01001000 01100101 01101100 01101100 01101111
My prediction for 2030 is that models will be coding in a brand new language that they create for themselves.
Today's programming languages are very much geared towards bridging the communication between humans and computers.
It's done in a way that allows humans to express themselves (and understand expressions of others) in natural language.
With AI taking over coding, we don't need this.
At the same time I don't think it will be machine code that will be generated directly, specifically binary code. It would be a massive waste of resources during a time when compute resources are not only limited but also expensive.
There are algorithmic innovations happening already with AlphaEvolve. Why stop there? Why not go after a new abstraction layer?
This naturally runs into the data question. If there is no data that the model can use to train itself on this new programming language, how would it become capable?
My gut feeling is that AI models today have enough capabilities and expertise across different programming languages to come up with their own version that is suited for their strengths and weaknesses.