Given that frontier models are rapidly consuming mathematical and scientific reasoning as core competencies, what prevents the Large Physics Model from becoming a lagging, under-funded shadow of a commercial foundation model that simply has a physics plugin? Isn’t the community betting on a static capability gap that doesn’t exist?
The paper champions openness and control, but building a model from (or heavily fine-tuning on) a corpus of physics data requires immense, continuous compute. If a commercial API is 90% as good and infinitely more scalable, won’t the astronomical cost of maintaining a truly competitive, sovereign LPM lead to the exact vendor lock-in the authors seek to avoid?
Can scientific understanding be effectively embedded in a model that is only trained on physics data? Or is that understanding intrinsically linked to the world knowledge, linguistic nuance, and reasoning strategies that only massive, general pre-training can provide? By building an isolated LPM, aren’t we at risk of creating a savant that is brilliant at textbook problems but useless for the messy, interdisciplinary work of real discovery?