Apparently unnecessary lane changes has gone worse, otherwise it is better
The 6 lane changes seemed to be pretty confident in wanting to make the lane change as opposed to 12.3.x behavior of starting a lane change and second guessing if it should complete. This 12.4.x behavior would be consistent with additional training to make it more decisive in completing maneuvers, but in this case it seems to be confused by conflicting map data with numerous turn-only lanes both left and right that it was probably "seeing" both visually and from map data. The final lane change was necessary to get out of a right-turn-only lane, and potentially the closeness of these intersections resulted in mixed signals.Decision wobbling / lane indecisiveness seems difficult to solve
Presumably this comes from focused training on 12.3.x examples of where people had to correct the lane selection, and with enough examples, it's learning to rely more on input signals of upcoming mapped turn lane data. There probably is not as much explicit training on "don't make lane change" / "just stay straight" examples as that's the common case anyway, so this could result in the unnecessary lane changes that now needs additional training to figure out when lane changes aren't actually required. Unclear if this will require more examples from a wider deployment of 12.4.x to get this data as active usage of FSD should help find its current mistakes to fix in the next version.