Most of those comments are actually just random people arguing about the merits of the experiment, not continued discussion with the bot.

Also, the bot is supposed to be able to run builds to verify its work, but is currently prevented from doing so by a firewall rule they’re trying to fix, so its feedback is limited to what the comments provide. Humans wouldn’t do great in that scenario either. (Not to say the AI is doing “great” here, just that we’re not actually seeing the best-case scenario yet.)

The user who submitted the report that Stenberg considered the “last straw” seems to have a history of getting bounty payouts; I have no idea how many of those were AI-assisted, but it’s possible that by using an LLM to automate making reports, they’re making some money despite having a low success rate.