I get your point but the highlighted part is incorrect. First of all, OpenAI is hardly a start up - it is more like a research foundation funded by billionaires like Musk, most actual start ups are more likely to focus on existing solutions and their adaptation for commercial application because they don't have large budgets to burn through. Large models like GPT-3 cost millions of dollars to just train from scratch, let alone going through all the research iterations. Their commercialization is also questionable - I am 100% sure that even with the paid subscriptions OpenAI is still losing money on their models.
Secondly, most of the key discoveries that transformed the DL scene into what we see today were made at big corporations or universities - from the first large-scale image classification dataset ImageNet which essentially kickstarted the "arms race" of models was created at Stanford; most fundamental image models like ResNet, Inception, EfficientNet, etc. all come from large corps; first LLMs (GPT-3 is a LLM) were developed in big corps (e.g. BERT); the transformer network which lies in the heart of GPT-3 and other modern LLMs was developed by Google team; etc. You are focusing too much on the hype around certain techs rather than looking at the progression - it is undeniable that large corporations are quite capable of innovating and most of the ground-breaking research advancements in fact often came from such entities. The explanation is rather simple - they have budgets that can sustain such R&D endeavours while start ups are essentially on borrowed time until they are able to commercialize their tech.
