298 New National Standards Released in Batch: AI Will Also Have Clear Measurable Benchmarks to Follow

The smart speaker at home cannot understand dialects, and the smart door lock fails to connect to the home hub; when a factory introduces an AI quality inspection system, it needs to be re-debugged for every new production line; ports use algorithms from different enterprises, but the data interfaces are like several mutually unintelligible "dialects".

AI can do more and more things, but when it is actually applied to production and daily life, the problem is often no longer "whether the model is smart", but another set of more trivial and more costly challenges:

Can the interfaces match? How can the recognition effect be considered qualified? Who is responsible for handling system errors? Can different devices work together?

On September 10, the State Administration for Market Regulation and the Standardization Administration of China approved and released 298 important national standards, including 27 mandatory national standards and 271 recommended national standards. In the high-tech field, there are 47 standards related to artificial intelligence, intelligent voice interaction, smart services and other areas, with applications extended to scenarios such as electric power, ports, petroleum, and home furnishing.

A noticeable change is taking place: in the first half of the AI industry, competition focused on who can run faster; after entering real scenarios, the competition also lies in who can deliver stably according to the same unified benchmark.

47 Standards: More Than Just "Setting Rules" for AI

First, we need to clarify a unified definition.

The 47 standards released this time are not all pure AI technical standards, but cover multiple related directions such as artificial intelligence, intelligent voice interaction, and smart services. The 298 standards are not only focused on the digital industry, but also cover fields including semiconductors, integrated circuits, new materials, work safety, agriculture and rural areas.

What is really worth noting is that AI standards are no longer limited to algorithm indicators in laboratories, but have begun to enter real scenarios such as electric power, ports, petroleum and home furnishing.

For a model in the laboratory, getting one question right counts as progress; after being deployed at a port, it may need to identify containers, schedule equipment, give risk warnings, and connect to existing management systems. After being integrated into the power system, in addition to accurate prediction, it also needs to take into account stable operation, data security and exception handling.

This is like a chef making a great dish in a competition, which proves their excellent cooking skills; but to run a chain of restaurants, it is necessary to formulate a set of replicable standards covering raw material specifications, operating procedures, hygiene requirements and product quality.

What AI industrialization is facing is exactly the leap from "making one great dish" to "stably operating one hundred restaurants".

Standards may not be as eye-catching as large model parameters, but they determine whether a technology can move from demonstration projects to mass delivery.

Mandatory and Recommended Standards Are Not the Same

Among the 298 standards, 27 are mandatory national standards and 271 are recommended national standards. The two types of standards have different legal attributes, and people should not assume that all enterprises must immediately implement them just because they are marked as "national standards".

In accordance with the Standardization Law of the People's Republic of China, mandatory standards must be implemented, and they are usually directly related to the protection of personal health, life and property safety, national security, and ecological environment safety. Recommended standards are encouraged by the state to be adopted, and are mainly used to unify technical requirements, testing methods and service specifications.

However, "recommended" does not mean "optional".

For an AI enterprise, whether it can comply with prevailing recommended standards may affect customer procurement, project acceptance, bidding and supply chain cooperation. If an enterprise promises to implement a certain recommended standard in contracts, product descriptions or public statements, the standard will also become part of the transaction relationship between the two parties.

Put in plain terms: mandatory standards are more like "red lines that must be crossed", while recommended standards are like the "common language" in industrial cooperation.

Enterprises can have different technical routes, but if the interfaces, testing methods and quality evaluation are completely different, every time a customer purchases products from a new supplier, they have to re-translate, adapt and verify. Before the project starts to generate benefits, the budget may already be consumed on "decoding secret signals".

What standards reduce is exactly this part of the less visible communication and collaboration costs.

When AI Enters Homes, "It's Smart" Is Far From Enough

Ordinary consumers rarely take the initiative to read national standards, but they are enjoying the results brought by standards every day.

Whether the charging interface can match, whether the home appliance size fits, how to read food labels, and how to determine liability when a product malfunctions, there is an invisible set of benchmarks behind all these scenarios.

After AI enters the home, it cannot avoid this logic.

For example, an intelligent voice device can answer questions fluently at the press conference, but after it is actually placed in the living room, it has to deal with situations such as unclear speech from the elderly, incomplete expressions from children, different accents of family members, and environmental noise interference. Enterprises cannot only demonstrate a successful demo once, but also need to specify the test conditions, how the recognition rate is calculated, and how to give reminders after a failure.

The problems of smart home are even more prominent.

Door locks, air conditioners, lighting systems, speakers and home robots may come from different enterprises. If each device uses its own set of data formats, what consumers bring home is not a "whole-house smart home", but a room full of appliances that are smart on their own but cannot communicate with each other.

Unified interfaces and interaction requirements can reduce repeated adaptation; clear safety and quality requirements help reduce risks brought by misoperation, data leakage and system failure.

Therefore, standards do not make all products look the same, but first define the width of the door, the shape of the plug and the language of communication, so that innovation does not have to start from "whether it can connect" every time.

For Enterprises, Standards Are a Cost Sheet

What AI enterprises are most likely to show is model capability, while customers are more concerned about whether the project can be delivered on schedule.

When a system is deployed in a factory, it usually goes through requirement analysis, data sorting, interface modification, model debugging, test acceptance and post-maintenance. In the absence of general standards, many tasks have to be redone for every new customer.

The result is that although the unit price of the project seems not low, the profit may be eaten up by a large amount of custom development and after-sales services.

The commercial value brought by standardization can be broken down into four accounts.

Unified interfaces can reduce the costs of system integration and repeated development; unified evaluation methods help reduce repeated disputes between suppliers and customers in the acceptance stage; unified quality requirements make it convenient for enterprises to replicate one project to more customers; clearer safety boundaries can also reduce compliance uncertainties after application deployment.

This is why standards are becoming increasingly important as AI enters the deep water zone of the industry.

In the past, enterprises might rely on an experienced team to build projects from scratch: solve problems on site when they are encountered, and re-develop for new customers. This method can produce demonstration projects, but it is difficult to form a stable scale.

A truly mature AI product should be like industrial components, with clear definitions of input, output, performance testing methods, and fault handling solutions. What customers buy is not just the experience of several engineers, but a set of verifiable, replicable and maintainable products.

The clearer the standards are, the more likely AI will evolve from a one-time project to a continuously sold commodity.

Chips and Materials Are Also Building the Same Shared Foundation

Among the standards released this time, there are 26 standards for semiconductor devices, integrated circuits and other related fields, 4 standards for information communication, data and software, and 6 new material standards for photovoltaic glass, fine ceramics and other products.

These fields seem scattered, but they are actually highly related to AI industrialization.

AI applications require computing power, which is supported by chips; when chips are installed in devices, communication, data and software collaboration are needed; when devices are deployed in factories, automobiles, homes and energy systems, materials, sensors and safety rules are also indispensable.

Whether a product can be mass-produced stably is never determined by a single model alone. Any link without a unified benchmark, including chip reliability, data format, software interface, material performance and complete machine testing, may affect delivery.

This also reminds the market that the number of national standards cannot be directly equated to market size, let alone used as proof that a certain technology has taken the lead.

Standards solve the problem of "how everyone tests, connects, and judges qualification"; whether an enterprise can make good products still depends on its R&D, manufacturing, cost control and service capabilities.

Having benchmarks does not mean everyone can get a high score in the exam. But without benchmarks, it is even hard to tell who performs better.

Next Step: Do Not Only Count the Number of Standards

The concentrated release of 298 national standards supplements a batch of industrial rules for AI, semiconductors and new materials. However, there is still a long way to go between the release of standards and their actual effect.

Some standards require enterprises to modify their products and processes, some need to be incorporated into procurement, testing and acceptance systems, and others require joint adoption by the upstream and downstream of the industrial chain. This is especially true for recommended standards: release is only the starting point, and the adoption rate determines their influence.

In the future, to judge whether AI standardization is truly implemented, five things can be focused on:

First, look at applicable scenarios. Are the standards only staying at the conceptual level, or have they entered specific businesses such as ports, electric power, and home furnishing?

Second, look at testing methods. Can the accuracy, response speed and security capabilities claimed by enterprises be reproduced under unified conditions?

Third, look at compatibility. Can repeated adaptation be reduced for devices and software from different enterprises?

Fourth, look at procurement and acceptance. Are relevant standards included in bidding documents, product contracts and project acceptance indicators?

Fifth, look at the implementation date. Mandatory standards and recommended standards have different natures, and the release date does not mean that the industry has completed the transformation.

In the early stage of AI development, competition focused on parameters, but after AI truly enters factories and homes, the competition often lies in another thing: whether it can work stably, whether it can be replicated in batches, and whether the responsibility can be clearly defined when problems occur.

Models determine how fast AI can go, while standards determine how steadily it can advance.

When new technologies enter the stage of "measurable by unified benchmarks", the competition of the AI industry has shifted from capability demonstration on the stage to long-term delivery in production lines, procurement orders and daily family life.

This article is from the WeChat official account "BT Tech & Business", written by Shuyan, and published with authorization from 36Kr.