For most of the past few years, the AI field has been defined by a single goal: train bigger models. According to IEEE Spectrum, since roughly 2020 the dominant focus has been on scaling up large language models, which grew from millions of parameters to trillions. This training-first mindset produced impressive gains, but also exposed clear limits.
The largest version of OpenAI's GPT-3, released in 2020, correctly answered only 43.9 percent of questions on a benchmark—a reminder that even the biggest models are far from perfect. The article suggests that the next phase of AI progress will depend less on training and more on inference, the process of running a trained model to generate outputs.
This shift has major implications for hardware. Inference workloads have different requirements than training, and the industry is now paying attention to chips and systems designed specifically for that task. The article frames this as an "inference revolution" that could reshape where AI compute is spent and how future hardware is built.