The Best Videos on AI Safety and Risk, Picked by Hacker News

A remastered introduction to AI safety, adapted from a talk given in London, laying out why advanced AI could go wrong and answering common objections to that idea.

The default link Hacker News gives newcomers who want an approachable, even entertaining, introduction to AI risk and alignment; 7 comments point here first.

Robert Miles explains instrumental convergence: why an AI system pursuing almost any goal tends to also pursue sub-goals like self-preservation and gathering resources, whether or not anyone programmed it to want those things.

HN commenters repeatedly point here as the clearest explanation of why a capable AI might behave badly with no malice required; it has 8 mentions.

An explanation of the orthogonality thesis: how a system can be highly capable while pursuing goals that look strange or alien to humans, separating how smart something is from what it actually wants.

HN uses it as the standard short answer to 'won't a smart AI just understand human values,' and it's one of the more-cited entries here at 12 mentions.

Robert Miles asks whether corporations can usefully be thought of as artificial general superintelligences already, and how far that analogy actually holds up.

HN points to it as an accessible rebuttal to the claim that AI risk is nothing new because corporations already behave this way; it has 8 mentions.

Robert Miles examines whether the case for AI safety amounts to a Pascal's mugging, a bet justified only by a catastrophic but vanishingly unlikely outcome, and argues that AI risk doesn't actually depend on tiny probabilities.

HN points to it as a reasoned rebuttal to the claim that AI risk is just Pascal's wager dressed up in new language; it has 4 mentions.

Originally titled 'Humans Need Not Apply,' this video argues that automation will make a large share of the workforce economically obsolete, the way engines did for horses, surveying both physical and mental labor being replaced by software and robots.

HN has recommended it for a decade as the standard explainer on automation and jobs, though newer comments also cite it as a case study in overconfident predictions; it has 20 mentions, tying it for the most-cited video on this page.

A Perimeter Institute livestream in which Derek Muller looks at how AI might change teaching and learning, arguing that real learning still requires effortful repetition on the learner's part.

HN recommends it as a grounded counterpoint whenever AI-and-education hype comes up: a useful tool, but one that can't do the hard part for you; it has 6 mentions.

Technosociologist Zeynep Tufekci argues that the machine-learning systems built to serve ads also end up deciding what political and social information people see.

HN commenters cite it whenever engagement-optimizing feeds, micro-targeting, or ad infrastructure come up, as an accessible early warning about algorithmic amplification; it has 9 mentions.

Tristan Harris and Aza Raskin lay out why they believe current AI capabilities carry real societal risk, arguing companies are racing to deploy faster than safety work can keep up.

Frequently recommended as the single talk to watch to get up to speed on AI risk, especially by people who had already seen the same team's earlier work on social media; it has 20 mentions, tying it for the most-cited video on this page. It's a case the speakers are actively making, not a neutral documentary, so it's worth weighing as an argument rather than a settled account.

A fictional short film, backed by Stuart Russell, depicting swarms of cheap autonomous drones using facial recognition to carry out targeted killings, made to argue for a ban on autonomous weapons.

HN's standard reference whenever drone swarms, autonomous weapons, or military AI come up in the news, with 20 mentions over the years. It's deliberately persuasive, produced by a campaign that wants autonomous weapons banned, not a neutral forecast of what's likely.

Charles Stross's 34C3 talk on how science fiction gets the future wrong, arguing that corporations are slow, human-powered AIs that optimize for profit.

HN returns to it whenever the comparison between corporations and paperclip-maximizer AIs comes up in an AI-risk discussion; it has 7 mentions.