AI: Striving to Become a Trusted ‘Future Advisor’

Can you imagine what predictive technology looks like? When the foundational capabilities of general large models, the precision of specialized predictive models, the practical value of external tools, and the assurance of trustworthy mechanisms are organically integrated, AI will gain a new insight into the future. It will become a trusted ‘future advisor’ for humanity in critical areas such as financial risk control, weather forecasting, public governance, and industrial production, providing intelligent support for understanding future trends and becoming a significant force in empowering social development and modernizing national governance.

Four Technical Paths for ‘Predicting the Future’

Faced with the increasingly complex predictive demands of the real world, researchers have developed two core lines and four specific technical paths around large model predictive technology. These paths are not competing alternatives but complement each other in different scenarios, collectively constructing a complete research framework for large model predictions.

The essential difference between the two core lines lies in whether a dedicated model is tailored for the prediction task: one is ‘borrowing a boat to go to sea,’ cleverly utilizing existing mature large language models for predictions; the other is ‘building a ship to sail far,’ reconstructing dedicated foundational models for predictions. Both paths advance simultaneously, adapting to diverse task requirements.

Directly invoking large language models is the easiest entry point for large model predictions. Researchers convert various predictive tasks into common natural language questions, providing historical information, event backgrounds, and constraints for the model to directly assess future trends and output predictions. This method has a low barrier to entry, requiring no significant modifications to the model; it merely changes the application of existing tools, performing well in open-world problems like news event analysis and business trend assessment. However, it is limited by the numerical computation capabilities of large language models and the potential for factual output deviations, making it challenging to meet the stringent requirements for high-precision numerical predictions in fields like meteorology and finance.

Time series tokenization modeling is a cross-domain ‘intelligent borrowing.’ It cleverly introduces classic natural language processing ideas into time series data analysis, using techniques such as discretization, scaling, and quantization to transform continuous time series data into token representations similar to words in language, and then trains using a language model architecture. The representative model, Chronos, maps time series to a fixed vocabulary, achieving probabilistic predictions and cross-dataset generalization, significantly reducing development costs by reusing mature language model architectures. However, this convenience comes at a cost, as the data transformation process inevitably leads to the loss of numerical details and quantization errors, akin to roughly polishing fine parts, which can affect prediction accuracy.

Building dedicated time series foundational models marks a shift from ‘borrowing strength’ to independent innovation in large model predictive research. Researchers no longer view time series simply as pseudo-text but design pre-training schemes and model architectures tailored to the essential laws of time series data and the core needs of predictive tasks. Google’s TimesFM employs a decoder architecture, demonstrating strong zero-shot prediction capabilities; Lag-Llama, developed by multiple universities and research institutions in the U.S., focuses on probabilistic predictions and cross-domain generalization; and Moirai, developed by an American AI company, boldly attempts to adapt to more scenarios using a unified training approach. These models act like ‘custom armor’ tailored for predictive tasks, closely aligning with the characteristics of the tasks themselves, achieving higher precision in numerical predictions and becoming the preferred choice for high-precision prediction scenarios.

Reprogramming large language models and multimodal integration provide a low-cost approach for large model predictions. Research related to Time-LLM confirms that without retraining massive time series models with hundreds of billions of parameters, aligning time series with textual prototypes through reprogramming allows ‘frozen’ large language models to participate in prediction tasks. This approach opens a feasible pathway for the general large model + specialized adaptation technical route, promoting the deep joint modeling of text, numerical, and contextual knowledge, allowing predictions to integrate multi-source heterogeneous information like human thinking, better fitting the complex and variable predictive scenario requirements of the real world.

These four technical paths do not have absolute advantages or disadvantages; they are like different keys fitting different locks. When prediction tasks require combining general knowledge and textual backgrounds for open trend assessments, routes related to large language models act like master keys with greater advantages; when tasks pursue high-precision numerical outputs and stable cross-domain generalization capabilities, dedicated time series foundational models become the customized keys for precise matching. They support and enhance each other under different resource conditions and actual task requirements, collectively advancing large model predictive technology steadily forward.

Moving Towards Real Application Scenarios

In the research arena of large model predictive technology, international research has started earlier and has a more systematic technical framework, delving deeper into basic research and frontier exploration; domestic research, though starting slightly later, has rapidly caught up with strong momentum, forming unique advantages in scenario adaptation, open-source ecology, and application implementation.

International academic research on large model predictions has evolved from text reasoning to multi-dimensional predictions. Early research primarily focused on applying large language models to text reasoning and event development assessments, akin to cultivating a small plot of land; in recent years, it has gradually broken boundaries, expanding into time series, spatiotemporal data, and even scientific predictions, entering a new phase of ’expanding territory.’ In the more complex field of scientific predictions, Microsoft’s ClimaX has pioneered the establishment of a foundational model framework for weather and climate tasks, while another Microsoft project, Aurora, extends foundational model ideas to the Earth system, capable of handling multiple predictive tasks such as weather, air quality, and wave forecasts, akin to equipping the Earth with an intelligent early warning system, showcasing the immense potential of scientific foundational models in complex system predictions.

Notably, the international academic community maintains a rational and prudent attitude towards the predictive capabilities of large models. Relevant studies have found that the excellent performance of large models in standardized tests does not equate to reliability in predicting future real-world events—GPT-4’s probabilistic predictions in open-world prediction competitions have been shown to be weaker than the median predictions of human groups. Addressing this core issue, international researchers have successively conducted competition studies, retrieval enhancement studies, and uncertainty detection studies, forming a distinctive characteristic of international research that emphasizes ‘model capability enhancement + prediction result verification + trustworthy mechanism construction,’ laying a solid foundation for the practical application of technology.

Domestic research, relying on the rapid development of general large models, has achieved impressive late-stage catch-up, gradually forming a positive development pattern of rapid iteration of general large models, systematic review research, and steady progress in application implementation. In the arena of general model ecological construction, various players showcase their strengths: Qianwen 3 has established a complete system for multilingual support and reasoning efficiency optimization, akin to building a multilingual intelligent bridge; DeepSeek-V3 has achieved a technological breakthrough in high-performance open-source models, making core technologies more accessible; and Wenxin 4.5 continues to refine multimodal integration and engineering deployment, increasingly aligning with actual application needs. Although these general large models are not solely focused on prediction, they provide a solid capability foundation for domestic large model predictive research, enabling researchers to stand on the shoulders of ‘giants’ and conduct more targeted studies.

At the application implementation level, domestic efforts are actively exploring ways to bring large model predictive technology out of the ‘ivory tower’ and into real application scenarios across various industries. Some studies deeply integrate expert knowledge with large language models for strategic warning, accurately realizing trend assessments and risk identification in complex situations; others closely combine large models with meteorological monitoring data, attempting to enhance the accuracy and timeliness of short-term precipitation predictions. Although these studies are not entirely equivalent to pure numerical time series predictions, they signify that domestic large model predictive technology is transitioning from theoretical discussions to practical applications, beginning to explore technical paths that meet local needs and align with industry realities.

Overall, international research has delved deeper into the development of dedicated foundational models for predictions and scientific predictions, akin to excavating extensive tunnels underground, forming a relatively complete technical system; domestic research, on the other hand, showcases distinctive features in adapting to Chinese scenarios, constructing low-cost open-source ecosystems, and implementing industry applications, akin to building high-rise buildings that fit local contexts above ground. With the continuous accumulation of high-quality time series data and industry-specific data in China, as well as the gradual improvement of dedicated evaluation systems, there remains significant room for improvement in domestic foundational models aimed at predictive tasks, which will undoubtedly contribute unique and valuable Chinese wisdom to the global development of large model predictive technology.

Bridging the Gap from ‘Powerful to Trustworthy’

Compared to traditional predictive methods, large model predictive technology has achieved a profound transformation from ‘point calculations’ to ‘comprehensive assessments,’ evolving from a cold mechanical computing tool into an intelligent entity capable of understanding contexts, weighing factors, and providing rational judgments. This unique ability stems from its inherent core advantages, yet like a growing star, it is steadily evolving towards ‘from powerful to trustworthy,’ striving to become a reliable ‘future advisor’ for humanity.

The core advantages of large model predictive technology are its innate exceptional capabilities, particularly prominent in practical applications. First, it has strong cross-task transfer capabilities. Traditional agricultural yield prediction models cannot be directly applied to stock market trend analysis; switching fields requires a complete overhaul. In contrast, large models, with their general representation capabilities from extensive pre-training, can quickly adapt to different domains like agriculture, finance, and industry with minimal samples. Second, it has great potential for handling complex dependencies. For instance, predicting river water levels during flood seasons is influenced by multiple factors such as rainfall, upstream discharge, and terrain, which traditional models struggle to capture. In contrast, time series foundational models can learn patterns within contextual ranges, akin to having ‘keen insight’ to see the connections behind the data. Third, it excels in multi-source information integration. Traditional meteorological predictions rely solely on numerical monitoring data, while large models can integrate multi-source content such as satellite cloud images, meteorological text reports, and geographic information, transforming predictions from ’narrow observations’ to ‘panoramic views.’ Fourth, it possesses excellent prediction interpretation and decision support capabilities. It can not only predict the trend of a specific stock but also explain the influencing factors like industry policies and market supply and demand, even providing risk control suggestions, becoming a professional intelligent partner for decision-makers.

Despite these significant advantages, large model predictive technology is not without flaws; there remains a ‘gap’ to be bridged from the laboratory to real application scenarios. First, the model’s generative and inferential capabilities do not equate to actual predictive capabilities. Some models perform excellently in simulated meteorological prediction tests but often ‘fail’ in real severe convective weather warnings, simply because the test answers are buried in the training data, while real predictions require comprehensive assessments of unoccurring events—talking on paper is easy, but ‘real combat’ is challenging. Second, retrieval enhancement addresses symptoms rather than root causes. While pairing models with information retrieval improves prediction accuracy, it also indicates that models rely solely on their memory of knowledge, akin to guarding an old library, struggling to keep up with real-world changes; acquiring up-to-date knowledge in real-time is crucial. Furthermore, hallucinations and factual instability pose core obstacles, akin to hidden time bombs. Additionally, constraints of cost, data, and evaluation systems make large-scale applications challenging. Training high-precision models requires massive computational resources, leading to high development costs; in reality, time series data is fragmented and lacks uniform labeling, making it difficult to produce high-quality outputs from poor raw materials. Existing evaluation systems often focus on numerical errors while neglecting factual stability, causing many models to appear excellent yet struggle to implement effectively.

Looking ahead, the development direction of large model predictive technology is clear, focusing on ‘from powerful to trustworthy’ to create a mature technical system that can reliably serve real decision-making. First, general large models will evolve into dedicated foundational models for predictions, demonstrating stronger competitiveness in high-precision demand scenarios like meteorology and finance. Second, tool enhancement will become an important direction, allowing models to autonomously call external tools like search and simulation, akin to equipping intelligent agents with a toolbox to better tackle complex scenarios. Third, trustworthiness, controllability, and interpretability will become research priorities; future prediction systems must not only be numerically accurate but also quantify risks and trace judgment bases, which is key for implementing high-risk scenarios. Fourth, accelerating low-cost deployment and industrialization will transform technology from exclusive assets of a few institutions into common tools across various industries as inference costs decrease and open-source ecosystems improve. Finally, domestic research will deepen localization adaptation, creating dedicated models that combine the Chinese context and local data, making large models more accurate, stable, and trustworthy in domestic financial risk control and governmental early warning scenarios.