DeepMind CEO Believes Human-Level AI May Be Possible Within a Decade

The quest to create artificial general intelligence (AGI) – machines with the full breadth of cognitive capabilities as humans – has been the guiding North Star of the field of AI since its inception. In recent years, the relentless march of progress in AI research has led some prominent experts to suggest this lofty goal may be achievable sooner than we think.

Most notably, Demis Hassabis, the co-founder and CEO of Google‘s AI division DeepMind, recently stated that he believes human-level AI could potentially be developed "within 10 years or so." Hassabis made the comment during an interview at the Wall Street Journal‘s Future of Everything Festival in May 2023.

As one of the most influential and respected figures in AI, Hassabis‘ prediction carried significant weight and sparked much discussion in the field. Hassabis has been at the forefront of AI research for over two decades, with a string of high-profile breakthroughs that have redefined what‘s possible with machine learning.

In 2010, he co-founded DeepMind with the mission of developing artificial general intelligence and using it to benefit humanity. The company made history in 2016 when its AlphaGo system defeated world champion Lee Sedol at the ancient Chinese board game Go – a feat previously thought to be at least a decade away. DeepMind has since produced groundbreaking work in areas like protein folding, nuclear fusion, and game theory.

So what led Hassabis to make such a bold prediction about AGI? And what would it take to actually achieve human-level AI in the next decade? Let‘s dive in.

The State of the Art in AI

To understand why AGI may be on the horizon, it‘s important to first examine the current state of the art in artificial intelligence and the rapid progress that has been made in recent years.

The field of AI has been transformed over the past decade by the resurgence of neural networks and deep learning. These techniques have allowed machines to achieve superhuman performance on a range of cognitive tasks such as image classification, speech recognition, and natural language processing.

Some of the most impressive AI systems developed in recent years include:

• GPT-4: OpenAI‘s latest language model can engage in open-ended conversation, answer follow-up questions, and even analyze and explain its own outputs. It outperforms most humans on a range of academic and professional benchmarks.

• PaLM: Google‘s Pathways Language Model is a massive 540-billion parameter language model that can perform tasks like code generation, mathematical reasoning, and question-answering at a near-human level.

• Chinchilla: DeepMind‘s 70-billion parameter language model matches the performance of much larger models, demonstrating improved parameter efficiency. It showcases few-shot learning abilities on a range of datasets.

• Gato: Another DeepMind model, Gato is a single transformer neural network that can perform over 600 distinct tasks including image captioning, robotic manipulation, and playing Atari video games. This type of multi-modal, multi-task learning is seen as a key stepping stone to AGI.

However, despite these impressive achievements, today‘s AI systems are still considered "narrow" or "weak" AI – they can only perform the specific tasks they were trained on and cannot generalize their knowledge to new domains like humans can. Artificial general intelligence, in contrast, would have the flexibility and adaptability to learn and reason about any topic, combining different skills and knowledge bases to solve novel problems.

The Challenges of Achieving AGI

So what stands between today‘s narrow AI and full-blown AGI? Experts point to a number of key challenges that will need to be overcome, both technical and philosophical.

One major hurdle is that we still lack a clear definition and criteria for what exactly constitutes artificial general intelligence. The famous Turing Test, proposed by Alan Turing in 1950, judges a machine‘s intelligence by its ability to engage in open-ended conversation indistinguishably from a human. But this narrow skill is not sufficient to capture the full scope of human intelligence.

Jose Hernandez-Orallo, an AI researcher at the Universitat Politècnica de València, has proposed a more comprehensive framework for measuring machine intelligence based on a set of over 100 cognitive tests spanning perception, reasoning, memory, planning, creativity, and more. Truly achieving AGI will likely require machines to match or exceed human-level performance across this full battery of tests.

There are also still significant limitations in the core techniques used in today‘s AI systems. Deep learning models are extremely data-hungry, require massive computational resources to train, and are often brittle and inflexible when presented with scenarios outside their training data. Cutting-edge language models like GPT-4 and PaLM still sometimes hallucinate false information, struggle with complex reasoning, and lack common sense understanding of the world.

Overcoming these limitations will likely require fundamental algorithmic breakthroughs beyond current approaches, as well as continued exponential growth in compute power. Some experts estimate that matching the computational capacity of the human brain to power an AGI could require over 1000x the compute that the largest AI models use today.

There are also immense challenges around goal-setting and control for AGI systems. How can we specify an advanced AI‘s objective function to ensure it behaves in alignment with human values? This is known as the "AI control problem" – the difficulty in maintaining control over an AI that may become more intelligent and capable than its human creators.

Techniques like inverse reward design and debate have been proposed to try to solve this problem, but a provably safe and reliable solution remains elusive. Philosophical questions also abound – would a sufficiently advanced AGI be conscious and deserve moral consideration? Could an AGI be considered a legal person with rights and responsibilities?

The Economic and Societal Impacts of AGI

If human-level AI is developed in the coming years as Hassabis and others suggest, it would undoubtedly be one of the most transformative events in human history, with immense consequences across every facet of society.

On the positive side, an AGI could help solve many of the greatest challenges facing humanity today. It could accelerate scientific discovery and technological innovation, finding cures for diseases, developing clean energy solutions, and expanding our knowledge and capabilities beyond what is currently possible.

However, the specter of mass technological unemployment looms large. A recent report by McKinsey Global Institute estimated that up to 800 million jobs worldwide could be displaced by automation by 2030, even with narrow AI. An advanced AGI could potentially obviate the need for human cognitive labor altogether, at least in its current form.

This has led many experts to advocate for policies to help manage this transition, such as universal basic income, job retraining programs, and expanded social safety nets. But the full societal ramifications of artificial general intelligence remain difficult to predict.

The Path Forward for Responsible AGI Development

With the immense potential benefits and risks of AGI on the horizon, many are calling for renewed focus and expanded efforts around the responsible development of artificial intelligence.

Key principles of responsible AI development include transparency, fairness, accountability, robustness, and ensuring the alignment of AI systems with human values. This will require embedding ethical considerations into every stage of the AI development process and expanding collaborations between AI researchers, policymakers, ethicists, and the broader public.

A number of initiatives are working to guide the responsible development of AGI and ensure it benefits humanity. OpenAI is a non-profit research company that has made AGI development its core mission, with a focus on ensuring the technology is deployed safely and equitably. The Future of Humanity Institute at Oxford University conducts research on the long-term impact of artificial intelligence and other emerging technologies. And the Machine Intelligence Research Institute is focused on developing the mathematical foundations for safe and beneficial AGI.

Google, Microsoft, and other major tech companies have also formed AI ethics boards and signed pledges to develop AI responsibly. But some worry that the competitive pressures of the AI arms race could lead to corners being cut on safety.

As Demis Hassabis said in his interview, achieving AGI is something we need to do slowly and carefully to reap the benefits while mitigating the risks. With further progress likely to only accelerate in the coming years, now is the time to expand these efforts and ensure we get it right. The future of humanity may depend on it.