The AI revolution didn’t happen overnight—here’s why | Michael Wooldridge | TEDxManchester

Main ideas, concepts, and lessons

• AI didn’t start recently

  • The talk argues that artificial intelligence has roots in the late 1940s / early 1950s, tied to the first digital computers built at the University of Manchester.
  • Early computers could perform large volumes of mathematics quickly and accurately, raising the question: could machines be truly intelligent?

• Alan Turing helped launch the field

  • The talk highlights Alan Turing as both a key computer inventor and a foundational figure for AI.
  • 1950: Turing Test
    • Concept: a machine is considered intelligent if it can convince humans it is indistinguishable from human intelligence.

• Early optimism gave way to slow progress

  • Within a decade, AI-like systems could do practical tasks (e.g., playing games like chess/checkers and performing complex math).
  • However, the overall historical pattern is described as “progress is slow” for much of AI’s history.

• Mid-1980s breakthrough: neural networks

  • A major early AI idea emphasized by the speaker is neural networks.
  • The talk attributes the neural-network revival/technology development to Jeff Hinton and colleagues (mid-1980s onward).
  • Core inspiration:
    • Human brains contain many interconnected neurons
    • Artificial neural networks aim to reproduce similar network behavior in software

• The “AI revolution” required more than the science

  • Even though key neural-network ideas existed in the 1980s, the talk says the major revolution didn’t occur then because:
    • Computers weren’t powerful enough
    • Sufficient training data wasn’t available
  • Around 2005, conditions changed:
    • Computing power increased
    • Data availability improved
    • Neural networks (rebranded as “deep learning”) began working at scale

• Lesson about scientific culture: rebranding can matter

  • The talk describes neural networks as unfashionable around the turn of the century (rejected as “pseudoscience”).
  • Key lesson: when a field is dismissed, giving it a new name can help it regain legitimacy.
  • “Deep learning” is presented as essentially neural networks under a more acceptable label.

• 2012: GPUs supercharged neural networks

  • In 2012, Jeff Hinton and others recognized that GPUs (graphics processing units) were ideal for neural networks.
  • Why GPUs mattered:
    • They enable high-throughput/high-performance computation
    • Using GPUs gave about 10× better cost/performance for training large neural networks
  • Anecdote: GPU demand led to a playful story of researchers buying up GPU cards from game shops (Toronto and New York), highlighting how quickly resources were needed.

• 2017: Transformer architecture (“Attention Is All You Need”)

  • In 2017, a Google research paper titled “Attention Is All You Need” introduced the transformer architecture.
  • Transformer purpose:
    • Given a sequence of words, predict the next word
  • Reception:
    • It was well received but not immediately a viral “game changer.”

• OpenAI + scaling strategy → GPT-3 → ChatGPT

  • The talk claims OpenAI saw potential and was backed by Microsoft.
  • Strategy described:
    • Scale up data and compute dramatically (10×, then 10× again)
  • June 2020: GPT-3
    • Presented as a large language model with a step-change in capability.
  • ChatGPT
    • Uses GPT technology and “went viral,” reaching 100 million users in about 6 weeks (as stated in the subtitles).
  • Result:
    • A massive corporate pivot (e.g., Apple, Microsoft Copilot mentioned) to embed AI widely.

• Current state: impressive but uneven capabilities

  • The speaker distinguishes:
    • Clever AI: can converse fluently on complex topics (quantum mechanics, Roman Empire history) and solve advanced math problems.
    • Gimmicky AI: flashy but shallow uses (cat ears, meme-like edits, brief “cute” effects).
    • Lack of useful everyday intelligence: AI still can’t do basic household tasks reliably (e.g., understanding a new room situation, locating a kitchen, clearing a table, loading a dishwasher).
  • Key “disparity”:
    • AI is extremely capable in some dimensions (language/reasoning-like tasks)
    • But fails at tasks that resemble what a low-wage human worker can do

• Big open question: will AI solve productivity?

  • The talk emphasizes large investments in AI and the expectation that AI will solve the productivity problem.
  • The speaker frames this as uncertain:
    • Are we in a bubble?
    • Will trillions of dollars pay off?
  • The likely “lived experience” over the next few years will be whether AI becomes truly useful in everyday work and life.

---

Methodology / “instruction-like” content (as presented)

• When a scientific area is unfashionable, rebrand it

  1. Identify that a field is dismissed as pseudoscience / not fundable
  2. Rename the work (e.g., “neural networks” → “deep learning”)
  3. Use the renamed identity to regain credibility and attract investment

• How scaling turned research into major products (described strategy)

  • Start from a capable architecture (transformers)
  • Increase:
    • Data size (e.g., 10×)
    • Compute / computer power (e.g., 10×)
  • Continue scaling to achieve next-generation performance jumps
  • Package the result into interactive systems (GPT-3 → ChatGPT)

---

Speakers / sources featured (as stated in the subtitles)

• Michael Wooldridge (speaker; TEDxManchester talk)
• Alan Turing (credited with founding AI concepts; Turing Test)
• Jeff Hinton (credited with neural network ideas/technology development; GPU push)
• Google Research Lab (source of “Attention Is All You Need” paper, 2017)
• OpenAI (developed GPT-3 and integrated into ChatGPT, per the talk)
• Microsoft (provided backing; mentioned as competitor to Google and as funder of OpenAI)
• Nvidia (associated with GPUs; described as the key GPU company)
• TEDxManchester (platform/context for the talk)
• Oxford (mentioned as a place where people could check Latin—contextual reference)