Exactly!!

Finally, there is a Div 3 in 3 days! :)

That is a very difficult question. Nowadays, many common areas such as array-like data structures, graph theory, number theory and dynamic programming optimizations are widely researched. There were many problems which basically said 'write this algorithm', and this type of problems (e.g. LCS with Hirschberg optimization) is now mostly extinct. The reason is that there used to be quite simple algorithms and data structures, such as prefix sums, segment tree and many divide-and-conquer-based optimizations, which gave a decent optimization factor, e.g. by $$$\mathcal{O}(n)$$$ for prefix sums, $$$\mathcal{O}(\frac{n}{\log n})$$$ for segment tree, or $$$\mathcal{O}(\sqrt{n})$$$ for sqrt optimization. Nowadays, all these algorithms are widely known, and new algorithms give smaller optimization factors, mostly logarithmic. It makes it next to impossible to separate an optimal solution in Python from a dumb solution in C++. If we keep algorithms-based problems, only programmers using C/C++/Rust/Java will succeed.

Even then, new 'secret' algorithms such as matroid intersection are quite hard to write during 2 hour contests, and it's next to impossible to come up with them yourself. There is very little intuition behind them, it's only about mathematics. So only those who have written the algorithm before or know where to find the source code will succeed. This would make CP too similar to industrial programming and make Codeforces just a very short hackathon.

We're running off simple effective algorithms, all new 'effective' data structures are only effective on large scale (terabytes of data). Now, we can only have ad-hoc problems and non-asymptotic optimizations. The former implies CP is slowly being transformed into a mathematic olympiad, the latter implies CP with low-level languages only, non-classic problems and requires longer rounds.