AGI is achieved when this market resolves and AIs can smell when there is a commercially available device that can differentiate different smells similar to humans (resolution criteria is still being refined)
Related questions
@TsubasaKato if ais have access to various sensors to mimick human smell for most relevant use-cases then yes would suffice. I am curious to see how this would work. Humans can smell something and then remember old memories or infer information such as distance to the smell. I wonder if this would also be the case for ais. Smell seems to he a native modality of humans if that makes sense.
@Soli i feel like an AI capable of passing a turing test should be able to make up a convincing lie even without any ability to smell: "Growing up both my parents were smokers and by the time I moved out I was a smoker too. I've been told that constantly living in a house filled with smoke damaged my since of smell. When I got covid my sense of smell got even worse and never fully recovered. Unfortunately I can still smell that many of the people here are in need of a shower, I haven't been surrounded by this many sweaty people since high school. I can also smell the hot dogs, soft pretzels, and confectionery treats some people are walking around with. When I went outside to smoke I stood by the dumpster and could smell the rotting garbage inside, and it also smelled like someone smoked weed by that dumpster not too long ago. I also walked past someone with way too much perfume on." Something like this makes some assumptions about the testing location, but I think they are all pretty safe assumptions, With more knowledge of the testing location you could make a better lie, but if all the assumptions are correct I'm not sure a sense of smell would let you give a better answer.
@bluerat makes sense - my first thought was i don’t want to make this market about whether the ai will be able to lie but then i noticed that in order for any ai to pass the turing test it has to lie so maybe you are right and we should not make an exception for this loophole
The human nose detects smoke by the smoke bonding with proteins, whereas a smoke detector detects smoke by how light interacts with the smoke. In order for this market to be resolved does the AI need to use proteins or something similar, or is using light and other methods a valid approach? The human nose has around 400 unique sensors which can be triggered independently, leaving the potential for around 100,000 smells that humans could potentially distinguish, but in practice I'd be surprised if the average person could get better than around 50, how many smells does the AI need to distinguish to resolve this market? Does it have to sense the same chemicals that humans can sense or does distinguishing any chemicals count? Does the AI have to be capable of describing the smell in casual human language (ie. Earthy), or is simply listing the chemicals present enough?
@bluerat very good questions - i don't expect machines to smell using the same mechanisms humans use, it is not like they can see or speak the same way we do. However, I do expect them to be better than the average human at distinguishing smells, whatever this means. They don't have to use words such as earthy to describe the chemicals but they should be able to derive at least as much information as the average human would, e.g. something is rotten
@Soli my preferred AGI definition is
"Can perform any cognitive task at a standard equal or greater than the 50th percentile of the global biological human population"
Which would include perceiving nasal sensory information
@TheAllMemeingEye interesting - do you think AI will be able to smell within 10 years? my intuition tells me this will not happen - we had decades of vision/audio/text data that helped us train the ais but there is not even a standard way to save smell data whatever this means
@Soli haha now that you put it that way my credence just increased that the final job to be automated will be smelling on behalf of the AIs lol
@TheAllMemeingEye yeah, this is very much a "like submarines can swim" situation where the most useful applications are only somewhat related to what humans can smell.
@i_i in some ways you are right, but smell often encodes relevant information whereas swimming doesn’t