Shane Legg on AGI, DeepMind, and the risks of machine intelligence

Executive overview

Human intelligence has shaped everything around us — and machine intelligence is about to do the same, at far greater scale. Shane Legg, co-founder of Google DeepMind, has been arguing since 2001 that artificial general intelligence is not a distant fantasy but an imminent reality, with a 50% probability estimate of AGI by 2028.

DeepMind's work on AlphaGo and protein folding shows how deep learning can crack problems that stumped researchers for decades. The same techniques, scaled further, point toward systems that may exceed human cognitive ability across every dimension.

The core tension: the people building AGI believe it poses existential risk — and also believe it cannot be stopped, making safety research the only responsible path forward.

From childhood computer to AGI research

• A VZ200 computer at age 10 sparked a lifelong obsession with programming and machine-made worlds.
• Early machine learning work in the late 1990s involved small, printable datasets and basic classification tasks (e.g. predicting disease from blood results).
• Reading Ray Kurzweil's The Age of Spiritual Machines convinced Legg that powerful general AI was coming within decades.
• Legg proposed the term "artificial general intelligence" around 2001, defining it as: a system that can perform all cognitive tasks humans typically can, and possibly more.
• His 2001 estimate — 50% chance of AGI by 2028 — remains roughly his current view.

How DeepMind was built

• Founded in September 2010 with a one-sentence business plan: "Build the world's first artificial general intelligence."
• AlphaGo combined Monte Carlo Tree Search (MCTS) with deep learning — one network to evaluate good moves, another to predict who would win.
• The system trained by playing against itself, using wins and losses as the learning signal.
• Go required this approach because the branching factor is far too large for brute-force search.
• Google acquired DeepMind in 2014 for a reported $500–650M when the company had roughly 75 employees.

AlphaFold and protein folding

• Proteins fold into 3D shapes that determine their biological function — but predicting shape from atomic sequence had stumped researchers for decades.
• Traditional methods cost ~$200,000 and took years per protein; DeepMind solved this computationally.
• AlphaFold folded every protein known to science and released the results publicly; 1.7 million researchers have used the database.
• The tool doesn't solve drug discovery directly, but illuminates which proteins interact and where to target — a major accelerant for pharmaceutical research.

The transformer and the scaling surprise

• Google invented the Transformer algorithm and used it initially for language translation.
• OpenAI recognized its scalability potential and scaled it aggressively — producing results that surprised even its inventors.
• Large language models outperformed expectations as model size and training data grew, prompting Google to merge DeepMind with its AI division in 2023 to consolidate resources.

AGI, ASI, and what comes next

• A modern supercomputer has ~1 million times the energy consumption, physical space, and signal frequency of the human brain — with signals traveling at the speed of light vs. 30 m/s in neurons.
• AGI is the entry point: machines matching human cognitive breadth. Artificial superintelligence (ASI) is what may follow.
• Legg believes some dimensions of capability will scale far beyond human levels; others may plateau — the limits are not yet known.
• The trajectory: AGI may not stop at human-level performance but keep improving, compressing decades of human intellectual progress into shrinking timeframes.

Safety, risk, and why Legg keeps building

• Legg signed the 2023 letter stating "mitigating the risk of extinction from AI should be a global priority."
• His view: the genie cannot be put back in the bottle — intelligence is too valuable for any global agreement to halt its development.
• The analogy he uses is the Industrial Revolution: transformative, impossible to fully anticipate in advance, with both enormous benefits and enormous harms.
• He led DeepMind's AGI safety group for years and consults governments including the UK on safety challenges.
• Individual prepping (bunkers, relocation) is pointless — the only meaningful leverage is advocating loudly for safety research and broad societal engagement.
• Governing AI is not a technical problem solvable by a few insiders; it requires input from policymakers, social scientists, ethicists, and the public.