TED & TEDx Talks
Ten talks, none longer than twenty minutes
Talks are the highest-reach format in this collection by a wide margin — free, short, and requiring no commitment before the first idea lands. Most of the landmark AI talks are main-stage TED rather than TEDx, which matters if you are looking for one or the other, so both are labelled below. The ordering favours talks that still teach something rather than talks that were merely popular.
TEDx
Independently organised events.
Number 1: The incredible inventions of intuitive AI
2016 · Maurice Conti · TEDxPortland · ~15 min
TED — The incredible inventions of intuitive AI
The most-watched AI talk there is, on design as a collaboration with machines.
Conti's framing of the 'Augmented Age' — tools that participate in the design rather than execute it — reached an audience larger than most of the documentaries and books here combined. The generative-design examples are the clearest visual demonstration of a machine producing solutions a human would not have drawn. Light on limitations, and best paired with something more sceptical.
Number 2: The wonderful and terrifying implications of computers that can learn
2014 · Jeremy Howard · TEDxBrussels · ~20 min
TED — The wonderful and terrifying implications of computers that can learn
The best explanation of why deep learning changed everything, given before it obviously had.
Delivered when deep learning was a research result rather than an industry, Howard walks through what changed and why the consequences would be economic rather than merely technical. It is the rare prediction talk worth revisiting, because the mechanism he describes is the one that actually played out. Still the clearest short account of how learning from data differs from being programmed.
On this siteBackpropagation RevivalNumber 3: The rise of artificial intelligence through deep learning
2017 · Yoshua Bengio · TEDxMontreal · ~17 min
YouTube — The rise of artificial intelligence through deep learning
One of the three Turing laureates explaining his own field's mechanism.
Bengio covers how machines learn representations from data on principles loosely modelled on human cognition, and — unusually for a talk by a principal — is direct about the obstacles still in the way. Worth watching for the source rather than the polish: this is the argument for connectionism from someone who made it when it was unpopular.
Number 4: How I'm fighting bias in algorithms
2016 · Joy Buolamwini · TEDxBeaconStreet · ~9 min
TED — How I'm fighting bias in algorithms
The 'coded gaze' talk that turned algorithmic fairness into a field.
Buolamwini describes discovering that face-detection software would not register her face until she put on a white mask, and generalises it into a research programme. Nine minutes, one demonstration, and a direct line from this talk to the Gender Shades study, Coded Bias, and a decade of audit work. Among the highest impact-per-minute items in this collection.
Main-stage TED
The flagship conference.
Number 5: Why AI is incredibly smart and shockingly stupid
2023 · Yejin Choi · TED2023 · ~16 min
TED — Why AI is incredibly smart and shockingly stupid
Superhuman on benchmarks, defeated by problems a child solves.
Choi sets the benchmark results against simple reasoning failures and asks what the gap tells us about what these systems are doing. It is the most useful corrective on this list for anyone whose picture of the technology comes from demos, and it manages that without dismissing the capability. The best single talk on where the frontier actually is.
Number 6: Can we build AI without losing control over it?
2016 · Sam Harris · TED2016 · ~14 min
TED — Can we build AI without losing control over it?
The ants analogy, delivered to a general audience before the argument was mainstream.
Harris argues that a sufficiently capable system need not be hostile to be dangerous — it need only be indifferent, the way we are indifferent to ants. It is rhetoric rather than research, and its value is as the version of the risk argument that actually reached people. Read Russell for the same concern stated as an engineering problem.
Number 7: 3 principles for creating safer AI
2017 · Stuart Russell · TED2017 · ~17 min
TED — 3 principles for creating safer AI
Machines should be uncertain about what we want — and that uncertainty is the safety mechanism.
Russell, co-author of the field's standard textbook, argues that the problem is not machines pursuing the wrong objective but machines being certain about any objective at all. His alternative — systems that treat human preferences as unknown and defer accordingly — is the most concrete safety proposal on this list, and it comes with the authority of someone who wrote the curriculum.
Number 8: The danger of AI is weirder than you think
2019 · Janelle Shane · TED2019 · ~10 min
TED — The danger of AI is weirder than you think
The risk is not rebellion; it is a machine doing exactly what you asked.
Shane's examples — optimisers that cheat, robots that fall over on purpose because falling scored well — make specification failure funny and therefore memorable. It is the most shareable single item in this whole collection, and it teaches the most important idea in AI safety without using the vocabulary. Start non-technical readers here.
Number 9: What happens when our computers get smarter than we are?
2015 · Nick Bostrom · TED2015 · ~17 min
TED — What happens when our computers get smarter than we are?
Superintelligence as 'the last invention we need ever make'.
The compressed version of the book, delivered a year after it. Parts have dated — the timelines especially — but this is where a large part of the public vocabulary around AI risk entered circulation, and it is worth hearing the argument in its original form rather than through a decade of paraphrase.
Number 10: The inside story of ChatGPT's astonishing potential
2023 · Greg Brockman · TED2023 · ~16 min + Q&A
TED — The inside story of ChatGPT's astonishing potential
OpenAI's co-founder demonstrating the system, then being questioned about shipping it.
The talk itself is a capability demonstration; the reason it is here is the Q&A that follows, in which Chris Anderson presses Brockman directly on the decision to release. It is one of very few recordings of a frontier lab being asked about deployment risk in public and having to answer without a script.