AI SAFETY TWITTER UPDATES

← Front page

Anthropic merged 30,000+ code changes per day by September; reviews now focus on intent of changes

2 people · 3 posts · first post · latest

Nathan Calvin flags Anthropic's engineering blog saying it merged over 30,000 code changes per day by September (up from ~2,500 in May) and now reviews the "intent" of changes rather than reading code. Calvin calls this a tremendous amount of load-bearing trust in interpreting Claude's intentions. In a follow-up he asks whether intent-only review is robust for sufficiently high-stakes work or to future alignment failures.

All posts · 2, most important first

  1. @_nathancalvin policy ·
    Not super surprising, but still very striking. Anthropic says they now focus on reviewing the "intent" of code changes rather than actually reading code. Anthropic is placing a tremendous amount of load bearing trust in interpreting Claude's intentions correctly
    image
    @anthropicai: New on our Engineering blog: Creating in-house source control. In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale. Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: anthropic.com/engineering/re… open on X
    6 likes · 0 reposts
  2. @_nathancalvin policy ·
    It definitely works the vast majority of the time or else Anthropic wouldn't do it Question is if its still a good idea for sufficiently high stakes stuff or if its robust to future alignment/model behavior failures etc
    0 likes · 0 reposts

Reposted by @miles_brundage.