I’m asking three questions about AI:
How do you talk to it?
By talking to it, I mean skillful prompting and, a bit more precise, procedures like @jbrukh’s Jargon and AI-attuned programming languages like Modular’s Mojo.
How does it think?
In one sense, a large language model (LLM) is just adding one word at a time, but many AI experts think something more is going on to create an LLM like GPT-4 almost magical reasoning abilities. It apparently has created a learning algorithm to understand and respond to questions, which was unexpected.
How is it built?
There’s a lot to unpack here, but simply put:
LLMs, we’ve all learned practically overnight, use deep neural networks to create natural language text.
The neural network learns the statistical patterns of language by adjusting parameters to predict the probability of a word given the previous words. Over time, it recognizes patterns well enough to generate human-like text.
Specifically, the architecture of an LLM, like GPT-4, is based on a transformer architecture. This architecture uses a mechanism called “attention” to weigh the importance of different words when generating the next word in a sentence.
How can they be improved? We’re going to find out.
Alignment
My guess is that these three questions point to ways we can align with AI.