It's incredible to me that "coding," as we previously viewed it, has mostly been solved at this point. I first heard of the concept of AI-generated code back in 2015 when a friend of mine started working at a startup called @trygigster. Gigster's whole premise was to be a platform for hiring freelance software engineers. Their plan was to use the code those freelancers wrote for clients to train their neural nets. The hope was that these models could eventually generate code and replace freelance developers altogether. At the time, deep learning was starting to really catch on (this was before LLMs were discovered), and the Gigster team believed they could use the latest research - recurrent neural nets (RNNs) to train a machine to write code on its own. The problem with RNNs is that they generate gibberish that looks like valid code but doesn't actually work. They lack the ability to maintain large contexts and are very slow to train, so the idea of auto-generating code simply wasn't feasible then. Fast forward 10 years, and everything Gigster was trying to do is now available to everyone. The progress Anthropic, Google, and OpenAI have made in code generation is astonishing if you zoom out. What I thought was a pipe dream is now reality.(cygaar)