4 ms·

> They keep time: if blown with cold air on the first of every hour, by the third hour, they’ll retract just before the cold snap.

smegcicle 5y ago
Looks like they chilled it three times, then just watched for a few hours, then chilled it again just to prank it https://eprints.lib.hokudai.ac.jp/dspace/bitstream/2115/33004/1/PhysRevLett_100_018101.pdf https://eprints.lib.hokudai.ac.jp/dspace/bitstream/2115/3300...
Jensson 5y ago
That is super quick learning! And for a formless blob! Imagine if we knew how to train ML models that could reliably perform feats like this (learn after training, not train to withdraw every hour), while still managing all other parts of being a slime mold like hunting for food, distributing nutrients etc.
- voldacar 5y ago
  Physical slime molds have already been used in an experimental model to optimize road networks: https://www.sciencedirect.com/science/article/abs/pii/S0303264711000608 https://www.sciencedirect.com/science/article/abs/pii/S03032...
- otabdeveloper4 5y ago
  That won't happen for a long time. First we'd need to let go of the idea that "intelligence" means "neural network", and there's a huge amount of ideological resistance to the idea that intelligence isn't made with composeable atomic computation blocks.
- krisoft 5y ago
  > Imagine if we knew how to train ML models that could reliably perform feats like this I imagined it and I don’t find it that special. Picking up a strong and regular periodic signal and predicting the next iteration is not a hard problem. > while still managing all other parts of being a slime mold like hunting for food, distributing nutrients etc. ML models don’t do those things because ML models are abstract constructs without appendages or any other physical embodiment. We cannot manufacture artifical slime molds because we cannot manufacture artifical cells. This doesn’t seems to be an ML problem. (Or if it is, then everything become an ML problem too and the term lost its meaning.)
  - Jensson 5y ago
    > ML models don’t do those things because ML models are abstract constructs without appendages or any other physical embodiment. I am talking about it doing this in a simulation. We don't have AI that are as smart as a slime mold at the moment, it will be an achievement when we get there. Also we can't discount the processing power required to process raw signals detected, our AI models needs to also learn the signal processing as well. The raw data a slime mold has to work with is extremely bare, if you get that input do you really think an AI could do the same things as it? And this is a single cell organism, this is the capability of a single biological cell. > Picking up a strong and regular periodic signal and predicting the next iteration is not a hard problem. Not if that is the only thing you do, no, but if you also do a lot of other things at the same time then it gets extremely hard. Edit: For example, how does the mold know that it just didn't move into a sunny spot? Or maybe this is just a cloud and it retracted unnecessarily? There are a lot of things to account for in a real world scenario that the slime mold has to understand that your naive signal processor can just ignore.
    - krisoft 5y ago
      > For example, how does the mold know that it just didn't move into a sunny spot? Why do you think that it does know that? > Or maybe this is just a cloud and it retracted unnecessarily? I bet that it does retract sometimes unnecessarily. Every organism has these kind of false positives. > We don't have AI that are as smart as a slime mold at the moment, it will be an achievement when we get there. When you say we don’t have this AI do you mean you can’t download it from Github right now, or do you mean it is something humanity couldn’t build if they tried? I assume you mean it in the second sense. Why do you think it is so hard? Let’s say someone, a research project or an alien inteligence gives you such an AI on a thumbstick. How would you evaluate it to test if it is indeed a simulation of a slime mold? In other words what do you expect from the simulation?
      - Jensson 5y ago
        > When you say we don’t have this AI do you mean you can’t download it from Github right now, or do you mean it is something humanity couldn’t build if they tried? "Could build if they tried" is a strange description, we haven't made an AI with similar capabilities as a slime mold so we don't know how to do it. Maybe a research team could do it if they tried, and we would learn things from it, but before we have done it I'll say that we can't do it. Similarly 5 years ago I'd say that we don't know how to make a super human GO AI at the time. There was a team in the process of making a GO AI in the works, but we still didn't know how to build it since we didn't know if their tech would be good enough to do it at the time. > I bet that it does retract sometimes unnecessarily. Every organism has these kind of false positives. Sure, but it can't get into an infinite loop or hang, or it would die a lot. It needs to be smart enough to survive in most real world scenarios. That is the hard part. It can't just do something like "if I get cold two times in a row I'll retract and stop moving in preparation for the third time", that would probably make it stop way too much. And we do know that slime molds remembers locations of food, so it has a proper memory, likely it has some sort of reasoning about temperatures and locations as well to avoid going to its death. This just has evidence that it even reasons about time and temperature as well, and even detects recurring time series events in addition to everything else it does, that is a lot of features that needs to interact well. Edit: > How would you evaluate it to test if it is indeed a simulation of a slime mold? Simple, you make another team set up some tests for it, and see if it behaves similar to a real world slime mold in those scenarios. Slime mold environments are very quick and cheap to set up and run, so it would be neither hard nor expensive to do.

smegcicle 5y ago

Looks like they chilled it three times, then just watched for a few hours, then chilled it again just to prank it https://eprints.lib.hokudai.ac.jp/dspace/bitstream/2115/33004/1/PhysRevLett_100_018101.pdf https://eprints.lib.hokudai.ac.jp/dspace/bitstream/2115/3300...

Jensson 5y ago

That is super quick learning! And for a formless blob! Imagine if we knew how to train ML models that could reliably perform feats like this (learn after training, not train to withdraw every hour), while still managing all other parts of being a slime mold like hunting for food, distributing nutrients etc.

voldacar 5y ago
Physical slime molds have already been used in an experimental model to optimize road networks: https://www.sciencedirect.com/science/article/abs/pii/S0303264711000608 https://www.sciencedirect.com/science/article/abs/pii/S03032...
otabdeveloper4 5y ago
That won't happen for a long time. First we'd need to let go of the idea that "intelligence" means "neural network", and there's a huge amount of ideological resistance to the idea that intelligence isn't made with composeable atomic computation blocks.
krisoft 5y ago
> Imagine if we knew how to train ML models that could reliably perform feats like this I imagined it and I don’t find it that special. Picking up a strong and regular periodic signal and predicting the next iteration is not a hard problem. > while still managing all other parts of being a slime mold like hunting for food, distributing nutrients etc. ML models don’t do those things because ML models are abstract constructs without appendages or any other physical embodiment. We cannot manufacture artifical slime molds because we cannot manufacture artifical cells. This doesn’t seems to be an ML problem. (Or if it is, then everything become an ML problem too and the term lost its meaning.)
- Jensson 5y ago
  > ML models don’t do those things because ML models are abstract constructs without appendages or any other physical embodiment. I am talking about it doing this in a simulation. We don't have AI that are as smart as a slime mold at the moment, it will be an achievement when we get there. Also we can't discount the processing power required to process raw signals detected, our AI models needs to also learn the signal processing as well. The raw data a slime mold has to work with is extremely bare, if you get that input do you really think an AI could do the same things as it? And this is a single cell organism, this is the capability of a single biological cell. > Picking up a strong and regular periodic signal and predicting the next iteration is not a hard problem. Not if that is the only thing you do, no, but if you also do a lot of other things at the same time then it gets extremely hard. Edit: For example, how does the mold know that it just didn't move into a sunny spot? Or maybe this is just a cloud and it retracted unnecessarily? There are a lot of things to account for in a real world scenario that the slime mold has to understand that your naive signal processor can just ignore.
  - krisoft 5y ago
    > For example, how does the mold know that it just didn't move into a sunny spot? Why do you think that it does know that? > Or maybe this is just a cloud and it retracted unnecessarily? I bet that it does retract sometimes unnecessarily. Every organism has these kind of false positives. > We don't have AI that are as smart as a slime mold at the moment, it will be an achievement when we get there. When you say we don’t have this AI do you mean you can’t download it from Github right now, or do you mean it is something humanity couldn’t build if they tried? I assume you mean it in the second sense. Why do you think it is so hard? Let’s say someone, a research project or an alien inteligence gives you such an AI on a thumbstick. How would you evaluate it to test if it is indeed a simulation of a slime mold? In other words what do you expect from the simulation?
    - Jensson 5y ago
      > When you say we don’t have this AI do you mean you can’t download it from Github right now, or do you mean it is something humanity couldn’t build if they tried? "Could build if they tried" is a strange description, we haven't made an AI with similar capabilities as a slime mold so we don't know how to do it. Maybe a research team could do it if they tried, and we would learn things from it, but before we have done it I'll say that we can't do it. Similarly 5 years ago I'd say that we don't know how to make a super human GO AI at the time. There was a team in the process of making a GO AI in the works, but we still didn't know how to build it since we didn't know if their tech would be good enough to do it at the time. > I bet that it does retract sometimes unnecessarily. Every organism has these kind of false positives. Sure, but it can't get into an infinite loop or hang, or it would die a lot. It needs to be smart enough to survive in most real world scenarios. That is the hard part. It can't just do something like "if I get cold two times in a row I'll retract and stop moving in preparation for the third time", that would probably make it stop way too much. And we do know that slime molds remembers locations of food, so it has a proper memory, likely it has some sort of reasoning about temperatures and locations as well to avoid going to its death. This just has evidence that it even reasons about time and temperature as well, and even detects recurring time series events in addition to everything else it does, that is a lot of features that needs to interact well. Edit: > How would you evaluate it to test if it is indeed a simulation of a slime mold? Simple, you make another team set up some tests for it, and see if it behaves similar to a real world slime mold in those scenarios. Slime mold environments are very quick and cheap to set up and run, so it would be neither hard nor expensive to do.