A reshared comment says Humans&'s Trickle Test studies a failure mode and calls the release's timing “very cool.”
A comment reshared on October 13 says Humans& had just released Persimmon and that its Trickle Test studies a failure mode. The commenter called the timing “very cool.”
539
2 Sources, first seen ago
—
Not ranked yet
—
Not ranked yet
@philotoralso very cool timing @humansand just released Persimmon, and their Trickle Test (@niloofar_mire) studies a failure mode that looks very similar to one we found independently: assistant models tend to overshare when they’re asked to play the user. feels like user modeling is becoming a real research problem, would love to chat with anyone working in user proxy research :)
@niloofar_mireRT @philotor: also very cool timing @humansand just released Persimmon, and their Trickle Test (@niloofar_mire) studies a failure mode th…@philotoralso very cool timing @humansand just released Persimmon, and their Trickle Test (@niloofar_mire) studies a failure mode that looks very similar to one we found independently: assistant models tend to overshare when they’re asked to play the user. feels like user modeling is becoming a real research problem, would love to chat with anyone working in user proxy research :)@niloofar_mireRT @philotor: also very cool timing @humansand just released Persimmon, and their Trickle Test (@niloofar_mire) studies a failure mode th…@philotoralso very cool timing @humansand just released Persimmon, and their Trickle Test (@niloofar_mire) studies a failure mode that looks very similar to one we found independently: assistant models tend to overshare when they’re asked to play the user. feels like user modeling is becoming a real research problem, would love to chat with anyone working in user proxy research :)@niloofar_mireRT @philotor: also very cool timing @humansand just released Persimmon, and their Trickle Test (@niloofar_mire) studies a failure mode th…@philotoralso very cool timing @humansand just released Persimmon, and their Trickle Test (@niloofar_mire) studies a failure mode that looks very similar to one we found independently: assistant models tend to overshare when they’re asked to play the user. feels like user modeling is becoming a real research problem, would love to chat with anyone working in user proxy research :)@niloofar_mireRT @philotor: also very cool timing @humansand just released Persimmon, and their Trickle Test (@niloofar_mire) studies a failure mode th…
