The New Paradigm of Large Models: How AI is Reshaping Industries

Digital Trinity and Paradigm Shifts

The article defines a digital trinity consisting of three subsystems that together form a stable structure for analyzing technological change:

Information subsystem : acquisition and storage of data. Historically, the cost of information shifted from marginal (pay‑per‑use) to fixed (large upfront investment) around 1995‑1996, enabling ubiquitous access (e.g., Google Maps).

Model subsystem : representation of knowledge. Starting in 2022‑2023, large‑language‑model (LLM) costs underwent a similar shift from marginal to fixed, turning models into a new production factor.

Action subsystem : planning and interaction with the environment based on model inference.

These three dimensions allow a systematic comparison of past inflection points (the rise of Google, Apple, Amazon) with the current AI inflection driven by OpenAI’s GPT series.

Evolution of the GPT Family

The progression of OpenAI’s models illustrates how the trinity is realized in practice:

GPT‑1 (2018) : Demonstrated that a single pre‑trained model could outperform task‑specific models on many NLP benchmarks.

GPT‑2 (2019) : Introduced fine‑tuning to transfer learned representations to downstream tasks.

GPT‑3 (2020) : Showed strong zero‑ and few‑shot generalisation via in‑context learning, eliminating the need for task‑specific fine‑tuning.

GPT‑3.5 (2022) : Added instruction fine‑tuning, enabling reliable conversational behaviour (ChatGPT).

GPT‑4 (2023) : Completed the engineering pipeline, supporting plugins, multimodal inputs, and a growing ecosystem of applications.

Key technical advances include unsupervised pre‑training on massive text corpora, instruction tuning, and Reinforcement Learning from Human Feedback (RLHF) for alignment.

OpenAI’s Platform Architecture

OpenAI’s ecosystem is split into a backend (the GPT‑N series, currently GPT‑4) and a frontend (ChatGPT). The backend provides APIs, a token‑based pricing model, and a rapidly expanding third‑party developer ecosystem. The frontend delivers a universal conversational interface and a plugin system that turns ChatGPT into a “killer app” for domains such as code interpretation, data analysis, and knowledge retrieval.

Economic and Societal Impact

According to OpenAI, 19 % of the global workforce will have at least 50 % of its tasks affected by LLMs ; Goldman Sachs estimates that 300 million jobs worldwide could be automated, with 80 % of U.S. workers seeing at least a 10 % impact . The article argues that knowledge, now compressed into models, becomes the dominant factor of production, accelerating innovation in knowledge‑intensive sectors such as healthcare, research, and education.

Technology‑Economic Paradigm (Carlota Perez)

The author applies Perez’s four‑phase model (eruption, frenzy, synergy, maturity) to the AI wave, predicting an early “bubble” phase followed by a post‑bubble era where new firms built on fixed‑cost models dominate the market.

Strategic Opportunities

Opportunities are organised around the trinity and the “human + task” framework:

Information layer : Infrastructure for compute, networking (RDMA, RoCE), tokenisation, and vector databases. Companies that provide the “golden shovel” (e.g., high‑performance interconnects, storage, embedding services) can capture outsized value.

Model layer : Open‑source model ecosystems (LLaMA, Dolly, MosaicML), parameter‑efficient fine‑tuning (PEFT), domain‑specific adapters, and multimodal diffusion models (Stable Diffusion, ControlNet). Early movers can shape standards for alignment, safety, and value‑alignment.

Action layer : Autonomous systems that combine perception, planning, and actuation—robotics, autonomous driving, and spatial computing. The author highlights the advantage of language‑first models for generalisation (e.g., a driver recognizing an elderly pedestrian).

Sector‑specific analyses include:

Search : Transition from keyword retrieval to conversational, task‑oriented agents (Bing, Bard). New monetisation models beyond ads are needed.

Content creation : Democratisation of high‑quality media (text, image, video) via diffusion models; opportunities in tooling, workflow automation, and intellectual‑property management.

Gaming : Procedural generation of assets, narratives, and NPC behaviour (e.g., Opus.ai, AI Dungeon, Unity Copilot).

E‑commerce : Context‑aware, scenario‑driven shopping assistants that combine product knowledge with personal preferences.

Social & Community : Niche professional networks built around proprietary models (e.g., specialised matchmaking, knowledge‑sharing).

Communication : AI‑enhanced meeting summarisation, action‑item extraction, and real‑time translation.

Healthcare : Augmented clinicians, AI‑driven diagnostics, and robotic assistance; the Chinese market offers scale advantages.

Education : Personalised, one‑on‑one tutoring powered by LLMs, enabling universal access to high‑quality instruction.

Developer tools : Natural‑language programming, code generation, and AI‑assisted debugging (e.g., GitHub Copilot, Microsoft Designer, Adobe Firefly).

Design : AI‑generated visual assets, 3‑D models, and rapid prototyping across graphic, industrial, and architectural design.

Research : AI‑augmented scientific discovery, hypothesis generation, and automated experimentation pipelines.

Enterprise software (ERP, OA) : Conversational interfaces for workflow automation, data entry, and decision support, moving from static rule‑based systems to adaptive LLM‑driven assistants.

Manufacturing & Production : Closed‑loop integration of LLMs with robotics, PLCs, and IoT for flexible, model‑guided factories.

Smart Cities & Infrastructure : Unified, language‑driven control of urban services, transportation, and utilities.

Energy, Life Sciences, Materials, Space : Model‑centric R&D pipelines that accelerate discovery, design, and deployment of new fuels, proteins, alloys, and spacecraft.

New Execution Environment

The author warns of a widening “Matthew effect”: organisations with large compute, data, and capital will dominate, while smaller players must focus on niche verticals, proprietary data, or innovative alignment techniques. Internationalisation, multilingual support, and vertical‑specific data ownership become critical competitive levers.

Human Capital in the AI Era

Technical skill alone is becoming less differentiating. Success will hinge on:

Unique insight and the ability to articulate a compelling long‑term vision.

Perseverance to pursue that vision despite rapid technological change.

Leveraging AI as a “copilot” to amplify personal productivity.

OpenAI is presented as a case study: a small, vision‑driven team out‑performing larger incumbents by focusing on a bold hypothesis about general AI.

Science and Commercialisation Converge

Historical science funding (Vannevar Bush’s 1945 report) separated basic research (universities, national labs) from application (industry). The article argues that the fourth/fifth‑generation scientific paradigm now requires commercial data and compute, blurring that line. Start‑ups are increasingly performing foundational research, attracted by venture capital that rewards rapid, data‑driven breakthroughs.

China vs. United States

Both countries possess the four pillars needed for AI leadership—talent, capital, compute, and market scale. China’s advantage lies in its massive domestic data, manufacturing base, and government willingness to invest in compute infrastructure, while the U.S. leads in venture capital and open‑source ecosystems. The race to achieve GPT‑3.5‑level capability is framed as a decisive competitive milestone.

Conclusion

The digital trinity provides a lens to understand the ongoing AI inflection, its economic ramifications, and concrete pathways for entrepreneurs, investors, and policymakers. By aligning infrastructure, model development, and action systems, and by focusing on vertical‑specific opportunities, stakeholders can capture value in the emerging era where knowledge—encoded in large models—becomes the primary engine of productivity.

Code example

1. 信息子系统（subsystem of information)，体系必须从环境中获得信息。
2. 模型子系统（subsystem of model)，用模型对信息进行表达。它必须充分有效地表达信息，这种表达方式让它可以做推理、做分析、做规划。
3. 行动子系统（subsystem of action) ，根据推理和规划与环境互动，来达到这个复杂体系的目的。