Jared Kaplan is a theoretical physicist and machine learning researcher known for leading the research that established empirical Scaling laws for language models, and as a co-founder and chief scientist of Anthropic.
Background and scaling laws
Kaplan held a faculty position in theoretical physics at Johns Hopkins University before moving into AI research, a background he has said shaped his approach to treating model behavior as an empirical, quantitative system to be measured rather than only engineered by intuition. In 2020, while affiliated with OpenAI, he led "Scaling Laws for Neural Language Models," a paper that showed loss on language modeling tasks follows smooth power-law relationships with model size, dataset size, and training compute. The paper gave labs a predictive framework for deciding how to allocate a fixed compute budget, and it directly informed the training of GPT-3. A widely cited 2022 follow-up from DeepMind, the "Chinchilla" paper, revised Kaplan's original conclusions about the optimal ratio of parameters to training tokens, but the underlying power-law framing he helped establish remained foundational to how frontier labs plan large training runs.
Anthropic
In 2021, Kaplan co-founded Anthropic alongside Dario Amodei, Daniela Amodei, Tom Brown, Chris Olah, and others who had worked together at OpenAI. As Anthropic's chief scientist, he has overseen research directions spanning pretraining, AI alignment, and the empirical study of model capabilities as they scale, continuing the scaling-focused research program he began at OpenAI. Anthropic's public safety framing, including its Responsible Scaling Policy (see Responsible scaling policies), rests in part on the premise that capability growth is predictable enough from scaling trends to be planned for and gated, an idea directly descended from Kaplan's earlier research.
Influence
Kaplan's scaling laws work is frequently cited alongside the GPT-3 paper as one of the two developments that convinced the wider field that continuing to scale Transformer (architecture) models predictably, rather than searching for fundamentally new architectures, was the most reliable path to further capability gains, a conclusion that shaped frontier lab strategy through the mid-2020s.
His background in theoretical physics is frequently mentioned in profiles of Anthropic's leadership as a distinguishing feature among AI lab founders, and he has continued to draw explicit analogies between the empirical methods used to study physical systems and the way frontier labs now study the behavior of large trained models before and after deployment.