We're running out of benchmarks to upper bound AI capabilities

We can definitely make harder evals, the problem is a good eval set is indistinguishable from good training data / market edge, so no one is incentivized to share their best eval sets publicly.

This is the least true thing ever. All LLMs are terrible at ARC-AGI-3. Every video game can be used as a benchmark. You could rank LLMs on how long they can keep a game of Dwarf Fortress running or how fast they can beat GTA5.

Start front loading the models with 5k, 10k, 50k, 100k tokens of messy quasi related context, and then run the benchmarks.

These models are ridiculously powerful with a blank slate. It's when they get loaded down with all the necessary (and inevitably unnecessary) context to complete the task that they really start to crumble and fold.