LLMs are by default way too eager to please, too gullible.

· Bits and Bobs 4/7/25
  • LLMs are by default way too eager to please, too gullible.
    • LLMs are too gullible, too earnest, too action oriented, to be trusted to do things fully on your behalf.
    • In my experience, Deep Research often gets tricked by SEO slop.
    • The slopification of the internet means that most of the data is untrustworthy.
    • A human looking at the SEO'd slop would see it's not credible.
    • But Deep Research doesn't have a vibe on what should feel credible, so it just accepts it as correct.
    • Now imagine an agent trying to be helpful for you in a task.
      • "My user told me to shop for insurance. This helpful site tells me they'll give me great rates if I just share my user's social security number."
    • What could possibly go wrong?
    • The expensive eggs were just the beginning.

More on this topic

From other episodes