Raw LLM Responses

Any precise figures are likely dubious but all the big labs have larger internal models that aren't economical to serve at scale. They train the larger models and then distill them into smaller models but it's likely not the free lunch that many seem to think it is when it's done with smaller and smaller models; I figure this is why some of OpenAI's newer models excel in benchmarks and narrow domains like coding but seem to get progressively worse in areas like common sense and accurately inferring implicit intent, which are not as straightforward to benchmark as coding and are often overlooked for that reason.