3 ms·

Bees are one of the reasons I think the current metrics for "smartness" of AI are misguided and that we are very, very far away from AGI. Bees are capable of le

Bees are one of the reasons I think the current metrics for "smartness" of AI are misguided and that we are very, very far away from AGI. Bees are capable of learning, planning, communication, prediction and at least some form of abstract reasoning. And it can operate sensibly in contexts very different from its normal environment. All of that with just 960,000 neurons. No artificial neural networks show anything close to this level of flexibility. Does it really matter whether a bee is better at any of those tasks than "average" or "best" humans? Clearly, once flexibility is achieved, other capabilities can be improved upon iteratively, because that's exactly what happened in nature. And yet I don't see any researchers looking at how many different tasks their specific AI can accomplish, instead of how well it can accomplish a particular one.

TrainedMonkey 8y ago
This is a simplistic view of what is going on. That complexity is not produced by 1e6 neurons alone. Rather it is a combination of a neural net embedded in a biochemical harness. The harness interacts not only with individual bee's systems, but also with environment and other bees (via pheromones, etc.) I do agree with you that we still have much to learn in neural network design (or more generally distributed algorithms). For example distributed systems are notoriously bad at state machines (or sequencing in general). One of the ways biological systems handle sequencing is via multiple signaling layers with different time constants (e.g. pheromones vs neurons). On the other hand deep learning systems I am familiar with just throw more parameters at the problem, obviously this scales poorly with the sequence length.