Sunday I spent most of the day in transit, so I used it to binge learn. I powered through the last few tutorials of crypto zombies and through the first few lessons of fast.ai.
I highly recommend anyone interested in tokens and dapps to do the crypto zombies tutorial/game. It really filled in some gaps in my token knowledge – like when to create a modifier, how to implement the approve interface (if you saw my blog post a few days ago you will see my puzzlement around this interface method), and the secrets of Emit. Also there are some best practices for creating webdapps, that I had to figure out the hard way, so save yourself some heartache and go through the tutorial.
One of the main takeaways was how to create non fungible tokens that are not really unique. This is under a specification called ERC732x (according to cryptozombies). The example is something like game assets such as a sword or a magic rock. Many people might have the same game assets – such as the same sword. But it is non fungible to a certain extent. This is the same thing with prayers. We may have the same prayers, similar to the way we might all have the same sword in an RPG.
The biggest revelation came with my binge learning on fast.ai. It is blowing my mind. I did the Andrew Ng course on coursera a few years ago, but fast.ai is amazing. The teacher is very inspirational and it is all about learning tools and then coming up with new ideas to apply the tools to. I finally understand what the different levels in a neural net do. Why you would want to freeze some and train others more. The first example is to create a CNN for recognizing cats from dogs. I replaced it to recognize mala beads from rosary beads. I am hitting a few snags with the fast.ai library finding the path for my images but if I bite the bullet and use keras I think I could fix it.
But really I am interested in the relation between things not the differences! So step two – which I learned in lesson two of fast.ai is to separate out the different CNN layers to see what the interpretation is at each layer. Then it is to use all manner of prayer tech, prayerwheels, tefillin, and maybe create an ai that creates new prayer tech – that would be kind of amazing.
But practically,I think it would be amazing to write an article around the tech behind the machine learning, and with the CNN visualizations this is really approachable.