Little Learning Machines Reviews

great game with great artstyle. if youre a machine learning enthusiast or interested by AI this is the game for you.

Really enjoying the collectables/aesthetic in addition to the AI/pet-raising mechanics.

Super cool and fun way to see in action deep learning :)

This is a very cute and chill game you can easily spend a lot of time experimenting different learning combinations. There's a couple things that take some getting used to, such as inability to modify controls, but overall this is one of the most unique games I've played! I really liked how from the very beginning it allowed me to think outside of the box and be creative in how I want to lead the bots how to do certain actions. I'm sure there's many possibilities I haven't even considered yet! There's pretty good humor that got me chuckling at times and the costumes are SO adorable!

A very charming game that provides an excellent environment to play around with neural net training. The developers are very receptive to input as well.

This game needs a much better tutorial and needs to be more readable. I managed to complete a few quests, but most of the time I don't know what I'm doing or why the bots are behaving as they are. Feels very unsatisfying. I think the game would be great, but since I can't read it it's just frustrating.

LLM is a delightful game that combines a cozy art style with the fascinating world of AI learning. While this is not the normal games I play the fun brain challenge of how to effectively train my little robot to complete the challenges instantly captured my interest. The serotonin you get after spending minutes pondering out of the box solutions and training your animo and finally achieving said challenge makes this game worthwhile to play.

The idea behind this game is super novel, and I love programming games, but this isn't enough of an engaging experience yet to warrant a recommendation as of typing. This is still in early access and receiving updates, so I fully expect the gripes I'm writing about to be fixed sometime soon, and likewise, I'm expecting to turn this review into a positive when they are fixed. Right now, the biggest issue for me with this game is the lack of player agency. It feels you're always too constricted in what's provided to you to fully have your bot do what it is you want them to do. The type of machine learning implemented here is a very hands off version, so you're very limited in what you're able to train. More manual supervision options would be very much welcomed, at least for a player like myself. The whole gist of all that follows is that you're extremely limited in what you're given to work with; It's all very oversimplified IMO. It's fair to have something like that in place for beginning players who have no experience with machine learning (which I was when I picked this game up), but there's no room to expand from that point. - The biggest limitation for certain is only having 5 reinforcement slots that you can use that will command everything you want your bot to do. I'm not trying to train a super advanced AI or anything, I just wanna make a basic program, but it always feels like you're fighting against the slot limit rather than working within them. In the quest that ended up getting me to put the game down for now, you have to free this bot that's enclosed by trees and stumps that first have to be destroyed in order to access and unlock it. In that mission, the bot was simply just stuck no matter what I did; there simply wasn't enough slots to reward what you wanted, and punish what you didn't want. To make an adjustment to your bot to fix the instance that it gets stuck on, would then undo a previous fix that you already had beforehand, and your bot gets stuck in that way instead of the other way. I think I tried like three or four different reinforcement variations, all of them ending up getting stuck in some manner. You for certain need two slots just for rewarding picking up an axe and punishing putting down one. To do otherwise is risking your bot getting stuck in a loop of just dropping the axe and picking it back up infinitely so it perpetually gains rewards on pickup. Two more slots need to be added to give an award for cutting down stumps and trees, because this is the main function of program you want to train. With four out of the five slots already taken up, you can't add the necessary negative reinforcements that would break loops of your bot getting stuck bumping into objects, or infinitely using action on an object while no tool is equipped. That would require more than five slots, which don't exist as of now. With no option of further clarify what you want, you're basically given the option of 'choose the type of stuck you want your bot to be' and that's it. - There may be a constrictively low limit to the test islands that are available as well. Up to the point I got to, you're only able to work with four testing islands, as opposed to having the ability to create and delete ones at will (for example, making a test island for each program you want to train). If that is the case, then please add an ability to manually add and delete training/cloud islands. - Additionally, You can only give reinforcements in sizes of 1, 10, and 100. You can't set a custom value. E.g. if I wanted the reward for picking up a tool to be +50, but the punishment to be -100 so the bot doesn't get stuck dropping and picking up the tool forever, that wouldn't be possible as of now. These hard limits are definitely something to be fixed before I pick this up again, that was the main deal breaker for me. I can post more suggestions for additions and tweaks to the training sections on the discussion boards if anyone's interested. Onto positive stuff: this game is super fantastic aesthetically speaking. It introduces technical concepts in a really not-scarily-technical looking way, which makes picking it up a lot more appealing than watching or reading something super technically dense like a Wikipedia article on it or something. Plus you learn hands on with it as well. I'm always a fan of coding games/educational games in general for that sort of reason👍 It's also worth noting that this game doesn't use any cloud computing at all, even though the training area is called the cloud. It's all computed locally on your machine IIRC, so if I'm thinking correctly you don't need an internet connection to play this, only on the initial first launch to install the Python dependencies (which isn't included in the base game files for some reason). Leave a comment on this review if any of what I mentioned gets resolved so I get a notification for it. This game is super promising, and I'm eager to see where it goes forward from this point.

Very unique and original concept, but currently some major issues: - Changes to training islands doesn't always save, there is no indication of an autosave and you can't manual save afaik, so very annoying if you want to exit the game without losing progress. Also I've had several training islands just reset fully so this might be more than a no recent save issue - Training cannot always be stopped for no apparent reason, making it a pain to change love/fear - There are objects you can't find/place in your training that exist in the normal islands (e.g. pumpkin on harvest hill) not sure if intentional or not, but afaik the robots don't recognise objects they haven't encountered in training - No reward can be made for being on empty ground for robots (e.g. if you want to reward them staying on empty ground so they don't step on flowers) or planting on empty ground (so they don't try planting into fences) - More love/fear conditions needed to speed up training (hopefully), but limiting it to 5 might be a design choice - Game doesn't seem to be able handle multiple robots training at once, a soft or hard guideline would be helpful Overall a really cool and unique game but there's so much annoying stuff I can't really enjoy it atm

As someone who does a lot of machine learning, this game fascinated me when I first saw it a couple of days ago. I love the idea and I think more games need to take a more open-ended approach to allowing players to try different strategies for training neural networks. As a game, it can be fun, and you can learn to be good at this game just through trial and error. But while you can learn with trial and error, you will find the "why" difficult when you fail to properly train your bots. Someone like me who is experienced with the game can sort of guess what goes wrong, either a killed network that is too negatively weighted or is overfit to one specific scenario, but if you're getting this and aren't familiar with neural networks, you'll struggle until you either figure it out or you look up a guide. Let me give an example to show the challenge and triumph: Let's say you want your bot to interact with a ball to throw toward a dog. You may start out direct, rewarding the bot for throwing the ball at the dog. You set up the training island (where you develop the brains of your bots) with a ball and a dog with your bot placed in a specific spot. You may figure out your single reward does little, because your bot is so random that there's very little chance it will do the one thing that gives it rewards. You then have to decide between two methods: #1, train your bot on how to move first or #2, give it a smaller reward for moving toward the ball, for interacting with the ball, and then the largest reward for doing what you want. Your bot finally does what you want. So, you go away from the training island, move your bot to the "challenge" area where it is required to perform the task, only to figure out that your bot is getting stuck on walls, because while it knows what a ball is, you never taught it how to deal with obstacles. So, you go back to the training environment, try to train it on obstacles specifically. In doing so, your bot loses the knowledge it gained to interact with the ball. And that's the balancing act of the game, understanding the delicate balance between teaching your bot to perform a task without overtraining and overspecializing it where it cannot perform other basic functions. You'll learn a lot of techniques just from trying out the game. TL:DR the game is a fun and unique experience. It is a very great introduction to neural networks and shows how difficult it is to train an AI. That's where this game is really fun.