• Sources: primary, discussion
  • Summary: In the interview, Blow says current models write passable small throwaway code, such as a 30-line Excel script that cut his time from about five hours to one, but do not reason well enough to maintain the constraints between parts of a growing program. He asks where the output is if AI has made programmers 5 to 10 times more productive, and says he sees no new software capability. He suggests an AlphaZero-style system trained on program rules might suit code better than LLMs. The Hacker News thread reached 253 points and 261 comments, more comments than points.
  • Why it matters: The thread shows practitioners still dispute whether agent-written code yields measurable gains beyond small, well-bounded tasks.