Model Training

GPT-2 is a model launched by OpenAI in 2019. Its training cost was once $256 per hour. So, five years later, in the GPT-4 era, with the advancement of software, hardware and data, will the time and cost required to train the same model be reduced? The answer is yes. According to Tom's Hardware today, Andrej Karpathy, former director of Tesla AI, co-founder of OpenAI and project developer, used llm.c to "reproduce" GPT-2.

According to Tom's Hardware today, Dario Amodei, CEO of AI startup Anthropic, said in a recent podcast interview that the current training cost of models like GPT-4o is about $100 million, and the training cost of large AI models currently under development may be as high as $1 billion (currently about RMB 7.294 billion). Of course, $1 billion is not the "end point". Dario Amodei predicts that in the next three years, AI...

Roboflow is a platform focused on computer vision that helps developers quickly build and deploy their own computer vision applications, regardless of their skills or experience. Roboflow.com was founded in January 2020 by a group of developers with extensive experience in computer vision projects. Their mission is to remove any barriers that may prevent developers from successfully applying computer vision and make computer vision tools available to every developer. Roboflow.com has received support from several well-known investment institutions and industry leaders, including…

Microsoft announced a technical partnership with Mistral AI, an open source large model platform, with a focus on strengthening cooperation in the field of artificial intelligence and large language models (LLM). The three key points of the cooperation include: Microsoft will provide Mistral AI with basic AI computing power through Azure cloud services to help it accelerate model training and development. At the same time, Microsoft will provide customers with Mistral AI's basic large models through Azure cloud services, providing users with more choices, including not only OpenAI's GPT series models, but also M...