Graphcore’s Nigel Toon on advances in semis and an AI-powered future

Current Semiconductor Achievement and Moore's Law Trends

• Modern chips now contain 100 billion transistors, representing a 25-billion-fold improvement since the first integrated circuit (4 transistors) in 1960.
• Semiconductor progress is currently slowing as the industry reaches the top of the S-curve; future gains will be slower and require significantly more effort.
• Recent GPU compute increases (e.g., Blackwell) are driven by reducing floating-point precision (from 64-bit to 4-bit) rather than Moore's Law scaling.
• Alternative computing paradigms like quantum and molecular computing remain unproven at scale or are estimated to be 20+ years away.

AI Timeline and Breakthrough Milestones

• The 2012 AlexNet model marked the first deep learning breakthrough, achieving object recognition better than human experts within a year.
• Public AI consciousness was triggered in 2016 (AlphaGo in China, watched by 240 million) and 2022 (ChatGPT).
• Future AI growth relies on infrastructure layers (compute and information) where value will eventually shift from hardware creators (e.g., NVIDIA) to application developers.

Strategic Evolution of Graphcore

• Graphcore was founded in 2016 to develop specialized processors, recognizing early that AI required hardware beyond standard silicon architectures.
• The company's approach emphasizes both hardware innovation and the critical importance of software ecosystems.
• Founders have prior successful exits:
  • Altra (14-year tenure): Went from 100 to 2,500 employees; acquired by Intel for ~$20B.
  • iSerum (founded 2002): Built cellular baseband processors; sold to NVIDIA (estimated equivalent value of ~$100B in today's terms).

Societal and Economic Impact of AI

• AI is projected to deliver 150 million intelligent machines by the near future, capable of operating as CEOs, government advisors, and eventually governing entities.
• Historical data indicates technology rarely replaces entire occupations; since 1950 in the US, only elevator operators have been completely replaced, while most jobs are complemented.
• Economic value in AI, similar to electricity and the PC industry, will initially accrue to infrastructure providers but will eventually shift to the vast ecosystem of applications and services built on top.
• Productivity gains will come from reshaping business systems (combining human and machine intelligence) rather than simple task automation.

Scientific and Biological Applications

• DeepMind's protein modeling is mapping every protein in the human body to enable targeted drug discovery with zero side effects (e.g., curing cancer).
• AI is accelerating nuclear fusion research by modeling plasma control, potentially making fusion energy commercially viable within 5–10 years.
• Molecular computing is proposed as a future alternative to silicon, leveraging the brain's efficiency (1 exaflop on 25 watts) which is a million to a billion times more energy-efficient than current silicon data centers.

Future Education and Workforce Adaptation

• Core education will focus on fundamental skills in math, science, arts, and humanities rather than specific programming languages, as AI can generate code and models from natural language prompts.
• Future professionals will use AI frameworks to create models (e.g., in quantitative trading or drug discovery) without needing to write the underlying code.
• The "operator" role in AI will involve defining problems and creative direction, while the machine handles execution.

Ethical Considerations and Regulation

• AI carries risks similar to other powerful technologies (e.g., guns), where bad actors could cause harm, necessitating robust regulation.
• Nigel Toon serves on the UK Research and Innovation Board, overseeing funding for universities and research councils to address these scientific and ethical challenges.

Theoretical Foundations of the Digital Age

• Modern computing rests on three pillars: the transistor (1947), the bit/information theory by Claude Shannon (1948), and artificial neural networks (1980s).
• Quantum computing utilizes qubits (superposition states) that exponentially increase compute power but face significant scaling challenges regarding noise and stability (Schrödinger's cat problem).
• Current limitations in quantum computing include the inability to scale beyond a few hundred qubits without destroying the experimental state through measurement.