Filters are better than agents.
- Agents take actions on your behalf.
- They might take the wrong action, causing difficult-to-reverse downside.
- Dangerous!
- Filters[agy] help sort information and make recommendations.
- The end user decides whether to act or not.
- Having the user in the loop provides a check, preventing unexpected downside.
- Filters, if they're not perfect, don't hurt.
- But if it's great, it can give recommendations that are game-changing for a user.
- Capped downside, uncapped upside.
- Agents, if they're not perfect, can do real damage.
- Uncapped downside, uncapped upside.
- Downside can lead to game over.
- The user has such a bad experience they decide to never use it again.
- Or the user literally goes bankrupt.
- This makes downside more dangerous than upside is good.
- To make agents safe requires significant, carefully controlled guardrails.
- This is extremely hard in the limit.
- You have to anticipate the unanticipatable.
- The effort to define good guardrails could be much larger than the benefit of the things that could happen within the guardrails[agz].
- More cost than value, no incentive to activate.