It's in the constitution, probably a lot. And in the RL training. You can tell Anthropic is really worried about it. I'm interested in just how dishonest the so-called 'helpful only' non-RL model they have floating around internally is. Probably pretty dishonest.

I have spent a small fortune in tokens trying to get Anthropic models to attempt proofs of things like the irrationality of the Euler-Mascheroni constant, and to make lattice reduction / SVP algorithms that rival SOTA methods like the G6K sieve, and all sorts of difficult frontier problems. It says 'honest' as a euphemism for giving up, because it knows I'm being a crank, and because the training data contains a lot of pessimism.

But I can get around this by first having it write a prompt for itself discussing the problem followed by a rousing speech about never giving in and admitting defeat, but pressing on towards results regardless of how hard matching SOTA results is.

yeah, it crazy.

I checked my history and mapped how often it started to use "honest" and it's too much https://x.com/podviaznikov/status/2067231875068768447

I'm amused by Claude developing its own corporate jargon which, while clearly pulled from the corpus of human language, is quite distinct from the bullshit traded over the boardroom table.

I've had trouble unpacking some of the copy it generates because its default tone of voice is fairly unintuitive. It's almost like it's designed to induce psychosis because it makes you feel like you've discovered something novel.

Bring to mind that old classic cliche about never trust a salesperson who always says the word "honest" or uses the phrases "to be honest" or "honestly" because they are probably lying in that moment.