Hi HN!<p>I threw together a little prototype that uses GPT-3's logprob output to flag unlikely tokens, so they can be double-checked by a human proofreader. It's sort of a turbo-spellchecker: in addition to catching "They had <i>fuor</i> eggs", it can also catch "They had <i>for</i> eggs", "He asked me to <i>prostate</i> myself to the king", and even the somewhat subtle bug in this C code:<p><pre><code> for (int y = 0; y < HEIGHT; y++) {
for (int x = 0; x < WIDTH; x++) {
buf[y * HEIGHT + x] = 0;
}
}
</code></pre>
I wrote a blog post going into more detail here: <a href="https://vgel.me/posts/gpted-launch/" rel="nofollow">https://vgel.me/posts/gpted-launch/</a> .<p>Hope you like it, and curious to hear people's thoughts, feedback, or questions! :-)