Friday, February 16, 2024

Meta-Prompting AI

I’ve discovered a juicy way to prompt AI that hasn’t been much discussed. It’s juicy because it’s super helpful, a bit unexpected, and viscerally shows that AI does possess a sort of self-aware intelligence. It's hard to go back to Siri or Alexa after using language model AI this way.

Say you've asked the AI to produce some result, and, as is often the case, it's not exactly what you asked for. There are flaws. You point out a flaw, and the AI, per its programming, fawns all over you with apologies, and offers a new version, which, ugh, has other, similar flaws. You point out a few. Rinse, repeat, rinse repeat. Lots of apologies, lots of incremental progress, but the piecemeal improvement begins to annoy. So here’s my trick:
"It took a lot of back/forth, above, to get what I wanted. Without apologizing or explaining the limits of AI, can you review it and advise me, now that you better understand what I wanted, how I might have prompted you in the first place to get to this result more quickly and easily? Where did I go wrong?"
Of course, people rarely ask an AI (or any other sort of intelligence) how they could do better. That’s not a direction humans are prone to choosing. But the AI, cognizant of the previous interaction, can evaluate it from a higher level to help you see how it might have been directed to the present point more efficiently. It can, in other words, coach you to use it better.

The AI will also reframe(!) to an even higher level. "Would that style of prompting also help in this somewhat related sort of hypothetical case?" The AI will hop effortlessly to the higher level, and analyze broader applicability. You can even go all the way and ask it to suggest far-flung applications for the suggested style of prompting. A whole vein of self-aware guidance is available! And the AI actually seems (it's not real) to find it refreshing to think/help in this way. It seems (it's not real) to enjoy it.

FWIW I access AI tools via TypingMind, a friendly interface for AI interactions (accessible on both desktop and mobile), with easy entry into various AI programs (language stuff, image stuff, specialized knowledge, etc), and stores searchable logs of your interactions. There are other such services, but TypingMind is popular and well-liked. You'll need to also create an account on OpenAI, but you're coached through that process.


Note that this trick works best once you've finally gotten a good result. But you can also interrupt the process to say, "I'm working hard to get you to fix a lot of fairly similar problems. Can you evaluate our interaction and try, freshly, to assess, at a higher level (now that you've gotten additional feedback), where I'm going with this? No apologies, caveats, or statements of intention, please." This actually works!

Then, if it does make a big leap, you can proceed with the prompt above (i.e. "How could I have prompted you to this point more efficiently from the start?"). And if you don't like the reply, just say "Give me a different answer!" Once, I thought of an even tighter prompt than the AI had suggested, as it affably admitted. It’s refreshing to interact with an ego-less intelligence!


1 comment:

  1. Very interesting! I will remember this when I take my daughter’s advice and start using AI to help my business

    ReplyDelete