The definition of intelligence¹ I use is the ability to model² complex systems. The more complex the system you can model in your head the more intelligent you are. The origin of intelligence is in tiny organisms that gained some primitive sensing organ and a model of the world just complicated enough to move towards food more often than they turned away from it. Intelligence has come a very long way since then but until recently the ultimate purpose of it remained the same. And frankly for many humans alive today that is really all they use their intelligence for.

Estimating the intelligence of another person is quite difficult. Especially if the person you are estimating is more intelligent than you. If they are less then you can get a pretty good sense for it just by talking to them about several topics. But to nail it down you need to talk to them about their specialty since that shows their true potential. And their model in that domain is likely to be more sophisticated than yours so you really just have to guess at how smart they are³. There is one reliable method for evaluating intelligence though and that's to get them to make a prediction and then check how close they were. The more complex someones mental model of the world is the more accurate the predictions they generate will be. And since comparing a prediction to a result requires relatively little intelligence this method enables you to accurately evaluate the intelligence of people much smarter than yourself and even compare them to see who is smarter.

I've never taken an IQ test but from what I understand they are mostly pattern recognition tests. The idea being that there is a strong correlation between the complexity of the patterns you can recognise and the complexity of the models you have of the world. This correlation obviously breaks fairly often though because many people who test high on IQ don't go on to become particularly rich. Which I think is evidence that their models of the world aren't all that complex because you only have to be smarter than other traders to make money in securities⁴.

Couples tend to end up matching on intelligence. Which is a little surprising since so few people explicitly value intelligence. However a mismatch in intelligence will eventually cause misunderstandings and those misunderstandings will lead to arguments and eventually to breakup. So even when people don't explicitly try and match on intelligence they will tend to experience relationship failures until they end up with an approximate match. It's easy to think that all misunderstandings that stem from intelligence differences can be resolved by talking. The problem is that intelligence differences are often much larger than they appear. If the person of lower intelligence is self aware and trusts their partner then it can work. Otherwise no.

AI has made very easy to see what intelligence is. LLMs essentially take a large math equation and fine tune all the parameters in it until it accurately models language. The way this fine tuning is done is by asking the computer to make a prediction of what the rest of the text will look like based on some initial prompt and then compare the actual text to the prediction. LLMs are the first good AI models because there is already lots of easy to use data available to do the fine tuning on. It's obvious that AGI will occur almost as soon as we can generate good training data for the real world. And if that turns out to be too difficult then we will just have to build a lot of humanoid robots and have them learn through teleoperation or examples. This is how Tesla's Robotaxi was trained. If the training data can be digitally provided to the training computer then it will be because it's way cheaper. But if it's just too hard then we will see that data generated by tele-operating humanoid robots for a while.