2 min read

ChatGPT: Living with a gnome in a box

I have been interacting intensively with GPT3, 3.5 and ChatGPT for about 6 months. It has been an exhilarating, bewildering experience. It has been an experience which has made me re-evaluate what it means to be human, what it means to be intelligent.

synesthesia - is when you hear music, but you see shapes. Or you hear a word or a name and instantly see a color. Synesthesia is a fancy name for when you experience one of your senses through another.

That's been my understanding of ChatGPT, colored from what I understand from both Tesla and Stable Diffusion.

ChatGPT feels like a gnome in a box, a partially intelligent creature, extremely limited, without a super solid understanding of the outside world, but weirdly competent and creative at some tasks.

It is excellent at style transfer, best illustrated visually.

Except it can do

  • scientific paper -> rap song
  • newspaper article. -> sonnet
  • legal document -> explanation to a 5 year old
  • code comments -> code
  • python code -> rust code
  • customer complaint -> legal notice

As you go down the list above, each item has more and more economic significance, and finally you realize that what a lot of what white collar work is.. for lack of a better word.. style transfer.

How does GPT achieve this style transfer? I know the technicalities behind word2vec etc.. but what it feels like when you interact with ChatGPT is that it understands underlying form well. It can manage iambic pentameter or customer service speak well.

It does this better than the 90th percentile human writer... so not superhuman or better than human but better than most humans.

Even this alone makes it exceptional in my view. Like we did math in our heads until we got the calculator.. and now no one does much multiplication without using a calculator. Similarly we used to write code based on commented definitions.. and in a few years we would just use GPT.

Think about that for a moment. Every single text style transfer task .. API documentation to code, travel plan to plane tickets, book to book review to exam question to exam answer,

  • Quarterly plan to monthly plan to weekly tasks
  • Task execution data to weekly summary to monthly report to quarterly financials

You can run a company this way. Some mix of GPT3 and heuristics and hooking up to external data sources for context.

That's the gnome in the box. A creation that's really good at a form of style transfer, and which we can hook up immediately with some rough hacks to do something interesting.