The Shift of Frontier AI
Frontier models will shift toward different benchmarks to conquer.It appears we are approaching an interesting point in time with frontier models. They seem to be plateauing on categories like coding and the reasoning that goes with it. Now, I know that the benchmarks still show improvment, but when you take into account the increase in cost, model size, etc. it means less.
It will be interesting to see what happens when these companies shift their focus towards spaces like biology. Part of the challenge will be that they need to collect data, have benchmarks for training, and the space has less competition which has positives and negatives.
Even if biology isn't the next area these companies shift toward, they'll still aim to dominate another field as their room for growth in coding capabilities dries up. Another interesting piece to think about will be how these models continue to adapt context-wise. How will they be able to keep up with entirely new tasks that differ significantly from coding?
The question I am left with is: Will we end up in a world with many specialized models or will large frontier companies still try to have one model able to do everything?