emergent abilities
In plain wordsSkills that seem to switch on suddenly as AI models grow: small versions fail at them, then larger ones past a certain size succeed. Some researchers argue the suddenness is partly an artifact of how skills are measured.
Capabilities of language modelslarge language model, n.: A computer program trained on enormous amounts of text to predict the words that come next. At vast scale, that simple skill lets it chat, answer questions, translate, summarize and write code. that stay at or near chance across the smaller members of a model family and then appear once models pass some scale, so that a trend fitted to the smaller models would not have forecast them. Reported examples include multi-step arithmetic, a number of BIG-benchBIG-bench, prop. n.: A huge collaborative test of 204 tasks, written by hundreds of researchers, built to probe what language models can and cannot do as they grow larger. tasks and the benefit of chain of thoughtchain of thought, n.: Getting an AI to write out its reasoning step by step before giving an answer, the way a student shows their working. This often makes answers to math and logic problems more accurate. prompting, which in 2022-era models appeared only at around 100 billion parameters. Emergence in this sense is defined by scale — of parameters, training compute or data — and its significance for safety is that dangerous capabilities might likewise arrive without warning. Critics showed that many apparent jumps are produced by all-or-nothing scoring such as exact-match accuracy; measured with a graded score such as per-token likelihood, the same models often improve gradually and predictably. Others reply that the capabilities that matter in practice are often all-or-nothing, so threshold effects are real for users even if underlying competence grows smoothly.
The press release called the coding skill emergent, but per-token loss on the task had been falling steadily for three model generations.
Origin
Popularized by Jason Wei and 15 co-authors in 'Emergent Abilities of Large Language Models' (TMLR, 2022), who called an ability emergent if smaller models lack it and larger ones have it, and cited Philip Anderson's 1972 essay 'More Is Different'. Rylan Schaeffer, Brando Miranda and Sanmi Koyejo's 'Are Emergent Abilities of Large Language Models a Mirage?' (2023; NeurIPS 2023) set out the main critique.
Usage note
Distinct from the philosophical sense of emergenceemergence, n.: When a whole shows properties or abilities that none of its parts has and that are hard to predict from them — like a traffic jam arising from individual cars, or new skills appearing in larger AI models. — wholes with properties their parts lack — and from 'emergent' as loose praise for any surprising behavior. See also inverse scalinginverse scaling, n.: When bigger or more heavily trained AI models do worse on a task than smaller ones — the opposite of the usual rule that scale brings improvement., where larger models do worse.
Related
- scaling lawsscaling laws, n.: Measured patterns showing that an AI model's error shrinks smoothly and predictably as the model is made bigger and given more data and computing power, so labs can forecast a giant model's performance from small, cheap experiments.
- emergenceemergence, n.: When a whole shows properties or abilities that none of its parts has and that are hard to predict from them — like a traffic jam arising from individual cars, or new skills appearing in larger AI models.
- inverse scalinginverse scaling, n.: When bigger or more heavily trained AI models do worse on a task than smaller ones — the opposite of the usual rule that scale brings improvement.
- grokkinggrokking, n.: When a model that has merely memorized its training examples suddenly, long afterward, figures out the underlying rule and starts getting new examples right.
- BIG-benchBIG-bench, prop. n.: A huge collaborative test of 204 tasks, written by hundreds of researchers, built to probe what language models can and cannot do as they grow larger.
- chain of thoughtchain of thought, n.: Getting an AI to write out its reasoning step by step before giving an answer, the way a student shows their working. This often makes answers to math and logic problems more accurate.
- dangerous capability evaluationdangerous capability evaluation, n.: A test of whether an AI system can do things that could cause severe harm in the wrong hands or on its own — such as helping make a bioweapon, hacking computers or copying itself.
- inverse scalinginverse scaling, n.: When bigger or more heavily trained AI models do worse on a task than smaller ones — the opposite of the usual rule that scale brings improvement.
Sources
- Wei, J. et al. (2022). Emergent Abilities of Large Language Models. Transactions on Machine Learning Research. arxiv.org/abs/2206.07682
- Schaeffer, R., Miranda, B. & Koyejo, S. (2023). Are Emergent Abilities of Large Language Models a Mirage? NeurIPS 2023. arxiv.org/abs/2304.15004v1
Mentioned in
emergenceemergence, n.: When a whole shows properties or abilities that none of its parts has and that are hard to predict from them — like a traffic jam arising from individual cars, or new skills appearing in larger AI models., scaling lawsscaling laws, n.: Measured patterns showing that an AI model's error shrinks smoothly and predictably as the model is made bigger and given more data and computing power, so labs can forecast a giant model's performance from small, cheap experiments., BIG-benchBIG-bench, prop. n.: A huge collaborative test of 204 tasks, written by hundreds of researchers, built to probe what language models can and cannot do as they grow larger., capabilitycapability, n.: Something an AI system is able to do, such as writing working code or solving math problems, and how well. Measuring capabilities is how researchers track progress and spot dangerous skills., scaling hypothesisscaling hypothesis, n.: The bet that today's AI methods, simply made bigger with more computing power and data, will keep getting smarter — possibly all the way to human-level intelligence and beyond — without needing fundamentally new ideas., grokkinggrokking, n.: When a model that has merely memorized its training examples suddenly, long afterward, figures out the underlying rule and starts getting new examples right., inverse scalinginverse scaling, n.: When bigger or more heavily trained AI models do worse on a task than smaller ones — the opposite of the usual rule that scale brings improvement., PaLMPaLM, prop. n.: A very large Google language model from 2022 that showed sudden jumps in reasoning, and even explained jokes, as it grew. Its successor powered Google's Bard chatbot before the Gemini models replaced it., Sparks of AGISparks of AGI, prop. n.: A 2023 Microsoft research paper arguing that an early version of GPT-4 showed the first glimmers of general, humanlike intelligence. Its claim provoked sharp debate about how to judge AI abilities.