The website does lean a little more marketing-heavy than my blog but I think that is just how Taylor writes and presents on the internet. A lot of the fancy-sou

The website does lean a little more marketing-heavy than my blog but I think that is just how Taylor writes and presents on the internet. A lot of the fancy-sounding claims are just very boring details. The dependency fixing thing is mostly a claim based on something like static linking, where you can do the moral equivalent of require(x) and have x downloaded and checksum'd and loaded into memory---and then you keep using that version. Like a Merkle DAG. I don't know that this approach fixes all problems, but the solution is a lot less interesting than it sounds in that description IMO. "scrap" is just an expression but you can reference any expression by its hash since it's all serializable. That's also what he means by the next bullet. "Optimized for AI" is, IMO, a bad misrepresentation because what he means (I think) is since you put the use before the definition (f x . f = ...), it gives more room for sufficiently advanced autocomplete to fill in the definition. If you can picture a language where every expression is a Git object and you can pull in objects from other repos at will and you can take arbitrary snapshots of your program (a tree-equivalent), you kind of get an idea of what's going on.