Very Large Language Models and How to Evaluate Them

look, everyone's throwing around 'evaluation,' but it's a total mess. you can't just look at perplexity or simple accuracy; those metrics are garbage when you're dealing with emergent reasoning. we need more rigorous, human-in-the-loop evaluation methods, like adversarial testing, not just simple benchmarks. the problem is that LLMs aren't deterministic; they're statistical parrots dressed up as thinkers. stop relying on public benchmarks; build your own stress tests.