All the definitions of intelligence I've seen online include LLMs, and any alternative I've seen proposed either still includes LLMs anyway, excludes humans anyone would say are intelligent, or smuggles in non-functional requirements like "must be biological" or "must be conscious" that aren't really about intelligence at all.

First, there's modern dictionaries:

Merriam-Webster: "The ability to learn or understand things or to deal with new or difficult situations."

Oxford Advanced Learner's Dictionary: "The ability to learn, understand and think in a logical way about things."

Compact Oxford English Dictionary: "The ability to acquire and apply knowledge and skills."

Next, expert/academic definitions:

Gottfredson et al., "Mainstream Science on Intelligence" (1997, 52 signatories, published in The Wall Street Journal and the journal Intelligence):

"Intelligence is a very general mental capability that, among other things, involves the ability to reason, plan, solve problems, think abstractly, comprehend complex ideas, learn quickly, and learn from experience."

APA Task Force, Intelligence: Knowns and Unknowns (1996) :

"Individuals differ from one another in their ability to understand complex ideas, to adapt effectively to the environment, to learn from experience, to engage in various forms of reasoning, to overcome obstacles by taking thought."

Legg & Hutter (2007, Minds and Machines), synthesizing 70+ definitions

"Intelligence measures an agent's ability to achieve goals in a wide range of environments."

The following are the traits I have extracted from these definitions that if I show an LLM possesses them all, then I've shown them as intelligent in line with mainstream dictionary and scientific definitions of intelligence. To change my view, you must show me that an LLM either does not posses one of these traits or that I am missing a trait, or that there is an alternative definition of intelligence that is mainstream but LLMs do not satisfy.

The traits are as follows:

Ability to learn:

LLMs learn in two ways. During training, they acquire patterns, facts, and strategies from massive amounts of data. An emergent property from just predicting text is the ability to learn how to answer questions factually, problem solve, write code, etc. Of course this process especially is very different from how humans learn (for one, it needs far more data than a human does), but it is learning nonetheless.

During inference, they learn within context. You can teach an LLM a new notation system, a made-up language, or a set of rules in the prompt and it will apply them correctly. This is called "in-context learning" and is extensively studied ( survey: Dong et al., 2024 ). If you show someone three examples of a pattern and they pick it up, that's learning. LLMs do exactly this, routinely, with zero weight updates. This learning will be gone if the facts are taken out of its context, but this is just a memory issue rather than an issue with learning (having…

为什么值得关注

提供了用户原本不知道的新信息；能改变理解方式，而不只是重复常识；它带来了新的事实、进展或信息，不是在重复旧内容

来源：reddit，领域：tech，保留分：0.73