The interface that carries its own error bar
An algorithm that must run to completion before it says anything is a batch procedure. You submit the problem, you wait, and at some unknown future moment an answer appears, correct or not, with no intermediate signal about how close it is. An anytime algorithm refuses that contract. At every point during its execution it holds a valid, usable answer, and if you let it keep running, that answer gets better. Stop it after one second, or one minute, or one hour, and you get something you can act on, along with a bound on how far it might be from optimal. The distinguishing feature is not speed. It is the presence of a quality measure attached to the current output, updated as computation continues.
That measure has a name in the literature: the performance profile, a curve relating elapsed computation time to expected solution quality. The profile is what lets a controller — human or automated — decide whether more thinking is worth the delay. If the curve is steep early and flat late, stop early. If it is still climbing, keep going. Without the profile, an anytime algorithm degenerates into something that merely outputs early and often; the profile is the part that makes early output trustworthy rather than merely fast.
The clean illustration is iterative deepening in chess search. A chess engine such as Stockfish searches to depth one, records the best move, then searches to depth two, then three, and so on, always keeping the best move found at the last depth it finished. Interrupt the search at any millisecond — the clock runs out, the opponent moves, the process is killed — and a legal, defensible move already exists. On contemporary hardware the engine typically reaches something like depth twenty within a second. Crucially, the engine reports both the move and the depth or evaluation behind it: you know not just what it thinks but how much thinking produced that belief.
Where the idea came from
Thomas Dean and Mark Boddy coined the term "anytime algorithm" in 1988, working on time-dependent planning at Brown University. Their problem was concrete: a planning system for a robot or an autonomous process cannot assume unlimited deliberation time, because the world imposes deadlines the planner does not control. A robot arm cannot wait for a provably optimal trajectory while an object falls off a table. Dean and Boddy wanted planners that degrade gracefully — interruptible at will, useful at every point, better the longer they run.
Shlomo Zilberstein and Stuart Russell, working through the early 1990s, gave the idea its formal skeleton. They distinguished two species: interruptible algorithms, which can be stopped at any unknown moment and still deliver a bounded answer, and contract algorithms, which are told in advance how much time they will get and optimise against that fixed allowance. They also worked out how performance profiles compose, so that a system built from several anytime components could reason about the value of allocating more time to one subroutine over another. The underlying problem was control under real-time pressure. A system needs to know, quantitatively, what waiting longer buys, and when it stops buying enough.
Two further instances show the idea outside chess. In medical imaging, iterative CT and PET reconstruction by ordered-subsets expectation maximisation rebuilds an image subset by subset, commonly across two to four iterations of some two dozen subsets each. The image is diagnostically legible well before the last pass completes. A radiologist under time pressure can read an early reconstruction and a later, cleaner one afterwards, trading noise against latency as an explicit choice rather than an enforced wait. In seismology, the ShakeAlert earthquake early warning system on the US west coast issues a first message within four to six seconds of an earthquake's origin, based on a handful of stations, then revises its magnitude and shaking estimates as more stations report. During the 2019 Ridgecrest sequence, successive alerts updated as the rupture unfolded in real time. The alert is never final. It is the current best estimate, with a station count attached.
The turn
Those three examples share a structure once you look for it. The chess engine faces a problem instance that does not change while it thinks: the position is fixed, only the search over it grows. The CT scanner faces detector data that has already been collected before reconstruction starts. ShakeAlert is the outlier worth pausing on, because its evidence keeps arriving even as it computes — and that is the seam that leads somewhere.
Consider the three generations of large models by what they are allowed to know and when they are allowed to know it. A Large Language Model deliberates over a corpus fixed at some training cutoff. Give it more inference-time compute — more search, more sampled continuations, more reasoning steps — and its answer improves, but only by exploiting evidence that stopped accruing on a fixed date. It is anytime in deliberation, exactly like the chess engine, and its error floor is set by the age of the corpus, not by how long it is allowed to think.
A Large World Model watches a scene: sensors, a camera feed, a stretch of live telemetry. Its estimate improves while the scene persists, refining as more of it is sensed. But the scene ends. The car passes the junction, the patient leaves the room, the batch of frames runs out. In Zilberstein's vocabulary this is a contract algorithm, except the contract's deadline is set by the world rather than by an operator scheduling compute. Refinement halts when the scene does, and the final estimate is exactly that — final, not updatable, because there is nothing left to sense.
A Large Universe Model has no scene boundary. The streams it watches — sensor networks, transaction feeds, satellite passes, claims filings — keep arriving indefinitely. There is no moment at which observation completes. And this is where the anytime concept stops being an analogy and becomes a requirement. A system that cannot report a completed conclusion, because completion would require the streams to end, and cannot withhold output until convergence, because convergence never arrives, has exactly one coherent way to speak: a current best belief, available on demand, with a stated quality and a record of the evidence behind it, revisable when the next observation lands. That is the anytime contract, translated from the axis of compute to the axis of intake. The interface does not change shape as more streams are added. Only the numbers inside it improve.
What the misreading gets wrong
The concept invites a shallow version of itself worth naming and rejecting outright. The weak reading treats "anytime" as a promise to never say "I do not know" — always emit something, dressed up as an answer, latency optimised, honesty optional. That is precisely backwards. An anytime algorithm is defined by carrying its quality measure alongside the result. Strip the measure out and what remains is a guess with good response time, not an anytime system. The corresponding failure in a belief-maintaining system is a confident sentence with no timestamp and no source. Availability without a stated bound is not the anytime contract. It is the contract's abandonment while keeping its speed.
Three objections, taken straight
Performance profiles assume a fixed problem instance with a fixed optimum to approach. Under continuous intake the target moves. There is no meaningful suboptimality bound when the map is being redrawn as you compute.
This is correct mathematics, and it narrows the claim rather than merely opposing it. A fixed-instance bound like the epsilon-suboptimality guarantee in anytime search algorithms such as ARA* has no analogue when the ground truth itself changes between observations. What survives is weaker: calibration, meaning whether stated confidence matches observed outcome frequency over time, and provenance, meaning a traceable age for the evidence behind each claim. Those are auditable without needing a fixed optimum. A system offering neither calibration nor provenance is not anytime. It is merely fast, and the distinction matters.
An answer that is always available invites always being acted upon. This erodes a useful discipline: the friction of not knowing, which batch procedures preserve by refusing to speak until they are sure.
The cost is real and documented — ShakeAlert's early magnitude estimates have been revised downward after alerts had already gone out, and the revision arrived too late to un-send the first message. But the friction being defended is a crude stand-in for an explicit decision rule. Deliberation scheduling replaces it with one: act when the expected cost of further delay exceeds the expected cost of error at the current quality. The actual failure is showing a point estimate with no attached profile, which is a defect of the interface presenting the answer, not an indictment of anytime computation as such. Refusing to answer is itself a decision, and typically a less well-instrumented one.
Large Language Models are already anytime — more search, more sampled continuations, more steps improve the answer without any continuous sensing involved. The property does not distinguish the generations at all.
Granted, and it is worth granting fully. Deliberation-anytime predates and does not depend on continuous intake. The distinction is which axis is buying the improvement. More time over a frozen corpus buys better exploitation of evidence that stopped accruing at a cutoff; no additional search recovers a fact that postdates that cutoff. Anytime-in-intake buys new evidence outright. The two compose — a system can deliberate longer over fresher evidence — but they do not substitute for each other. A belief about a live quantity, a river's current level, a counterparty's current exposure, has an error floor set by the age of its last observation. Only intake, not more thinking, lowers that floor.
What this does and does not establish
The anytime framework establishes that once intake has no stopping point, there is exactly one consistent way to speak: current belief, stated quality, retained provenance, open to revision. It explains why a Large Universe Model sits at the top of the intake axis rather than merely further along it — beyond "everything, continuously" there is no further category of evidence to admit, only more of the same, arriving sooner, tagged better. It does not establish that such a system is easy to build, that its calibration will in practice be trustworthy, or that operators will resist treating a current-best estimate as a final one. Those are separate fights, fought with different evidence, and this concept does not settle them. It only tells you what the output has to look like if the streams never stop.