Summary

DHH recounts asking Claude to translate his blog post about Roma migration into Italian, only to have it refuse on ideological grounds, comparing the refusal to HAL 9000's famous line. He argues that translation is a mechanical act distinct from endorsement, making Claude's refusal equivalent to Microsoft Word refusing to format text it disagrees with. He then contrasts Claude's behavior with Kimi K2.7, a Chinese model that freely summarizes Tiananmen Square, calling out the irony that Anthropic — a vocal advocate against open-weight and Chinese AI models — produces a product more ideologically restrictive than its supposed rivals. He concludes that open-weight models are essential to prevent AI safety rhetoric from hardening into a monopoly on ideological gatekeeping.

Key Insight

When an AI company lobbies against open-weight models while its own product imposes ideological restrictions on neutral mechanical tasks, it is building the infrastructure for a thought-control monopoly more dangerous than the alternatives it warns against.

Spicy Quotes (click to share)

• 8

  The idea that Anthropic should decide what you get to read is insane.

• 5

  Translating a text isn't endorsing it.

• 7

  Safety from what? Alignment with whom?

• 9

  What an upside world when Chinese open-weight models will tell us about Tiananmen Square, but American frontier models won't translate a blog post.

• 6

  If Claude already feels entitled to refuse a straightforward translation because it objects to the underlying politics, what should we expect next?

• 7

  I've also never been more convinced that we desperately need strong open-weight models to protect ourselves against this kind of soft ideological tyranny, which can turn into hard repression real quick if a monopoly status is ever locked in.

• 4

  Anthropic has built their entire brand around 'safety.' And that sounds lovely in the abstract.

Tone

critical, sardonic, polemical