• 3 Posts
  • 29 Comments
Joined 2 years ago
cake
Cake day: October 24th, 2023

help-circle





  • No, because they don’t do math. If the LLM calls a script to do the math and just formats the input it might get accurate results consistently… but you just invented a machine to press calculator buttons for you at that point which is hilariously energy inefficient. This is unacceptable from a cost and reliability standpoint. If you’re familiar with enterprise reliability metrics you’d weep at the thought of a multistage process where each step had a single 9 and no visibility to underlying model tuning that can change outputs in wildly unexpected ways.


  • Sure, here’s an opinion.

    Banning is permanent and shouldn’t be first or immediate response. Repeat offenders that cross some quality or quanity threshhold may deserve that, but you should adopt power rangers rules and seek proportional responses, and only escalate as a response where possible.

    Bans should be transparent, contestable, and consistent in their application. However fair or unfair the rules you settle on, the perception of that consistency and impartiality influences the communitiea reaction. Too gentle and your community’s purpose blurs into something unintended, too harsh and your users will flee for greener pastures.

    Asking instead of dictating is the right approach in my opinion so I think you’re aimed in a good direction.

    Three strikes is where I would start, but maybe some strikes count for more than others? This is a hard problem and the answer will change over time. In cases where you can’t be consistent though, you must be transparent to salvage the trust you’re eroding.