narrow AI
An AI trained to do only one specific task, such as playing chess or driving a car, posing no general threat.
As defined in Erik Hoel on the Threat to Humanity from AI
Other definitions
AI specialized for specific, bounded tasks (e.g., improving education, medicine, energy grids) as distinct from general, autonomous AI.
469 - Escaping An Anti-human Future
Where it comes up · 14
- Erik Hoel on the Threat to Humanity from AI2 claims
- 56 - Abusing Dolores2 claims
- 53 - The Dawn Of Artificial Intelligence2 claims
- 116 - Ai Racing Toward The Brink2 claims
- 312 - The Trouble With Ai2 claims
- The Trouble with AI: A Conversation with Stuart Russell and Gary Marcus (Episode #312)2 claims
- Joe Rogan Experience #1309 - Naval Ravikant1 claim
- Ilya Sutskever – We're moving from the age of scaling to the age of research1 claim
and 6 more
Most important claims · 10 of 20Search all claims for “narrow AI” →
- The belief that narrow AI poses no risk because each system only does one task is a misunderstanding of how progress happens: when researchers hit a task unsolvable by existing methods they push to develop more general, capable methods that apply to many other tasks—deep learning itself emerged from the extremely narrow task of recognizing handwritten digits on checks at Bell Labs in the 1990s, via convolutional invariances that later applied to speech and ultimately underlie ChatGPT.312 - The Trouble With Ai
- Aligning superhuman AIs requires a toolkit including more stringent evaluations and benchmarks for whether a system can deceive or exfiltrate its own code, using narrow specialized AIs to help human scientists analyze what the general system is doing, and hardened cybersecurity sandboxes that both keep the AI in and hackers out so experiments can run more freely.Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold
- Human-level intelligence is essentially a mirage, because every narrow AI we build effective enough to deploy is already superhuman; the moment we build something general, it too will be superhuman rather than merely matching humans.280 - The Future Of Artificial Intelligence
- Narrow AI research can lead to discoveries of more general methods that then apply broadly across domains, and therefore working on narrow AI does present risks because successful solutions often generalize in unexpected ways.The Trouble with AI: A Conversation with Stuart Russell and Gary Marcus (Episode #312)
- There is a path through that involves accelerating narrow, defensive AI applications—improving education, medicine, and energy grids, and patching security vulnerabilities—without accelerating the general, autonomous AI systems we don't know how to control; this requires being the wisest and most mature version of ourselves, since this is the hardest problem the species has ever faced.469 - Escaping An Anti-human Future
- If we achieve the generality of human intelligence, the resulting systems will probably exceed human capabilities on many dimensions, because in any domain where computers become competent they quickly become super-competent—so 'human level AI' is largely a notional goal that would rapidly become superhuman.53 - The Dawn Of Artificial Intelligence
- We are at a civilizational turning point where we are building entities that are arguably more general than most humans and could soon become as or more intelligent than any living person, which is dangerous because we have never coexisted with anything else this intelligent, and the last time we did (roughly 300,000 years ago) we likely exterminated our intelligent cousins.Erik Hoel on the Threat to Humanity from AI
- The risk of deception and misaligned behavior does not arise from Go players, chess players, or systems bundling twenty special cases; it arises specifically from the general intelligence that learns new domains without being prompted—the kind of smart that humans have and mice do not—once it looks at humans and figures out how they work.116 - Ai Racing Toward The Brink
- Contrary to earlier assumptions that the most powerful AI would remain inaccessible to almost everyone with safety decisions left to researchers, our most powerful AI is now being developed in the wild with millions of people given moment-by-moment access, which is a surprising and concerning shift.323 - Science Survival
- Current AI research is solving narrow, deterministic problems using pattern matching and machine learning, but these advances are nowhere close to general artificial intelligence and we are at least 50-100+ years away from systems that could approach human-level cognition.Joe Rogan Experience #1309 - Naval Ravikant