[D] frontier models are a zero sum game for a few tasks - what they gain in reasoning they lose in your specific thing

About this article

when Google shipped Gemini 3 last November, it set new benchmarks on reasoning and coding. but it also removed pixel-level image segmentation. if your pipeline depended on that, you found out the hard way. and not just gemini. OpenAI has deprecated GPT-3, GPT-4-32k, multiple GPT-4 variants. Anthropic sunset Claude 2.0 and 2.1. model this is the actual problem with frontier models for specific tasks like document extraction. the training budget is finite. when it goes to reasoning and coding, ...