I added a bit of OpenAI's press conference.
I have a small feeling that, as the model scale desperately enlarges, the position of algorithm engineers is becoming more and more awkward.
The specialization level of ML/DL jobs is getting higher and higher. MLops doesn't need algorithm engineers, model inference acceleration and deployment don't need algorithm engineers, and the development of inference engines, training frameworks, and other infrastructure doesn't need algorithm engineers.
Algorithm engineers only need to call the APIs provided by these specialized jobs in front. Most algorithm engineers don't even have research tasks, they just call the APIs packaged by a few algorithm engineers.
Who would have thought that deep learning algorithm engineers would become the "front-end" of deep learning in the era of large models.
Indeed, this trend began with Transformers dominating the NLP field. It was then discovered that a general model would surpass domain-specific models, and this general model also showed better performance in computer vision and multimodal tasks. As large models increase the training cost to a level that ordinary players cannot reach, the products on the market have become more homogeneous. Deep learning algorithm engineers have few opportunities to showcase their abilities without attaching themselves to a specific application domain, making it difficult for them to have their own unique competitiveness.