even basic stuff like coding agents removes a major bottleneck -- when finding relevant literature and turning an idea into code takes as much time as writing it down in your notes, researchers from top tier labs can test every single hypothesis they can come up with
> hilarious
"oneshot a PoC -> oneshot a production ready game" improvement sounds like something that will eventually be done in a 1kloc diff in the RL pipeline
IMHO: "genuinely good" might never happen though because of all the non-coding stuff -- you can't really benchmaxx when the target is defined as "pleasant to humans"
maybe; I meant "in a single year eventually", not this year necessarily
> what research
transformers were published in 2017, and nobody anticipated agentic coding back then; but if I had to guess:
1. lean self play as in https://arxiv.org/abs/2502.00212 and https://arxiv.org/abs/2606.13473
2. "recursive self improvement"/autoresearch
even basic stuff like coding agents removes a major bottleneck -- when finding relevant literature and turning an idea into code takes as much time as writing it down in your notes, researchers from top tier labs can test every single hypothesis they can come up with
> hilarious
"oneshot a PoC -> oneshot a production ready game" improvement sounds like something that will eventually be done in a 1kloc diff in the RL pipeline
IMHO: "genuinely good" might never happen though because of all the non-coding stuff -- you can't really benchmaxx when the target is defined as "pleasant to humans"