1% improvement is much different then a 50% improvement. There are diminishing returns when you start getting towards optimized coding or even something like scrypt, which is running into a utilization envelope.
Did 15% on scrypt-jane. Not 50%, but still substantial, and there's more in it.
Which is not substantial compared to 50%. We're talking about improvement per hour of time put into it. 15% gap can be closed and made up for a lot more then a 50% gap.
If you get a highly optimized public miner from the start, then private kernals will die off, which is good. It's actually sorta funny, it turns out scrypt was still the best algo to mine on. Reduced power usage was great, but that gave way to private kernels which destroyed that.
No worries, I'll have fun with more reduced power usage algos later