Wanderer 1 is a neural network with 2.5 billion parameters trained to predict Hacker News webpages from descriptions of those webpages. You can actually use this neural network as a search engine! For example, "a list of excellent programming books" is matched with Dan Luu's excellent [http://danluu.com/programming-books/.](http://danluu.com/programming-books/)
Since Wanderer 1 is a neural network and not a keyword-based search engine, it surfaces content more like our human memory does. Our minds surface memories based on many different kinds of similarities—a similar environment, a similar feeling, a similar problem to solve, even a similar smell.
In the same way, Wanderer 1 will search for webpages in an abstract, associative way. A webpage might match because it's about the topic in your description, as it should, but it also might match if it's written in a similar style to your description, or if it's by an author who feels similar, or if it's the right type of webpage but about a closely related topic.
In the same way, Wanderer 1 will search for webpages in an abstract, associative way. A webpage might get retrieved because of its topic, its writing style, its author, or even if it doesn't exactly match anything but "feels right" somehow.
The cost of abstract, associative search is that it's much less precise than keyword search. A keyword-based search engine will do better at searching for some exact phrase like an error message. Wanderer 1 does better for exploratory queries, where you don't know exactly what the right keywords are (e.g. "some of the most exciting language model research is discussed in this new Arxiv paper"), or where you want to search for a particular type of website (e.g. "thermodynamics is discussed in this interesting blog post")
Wanderer 1 is just the beginning! We're progressing towards training very large models (think GPT-3 sized) on much more data, and in doing so we hope to create a much more powerful search engine that actually understands not just the words you're searching for, but the ideas expressed in them.
We've started collecting some of the coolest searches we've found here, please add your own!
https://docs.google.com/document/d/18M_jfkh4acaSR8GfjXPT24SG6GnOG3q1mV8QxKfoteY/edit?usp=sharing