Every few years I start writing a programming language. None of them ship. None of them are supposed to.
The thing people miss about toy languages is that they're not failed startups - they're crossword puzzles. You sit down on a Sunday morning, you decide what if should mean this time, and by evening you've discovered something small about how meaning gets built. The language gets thrown away. The discovery doesn't.
Right now I'm on something I call filament - it's Lispy, but every value remembers which expression produced it. Errors print a graph instead of a stack trace. Will it be useful? No. Has it taught me something about why Smalltalk debuggers feel kinder than Java ones? Yes. That's the trade.