Conversation

Notices

1. Embed this notice
   mekka okereke :verified: (mekkaokereke@hachyderm.io)'s status on Friday, 01-Nov-2024 21:35:35 JST mekka okereke :verified:

   100%! We did it! 🏆🥇

   Great job everyone!

   https://www.geekwire.com/2024/ai-overwhelmingly-prefers-white-and-male-job-candidates-in-new-test-of-resume-screening-bias/

   • Rich Felker repeated this.
   • Embed this notice
     Tom Bellin :picardfacepalm: (tob@hachyderm.io)'s status on Friday, 01-Nov-2024 21:41:04 JST Tom Bellin :picardfacepalm:

     in reply to

     @mekkaokereke I'm sure this is an easy fix. Maybe put "don't be racist" in the prompt.

   • Embed this notice
     mekka okereke :verified: (mekkaokereke@hachyderm.io)'s status on Friday, 01-Nov-2024 21:45:55 JST mekka okereke :verified:

     in reply to

     • Tom Bellin :picardfacepalm:

     @tob 🤣

   • Embed this notice
     Rob Ricci (ricci@discuss.systems)'s status on Friday, 01-Nov-2024 21:49:13 JST Rob Ricci

     in reply to

     @mekkaokereke Turning Test but for racism

   • Embed this notice
     Dana Fried (tess@mastodon.social)'s status on Friday, 01-Nov-2024 23:53:34 JST Dana Fried

     in reply to

     @mekkaokereke didn't Amazon try this - and immediately shut it down for the same reason - like a decade ago?

   • Embed this notice
     mekka okereke :verified: (mekkaokereke@hachyderm.io)'s status on Saturday, 02-Nov-2024 00:00:41 JST mekka okereke :verified:

     in reply to

     • Dana Fried

     @tess

     No, that was completely different! 🤡

     That was a primitive model, using random forests or something, that accidentally quantified and exposed bias. Old school. Remedial. An inch above trying to make fire by rubbing sticks together, or slapping flint rocks together above a clump of moss. Basic.

     This is cutting edge LLMs, based on Transformer architectures... that accidentally quantified and exposed bias. High tech! Much innovation!

     The future is now(tm)!

   • Embed this notice
     Galbinus Caeli 🌯 (skiphuffman@astrodon.social)'s status on Saturday, 02-Nov-2024 00:09:32 JST Galbinus Caeli 🌯

     in reply to

     @mekkaokereke

     Generally unsurprising. Use biased data to produce a biased algorithm, get biased results.

     But I am surprised that "name" is included in the evaluation factors. I can see no use for including that.

     Rich Felker repeated this.
   • Embed this notice
     Dieu (hllizi@hespere.de)'s status on Saturday, 02-Nov-2024 03:19:29 JST Dieu

     in reply to

     • Galbinus Caeli 🌯

     @SkipHuffman @mekkaokereke That's part of the secret sauce! If half your workforce is named "John", another John cannot hurt. And that's the kind of invaluable functionality that's promising to convince droves of CEOs they have hit pure gold.