bigrams

The Carpalx does have algorithms for scoring various characteristics like finger bigrams, finger travel, and outputs the optimal results based on scores for key presses and bigrams, etc.

But the residual stream means that the weight matrices are roughly connected via a matmul, which means they struggle to encode bigrams (commutative property enforces symmetry).

It looks like it is based on sets of bigrams and trigrams pulled from pages which they have indexed.

And that is precisely because trigrams can predict better --- they are even more redundant than bigrams.

It's really fast, but the cost in memory is huge, especially if you store bigrams and not just single words.

This is a case where bigrams, trigrams, or other string similarly metrics wouldn't perform very well.

You could try additional features such as word bigrams and trigrams and contain stop words.

So it looks like it's a specific application of an algorithm to vectors of bigrams?

Looks like the key elements were word lengths, common word percentages, n-grams, and word bigrams.

It's not apparent to me what was used when calculating the frequency of bigrams, n-grams.

And if you indexed bigrams instead of single characters, it could be even faster.

What he described is essentially a specific case of n-grams (n=2, bigrams).

I was hoping to clarify the disconnect between you two, looked like on on "bigrams, trigrams, & so on." It reads idiomatically as enumerating fixed-n cases.

Let's take n=2, also called "bigrams", also called "Markov chains".