The Beautiful Intelligence of Bacteria and Other Microbes

Intelligence is the ability to take decisions that lead to maximum utility, not just "good decisions". Utility is the main currency of decisions.

I found this article very interesting because it showcases the exploration/exploitation dilemma of reinforcement learning. Even a bacteria colony is a RL agent, reacting to external events in a useful way.

It's easy to get lost in the hype of statistical methods and neural networks, and in doing so forget that these do not make up the field of AI as a whole. Making predictions is, in my opinion, a subset of what AI generally addresses. In AI you often want to behave intelligently in a given system; that is, you want to observe a system and not only characterize it in a computationally amenable format, but use that characterization to decide on an action. You may make "predictions" in determining this characterization but what you're really looking for is the action, or rather a policy on your problem instance.

For example, in AlphaGo, the "predictions" are some heuristic evaluation of a Go game-state. The actual desired output of the AlphaGo system is not how good a state of the game is, but how to choose the best move at a given stage of the game. This is done through Monte Carlo Tree Search, an AI algorithm separate from the choice of algorithm used to make the prediction of "goodness" of a state, which does not in itself make predictions but uses the predicted heuristic values in order to determine the best move. In fact, MCTS can be used in the absence of a learned heuristic prediction altogether if you can derive a good heuristic to evaluate a state given the problem domain.

I think a better way to consider AI is as a set of tools for solving general problems. There is of course a bit of ambiguity here in what qualifies as general and what does not. Making a prediction given some input vector or data (classification/regression/dimensionality reduction) is a pretty general problem, but so is the problem of making optimal moves in a game. There are also general problems such as how to extrapolate from ground truths to prove a theorem (automated problem solving), or how to plan actions to arrive at some final state given some initial state and a set of possible transitions (planning), which are generally considered as part of AI.

Similarly, these biofilms are solving a complex problem with many different components. The dendritic swarming is essentially a method for solving the multi-armed bandit problem. The film needs to budget food consumption when it finds new sources so that it does not colonize that food source too heavily (expending too much of the food itself) so that there is enough remaining food to give to the dendrites, so there is a notion of planning with a temporal component (e.g. the metabolic rate of the colony). And the film's ultimate goal can be considered to maximize the total food consumed by the film and its member's descendants. To me, that makes the "intelligence" capabilities of the film pretty clear