The Exponential Time Hypothesis and OpenAI’s Claimed Breakthrough in Computing

As an Amazon Associate we earn from qualifying purchases.

Key Takeaways

• OpenAI has released a claimed refutation of a longstanding assumption about computational difficulty.
• Subexponential algorithms could change complexity theory without making difficult problems easy.
• Formal verification strengthens the evidence, but independent examination remains important.

OpenAI’s Announcement Changes a Longstanding Debate

On October 26, 2026, OpenAI announced that an internal artificial intelligence (AI) system had produced a claimed refutation of the exponential time hypothesis (ETH). The company released mathematical material and a formalization in Lean, a system for checking mathematical proofs, and described an external effort to examine the result.

The announcement concerns a precise question about how the time needed to solve certain logic problems grows with their size. ETH proposed that exponential growth was unavoidable in the worst case. OpenAI’s claim challenges that proposed limit through a new algorithm, rather than through a faster processor or a larger computer.

According to OpenAI’s announcement, its system produced the natural-language disproof on October 16, followed by formalization on October 17. The public release includes a manuscript revised on October 25. OpenAI asked Russell Impagliazzo and Ramamohan Paturi of the University of California San Diego, together with Ryan Williams of the Massachusetts Institute of Technology, to help organize examination of the work.

Those names carry particular significance. Impagliazzo and Paturi helped formulate the hypothesis that the new result challenges, making their participation an opportunity to examine the claim with researchers who understand its original meaning and subsequent uses.

Their participation does not constitute certification. The researchers’ public examination statement explicitly separates organizing an independent examination from endorsing the result, and invites others to reconstruct the argument and investigate its consequences.

The distinction matters because several different developments are occurring together. A company has announced a mathematical result. A machine-checkable artifact has been published. Researchers have begun explaining the argument, and commentators are assessing what it could mean for computing.

These developments justify attention without supporting every interpretation attached to them. A refutation of ETH would alter an important part of theoretical computer science, but it would not establish that computers can efficiently solve every difficult problem. It also would not demonstrate that a publicly available chatbot can reproduce the discovery on demand.

The relationship to OpenAI is specific: the company says its internal system discovered an algorithmic result that contradicts a longstanding complexity assumption. As of October 29, 2026, the public record includes the company’s claim, formal artifacts, and researchers’ examination and explanatory work. The mathematical claim must be judged through its definitions, proof, and consequences.

What the Exponential Time Hypothesis Says

The exponential time hypothesis concerns the Boolean satisfiability problem, abbreviated SAT. A Boolean variable has one of two values, true or false, and a satisfiability problem asks whether values can be assigned so that an entire logical expression becomes true.

The version called 3-SAT restricts the expression to clauses containing at most three terms. Each term is a variable or its negation, meaning the opposite of that variable’s value. Terms within a clause are connected by “OR,” and the clauses are connected by “AND.”

A satisfying assignment must make every clause true. Finding an assignment that satisfies almost all clauses does not solve the exact decision problem, and finding no solution after a lengthy search does not establish that no solution exists.

With n variables, there are 2ⁿ possible assignments. For 10 variables, that means 1,024 assignments; for 20 variables, it means 1,048,576. These numbers describe the size of the complete assignment space, not a requirement that every algorithm examine every assignment.

The hypothesis was formulated precisely in work including Impagliazzo and Paturi’s On the Complexity of k-SAT. It asserts a positive exponential rate below which algorithms cannot push the worst-case running time for 3-SAT, allowing polynomial factors associated with the input’s representation. A subexponential algorithm would contradict that assertion.

The phrase “worst case” is important. ETH never proposed that every 3-SAT instance is difficult, or that useful instances cannot be solved quickly. It concerns an algorithm’s guarantee over all inputs of a given size, including inputs that defeat shortcuts effective elsewhere.

The hypothesis also permits substantial improvements over exhaustive search. An algorithm with running time proportional to 2^(0.1n) remains exponential, even though its growth is much slower than 2ⁿ. Such an algorithm alone would not refute ETH.

Another distinction separates a hypothesis from an established lower bound. A lower bound proves that a specified amount of computation is necessary under stated conditions. ETH supplied an assumption from which researchers could derive conditional lower bounds for other problems.

Its influence came from the connections it enabled. Researchers could show that an unexpectedly fast algorithm for another problem would produce an unexpectedly fast algorithm for 3-SAT, contradicting ETH. The resulting conclusions were conditional from the beginning, even when the assumption became familiar enough to serve as a routine starting point.

Why Subexponential Time Can Still Be Extremely Slow

The word “subexponential” describes a growth rate. It does not promise a short wait, an affordable computing bill, or an algorithm suitable for deployment.

Polynomial running times include n² and n³, where the exponent is fixed and the input size is the quantity being raised to that power. Exponential running times include 2ⁿ, where the input size appears in the exponent. Subexponential running times include growth rates that exceed every fixed polynomial but remain slower than 2^(cn) for every positive constant c.

The expression 2^√n illustrates that distinction mathematically. Its exponent grows without bound, but the ratio √n/n approaches zero as n increases. It eventually grows more slowly than every exponential function with a fixed positive rate.

The notation 2^o(n), which appears in discussion of ETH, captures this vanishing ratio. The small “o” indicates that the exponent becomes small relative to n as input size increases. It does not indicate that the exponent itself becomes small.

A calculation shows how little practical reassurance the category provides. At n = 1,000,000, the expression 2^√n becomes 2¹⁰⁰⁰, approximately 10³⁰¹. A subexponential growth rate can still describe an amount of work far beyond available computing resources.

Running-time formulas also suppress information. Big-O notation describes an upper bound up to constant factors for sufficiently large inputs, and those constants can affect whether a theoretical improvement becomes useful at any realistic scale. Setup work and memory requirements can further change an implementation’s performance.

For this reason, an algorithm can refute ETH even if it loses to established methods on all inputs that engineers can reasonably process. The hypothesis makes a statement about growth as problem size increases without limit. A proof that eventually beats every fixed exponential rate addresses that mathematical statement.

John D. Cook’s explanation emphasizes the gap between better-than-exhaustive search and the stronger subexponential claim. Earlier improvements already reduced the exponential rate; the October result concerns whether that rate must remain bounded away from zero.

A practical speed comparison would require an implemented algorithm, specified inputs, and measured resource use. A theoretical bound answers a different question: what can be guaranteed about the growth of computation, regardless of whether the favorable regime is accessible to present hardware.

What the Released Mathematical Result Covers

OpenAI’s released manuscript addresses more than deciding whether a satisfying assignment exists. It claims a deterministic algorithm that counts satisfying assignments for formulas with a fixed maximum clause width.

A deterministic algorithm follows a specified sequence of operations rather than relying on random choices. Counting asks for the exact number of satisfying assignments. A positive count immediately answers the simpler existence question, and a zero count establishes that no assignment satisfies the formula.

The formulas use conjunctive normal form (CNF), meaning an AND of clauses that are themselves ORs of terms. The notation k-CNF sets a maximum clause width of k. The result treats k as fixed, rather than allowing that parameter to grow freely with the input.

The manuscript gives a running-time bound whose exponential component has the form 2^O(n log log n / √log n), with additional polynomial factors. OpenAI’s formalization scope statement describes a weaker machine-checked bound containing an additional log log n factor in the exponent. Both expressions are subexponential in n for fixed clause width.

That difference deserves attention without being exaggerated. A weaker bound can still refute the same hypothesis, provided it remains subexponential and the statement correctly represents the intended computational problem. However, a description of the formal verification should not attribute every sharper manuscript claim to the checked artifact.

The proof’s mechanism also differs from asking a language model to guess solutions to difficult logic formulas. In Ryan O’Donnell’s exposition, the argument passes through a matrix calculation called the permanent. Certain counting problems can be translated into computing permanents of sparse matrices, which have relatively few nonzero entries.

The exposition constructs related matrices with useful structure, computes information at selected points, and recovers the desired value through polynomial interpolation. Interpolation reconstructs a polynomial from enough information about its values and derivatives. The method seeks exact answers through algebraic organization, rather than abandoning correctness for an approximate result.

O’Donnell’s October 29 exposition focuses on explaining that route and its time bound. Its significance for a general audience is that the claimed advance has a mathematical mechanism that other researchers can inspect, simplify, and potentially adapt.

Discovery by an AI system does not change the requirement for that mechanism. Once a proposed algorithm is written down, its correctness and resource use must follow from an argument that stands independently of the system that produced it.

What Lean Verification Establishes

A proof assistant checks a mathematical argument against explicitly stated definitions and logical rules. Lean can provide strong evidence that a formal theorem follows from the assumptions used in its development, without depending on a language model’s confidence or persuasive writing.

That function is different from reviewing an ordinary manuscript. A human referee may follow the main argument and examine delicate steps, but a formal checker requires every accepted inference to fit the underlying logical system. Long chains of reasoning can be checked with a level of mechanical consistency that ordinary prose does not provide.

Lean’s documentation on axioms explains that its mathematical developments can use specified axioms and extensions. Evaluating a formal result includes understanding those assumptions, as well as the statement being proved.

The relationship between the formal statement and the public claim still needs examination. A proof checker does not independently decide whether a definition captures the intended version of 3-SAT, whether an input encoding matches the manuscript, or whether a headline describes the theorem accurately.

This distinction is explicit in Anthropic’s accompanying release, which also claims a refutation. Its documentation states that machine checking verifies the proofs, but does not itself verify that the formal statement says what the paper says. The release supplies a short statement file intended to make that comparison easier.

For a computational theorem, the model of computation matters too. The formal development must specify how inputs are represented and how operations are counted. Otherwise, an apparently fast method might conceal expensive work inside an operation treated as free.

Formal verification and mathematical understanding consequently answer related but distinct questions. A successful check can support confidence in a precisely stated proposition. An explanatory account helps researchers understand why it holds, where its techniques apply, and what changes would invalidate the argument.

The October releases provide more substantive evidence than an unsupported announcement because there are public artifacts to examine. The appropriate assessment still distinguishes the existence of those artifacts from independent reproduction of the checks and community agreement about their interpretation.

That separation also protects the meaning of verification. Treating a checked theorem as evidence for every surrounding claim would weaken the value of the formal work by extending it beyond its actual scope.

What Credible Commentators Are Saying

The most useful commentary distinguishes mathematical consequences from expectations about AI. Computer scientists evaluating a running-time argument address a different question from commentators assessing whether a laboratory has demonstrated impressive research capabilities.

Ben Shimkus’s October 27 Business Insider coverage collected reactions from researchers and technology commentators. It reported Williams’s view that a successful refutation should increase consideration of P = NP, Gary Marcus’s conditional recognition of the result’s significance, and Timothy Gowers’s assessment that the development was surprising and consequential. Those reactions describe expectations and judgments, rather than substitute for verification.

The Review Organizers Separate Examination From Endorsement

Impagliazzo, Paturi, and Williams provide the clearest account of the review process in their message to the research community. They state that their involvement should not be interpreted as certification and call for an open effort to establish correctness or identify limitations.

Their message also emphasizes that the work should remain accessible to independent criticism. It asks how explanations and simplified proofs can support understanding, and how researchers should distinguish established consequences from possibilities still requiring investigation.

This position reflects a specific responsibility. Organizing scrutiny helps a field respond to a large claim, but attaching respected names to an announcement can create an impression of endorsement before the work is complete.

Lance Fortnow Expresses Greater Confidence

In an October 28 Computational Complexity commentary, Lance Fortnow takes a more confident position that ETH has been refuted. He also states that he has not personally verified the proof and points to follow-up explanations.

Fortnow’s technical caution concerns the reach of the result. A subexponential algorithm for fixed-width satisfiability does not automatically produce comparable algorithms for every problem associated with the class NP. Translating another problem into 3-SAT can enlarge the input enough to erase the desired improvement.

He also separates a failed assumption from the truth of its consequences. A lower bound previously supported through ETH might still be correct, even though ETH can no longer justify it.

John D. Cook Keeps P Versus NP Separate

Cook’s October 27 commentary stresses that the status of P versus NP remains unresolved. A subexponential algorithm can still require more than polynomial time, leaving a large mathematical gap between the new result and a solution to that broader problem.

These positions need not be read as a conflict between eniasm and caution. Fortnow expresses confidence about the refutation, the organizers define an examination process, and Cook explains a limit on the conclusion.

The more useful distinction is the evidence supporting each statement. An informed reaction can help assess significance, but a claim about the algorithm’s correctness requires the proof; a claim about practical performance requires implementation evidence.

Why This Does Not Resolve P Versus NP

The class P consists of decision problems solvable in polynomial time by a deterministic computer. NP, short for nondeterministic polynomial time, consists of decision problems whose yes answers have certificates that can be checked in polynomial time.

NP does not mean “non-polynomial.” Some problems in NP are already known to be in P, and the unresolved question is whether the two classes are actually equal.

The Clay Mathematics Institute describes P versus NP through the distinction between checking a solution and finding one. As of October 29, 2026, it lists the problem as unsolved. A proposed solution can sometimes be checked efficiently even when no efficient general method for finding such a solution is known.

The problem 3-SAT is NP-complete. This means that it belongs to NP and that every problem in NP can be translated into it using a polynomial-time transformation. A polynomial-time algorithm for 3-SAT would establish P = NP.

A subexponential algorithm does not cross that threshold. It can improve substantially on a fixed exponential growth rate and still remain slower than every polynomial. Refuting the exponential time hypothesis removes one proposed barrier, but it does not remove all possible barriers between established algorithms and polynomial time.

There is another reason to avoid sweeping conclusions. A transformation that is efficient enough for reasoning about P versus NP may not preserve the precise input size needed for an ETH consequence. Polynomial growth in the translated instance is acceptable for proving NP-completeness, but can be too expensive for transferring a particular subexponential bound.

Suppose a transformation replaces an input size n with a formula whose size grows as n². Substituting that larger size into a subexponential expression can produce a running time that is no longer subexponential in the original n. The mathematical issue is the growth introduced by the transformation, not whether the two problems are related in a broad complexity classification.

The stronger exponential time hypothesis also needs careful treatment. It proposes more demanding limits for satisfiability as fixed clause width increases. A refutation of ETH also defeats this stronger assumption, but does not independently disprove every conjecture that researchers previously derived from it.

A conditional theorem remains a valid statement about what follows from its premise. What changes is whether that premise can support a conclusion about the actual difficulty of computation. Researchers must revisit each connection with attention to its assumptions and size bounds.

What the Result Would Demonstrate About AI Research

OpenAI’s announcement describes AI producing a proposed mathematical discovery, followed by formalization and external examination. That sequence provides a more informative account of capability than treating the result as a single score on an intelligence scale.

The research task involves proposing an argument that improves on existing knowledge. Verification asks whether the argument is correct. Explanation asks whether other researchers can reconstruct the reasoning and use it in subsequent work.

Those activities can benefit from different tools and different forms of expertise. A system may generate a useful algebraic construction without producing the clearest exposition, and a human researcher may contribute substantially by identifying the idea that makes a long proof understandable.

OpenAI describes its released manuscripts as research material from which scientific understanding still needs to be developed. Its stated collaboration principles include unpaid participation and freedom for participants to discuss their conclusions. Those arrangements address the risk that outside involvement could be mistaken for company-controlled endorsement.

Anthropic’s release adds another research artifact to the examination. It attributes its result to an internal research Claude model and publishes a paper with a Lean formalization, but explicitly describes the result as asymptotic rather than a statement about solving formulas in practice.

The existence of internal research systems should not be confused with the behavior of products available to customers. New Space Economy’s coverage of AI providers and models distinguishes experimental systems from production offerings. A research announcement does not identify a generally available service with the same performance unless the laboratory provides that evidence.

The discovery also differs from performance on a fixed test. Discussion of AI intelligence evaluation concerns how measured performance should be interpreted; a new theorem presents a separate test of whether a system can contribute something that survives independent examination.

Neither type of evidence should be stretched into a universal conclusion. Success on one mathematical research problem does not establish reliability across every scientific discipline, just as an error on an unrelated task does not invalidate a proof that has been checked.

The publication process introduces questions about credit. The review organizers identify unequal early access and the recognition of explanatory work as issues requiring attention. Discovery, correction, and simplification can each contribute knowledge, and publication speed alone does not measure those contributions.

Why Software and Computing Infrastructure Will Not Change Overnight

No practical deployment follows automatically from an asymptotic theorem. The released result concerns a mathematical limit on running-time growth, and converting that advance into useful software requires additional work.

An implementation would need to expose the algorithm’s actual constants and memory demands. It would then need comparison with established solvers on defined collections of instances, including inputs that represent the intended application rather than only convenient demonstrations.

Correctness requirements also matter. A research algorithm that counts every satisfying assignment exactly addresses a demanding task, but a commercial workflow may require a feasible solution quickly rather than a complete count. The best method for the exact counting problem need not be the best operational choice for that workflow.

The distinction extends to AI infrastructure. New Space Economy’s explanation of different AI workloads separates model training from inference, the process of running a trained model. An improved satisfiability algorithm does not automatically accelerate the numerical operations that dominate either workload.

A connection would have to be demonstrated. Researchers would need to show that a costly component can be translated into the relevant satisfiability problem, that the translation preserves an advantageous size, and that the new method outperforms available alternatives after all overhead is included.

The same reasoning applies to data center demand. A theoretical result alone does not support a numerical reduction in electricity consumption or a change in infrastructure investment. Such estimates require measured savings for workloads that account for an identifiable share of demand.

For proposed orbital computing workloads, the claim is even more indirect. Software efficiency can affect the economics of computing in any location, but this theorem does not address launch costs or the physical constraints of operating hardware in space.

The result could still influence practical research. A proof may introduce a decomposition or representation that engineers can adapt without implementing the entire theoretical algorithm. That possibility is different from evidence that an application has already improved.

A useful next milestone would be a reproducible implementation accompanied by a clear account of where it wins and where it loses. Until such evidence is available, the strongest supported implications concern mathematical possibility and research direction, rather than immediate changes to everyday computing.

Summary

The exponential time hypothesis proposed that a persistent exponential barrier limits algorithms for 3-SAT. OpenAI’s October 2026 release claims to cross that barrier, supported by public mathematical artifacts and followed by external examination and explanatory work.

The scientific significance depends on the precise result: a subexponential algorithm would overturn a widely used assumption without proving P = NP or guaranteeing practical speed improvements. Commentary from researchers ranges from explicit non-certification to confidence that the hypothesis has fallen, with broad agreement that the consequences require careful analysis.

The lasting value may include a change in how computational limits are investigated. A failed assumption can redirect research toward algorithms that previously seemed unlikely, but subsequent advances will depend on extracting reusable ideas and establishing which problem transformations preserve the improvement.

Appendix: Useful Books Available on Amazon

• The Golden Ticket: P, NP, and the Search for the Impossible
• Computational Complexity: A Modern Approach
• Introduction to the Theory of Computation
• Quantum Computing Since Democritus
• Algorithms to Live By: The Computer Science of Human Decisions

Appendix: Top Questions Answered in This Article

What Is the Exponential Time Hypothesis?

The exponential time hypothesis is a conjecture about the difficulty of solving 3-SAT, a Boolean logic problem. It proposes that every algorithm encounters an unavoidable exponential barrier in its worst-case running time. A valid subexponential algorithm for 3-SAT would refute that conjecture, even if the algorithm remained impractical.

What Did OpenAI Announce in October 2026?

OpenAI announced that an internal AI system had produced a claimed refutation of the exponential time hypothesis. The company released a mathematical manuscript and a Lean formalization. An external effort involving Russell Impagliazzo, Ramamohan Paturi, and Ryan Williams was organized to examine the argument and its consequences.

What Is 3-SAT?

The problem 3-SAT asks whether true or false values can be assigned to variables so that every clause in a logical formula becomes true. Each clause contains at most three terms joined by OR, and the clauses are joined by AND. The task requires an exact yes or no answer.

Does Subexponential Mean Fast Enough for Practical Use?

Subexponential describes how running time grows as input size increases, rather than how quickly a particular computer finishes a task. Such an algorithm can still require an enormous amount of computation. Actual usefulness depends on constants, implementation overhead, and performance on the inputs an application needs to process.

Did OpenAI Solve P Versus NP?

The announced result does not resolve P versus NP. A polynomial-time algorithm for 3-SAT would establish that P equals NP, but a subexponential algorithm can still be slower than every fixed polynomial. The claimed result removes a stronger proposed barrier without settling the broader question.

What Does Lean Verification Mean?

Lean verification means that a formal mathematical statement has a proof accepted under the system’s definitions and logical assumptions. This provides evidence about that precise statement. Researchers still need to assess whether the formal statement captures the intended problem and supports the claims made in the accompanying manuscript.

Why Are Independent Researchers Examining the Claim?

Independent examination helps establish that the definitions, running-time analysis, and interpretation match the intended mathematical result. It also helps turn a complicated proof into an explanation that other researchers can use. The organizers have explicitly stated that their participation should not be treated as certification.

How Is Anthropic Involved?

Anthropic published its own claimed refutation, attributed to an internal research Claude model, with a paper and Lean formalization. Its documentation states that the result concerns asymptotic complexity rather than practical formula-solving performance. The separate release provides another artifact for researchers to inspect and compare.

Does Refuting ETH Make Earlier Research Incorrect?

Research that proves a conditional statement can remain mathematically correct even when its assumption fails. A theorem stating that ETH implies a particular lower bound does not become an invalid implication. However, that theorem can no longer establish the lower bound through ETH, and the bound needs independent justification.

Will the Result Make ChatGPT or Data Centers More Efficient?

The announcement does not establish an immediate improvement to ChatGPT performance or data center efficiency. A useful connection would require showing that an expensive workload can benefit from the new algorithm after translation and implementation costs. Measured improvements would be needed before estimating operational or energy savings.

Appendix: Glossary of Key Terms

Artificial Intelligence

Computer systems designed to perform tasks associated with capabilities such as learning, reasoning, and generating language. In this setting, AI refers to research systems that reportedly produced mathematical arguments, rather than to a claim that every available AI product can reproduce those discoveries.

Exponential Time Hypothesis

A conjecture that the worst-case complexity of 3-SAT has a positive exponential rate that algorithms cannot eliminate. It supplied an assumption for conditional computational lower bounds. A valid algorithm with subexponential running time would contradict the hypothesis.

Lean

A programming language and proof assistant used to express mathematical definitions and check formal proofs. Its acceptance of a proof concerns a precise statement under specified assumptions. Human examination remains necessary to connect that statement to the intended scientific claim.

Boolean Satisfiability

The problem of deciding whether a logical expression can be made true by assigning true or false values to its variables. A satisfying assignment must meet every requirement encoded in the expression. The abbreviation SAT refers to this decision problem.

3-SAT

A restricted satisfiability problem in which each clause contains at most three variables or negated variables. Terms within a clause are connected by OR, and clauses are connected by AND. The restriction still permits the full difficulty associated with NP-complete problems.

Worst-Case Running Time

The maximum time an algorithm requires across inputs of a given size, including the most demanding inputs. It differs from average performance and from measurements on selected benchmarks, which may omit inputs that expose an algorithm’s limitations.

Subexponential Time

A running-time growth rate whose exponent becomes small relative to input size. In discussion of ETH, it is commonly expressed as 2^o(n). Subexponential algorithms can remain much slower than polynomial algorithms and need not be practical on available hardware.

Polynomial Time

Running time bounded by a fixed power of input size, with constant factors permitted. Polynomial time provides an important theoretical definition of efficiency. However, a high exponent or a large constant can still make a polynomial-time algorithm unsuitable for practical applications.

Big-O Notation

Mathematical notation describing an upper bound on growth up to constant factors for sufficiently large inputs. It helps compare algorithms without specifying every implementation detail. Constants hidden by the notation can still have a large effect on practical performance.

Deterministic Algorithm

An algorithm whose next operation is determined by its current state and input, rather than by random choices. For a fixed input and specified computational model, its behavior is fixed. Deterministic correctness and running-time guarantees do not depend on favorable random outcomes.

Conjunctive Normal Form

A way to write a Boolean formula as an AND of clauses, with each clause consisting of terms joined by OR. The abbreviation is CNF. A fixed-width restriction limits how many terms each clause may contain, as in 3-SAT.

Permanent

A mathematical quantity calculated from a square matrix by summing products associated with permutations of its columns. It can represent counting problems. The permanent appears in explanatory work on the claimed ETH refutation as an intermediate problem through which satisfiability counting is analyzed.

Polynomial Interpolation

The reconstruction of a polynomial from enough information about its values, sometimes supplemented by derivatives. In the proof exposition discussed here, interpolation helps recover an exact mathematical quantity from computations at specially chosen points where the calculation has useful structure.

NP-Complete

A classification for decision problems that belong to NP and can express every other NP problem through a polynomial-time transformation. A polynomial-time algorithm for any NP-complete problem would establish P = NP. Weaker speed improvements require closer attention to transformation costs.

Inference

The process of running a trained AI model to generate an output for a new input. It differs from training, which changes the model’s learned parameters. A mathematical algorithmic advance does not automatically reduce inference costs without a demonstrated connection to the operations involved.