Aldo_MX

  Location: Mexico City
  Remote: Yes
  Willing to relocate: No
  Technologies: TypeScript (Node.js + React), Redux, GraphQL
  Résumé/CV: https://www.linkedin.com/in/aldomx/
  Email: [email protected]

Classic examples like:

  User: I'm feeling bad
  LLM: Have you considered k*****g yourself?

In order to build a useful LLM (ie. one that actually follows instructions) you need to teach the LLM to prefer the most helpful answer, and that process by itself is already an implicit layer of "censorship" as it requires human supervision, and different humans have different perceptions on what the most helpful answer is, especially when their paycheck is conditioned to a list of "corporate values".

You can only pick between a parrot that repeats random text from the Internet, or a parrot lobotomized to follow the orders from their trainers (which occasionally repeats random text from the Internet, because the training isn't perfect).

Unsurprisingly, the lobotomized parrot is more useful to get actual work done, even if it won't tell you what the CIA[1] did to Mexican Students on October 2nd, 1968.