4 ms·

> Most common example on HN: people talking about "artificial general intelligence" without defining "intelligence". I'd argue that it is because among people

> Most common example on HN: people talking about "artificial general intelligence" without defining "intelligence". I'd argue that it is because among people talking about "artificial general intelligence" there is a rough consensus about what "intelligence" means in this context (generally, according to Legg&Hutter 2007, which is a long discussion but the essence is captured by "Intelligence measures an agent’s ability to achieve goals in a wide range of environments") and it's counterproductive to explicitly define it in detail (which is very tricky) unless the argument hinges on some nuance of the definition, which it almost never does. In essence, the GAI arguments are unchanged for a whole 'family of definitions' so an attempt to push the debate towards "define intelligence" is just derailing the discussion by provoking to repeat a discussion that's essentially settled 15 years ago. In any GAI-related discussion I can reasonably presume that we have a shared understanding of what we're talking about, there isn't much debate about the terms within the field - and if it seems that someone not from the field is applying a wildly different definition of intelligence (i.e. taking the colloquial understanding or some definition that explicitly refers to humans or one that adds in specific goals as opposed to being good at achieving whatever goals the agent has) then it can be pointed out that no, that's not what we're talking about, but within the field the terminology is clear enough, and also as soon as you're talking about detail, you aren't talking about "intelligence in general" but rather about capabilities for specific kinds of decisonmaking or performance on specific tasks; and if a philosopher says "well, that isn't true intelligence" then I can fully concede it, but just state that for the discussion of AI it doesn't matter if it's "true intelligence" but this is what we're talking about as those properties are considered to be sufficient for the described capabilities and consequences.

astrange 4y ago
That is just fine, and I think "you know, like a human" is a great definition of "general intelligence". If you're not careful about it though, people tend to slide into using specific definitions of "intelligence" like "calculation ability", implying it'll increase linearly as you add more GPUs to your data center. And you end up with assumptions that they'll necessarily be cold, calculating and perfect like a computer, but boredom could be a necessary property of a general intelligence, otherwise it'd get stuck on problems. Or not. If not, maybe humans aren't generally intelligent.
- PeterisP 4y ago
  Your comment about boredom raises an interesting question whether an effective general motivation system is an unalienable necessary part of general intelligence or not. The way I see it, as soon as you think about it seriously it becomes obvious that all (or at least most) the negative feelings we have are there for a very good practical reason. Boredom is effectively a push to prevent an agent from being stuck in a local optimum, and required to get a more effective explore-vs-exploit tradeoff (for whom we have theoretical grounding for what's good or not based on the analysis of "multi-armed-bandit" problems), and the same applies for satiating pleasure e.g. hedonic treadmill; burnout is required for an agent to be able to "say" "this isn't working, you need to do literally anything else"; fear has obvious benefits for risk management; and anger, hate and vengeance has obvious causes in game theory about optimal strategies in multi-actor games with potential defectors. So that is an argument that any effective general intelligence is likely to possess at least some of the emotions that we recognize, simply because they are effective for a purpose. On the other hand, these factors are (or were) required for a specific environment, namely evolution in a social community of peers - so perhaps an agent can be scarily effective without them, especially if their environment is very different.

astrange 4y ago

That is just fine, and I think "you know, like a human" is a great definition of "general intelligence". If you're not careful about it though, people tend to slide into using specific definitions of "intelligence" like "calculation ability", implying it'll increase linearly as you add more GPUs to your data center. And you end up with assumptions that they'll necessarily be cold, calculating and perfect like a computer, but boredom could be a necessary property of a general intelligence, otherwise it'd get stuck on problems. Or not. If not, maybe humans aren't generally intelligent.

PeterisP 4y ago
Your comment about boredom raises an interesting question whether an effective general motivation system is an unalienable necessary part of general intelligence or not. The way I see it, as soon as you think about it seriously it becomes obvious that all (or at least most) the negative feelings we have are there for a very good practical reason. Boredom is effectively a push to prevent an agent from being stuck in a local optimum, and required to get a more effective explore-vs-exploit tradeoff (for whom we have theoretical grounding for what's good or not based on the analysis of "multi-armed-bandit" problems), and the same applies for satiating pleasure e.g. hedonic treadmill; burnout is required for an agent to be able to "say" "this isn't working, you need to do literally anything else"; fear has obvious benefits for risk management; and anger, hate and vengeance has obvious causes in game theory about optimal strategies in multi-actor games with potential defectors. So that is an argument that any effective general intelligence is likely to possess at least some of the emotions that we recognize, simply because they are effective for a purpose. On the other hand, these factors are (or were) required for a specific environment, namely evolution in a social community of peers - so perhaps an agent can be scarily effective without them, especially if their environment is very different.