Jerry Murdock is betting that every AI model alive today will be dead within a decade
Continuous learning models, if they arrive, do not merely improve on today's architectures. They replace them entirely. Jerry Murdock has put a 10-year horizon on that replacement, and the reasoning behind the call is worth examining before the clock runs out.
Every model that exists today will eventually go away. That is not a general observation about technological obsolescence. It is the specific, checkable claim Jerry Murdock has put his name to, with a 10-year horizon attached to it.
The mechanism Murdock points to is continuous learning. Today’s models are trained on a fixed corpus, deployed, and then frozen. They do not update from what they encounter after deployment. Continuous learning models, by contrast, would compress, abstract, and update from experience in the way that made them powerful during training, but without stopping. Murdock’s argument is that once that architecture exists, the current generation of static models does not merely become outdated. It becomes obsolete in kind, not just in degree. As he puts it: “Every generation, every model we have today dies, goes away.”
He is careful about the timeline. Ten years is his outer bound. He also notes that some researchers place the arrival of continuous learning models at two or three years out. That spread matters. A two-year horizon is a near-term engineering bet, the kind that can be falsified quickly by a product announcement or a conspicuous absence of one. A 10-year horizon is a structural call about where the field is headed, harder to falsify early and harder to dismiss as hype.
Within 10 years, I believe we and I think some people are thinking two or three years, continuous learning models will come into existence. That means that every generation, every model we have today dies, goes away Jerry Murdock
The underlying challenge that makes the timeline genuinely uncertain is not compute or data. It is stability. Training a model continuously without letting it catastrophically forget what it already knows is an unsolved problem, and has been for years. Public writing from researchers working on the problem describes the goal as moving from models that are effectively amnesiac after deployment to ones that accumulate something like experience. That framing captures why the gap between two years and ten years is so wide: the research community knows what it wants but not how hard the last mile will be.
What rides on Murdock’s call being right, and roughly on schedule, is considerable. Current AI infrastructure is built around the assumption of monolithic models served at scale. A world of continuously updating models looks different in almost every dimension: routing, versioning, rollback, and serving thousands of variants rather than one canonical model at massive scale. The companies and teams that have built around the current architecture would face not an upgrade cycle but a structural rebuild. That is a different kind of pressure than any prior generation of model improvements has imposed.
What rides on Murdock being wrong, or significantly early, is also significant. The field has a pattern of underestimating how long fundamental architectural shifts take to move from research demonstration to production reliability. If continuous learning remains a research problem for 15 or 20 years rather than 10, the investment theses and competitive strategies built around its arrival will have been priced in too early.
The honest read on the evidence is that the directional claim, that continuous learning will eventually replace static models, is widely shared among researchers. The falsifiable part is the timeline. Murdock has committed to a decade. That commitment is now on the record, and the clock is running.