov, were pre-programmed with knowledge, thus placing the intelligence in the programmers rather than the system itself.

Reinforcement Learning and Breakthrough in Atari Games:

DeepMind’s initial forays into AI were marked by significant success in developing agents capable of playing classic 8-bit Atari games. Their approach, combining deep learning with reinforcement learning in what’s known as Deep Q-Networks (DQM), enabled these agents to master games to an impressive degree. Notably, an agent designed to play Breakout learned the game by simply observing the raw pixels on the screen, with the objective of maximizing the score. Through practice, the agent developed sophisticated strategies, such as creating a tunnel to send the ball behind the wall.

The Go Challenge and AlphaGo’s Solutions:

In their quest to advance AGI, DeepMind took on the ancient and notoriously complex game of Go. The traditional brute-force search approach, effective in chess, proved inadequate for Go due to its immense search space and the intuitive nature of gameplay. DeepMind’s solution, AlphaGo, incorporated two neural networks: a policy network for suggesting promising moves and a value network for evaluating positions. This innovative approach empowered AlphaGo to make strategic decisions, exemplified by the now-famous move 37 in Game 2 against Lee Sedol, which defied conventional Go strategies.

AlphaGo’s Impact on Go and Beyond:

AlphaGo’s match against Lee Sedol in Seoul not only captured global attention but also revolutionized the game of Go. The event, drawing immense online viewership and extensive press coverage, played a pivotal role in popularizing Go in the West, as evidenced by a dramatic increase in online board sales. Professional players, including Lee Sedol, found new inspiration in AlphaGo, considering their encounters with the AI as among the greatest experiences of their lives. The machine’s novel approach to the game challenged traditional strategies and highlighted the potential of AI-assisted creativity. This milestone in AI development spurred widespread interest in the game and influenced the adoption of AlphaGo’s techniques in various fields.

Evaluation Function in Go:

In contrast to chess, where the concept of materiality aids in evaluating positions, Go presents a unique challenge due to its constructive nature. The game begins with an empty board, and as it progresses, predicting the future becomes increasingly challenging. Small positional changes can have significant implications, making the evaluation of positions in Go more complex than in chess.

AlphaGo’s Creative Move:

AlphaGo’s move 37 in Game 2 against Lee Sedol stunned the Go community, showcasing the AI’s creative capabilities. This move, which challenged traditional beliefs about the balance between influence and territory, highlighted possible oversights in human play. The objective of winning in Go provides a framework for evaluating such innovative moves.

The Union of Human and AI for Scientific Advancement:

The collaboration between human ingenuity and AI capabilities, as seen in AlphaGo’s achievements, offers promising prospects for scientific and technological advancement. This synergy is crucial in tackling complex global challenges and enhancing scientific exploration. AlphaGo, along with its advanced version, AlphaGo Master, exemplifies the potential of AI in contributing strategically to various fields, including material design, drug discovery, and energy optimization.

Ethical Considerations and Future Prospects:

As AI continues to evolve, addressing ethical considerations becomes increasingly important. DeepMind emphasizes the responsible use of AI technologies to ensure societal benefits and prevent misuse. The application of AI in real-world scenarios, such as optimizing energy consumption in data centers, demonstrates a future where AI can assist in diverse domains, including healthcare, education, and environmental sustainability.



DeepMind’s journey, from mastering Atari games to revolutionizing Go with AlphaGo, reflects its broader mission to understand and recreate intelligence. This quest extends beyond gaming, encompassing the potential to unravel mysteries of the human mind and address global issues. As DeepMind continues to push AI boundaries, it paves the way for a future where the synergistic relationship between humans and AI leads to unprecedented advancements and discoveries.

Supplemental Information:

AlphaGo’s impact on Go and the inspiration it provided to players was immense, drawing 280 million online viewers and over 35,000 press articles. The match popularized Go in the West, significantly increasing online board sales and inspiring players like Lee Sedol, who considered their experiences with AlphaGo as life-changing.

Defining intuition and creativity in AlphaGo, we see that intuition is an implicit knowledge gained through experience but not easily expressible. AlphaGo’s performance in Go demonstrated this kind of human-like intuition through its moves and evaluations. Creativity, on the other hand, involves synthesizing knowledge to generate novel ideas, which AlphaGo exemplified in its innovative strategies and tactics during matches.

AlphaGo Master, the successor to AlphaGo, aimed to perfect its Go skills and fill knowledge gaps. Its 60-match winning streak on Go servers introduced new ideas that professional players have since adopted, indicating the start of a new era in the game.

Comparing AlphaGo with top chess programs, we find differences in their strengths. While chess programs excel in tactical play and calculation, AlphaGo shows strengths in strategic thinking and innovative ideas. The union of human and AI players in both games suggests the potential for discovering even better strategies.

AlphaGo is viewed as a general-purpose tool for scientific exploration, akin to a microscope or telescope. Its potential applications in material design, drug discovery, and other areas signify vast possibilities, with its techniques already being successfully applied in optimizing Google data centers, resulting in significant energy savings.

DeepMind’s mission is to develop AGI as a meta-solution to complex challenges like information overload, climate change, and diseases. AGI can enable AI-assisted science to accelerate scientific breakthroughs. The ethical and responsible use of AI is vital for ensuring societal benefits, a principle central to DeepMind’s philosophy.

Demis Hassabis has a personal dream of making AI scientists or AI-assisted science a reality. His interest in understanding the human mind drives him to explore both artificial systems and neuroscience. This pursuit may provide insights into the mysteries of our minds, including consciousness, dreaming, and creativity.

The journey to build general intelligence could illuminate our understanding of consciousness and its relation to subjective experiences, thoughts, and feelings. It might also shed light on dreaming, its purpose, and connection to our conscious minds.

Exploring creativity and general intelligence, DeepMind’s work could deepen our understanding of the mechanisms of creativity and how it can be nurtured and developed. Ultimately, the development of general intelligence may serve as a powerful tool for comprehending our own capabilities, limitations, and the complexities of the human mind.

Notes by: OracleOfEntropy