Michael Wooldridge, a veteran AI researcher at Oxford and pioneer of agent-based AI, argues that the history of AI—from Turing to today’s large language models—is essential for cutting through hype, anticipating the future, and rediscovering overlooked ideas that could inspire the next breakthroughs.
He is deeply skeptical of the “Singularity” narrative—the idea that AI will recursively self-improve beyond human control—calling it implausible and a distraction from real, present-day risks like AI-generated fake news, surveillance, and the erosion of shared reality.
The history of AI reveals repeated cycles of boom and bust, where over-optimism leads to disappointment and funding collapse (“AI winters”), and where seemingly failed paradigms often contained valuable ideas that were simply ahead of their time.
The Singularity Is Bullshit
Wooldridge dismisses the Singulative narrative as “deeply implausible,” arguing it distracts from tangible risks and is driven more by primal, almost religious psychology than rational assessment.
The two main arguments for existential risk—the “paperclip maximizer” (AI follows literal instructions with catastrophic unintended consequences) and AI developing its own misaligned goals—both require us to hand over significant control, which he sees as unlikely and avoidable.
He proposes “existential risk risk”: the danger that focusing on speculative future catastrophes diverts attention from real harms happening now, such as AI-generated disinformation drowning elections and fragmenting society’s shared sense of reality.
On regulation, he opposes blanket “neural network laws” (since neural networks are just mathematics) and instead advocates for sector-specific laws targeting harmful uses of AI, especially surveillance, rather than the underlying technology.
Alan Turing: Foundations of Computing and AI
Alan Turing’s work in the 1930s solved a fundamental mathematical problem (the Entscheidungsproblem) by inventing the Turing machine—a mathematical abstraction that, with practical tweaks, became the blueprint for all modern digital computers.
This was a “great irony of mathematical history”: computers were invented as a byproduct of solving a pure math problem, not as a goal in themselves.
By the 1950s, early computers could perform intellectual feats (like complex math) far beyond human capability, sparking public debate about machine intelligence.
Turing formulated the Turing Test to cut through dogmatic debates about whether machines could “really” think: if a machine’s responses are indistinguishable from a human’s, the debate becomes pointless.
Wooldridge interprets this as Turing collapsing the metaphysical question (“does it understand?”) into a practical, behavioral one (“can we tell the difference?”).
He distinguishes between “strong AI” (machines that truly understand and experience) and “weak AI” (machines that simulate intelligence), noting that most AI researchers are pragmatically focused on weak AI.
He rejects efforts to build “moral AI,” arguing that moral responsibility must remain with the humans who design and deploy AI systems, not be abdicated to machines.
The Golden Age of Symbolic AI (1956–1974)
The first boom in AI, called the “Golden Age,” saw rapid progress from nothing to machines that could solve problems, play games, and do mathematics at levels impressive for the time.
The dominant approach was “symbolic AI” (or “modeling the mind”): explicitly programming rational thought processes, search algorithms, and logical rules into machines.
Researchers used a “divide and conquer” strategy, tackling simplified “microworld” versions of real problems (like the Towers of Hanoi or simulated warehouse robots).
Progress stalled in the early 1970s when researchers hit the wall of “combinatorial explosion”—many core AI problems were proven to be NP-complete, meaning the number of possible solutions grows so astronomically that brute-force search becomes physically impossible, even for computers.
The First AI Winter
By the mid-1970s, the combination of over-hyped promises and fundamental technical barriers led to the first “AI winter”: funding dried up, public trust collapsed, and AI researchers were sometimes seen as charlatans.
Wooldridge notes that AI had a poor reputation for decades; as recently as the early 2000s, colleagues warned him that working in AI was bad for his career, and neural networks were regarded as a dead end.
He reflects that working in a “winter” has advantages: fewer researchers mean more intellectual space to explore ideas quietly, without the pressure of a crowded, hype-driven field.
Expert Systems and Logic Programming
The 1980s saw a revival with “expert systems,” based on the idea that intelligence is primarily a problem of knowledge, not just search.
Developers interviewed human experts (e.g., doctors diagnosing blood diseases) and encoded their knowledge as discrete “if-then” rules. The classic example is MYCIN, which could diagnose infections.
This led to “logic programming” (e.g., Prolog), where knowledge is expressed in formal logic, and the system uses built-in logical deduction to derive answers. A planning program called WARPLAN could solve complex problems in just 100 lines of Prolog code.
The ultimate expression of this vision was Douglas Lenat’s Cyc project, which attempted to manually encode all of human knowledge into a vast logical structure. Lenat predicted Cyc would soon be smart enough to write its own rules.
Cyc was widely ridiculed as a failure and became a symbol of AI hype (the unit of “bogusness” was jokingly called a “microLenat”).
However, Wooldridge notes that the intellectual DNA of Cyc lives on in the knowledge graphs behind modern search engines like Google.
Behavioral AI
In the late 1980s, roboticist Rodney Brooks launched a rebellion against symbolic AI, arguing that intelligence does not arise from abstract symbolic reasoning but from the interaction of simple, reactive behaviors in the real world.
His “behavioral AI” approach was bottom-up: start with basic behaviors (like obstacle avoidance), then layer on more complex ones (exploring, picking up objects), letting intelligence emerge from their interaction.
Brooks insisted on embodiment: real intelligence requires a body interacting with the physical world, not a disembodied program in a simulation.
His ideas led to successful real-world robots, most famously the Roomba vacuum cleaner, which uses simple reactive rules rather than complex top-down planning.
The approach hit its own limits as the number of behaviors grew and managing their interactions became combinatorially complex.
Agent-Based AI and Multi-Agent Systems
In the early 1990s, Wooldridge’s own field—agent-based AI—emerged as a synthesis: combining the goal-directed proactivity of symbolic AI with the reactivity of behavioral AI, and adding a third dimension: sociality.
The idea is to change our relationship with software: instead of giving low-level instructions (like using Microsoft Word), we interact with autonomous agents that act on our behalf, cooperating with us and with other agents.
The most visible descendants of this vision are Siri, Alexa, and Cortana. The deeper vision is multi-agent systems: your AI agent negotiating directly with someone else’s AI agent to arrange a meeting, book a trip, or trade stocks.
Wooldridge believes a multi-agent future is inevitable, given the history of computing, but the exact form it will take is still an open research question.
Machine Learning: From Obscurity to Dominance
Machine learning (ML), particularly neural networks, developed largely separately from symbolic AI and was long dismissed as a fringe, “homeopathic” approach.
Key milestones include the development of backpropagation in the 1980s (enabling training of multi-layer networks), the 2000s advent of deep learning, the 2012 realization that GPUs could dramatically accelerate training, and the 2017 invention of the Transformer architecture.
The success of ML represents a poetic full circle: what ultimately worked was not explicitly programming rational thought (the symbolic approach) but imitating the brain’s physical architecture and letting intelligence emerge from data.
Large Language Models and Their Limits
Large language models (LLMs) like GPT-3 and GPT-4, built on the Transformer architecture, are “foundation models” trained on vast amounts of digital text to predict the next token (word) in a sequence.
They are genuinely dazzling: capable of conversation on any topic, generating text, and performing a wide range of language-related tasks. Wooldridge admits they took him by surprise.
However, he argues they are not “the end of the road” for AI because they are fundamentally disembodied: they exist as paused programs, not agents embedded in and interacting with the real world.
We can chat with ChatGPT about quantum mechanics, but we don’t have robots that can clear a dinner table and load a dishwasher. Real-world robotics remains extremely hard.
He is skeptical that LLMs truly reason: he distinguishes between genuine problem-solving and sophisticated pattern recognition.
Evidence: if you rephrase a planning problem using novel words the model has never seen, it fails, suggesting it is matching patterns from its training data rather than reasoning from first principles.
He lists logical reasoning, abstract reasoning, planning, and (until recently) arithmetic as capabilities still beyond LLMs’ true reach, even if they can sometimes produce plausible-looking outputs.
He sees the Transformer architecture as an engineering hack, not a deep model of mind, and doubts that simply scaling up data and compute will yield true reasoning or robotic intelligence.
The Bitter Lesson and the Future
Wooldridge acknowledges a “sobering” and slightly “melancholy” lesson from AI’s history: the biggest advances have come not from clever new architectures or deep scientific insights, but from brute-force scaling of compute and data (what Rich Sutton calls “The Bitter Lesson”).
Despite this, he finds the present moment “mind-boggling” and “genuinely exciting”: questions once reserved for philosophers (like “can machines think?”) are now experimental science.
He believes the path forward will require more than just scaling Transformers: it will involve understanding embodiment, the functional structure of the brain, and the evolutionary context that shaped human intelligence (interacting with the physical world and with other humans).
The success of black-box neural networks over explicit symbolic systems like Cyc may also teach us a humbling lesson about our own intelligence: perhaps human reasoning is less rational and more pattern-based than we like to believe.
