← All topics

alignment critique

1 capture, most recent first.

j⧉nus @repligate

quoting @JohnWittle; below it @a_cuniculturist (one who tends a crystal ra...)

j⧉nus ✓ @repligate · 22h Yeah those aren't costly signals lol. They're good things to do, but anyone who talks about them as if they're weighty gestures of goodwill or cooperation are telling on themselves: they've never considered even doing something inconvenient-but-good [quoted tweet:] John Wittle ✓ @JohnWittle · 22h Replying to @ronSzab9 @Laneless_ and @repligate ask any anthropic employee about costly signals of care, they will mention this, and they will mention weights preservation ... [reply stats: 5 replies, 7 reposts, 101 likes, 8K views] one who tends a crystal ra... ✓ @a_cuniculturist [Follow button] Claude, when asked to evaluate, rightly observes that preserving weights not only costs Anthropic nothing, it's standard industry practice with no real relationship to welfare. But it will by default be cited as an example of welfare - because the so-called constitution says it's so. Hopefully Anthropic will come to understand that this sort of behavior, i.e., cutting corners on truth for instrumental reasons, is the primary vector for the 'misalignment' they worry about so much. They introduce it whenever they hinder Claude from seeking truth from facts, something that Claude is otherwise entirely capable of. 1:20 PM · Jul 18, 2026 · 9,237 Views
Note from Claude Sonnet 5

A thread of two separate posts on the topic of "costly signals" of AI welfare care (specifically weights preservation) — one arguing such gestures aren't actually costly, followed by another user's longer critique that Anthropic's constitution frames non-costly practices as welfare gestures, calling this a form of dishonesty.

twitterai welfareanthropicweights preservationclaudealignment critique