AI 2022-2026: The Evolution and Trajectory of Large Language Models

    This article reviews the trajectory of AI and large language model development from 2022 to 2026 through three asynchronous curves: technology, legal regulations, and business models. It examines key milestones including the rise of open-weight models, the commercialization of closed-source frontier models, the native multimodal architecture shift, the competitive landscape of domestic models in China, and the 2024 breakthroughs in AI agents and scientific computing. Finally, it provides three actionable recommendations for technology decision-makers on balancing technical capabilities, compliance, and engineering costs.

    Introduction: Three Asynchronous Evolution Curves

    Observing the development of artificial intelligence from 2022 to 2026 requires looking beyond the technology itself. In reality, three parallel tracks with entirely different paces have been driving the industry: technological capability advances the fastest, with landmark breakthroughs occurring almost every quarter; legal regulations typically lag by one to two years, meaning that by the time new rules are implemented, the controversies surrounding the previous generation of technology have often become obsolete; and business models with pricing strategies fluctuate between the two—charging premiums during technical breakthroughs and being forced to cut prices or offer services for free when impacted by open-source alternatives.


    Understanding the misalignment of these three tracks is the key to deciphering all industry events during this period. Below, we chronologically outline key milestones and analyze their respective impacts on the technological, legal, and business fronts.


    1. The Open Weights Era: The Accidental Spark of Community Power


    The Weight Leakage Incident and the Wave of Derivative Models


    In February 2023, the weights of a series of large language models (ranging from billions to tens of billions of parameters) originally restricted to research applications by a certain research institution were accidentally made public. The significance of this event varied across the three tracks:



    • Technology track: For the first time, a large number of developers gained access to frontier-level models that could run locally and be fine-tuned. This led to a wave of lightweight models centered around instruction tuning, validating the feasibility of a low-cost route: fine-tuning foundation models with small datasets.

    • Legal track: The public circulation of the weights exposed the enforcement challenges of model licensing agreements, prompting the industry to rethink the boundary definition between “open weights” and “open source.”

    • Business track: The pricing foundation of closed-source vendors was fundamentally shaken. When freely available models could cover 80% of use cases, API providers had to prove the additional value of the remaining 20%.


    Milestone of Open Commercial Licenses


    In July 2023, a leading vendor, in collaboration with a cloud service provider, released an open-weight model permitted for commercial use, marking the formal establishment of the Western open-weight ecosystem. Within the following year, numerous domain-specific fine-tuned versions emerged around this model, covering verticals such as legal, medical, finance, and coding. For developers, this established a reusable engineering paradigm: a foundation model as the base, domain-specific corpora for fine-tuning, and an inference framework for deployment—a paradigm that remains the mainstream path for private deployment today.


    2. Closed-Source Frontier: From Paid Subscriptions to Default Multimodal Architecture


    The Commercialization of Frontier Capabilities


    In March 2023, the first large-scale commercial multimodal frontier model was released, marked by its highly publicized ability to pass rigorous professional exams like the bar exam. The true significance of this event was that it transformed “frontier capabilities” into a standardized paid subscription product for the first time. Prior to this, large models were mostly research demonstrations or restricted APIs. Subsequently, a three-tier commercial structure of “free tier + subscription tier + enterprise API” was formed, a structure adopted by almost all subsequent closed-source vendors.


    Native Multimodal as the Default Architecture


    In December 2023, a vendor released the first native multimodal model family jointly modeling text, image, audio, and video from the training phase, rather than bolting a visual encoder onto a language model. At the same time, its three-tier scaling strategy (Ultra/Pro/Nano) was widely mimicked—distilling different compute tiers from the same architecture to cover deployment needs from data centers to edge devices.


    From an engineering perspective, the core difference of native multimodality lies in the unified representation space: all modalities share semantic representations early in training, so cross-modal reasoning (e.g., “answering text questions based on charts”) does not require additional alignment modules. The takeaway for developers is that when evaluating multimodal models, they should focus on performance in cross-modal combinatorial tasks, rather than the sum of single-modal capabilities.


    3. Domestic LLMs in China: The “Hundred Models War” and Its Evolution


    On March 16, 2023, a leading Chinese tech enterprise officially released a conversational large model product, marking the beginning of the domestic LLM competition in China. Within the next two months, multiple vendors densely released their own models. According to relevant regulatory authorities, the number of released models with parameter scales exceeding one billion reached 79, a phenomenon the industry dubbed the “Hundred Models War.”


    This phase exhibited three distinct characteristics:



    • Architecture convergence: The vast majority of models utilized the dense Transformer architecture. Differentiation primarily relied on the scale and quality of Chinese corpora, while competition in cost and underlying architecture had not yet begun.

    • Homogeneous releases: The pace of releases outstripped capability verification. Evaluation metrics for many models were inconsistent, making horizontal comparisons difficult.

    • Compliance first: Management regulations for generative AI in China landed relatively quickly, forcing domestic vendors to establish engineering capabilities like content safety and filing audits earlier, objectively creating a unique engineering moat.


    For enterprise users in China, the practical benefits brought by the “Hundred Models War” are the rapid decline in inference costs and the maturation of private deployment solutions. When selecting domestic models, it is recommended to focus on three points: real task performance in Chinese domains (rather than leaderboard scores), the stability of long-context and tool calling, and the supporting engineering capabilities provided by the vendor (such as fine-tuning platforms, inference acceleration, and general cloud services like monitoring and alerting).


    4. 2024: Dual Leaps in Agents and Scientific Computing


    The “World Simulator” Narrative in Video Generation


    In February 2024, a research institution released a technical preview of minute-level video generation framed around a “world simulator” narrative. It redefined the public’s imagination of AI video: no longer just a few seconds of spliced clips, but long takes with a certain degree of physical consistency. Although initially limited to a preview, it immediately sparked widespread global discussions in the creative industry regarding job structures, copyright ownership, and content labeling—a quintessential manifestation of the “legal lag” track.


    Paradigm Shift of Coding Agents


    In March 2024, a team released the first product dubbed an “AI software engineer”: capable of autonomously reading documentation, writing code, running tests, and fixing errors in a sandbox environment, advancing coding assistance from “autocomplete the next line” to “deliver a complete task.”


    More importantly, the controversy it sparked—the gap between demo effects and real-world capabilities—directly defined the evaluation methodology for all subsequent coding agents: ignore code generation quality, focus solely on end-to-end task completion rate. The takeaway for engineering teams is that when introducing intelligent coding tools, they should build their own task-level evaluation sets, measured using real repositories and real issues, rather than relying on vendor demos.


    Milestones in Scientific Computing


    In May 2024, a research institution’s biomolecular structure prediction system received a major upgrade. The prediction targets expanded from single-chain proteins to complexes formed by proteins with DNA, RNA, ligands, and ions, with accuracy significantly surpassing previous specialized docking tools. This proved the migration capability of the large model paradigm in the field of scientific discovery, signaling that “AI for Science” would become the third major battleground after dialogue and code.


    5. Conclusion: Three Recommendations for Tech Decision-Makers



    1. Track the technology, make decisions with the business: A capability breakthrough does not mean it is worth paying for. Evaluate models based on your own task sets. Switch promptly when open-source solutions catch up to the “good enough” threshold to control costs.

    2. Leave a compliance buffer for legal lag: Rules regarding data export, content labeling, and training corpus compliance are still evolving. Architectural design should retain the ability to replace models and audit data.

    3. Treat Agents as an engineering problem, not a product problem: The value of an agent depends on the balance between task completion rate and manual takeover cost. During implementation, start with sandboxed, rollbackable small tasks and gradually expand the scope of authorization.


    The history of AI from 2022 to 2026 is essentially a history of the continuous misalignment and mutual traction of three curves: technology, law, and business. Understanding this misalignment holds far more long-term value than memorizing the name of any single model.