elsehow / aaronson-oracle Goto Github PK

I downloaded, unzipped and ran the index.html file from the example directory, but it seems that it cannot find bundle.js.

I am sure this program will fail badly

I'm assuming 'ffffg' is a typo and this should be 'ffffd'?

Are there any plans to circumvent the 2^5 "attacks" such as ddddd ffffff d ffff dd fff d f d fff ddd ff, which always yield a 0% prediction rate? Or is this project specifically aimed at implementing a 5-gram prediction with no enhancements?

I've been having fun with this, thanks!

Currently you cannot use the demo on mobile devices. as you - well - have no keyboard there.

And even if you would make the software keyboard showing up, this is still very inconvenient.

It would be nice to see the record of what the oracle was guessing against what the user provided.

Or maybe it's the same if you didn't deploy since 2016.

ff dd f dd f dd f d ff ddd f dd f dd f d f

and if you are lazy as I am:
sleep 4s; xdotool type 'ff dd f dd f dd f d ff ddd f dd f dd f d f' and focus your browser in 4 seconds afterwards.

I want to check a batched input.

For example, I want to check GPT 3.5 and GPT 4 outputs.

Is it a problem or intentional that it always guesses d first regardless of the first five letters entered?
Also if you hold down f forever* it seems to only get to 99% accuracy but holding down d gets to 100%

*more than 201 guesses

Or instead, reorder the way the guesses are displayed to reflect the text (or perhaps change "most recent" to "least recent"; but that would sound awkward).

The project has implications in terms of both psychological and commuter science research. I am wondering if you have made any latex summaries of the method you use to make predictions, and also if you have investigated more generally terminology regarding people's reaction to various results when told a computer predicts their actions.

Also just a small bug, there is no way that I guessed enough times for that trailing 1 to be meaningful. make sure you handle floating point error appropriately.

Interesting results...

I would like to port this to C. And then put it into Arduino board. Is it possible to provide pure JS implementation without using external libraries? Or a c/c++ implementation?

Should we make some "training set size," collecting a certain number of keypresses before we start displaying the Oracle's guesses, and calculating a rolling mean of its accuracy? The first 10 guesses, for example?