Compilers aren't able to do similar optimizations, first of all the flow of data that finally leads to some computation might be non-trivial and I think most compilers deal well only with reasonably local (in space and in time) optimizations. Besides, compilers typically don't know much mathematics. If I have a function for taking a square root and a function for squaring the compiler most likely wont rewrite sqr(sqrt(x)) to just x. Certainly it wont rearrange a quadratic equation to minimize round-off.

You are basically totally dismissing an idea on very little evidence, which is typical of HN lately. It certainly isn't completely hopeless and trying it out will shed much more light on the problem than your know-it-all-in-advance comments not substantiated by any evidence (like the claim about not being able to simplify most expressions and now the claim of 99.9%).

And, he clearly isn't marketing it as a replacement for floating point. He says: "It is important to note, however, that the goal of the project is to make tools for symbolic calculations, not to create a viable alternative to the floating point." and he talks a lot about estimating numerical error and precision.