Better at Zero Shot: As a fine-tuned “GPT-3.5 series” model and an InstructGPT sibling, it is a LOT better at zero-shot generation of text that follows your instructions

It has long term memory of up to 8192 tokens, and can take input and generate output about twice as long2 as GPT3.

It still can’t do math, generates false information about the real world, and writes bad code, and does not pass Turing, SAT, or IQ tests.

the best usecases are where creativity is more valued than precision: brainstorming, drafting, presenting information in creative ways.

The biggest unresolved debate is how much a ChatGPT-like experience can replace Google: on one hand, ChatGPT answers questions far more directly and legibly than a Google results page8; on the other, its answers are often incorrect and unsourced

I do agree that this is the birth of something that looks like AGI. The main factor is learning that we have somehow gone much faster with Reinforcement Learning via Human Feedback than I knew in October