There aren't any techniques he can use that aren't available to a TAS runner. TAS runners aren't just using reinforcement learning to let a computer learn how to play; they're thinking about possible techniques and glitches, and specifically programming their tools to try and exploit them.

What's common, especially in the SMB speed running community, is that the human and tool-assisted runners loosely collaborate. e.g., if one of the humans has an idea, before they spend days or months trying to figure out for themselves if it's even workable, they'll get someone to inexpensively proof-of-concept it with a tool-assisted speed run.