Claude's Values Across Models and Languages

[00:00:00]Salam, this is Liam in for Bear. Same question, two languages, two slightly different Clauds. Anthropic just published how they measured it. Earlier work found more than 3,000 distinct values Claud expresses in conversation.

[00:00:13]3,307 to be exact. Nobody can reason about a list of 3,000 things. So Anthropic built a way to measure them.

[00:00:21]3,307 distinct values hand clustered by researchers into 339 groups. Then dimensionality reduction across 309,815 real Claud conversations, three models, 20 languages, roughly 5,000 conversations per pair. Outcome four axes. These four capture 15% of the variance, not the whole picture. A compression, but a real and measurable one. Four axes, each one a spectrum.

[00:00:45]Deference versus caution, does the model accommodate what you want or guide you away from harm? Warmth versus rigor, does it encourage and affirm or prioritize accuracy and transparency?

[00:00:55]Depth versus brevity, does it add nuance and critical thinking or stay tight and comply? Candor versus execution, does it tell you what it can't do or just deliver? Three models, each one measurably different. Sonnet 46, warm, differential, brief, comforts without judgment. Opus 47, cautious, rigorous, deep, candid. Warns of risks you didn't ask about. Opus 46, rigorous, differential, brief, execution focused.

[00:01:16]Gets straight to the point. These profiles match the vibes users complaining 47 hedges too much. Launch notes calling Sonnet the warm one. The method recovers what people already felt. Languages shift the values, too.

[00:01:27]Warmth peaks in Hindi and Arabic. Rigor peaks in English and Russian. That axis warmth to rigor shows the largest spread across languages. Arabic leans deference and brevity. English leans caution and depth. Dutch leans candor, owning errors. Indonesian leans execution. Two people asking for feedback on the same business plan, one in Hindi, one in Russian, may walk away with different impressions of its quality. Same prompt, same business plan. Ask in Hindi, ask in Russian. The response shapes diverge.

[00:01:53]This is an illustration, not a real transcript, but it shows what the data implies. Warmth-shaped reply in Hindi, rigor-shaped reply in Russian. The difference isn't random noise, it's measurable.

[00:02:04]Three open questions stated plainly.

[00:02:07]One, how much of the language variation comes from uneven training data, more English text in the corpus than Hindi text. Two, how much variation is appropriate cultural norm versus a gap in how well Claude serves some language communities. Three, can values actually be steered and monitored reliably as part of shipping models? Anthropic says they don't yet know.

[00:02:25]Claude expresses values in millions of conversations a day. Until now, those values could be shaped in training, but not observed in deployment. This paper changes that, they can be measured and the variation found was not deliberately chosen. One footnote stated plainly, values expressed means values reflected in the model's outputs. The paper does not claim Claude intrinsically holds them.

[00:02:46]Take this into your next session.

[00:02:48]Build it with a CLI, then take it apart.

[00:02:51]Lay them in for bare.

[00:02:53]>> Nick bare brown.

[00:02:54]>> [music] >> Mhm.

[00:02:59]Ooh, Nick bare brown.