At Rise of AI 2018 in Berlin, Peter Bentley, CTO at Braintree, argues that most of what is called AI today is "weak AI": systems good at handling variety within narrow functional limits, such as reading license plates, but unable to do anything outside that task. He contrasts this with "strong AI", which has genuine intelligence, can tackle a broader range of applications, and can learn to do things outside its current scope. Bentley tests the audience with examples including Siri, which he shows misunderstanding a live query through statistical guessing rather than understanding, Amazon Echo, the Never-Ending Language Learner, the Cyc project, and DeepMind's AlphaZero, arguing each is a form of weak AI because it manipulates symbols or statistics without real understanding, echoing the philosophical "Chinese Room" argument. He then outlines what building strong AI would require: not imitating human intelligence, growing a broad range of tasks, allowing new algorithms to emerge, and doing far more testing than today's systems receive. He closes by arguing that strong AI may eventually be needed as companions and advisors to help people manage an increasingly complex world, while cautioning that simpler, explainable systems remain preferable for most applications, especially safety-critical ones.
Key takeaways
- Weak AI handles variety only within a narrow, specific task, while strong AI can tackle a broader range of applications and learn beyond its original scope.
- A recorded demo shows Siri repeatedly misunderstanding a question, illustrating that it guesses statistically rather than understanding meaning.
- DeepMind's AlphaZero is still classed as weak AI because it handles one task at a time, needs heavy developer input, and suffers catastrophic forgetting.
- The no-free-lunch theorem means no single algorithm works best for every problem, so strong AI likely needs many different specialized methods, like the brain's varied circuits.
- Giving an AI more capability requires exponentially more testing, which is why driverless cars still underperform after millions of test miles.
Detailed notes
AI's ups and downs, and a career in it
Bentley describes starting in AI from childhood, building robots and programming computers, and earning a Bachelor of Science in artificial intelligence in the late 1980s or early 1990s, when only three UK universities taught AI because the field was considered discredited; his own university, University College London, avoided the term and called its department "intelligent systems" instead. He notes AI is as old as computer science itself, tracing it to pioneers including John von Neumann, Alan Turing, Claude Shannon and Marvin Minsky, and to the first AI conference in the 1950s. Having lived through AI's unfashionable periods, he expects that if current hype outruns results, the field could dip again, though people will keep working on the underlying technology regardless of the label.
Defining weak AI and strong AI
Bentley defines weak AI as AI that handles variety and noise well but only within a narrow, specific application, such as reading vehicle license plates despite variation in lighting, plate style and size; asked to drive a car or play chess, such a system simply cannot. He calls this a simulation of a part of human intelligence, not the real thing. Strong AI, by contrast, handles variety across a broader range of applications, has wider expertise, and critically can learn to do things outside its current scope; it need not resemble human intelligence to genuinely possess intelligence within its own framework. He frames the distinction with the idea that a weak AI can look and act intelligent while being, in his words, a cardboard cutout with a tape recorder behind it, rather than genuinely intelligent.
Testing the audience: Siri, Amazon Echo, NELL, Cyc, AlphaZero
Bentley runs an audience quiz on several well-known systems. He plays a recording of Siri repeatedly failing to answer a question about the psychologist Hans Eysenck, misreading the query as unrelated searches, to argue Siri is weak AI that uses statistics to guess the most likely meaning of speech rather than understanding it, which is why it sometimes changes its answer mid-response. He argues Amazon Echo suffers the same limitations. The Never-Ending Language Learner, which mines the web to build a database of facts with confidence scores, is shown producing both plausible and clearly wrong facts (for example, classifying vitamin A as a video game with 100 percent probability), which he says shows it manipulates symbols without knowing what they refer to. The Cyc project, which collects common-sense rules of thumb, is described the same way: it holds correct-sounding rules without understanding what the underlying concepts mean. DeepMind's AlphaZero is treated as the strongest example shown, a neural, sub-symbolic system that can be retrained across different games such as Go and chess, but Bentley still calls it weak AI because it only handles one task at a time, needs heavy developer input to acquire new knowledge, and, like many neural networks, suffers from catastrophic forgetting.
The Chinese Room argument
Bentley introduces John Searle's Chinese Room argument to explain why symbol manipulation is not the same as understanding: a person inside a room who does not know Chinese can still produce correct Chinese answers to Chinese questions by looking them up in filing cabinets, appearing to converse fluently while understanding nothing. He argues this is what most chatbots and similar systems, including Siri, are doing today: manipulating symbols convincingly without understanding meaning.
What it would take to build strong AI
Asked what the world's strongest AI is, Bentley says there is no answer yet, only systems that are relatively stronger because they generate their own solutions with less developer input. He lists what he believes strong AI requires: first, not trying to simulate human intelligence specifically, since true intelligence takes many forms in nature, from dolphins and ants to the immune system; second, giving an AI a growing range of tasks to solve, since intelligence emerges from the need to handle difficulty and generality rather than specialization; third, allowing the AI to build new algorithms for new tasks rather than relying on one method such as deep learning, citing the "no free lunch" theorem that no single algorithm is best for every problem, mirrored by the brain's many different neural circuits; and fourth, far more testing, since giving an AI more capability requires exponentially more testing to verify, which is why he says driverless cars still do not work as well as hoped after millions of test miles, and why he is skeptical of rapid singularity-style progress.
Why we need strong AI
Bentley argues most applications do not need strong AI and that simpler, explainable systems are preferable, especially for safety-critical work such as an autopilot system his company is building for airliners, which must be explainable, testable and certifiable. But he argues strong AI may eventually be needed as the world becomes too complex for people to manage alone, comparing this to struggling to operate his own home television, and suggests strong AI could act as companions, teachers and advisors that genuinely understand people's needs, emotions and surroundings rather than merely simulating understanding.
Quotes
It's a complete red herring to try and simulate human intelligence.
Peter Bentley
There is no one-size-fits-all for intelligence.
Peter Bentley
Do not underestimate the need for testing.
Peter Bentley
Questions answered in this talk
What is the difference between weak AI and strong AI?
According to Peter Bentley at Rise of AI 2018, weak AI handles variety and noise well but only within a narrow, specific task and does not really understand what it processes, while strong AI has genuine intelligence, can tackle a broader range of applications, and can learn to do things outside its original scope.
Is Siri an example of strong AI?
According to Peter Bentley at Rise of AI 2018, Siri is weak AI: a live demo shows it repeatedly misunderstanding a question about a psychologist because it statistically guesses the most likely meaning of speech rather than genuinely understanding it.
Why is it hard to build strong AI today?
According to Peter Bentley at Rise of AI 2018, strong AI requires a growing range of tasks, new algorithms rather than a single method, and exponentially more testing as capability increases, which is why even heavily tested systems like driverless cars still fall short.