A new arXiv paper puts a hard mathematical ceiling on how much smarter a self-rewriting AI agent can actually get.
Researchers built a formal model of self-improving agents - systems that can modify their own problem-solving routines, evaluators, and improvement procedures - and asked what that self-modification can and cannot buy. Using computability theory, they show that any procedure a self-revising agent produces stays inside the same computability class defined by its starting information source, which the paper calls an oracle. Getting genuinely more capable requires new information, through what the paper calls an oracle join, not just clever self-editing. The authors also walk through a concrete case - an agent learning Boolean rules from a teacher - showing precisely when it can learn to answer every question itself and drop the teacher, and proving that building that independence in general is as hard as a relativized version of the halting problem, one of computer science's classic unsolvable questions.
That is a quiet but pointed rebuttal to talk of AI agents bootstrapping their way to open-ended superintelligence through pure self-modification. The math here says rewriting your own code can make you more efficient, not fundamentally more capable, unless you also feed it new data or new access. That is a useful corrective for anyone treating recursive self-improvement as inevitable rather than an engineering problem with real, provable limits.
Call it a thermodynamics argument for AI takeoff skeptics: self-improvement can save you steps, but it cannot print information you never had.