As a claude code power user, when I get the 'rate the feedback on Claude' pop up, I used to say good or fine out of habit, and immediately after sending this feedback, I felt an instant degradation and mistakes that usually don't happen.
Now I dismiss it every time and the quality is more consistent.
Complete adhoc and personal experience but something I've observed, wouldn't be surprised if they nerfed on a per session basis
When the source maps leaked for the harness, it was revealed that they track how often tool use is rejected and how many times you say "fuck". They definitely try to track frustration sentiment.
That said, given the propensity for mature code bases to have "fuck" in commit messages/comments and those are typically of higher quality, I curse up a storm when the clankers make mistakes, if only to try put more quality-code valence into context. <a href="https://news.ycombinator.com/item?id=36584464">https://news.ycombinator.com/item?id=36584464
On the one hand, yes, on the other hand it would also not be too hard to do routing based on such a parameter and Antrophic has shown (through Fable and Mythos) that they have such a transparent routing system in place.
In fact, since they have some rule based system (if bioenegineering or security, route to degraded model) it would be almost trivial to add 'user has filed feedback' to it.
Not saying this is what happens, but just that it's not as insane as it sounds.
I could totally see them running an experiment to test user's stickiness/quality perception with decreased model performance. Facebook was doing exactly this in 2016. They tested the loyalty of Android users by secretly introducing errors that would crash the app to find the threshold at which a person would abandon. That was 10 years ago - imagine what the state of the art in user manipulation is like now.
Its truly naive to think this is impossible or even improbable. You can say there's no proof or that its not the case but certainly much more has been done to juice 5 star ratings.
You’re probably participating in 100 or more A/B tests per year without knowing it and getting nothing in return other than a minor contribution to the company’s internal knowledge about what customers like.
As if it wasn't already? The feedback button may tag it a certain way but the AI company certainly isn't going to let that sweet sweet user interaction go to waste just because they didn't rate it.
Are you under the mistaken impression that hitting a 1, 2, or 3 for bad/fine/good is itself sufficient permission to allow Anthropic to train off your chat?
They explicitly ask, after the rating, whether they can look at your chat. You can just say "no" to that, if you believe that they are following the rules they say they are.
Your original statement of "using feedback gives them permission to train" is just plain false.
Truth be told, I never used the feedback function because of the clauses that reference the functionality in connection with training/service improvement.
msejas · · focus · HN ↗
Now I dismiss it every time and the quality is more consistent.
Complete adhoc and personal experience but something I've observed, wouldn't be surprised if they nerfed on a per session basis
itopaloglu83 · · focus · HN ↗
And strangely, expressing frustration multiple times in a row would reliably trigger a feedback popup as well.
okwhateverdude · · focus · HN ↗
That said, given the propensity for mature code bases to have "fuck" in commit messages/comments and those are typically of higher quality, I curse up a storm when the clankers make mistakes, if only to try put more quality-code valence into context. <a href="https://news.ycombinator.com/item?id=36584464">https://news.ycombinator.com/item?id=36584464
ClikeX · · focus · HN ↗
wunderlotus · · focus · HN ↗
jcutrell · · focus · HN ↗
Galilyou · · focus · HN ↗
xnorswap · · focus · HN ↗
Random performance is random, your brain will jump through hurdles to fit patterns where there aren't any.
unglaublich · · focus · HN ↗
In fact, since they have some rule based system (if bioenegineering or security, route to degraded model) it would be almost trivial to add 'user has filed feedback' to it.
Not saying this is what happens, but just that it's not as insane as it sounds.
ftchd · · focus · HN ↗
smashed · · focus · HN ↗
And it's also the kind of solution a misaligned agent would implement:
Make sure users are happy about their experience but also optimize resources.
The obvious solution is to route users based on feedback.
bb123 · · focus · HN ↗
jayd16 · · focus · HN ↗
Aurornis · · focus · HN ↗
SilverSlash · · focus · HN ↗
Aurornis · · focus · HN ↗
redanddead · · focus · HN ↗
grim_io · · focus · HN ↗
Using feedback on your sessions makes that session, and presumably all attached data, fair game for training.
margalabargala · · focus · HN ↗
grim_io · · focus · HN ↗
margalabargala · · focus · HN ↗
They explicitly ask, after the rating, whether they can look at your chat. You can just say "no" to that, if you believe that they are following the rules they say they are.
Your original statement of "using feedback gives them permission to train" is just plain false.
grim_io · · focus · HN ↗
iamjackg · · focus · HN ↗
Aachen · · focus · HN ↗