@enoch_exe_inc@mastodon.social
2024-06-15 21:16 UTC
@dreamwinder@mstdn.social @enoch_exe_inc@mastodon.social @jensorensen@mastodon.social It isn’t quite that simple. Remember Microsoft Tay back in 2017? It was released to the public for about a day because Microsoft’s researchers made the mistake of allowing it to learn from anyone on the Internet who talk to it, and in a few hours learned every racial slur and defence of Hitler there was.
Replies (2)
-
@dreamwinder@mstdn.social 2024-06-15 21:19
@enoch_exe_inc@mastodon.social @jensorensen@mastodon.social Right, but the point of Tay was that the interactions it had from initial release was in effect its training data. Its opinions were changing in real time based on who spoke to it. That setup is much more easily abusable than ChatGPT, in spite of the inherent flaws of both systems.
-
@enoch_exe_inc@mastodon.social 2024-06-15 21:25
@dreamwinder@mstdn.social @enoch_exe_inc@mastodon.social @jensorensen@mastodon.social It took a fairly huge effort on OpenAI’s part to make ChatGPT as inoffensive as possible—*mainly* so that it doesn’t attract (as much) negative attention. LLMs with fewer (or without) such restrictions will spew as much nonsense as any human.