🏅 Top traders

the definition of "intelligence" and maybe even "agi" has been moved several times in the last 100 years. If you asked alan turing, if eliza was intelligent (, i.e. solves his turing test) he may have answered "yes". if you asked him the same about GPT4 he would have said yes for sure.

as for agi, that would involve several things, today's AI can do separately, but we have nothing that can combine all of them, e.g.:

• language (GPT3 can do that)

• vision (tesla can do that)

• world model (https://worldmodels.github.io/ )

• agency (many virtual robots can do that)

• learning (https://www.youtube.com/watch?v=A2hOWShiYoM )

• strategy (alpha go can do that)

I don't think, we will see something this year, that will combine all of this. But we may see something, that combines the right things to get close to AGI (i.e. I think if they can create a language model, that has a world model and agency, AGI may be solved).

I don't think the market will resolve to true, but 1:20 odds are really good.

@BjornJurgens

"If you asked alan turing, if eliza was intelligent (, i.e. solves his turing test) he may have answered "yes". if you asked him the same about GPT4 he would have said yes for sure."

You obviously did not read his paper. The chance he we would have said Yes about Eliza is zero, and the chance he would have said yes about GPT-4 is extremely low.

My resolution to this market is if the Wikipedia page for AGI claims that AGI is reached in the year 2023

@CameronJohnson Are you planning to check it at any particular moment? This market isn't huge, so I'm not too worried, but your resolution source being something that anyone can edit at any time doesn't seem great.

@CameronJohnson You should clarify how you'll resolve this. This is your only market so far, and I'm hesitant to participate given the ambiguity.

@kenakofer that is a fair point. i will write something up shortly.

YES because I think the market's probability is too low. I give it 15-20%.

@connorwilliams97 never mind. I misread this as 2025. Oops.

alright, competitors, pick - your - definitions!

• LSTMs are agi: they can learn anything (badly, but it works if you've got a patient enough teacher)

• transformers are agi: they can learn anything, and practically (but they're just an architecture and you gotta train them)

• gpt3 is agi: it can talk usefully about anything (but isn't that smart)

• gpt4 is agi: it matches human level in a wide variety of tasks (but is not superintelligent and didn't grow up in a natural environment or get robotics training)

• gato is agi: it can rl in many games

• palm-e is agi: it not only knows language, it can move around

• a palm-e successor would be agi if able to move as fluently through the world as humans

• only an ai able to learn as well as humans in-context would be agi

• only an ai able to do lifelong learning with single history would be agi

• only an ai able to do everything any human can do better would be agi

• only an ai unrecognizeably smarter than all of humanity combined would be agi

How will you resolve the market?

Even if, we would not notice it on 1st of Jan 24. So my bet is no with a twist.