Prompt engineering < Prompt Spellcasting

•

Given that new models of LLMs and Gen. AIs are getting released often (GPT-3.5, chatGpt, GPT4, Carper, SD2, Dalle) And the older prompts sometimes no longer work, it's no longer "prompt engineering" Engineering means stability This is more about re-discovery of magic - BIG 🧵

Folks like @goodside, @hardmaru and @tszzl are all crafting incantations, and sometimes sharing publicly on twitter, to great success, are a new class of prompt wizard elders, the ones who mastered the early days of LLM incantation. twitter.com/karpathy/status/1600031572442218497?s=20&t=KaH98EUMqwJJ9l9wirnP5A

Just look at this piece of code I wrote for chatGPT-Telegram bot I created. The textual incantation 🪄 part took me way longer to "figure out" than the code ⚙️ (engineering) part How to get chatGPT to always answer with "just" the thing I want.

The same goes for startups like @copy_ai and @jasper They are experts at pre-baking incantations into UX for consumers, just like... bottling potions that took a long time to craft. Their potions are reliable for the consumers, but making them took a lot of iteration.

For image synthesis brand of magic, we have a whole host of wizards, many of whom found jobs at @StabilityAI @scale_AI and @OpenAI Folks like @KyrickYoung @pharmapsychotic and @KaliYuga_ai among others, are leading the charge into magical latent space to extract wisdom

These companies are hiring wizards because sometimes they release things, they themselves don't understand, a lot of the understanding comes from iterative tinkering, taking notes, trying again, and chatting with your 🧙‍♂️ friends. The word Almanac comes to mind.

And let's be honest, if you tell a 10yo kid, that he can be a "software engineer" when he grows up, or he can be a "spell-casting wizard" 🧙‍♂️ Just imagine which reactions you'll get for which option!

chatGPT is getting tons of excitement, as folks discover its possibilities, but they also received an imagination unlock. Maybe due to the UX, maybe due to quality, chatGPT was able to shatter the glass in many folks heads of understanding the potential now and in 10 years

Nobody knows which other techniques will unlock which possibilities, but we're sure as heck going to find out. We're at the beginning of a double exponential growth, software and hardware. twitter.com/elonmusk/status/889749519021121536

Don't be discouraged by nay-sayers, the "ha ha chatGPT is so dumb, look" dunk tweets, the "it can't even draw hands" ilk. Some of them are scared of these changes! The more models are released as open source, the less those ideas are effective at blocking progress.

My advice? Get on the "I'm excited about progress" side of the track! It's fun over here, we're doing magic! (and it's safe for now) Why sulk at the inevitable progress when you can be like "what a time to be alive!" on this side of the fence? Become a novice prompt magician!