Oxford researchers just published a paper that quietly dismantles a lot of the AI-will-replace-everyone narrative. Teppo Felin and Matthias Holweg (Saïd Business School, Oxford) argue in Theory Is All You Need that large language models cannot invent anything genuinely new. Not because they aren’t impressive, but because of how they work. An LLM is trained to predict the next token from existing data. It is structurally backward-looking and imitative. It cannot hold a belief the training data says is wrong. Every real scientific or technological breakthrough in history started with exactly that kind of belief. The Wright Brothers are the paper’s central example. In 1903 the prevailing data and expert consensus said human flight was one to ten million years away. Nine weeks later they flew. They did not have better data. They had a theory the data had not yet caught up to. They broke the problem into lift, propulsion and control, built their own wind tunnels, and generated the evidence that did not previously exist. That is the data–belief asymmetry humans can live with and machines cannot. Humans start with a causal theory, then intervene in the world and run experiments to create new data. LLMs start with the data and stay inside its statistical envelope. This is why the current hysteria is mostly noise. The models are extraordinary at compressing, remixing and accelerating what we already know. They are not going to independently invent the next physics, the next energy source, or the next medical paradigm. Those leaps still require a human who is willing to treat the existing record as incomplete or wrong. Use the tools. Just stop pretending the tools are the source of the next idea. Paper: pubsonline.informs.org/doi/1…