Re:LLMs trained on “A is B” fail to learn “B is A”
by ResearchCode on Saturday September 23, 2023 @07:12AM
(#37622171)
LLM engineers rushing to hard code B is A in their "AI". They call it reinforcement learning with human feedback.
[ Parent
| on HN ]
Re:LLMs trained on “A is B” fail to learn “B is A”
by throwaway02y on Saturday September 23, 2023 @07:16AM
(#37622197)
What do you do in your brain when you get something wrong?
Re:LLMs trained on “A is B” fail to learn “B is A”
by ResearchCode on Saturday September 23, 2023 @07:20AM
(#37622228)
Learn the concept better, not an exception to an incorrect rule. It seems like the model architecture just isn't good enough.
Re:LLMs trained on “A is B” fail to learn “B is A”
by adhesive_wombat on Saturday September 23, 2023 @07:54AM
(#37622442)
Sounds like learning irregular verb forms, or integration rules!
Re:LLMs trained on “A is B” fail to learn “B is A”
by Jensson on Saturday September 23, 2023 @08:46AM
(#37622773)
Hardcoding natural language rules doesn't work, we know that from decades of trying. Therefore hardcoding these natural language rules in the model wont work either, it will fail for the same reason, you can't encode enough rules to handle things as well as a human.
Re:LLMs trained on “A is B” fail to learn “B is A”
by throwaway02y on Saturday September 23, 2023 @11:10AM
(#37623954)
Are we arguing ChatGPT is not AGI? I can agree with that.
As for encoding enough rules to handle things as well as a human, what things? Humans don't work with infinite knowledge so there must be a set of rules (however large) that would be "enough".