Open dqian3 opened 2 years ago
Ended up trying to use wiki2vec anyways, got some interesting results. Essentially, I just tried going to the the best link in terms of cosine distance compared to the end article. Here are a few results:
Walt Whitman
to Walt Disney
:
Moving to Ray Bradbury, cos dist = 0.4334975526113556
Moving to Chuck Jones, cos dist = 0.3116722861072083
Moving to Dick Huemer, cos dist = 0.3034736034202481
Found link in Dick Huemer!
Navigation
to Barn owl
Moving to Spatial cognition, cos dist = 0.5895543313800581
Moving to Animal navigation, cos dist = 0.45084727252013634
Moving to Wood mouse, cos dist = 0.2966515170916927
Moving to Yellow-necked mouse, cos dist = 0.2724900744214668
Moving to Wood mouse, cos dist = 0.2966515170916927
Moving to Yellow-necked mouse, cos dist = 0.2724900744214668
...
Got stuck in a loop! After this I started keeping track of "visited", but thought it was interesting that there are loops.
Betty White
to IKEA
Moving to Mars, Incorporated, cos dist = 0.5046801705200095
Moving to Panera Bread, cos dist = 0.3983350386764827
Moving to Starbucks, cos dist = 0.35876893737395177
Moving to Costco, cos dist = 0.3631666245500923
Found link in Costco!
Could use something like this to generate "hints"
https://beta.openai.com/playground/p/9jY0NRB3SHQY8svpLodr5iQW?model=text-davinci-002
Added a visited set to solve the cycling, but we can probably make this better
Moving to Animal navigation, cos dist = 0.45084727252013634
Moving to Wood mouse, cos dist = 0.2966515170916927
Moving to Yellow-necked mouse, cos dist = 0.2724900744214668
Moving to Small Japanese field mouse, cos dist = 0.39043943855444097
Moving to Chevrier's field mouse, cos dist = 0.3917216049221972
Moving to Large Japanese field mouse, cos dist = 0.394882333691858
Moving to Striped field mouse, cos dist = 0.3975056767120768
Moving to Ward's field mouse, cos dist = 0.4050401093075543
Moving to Steppe field mouse, cos dist = 0.40844890464237826
Moving to Rodent, cos dist = 0.4109964176361306
Moving to Bank vole, cos dist = 0.26984503824414297
Moving to Tawny owl, cos dist = 0.22154424337359546
Found link in Tawny owl!
Started a new repo to keep track of some of the experiments: https://github.com/wikispeedruns/wikispeedruns-bot
We want to try making a bot that only uses the information available to users, as well as pretrained models.
We thought about using wiki2vec but it would be somewhat cheating, since it uses the link graph information in it's embeddings already.