Breaking the dimensional wall between the virtual and the real, the Pan-energy Network has created the "physical AI" in the energy and carbon field.

This is a prediction about the implementation of AI in the real world, which has been bet on by tech giants such as NVIDIA and Google. Physical AI means that AI can understand the operating laws of the physical world: from the grand natural laws to the motion trajectories of objects under inertia. When AI understands all this and can make corresponding plans, all its capabilities can be integrated into the physical world.

The rise of Physical AI as the protagonist of the new - generation technology is driven by the real - world needs of traditional industry upgrading and emerging industry development. It also marks a turning point for AI from technological showmanship to practical application. As AI has developed, people have shifted from technological fanaticism to a greater focus on AI's ability to solve real - world problems and pain points.

Especially in the industrial sector, technology is not a castle in the air. AI must ultimately be applied in reality to solve the real problems and needs faced by industrial development. The industrial sector urgently needs AI to understand the operating rules of the physical world and the characteristics of industrial laws - Physical AI meets this need.

With the goal set, the path to Physical AI involves choosing a technical route.

To some extent, large language models are not the ideal choice. On the one hand, large language models focus more on text - processing capabilities and struggle to handle multi - modal information flows such as audio, images, and videos in the physical world. More importantly, the core limitations of large language models, namely hallucination and forgetting, are fatal flaws that prevent them from truly supporting Physical AI.

Let's start with "hallucination". Since large language models generate content by statistically analyzing word associations in a vast amount of text and rely on text pattern matching, they are essentially "predictors" of words and cannot verify the authenticity of information through sensory perception or logical reasoning. Therefore, LLMs may produce content that seems reasonable but does not conform to the facts, which is often criticized as "nonsense".

"Forgetting" occurs when large language models learn new knowledge, irreversibly overwriting or weakening old knowledge, leading to the loss or distortion of previously mastered information. Whether it is context forgetting or knowledge forgetting, it reduces the coherence and accuracy of multi - round responses.

Obviously, these limitations of LLMs are far from what Physical AI needs to understand the real world.

The way out lies in another type of model: reasoning models that can perceive and understand the world, namely world models. This is also the technical solution advocated by Yann LeCun, a Turing Award winner and the chief AI scientist at Meta.

In terms of specific solutions, how can a model gradually understand the world and make reasoning and planning? According to 36Kr's observation, in the AI practice of the energy industry, the combination of "simulation + mechanism" is being used to enable AI to understand the physical environment and plan actions, creating a "world model" for the energy industry.

The crucial step towards Physical AI is taking place in the energy sector first.

More specialized than Physical AI is Energy AI

Compared with the grand proposition of "understanding the physical world", starting from a specific field and developing vertical - industry Physical AI is a more feasible and professional implementation plan.

Take the energy industry as an example. "Energy + AI" mainly refers to AI as a technological variable driving and empowering the transformation of the energy industry and upgrading energy management systems. However, the AI that can truly understand the complexity and operating laws of the energy industry and make corresponding reasoning, planning, and decision - making is "Energy AI". Different from a simple additive relationship, "Energy AI" is an organic whole and a complete new paradigm of vertical AI.

It can be said that Energy AI, as a new type of intelligent system, redefines the interaction between energy infrastructure and the real world, especially in the industrial sector.

Just as Physical AI requires the support of world models, building Energy AI requires a "world model" in the energy field. In terms of implementation paths, a trend - setting solution has emerged: simulation + mechanism.

Simulation is not new and is not unique to the energy industry. There is a technical consensus in the industrial sector: using simulation technology and digital twins to simulate real - world systems, processes, or environments, and analyzing and verifying the feasibility and effectiveness of a technology or solution in this "recreated world" to help humans understand complex environments, optimize designs, and reduce risks and trial - and - error costs.

Simulation is widely used in many fields such as industrial manufacturing (e.g., product design of automobiles and spacecraft), healthcare, and power management.

This model can be transferred to the energy application end. By simulating the real energy - carbon operating environment through simulation technology, model algorithms can be trained to enable the model to quickly master the production and operation laws of the industry.

Simulation is just the first step. To enable the model to truly understand the energy - using logic of the industry, form professional knowledge, and make behavioral decisions, the model algorithm needs to understand the interaction principles and operating laws of various elements within the energy system (such as different process equipment and sensors in a factory), that is, the energy - using mechanism, and continuously optimize based on this mechanism.

Through simulation + mechanism, the model can be well - prepared before entering the real energy world.

This is a reflection of technology being applied in reality. The trend of technological evolution is predictable. The key to differentiating a technology and determining whether it can release its true value lies in moving from the laboratory to the production line, being tested in reality, and exploring a feasible implementation path.

The biggest barrier in the implementation path of Energy AI lies in energy industry data and know - how. To enable AI to understand the energy industry, the first step requires a large amount of real and high - quality industry data to support model training. On this basis, industry mechanisms are also needed for fine - tuning and correction.

This undoubtedly raises the threshold for exploring Energy AI. However, some pioneers have ventured into the "uncharted territory" of building Energy AI, and ENN Fanneng Network is one of them.

Relying on ENN Group's more than 30 - year professional accumulation and private - domain data in the energy industry, Fanneng Network has developed an energy - specific tool called "Fanneng Simulation". Through technologies such as RAG (Retrieval - Augmented Generation) and enhanced fine - tuning, it is deeply integrated with the basic large model to build an AI that can understand the knowledge of the energy industry and the operating laws of the energy - using world, namely "Energy AI".

Smooth crossover of "Autopilot" from transportation to energy

On the exploration path of building Energy AI, Fanneng Network first proposed a core concept in 2024: Autopilot in the energy field. After a year of thinking, precipitation, and practical experience, "Autopilot in the energy field" has formed a systematic architecture.

Those who are first exposed to the concept of "Autopilot in the energy field" may wonder: What does "autopilot", a hot term in the transportation field, have to do with energy?

This is also a question that Cheng Lu, the CEO of Fanneng Network, is often asked. In his view, the current initial stage of transformation in the energy industry is very similar to the "intelligent driving" of vehicles.

Specifically, they also have corresponding similarities in the hierarchical logic of the system architecture.

Vehicle autopilot can be broken down into three major centers: First, the perception model of the world, including sensors, cognitive models, and decision - making systems, which is the brain center of intelligent driving; Second, the main system represented by the intelligent cockpit, which can interact with the outside world and make refined adjustments and decisions based on environmental variables; Third, all control and execution units, such as vehicle steering, acceleration, and deceleration.

The autopilot system in the energy field built by Fanneng Network is also based on the above three - layer logic.

First is the brain center: the energy - carbon large model. This is a professional large model that integrates Fanneng Network's years of energy - specific knowledge and private - domain data accumulation. Using AI technologies such as the Fanneng Simulation platform tool and RAG, it can understand the operating laws of the energy - carbon field. It can be said that this is a "world model" for the energy - carbon field.

The main system of the intelligent cockpit corresponds to Fanneng Network's professional Agents. These Agents are like a virtual expert team under the command of the brain center. From task planning and decomposition, decision - making execution in specific vertical scenarios, to industry knowledge and mechanisms, this Agents system architecture can decompose and efficiently execute complex energy management tasks in different scenarios.

Finally, the vehicle's control unit corresponds to energy - carbon intelligent control. Facing the specific needs and tasks of the energy management system, Fanneng Network's energy - carbon intelligent control is responsible for implementation, completing the closed - loop from perception, cognition to decision - making behavior.

Based on these similarities, Fanneng Network has proposed a grading system from L1 to L5 for energy autopilot.

Cheng Lu admitted that at present, many AI products in the energy industry focus on the L1 to L2 stages, that is, a small amount of automation and little human participation. Fanneng Network is gradually breaking through L3, that is, local autonomy, where the energy system can ensure safety, economy, efficiency, and convenience without human decision - making. By the L5 stage, it will achieve intelligent operation across multiple fields of water, gas, electricity, heat, and cold, realizing global autonomous collaboration.

Just as the ultimate goal of vehicle autopilot is not only to optimize the performance of a single vehicle but also to intelligently dispatch the entire transportation system through vehicle - network - cloud collaboration to make the overall operation more economical and efficient, Fanneng Network's vision for the ultimate goal of energy autopilot is to have both edge and cloud computing capabilities in different application scenarios and customer groups, possess a global operation decision - making system, and achieve collaborative optimization between the energy system and the Internet of Things.

From vehicle transportation to the energy system, "Autopilot" bears the future mission of global intelligent autonomy.

Energy AI grows "hands and feet" and enters industrial scenarios

Energy AI is the destination, and each step along the way requires a practical implementation carrier. The new - generation energy - carbon intelligent control all - in - one machine launched by Fanneng Network undertakes this role.

Compared with the first - generation product launched last year, the new - generation energy - carbon intelligent control all - in - one machine has been upgraded in both hardware appearance and solution capabilities. In terms of hardware, the new product integrates sensing devices such as cameras and NFCs, with a more technological appearance. In terms of product capabilities and user experience, through the edge - cloud system architecture, the product form presented to customers is not obscure tools and files, nor difficult - to - understand algorithms, but energy management programs, professional intelligent assistants that can be used immediately upon arrival.

This means that with the energy - carbon intelligent control all - in - one machine, Energy AI not only has a smart brain but also grows "hands and feet", entering various industrial scenarios in daily life in the form of embodied intelligence.

Taking the "clothing" industry as an example, last year, Fanneng Network applied the first - generation energy - carbon intelligent control machine to the dyeing vat scenario in the printing and dyeing industry. After a year of upgrading and iteration, the application scenarios of the second - generation machine have become more diverse, expanding from dyeing vats to stenter machines.

Tenter framing and setting is a key process in printing and dyeing. After the cloth is dyed in the dyeing vat, the setting process directly determines the size and texture of the fabric. From dyeing to setting, Fanneng Network has in - depth built multi - process Agents for the printing and dyeing industry, reducing the annual cumulative cloth loss enough to make more than 5 million T - shirts.

From building a world model in the energy - carbon field, to exploring energy autopilot, and finally to real - world implementation in daily life, Fanneng Network's Energy AI connects the digital world and the physical world, reconstructing the AI - driven transformation of energy infrastructure.

When AI shines into the energy world and physical factories, the industrial ecosystem and industry customers will also witness a new round of technological dividend release and value upgrading.