A new paper tries to formalize something engineers have long done by instinct: figuring out how much information actually matters, not just how much arrives intact.
The paper, posted to arXiv, proposes a mathematical theory of "pragmatic information" that extends Claude Shannon's classical framework beyond symbol-level accuracy. Its core idea is something the authors call an isoteleia mapping: if two different messages lead an agent to the same optimal decision, the theory treats them as equivalent, even if they look nothing alike. That lets the framework build a three-level stack, syntactic, semantic, and pragmatic, where each layer strips out details that don't change the outcome. The authors then derive pragmatic versions of entropy, channel capacity, and rate-distortion, and prove coding theorems that mirror Shannon's originals but score by decision quality rather than bit-for-bit fidelity. They also define a "value of information" and "cost of information" pairing, an efficiency bound for how much net benefit a resource-limited system can extract, and extend the math to continuous signals and multi-step decisions via a Bellman equation.
This matters most for systems that don't need perfect reconstruction, like self-driving cars, robots, or networked control systems that only need enough data to pick the right action. Standard information theory optimizes for transmitting symbols faithfully; this framework optimizes for what an autonomous system actually wants, which is to act correctly under a compute or bandwidth budget.
Worth remembering: Shannon's 1948 paper sat as pure math for years before it reshaped every network on Earth. This one is still in that phase, with no working system built on it yet, so treat the "extends Shannon" framing as a promising hypothesis, not a shipped result.