Still underwhelmed by the “improved” autocorrect, at least as it’s implemented when using the QuickPath (swipe) keyboard. Fixing that may require something other than an LLM implementation. Probably requires more learning about how people actually swipe and then make adjustments to the algorithm based on that physical input. Much the same way that Ken Kocienda had to “secretly” enlarge the desired key so that touches in its vicinity would trigger that key rather than adjacent keys. Like “knowing” that if I type the letter “T”, I am about 100% more likely to want “H” to be the next letter rather than “G”, “J”, “B”, or “N”. And substantially more likely than “U” or “Y”.