A better example than phi-2 might be OpenChat-3.5, a 7 billion parameter model that scores above Llama-2 70B on the lmsys leaderboard, while using 10% of the pa

A better example than phi-2 might be OpenChat-3.5, a 7 billion parameter model that scores above Llama-2 70B on the lmsys leaderboard, while using 10% of the parameters: https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboar... I've spent a while poking at it. It's not as strong as the best chatbots, but it's very surprising it works at all. If someone cracked open your skull and scooped out 90% of your brain, would you score the same on IQ tests?