Large Model Era

The setting

Hardware: GPU compute became a strategic resource; export controls and energy costs pushed efficiency and alternative accelerators.
Networks and infrastructure: AI data centers and high-bandwidth clusters became part of national infrastructure.
Capital: capital concentrated on foundation models and AI applications; open-source models sharply lowered prices and changed the cost curve.
Users: natural language became a new way to interact, and ordinary users used conversation directly to write, search and program.
Policy: the EU AI Act (2024) and the Council of Europe framework (2024) set risk-based rules; content labeling and evaluation requirements expanded.

The pressure

What needed to change

Large models spout nonsense with a straight face (hallucinations), and they do not know a company's internal or latest information.
Training and inference are costly and high-latency, making it hard to launch at scale.
A model can only "talk", not "do": it cannot connect to business systems or complete multi-step tasks.
New attack surfaces: prompt injection, jailbreaks, and models being used to generate scams and fake content.
The copyright and personal information of training data, and the difficulty of telling real generated content from fake, mean both regulators and users demand traceability.
The way software roles work is being rewritten: code can be generated, so how do you define quality, review and responsibility?
Dependencies grow ever more concentrated: a problem at a single model vendor, in one open-source component, or in a single security-software update can cause widespread cascading failures (the 2024 xz backdoor, the CrowdStrike outage).

The building blocks

Retrieval-augmented generation (RAG), vector databases and knowledge bases -> pain point 1: retrieve trusted material first, then have the model answer based on that material and cite its sources.
Open-source models, quantization, LoRA/QLoRA fine-tuning, inference frameworks (such as vLLM's PagedAttention) and mixture-of-experts (MoE) architectures -> pain point 2: prices and barriers dropped dramatically.
Tool calling, AI agents and the Model Context Protocol (MCP, released in 2024) -> pain point 3: models can call APIs, operate browsers, and write and run code.
Alignment training (RLHF), red-teaming, guardrails (input/output filtering) and least-privilege tool authorization -> pain point 4.
A filing system, explicit and implicit labeling of generated content (in effect 2025-09-01), watermarking, and compliance review of training data -> pain point 5.
AI coding assistants and code agents, together with benchmark sets, automated testing and code review standards -> pain point 6.
Software supply chain security (SBOM, dependency signing), multi-model redundancy and canary releases -> pain point 7.

The people

FrontendFrontend Engineer
Using AI to generate and rewrite code, building conversational interfaces (streaming output, generative UI), and building operating interfaces for agents.
BackendBackend Engineer
Wrapping business capabilities into tools the model can call (API/MCP), and handling asynchronous long-running tasks, cost and rate limiting.
ProductProduct Manager
AI product managers who define "the boundary of model capability plus a fallback experience", using evaluations to replace part of the requirements document.
DesignInteraction Designer
AI-native interaction (conversation, multimodality, presenting uncertain results), with AI-generated assets entering the design workflow.
OpsDevOps Engineer
LLMOps and GPU cluster operations, responsible for model versions and for the stability and cost of inference services.
SecuritySecurity Engineer
AI security (prompt injection, jailbreaks, the model supply chain), content security and filing compliance, and labeling of generated content.
Full-stackFull-stack Engineer
Builds the interface, retrieval and tool-calling layer around a model, and owns the evaluation loop.

Trust & safety

GlobalThe Digital Services Act (2022) set platform duties for illegal content, transparency, and risk assessment. Source
GlobalThe EU AI Act (2024) introduced risk-based obligations for AI systems, including general-purpose models. Source

Selected moments

2023-01-10
The Provisions on the Administration of Deep Synthesis of Internet Information Services took effect
Sources: www.cac.gov.cn
2023-02-24
Meta released LLaMA, starting the open-source large model wave
Sources: arxiv.org
2023-03-14
OpenAI released GPT-4
Sources: arxiv.org
2023-07-10
The Interim Measures for the Administration of Generative Artificial Intelligence Services were issued (in effect 2023-08-15)
Sources: www.cac.gov.cn
2023-08-31
The first batch of domestic large models passed filing and opened to the public
Sources: www.news.cn
2023-10-25
The National Data Administration was unveiled
Sources: www.gov.cn
2024-02-15
OpenAI released the Sora video generation model
Sources: openai.com
2024-03-29
The xz Utils backdoor was discovered
Sources: en.wikipedia.org
2024-07-19
A CrowdStrike update triggered a global outage
Sources: en.wikipedia.org
2024-08-01
The EU's Artificial Intelligence Act entered into force
Sources: eur-lex.europa.eu
2024-11-25
Anthropic released the Model Context Protocol (MCP)
Sources: www.anthropic.com
2025-01-01
The Regulations on the Administration of Network Data Security took effect
Sources: flk.npc.gov.cn
2025-01-20
The second 2024 revision of the Measures for the Administration of Internet Information Services took effect
Sources: flk.npc.gov.cn
2025-01-20
DeepSeek-R1 was released, an open-source reasoning model that sharply lowered costs
Sources: arxiv.org
2025-09-01
The Measures for Labeling AI-Generated and Synthetic Content took effect
Sources: www.cac.gov.cn
2025-10-28
The Standing Committee of the National People's Congress passed the decision to amend the Cybersecurity Law (in effect 2026-01-01)
Sources: flk.npc.gov.cn · www.npc.gov.cn
2026-07-27
The EU's Regulation (EU) 2026/1744 entered into force, amending the Artificial Intelligence Act
Sources: eur-lex.europa.eu

Looking forward

This era is still underway. The unsolved pain points that may drive the next segmentation include: whether agents can be reliably authorized to "get things done on a person's behalf", the cost of compute and energy, the trustworthiness and copyright of AI-generated content, and on-device models and embodied intelligence (robots). The next era is not presumed for now.

Dates and scope

Era boundaries

Boundary start: 2023 GPT-4 and open-source models; the EU AI Act takes shape
Boundary end: to the present

The setting

Hardware: GPU compute became a strategic resource; export controls and energy costs pushed efficiency and alternative accelerators.

Networks and infrastructure: AI data centers and high-bandwidth clusters became part of national infrastructure.

Capital: capital concentrated on foundation models and AI applications; open-source models sharply lowered prices and changed the cost curve.

Users: natural language became a new way to interact, and ordinary users used conversation directly to write, search and program.

Policy: the EU AI Act (2024) and the Council of Europe framework (2024) set risk-based rules; content labeling and evaluation requirements expanded.

What needed to change

Large models spout nonsense with a straight face (hallucinations), and they do not know a company's internal or latest information.

Training and inference are costly and high-latency, making it hard to launch at scale.

A model can only "talk", not "do": it cannot connect to business systems or complete multi-step tasks.

New attack surfaces: prompt injection, jailbreaks, and models being used to generate scams and fake content.

The copyright and personal information of training data, and the difficulty of telling real generated content from fake, mean both regulators and users demand traceability.

The way software roles work is being rewritten: code can be generated, so how do you define quality, review and responsibility?

Dependencies grow ever more concentrated: a problem at a single model vendor, in one open-source component, or in a single security-software update can cause widespread cascading failures (the 2024 xz backdoor, the CrowdStrike outage).

The building blocks

Retrieval-augmented generation (RAG), vector databases and knowledge bases -> pain point 1: retrieve trusted material first, then have the model answer based on that material and cite its sources.

Open-source models, quantization, LoRA/QLoRA fine-tuning, inference frameworks (such as vLLM's PagedAttention) and mixture-of-experts (MoE) architectures -> pain point 2: prices and barriers dropped dramatically.

Tool calling, AI agents and the Model Context Protocol (MCP, released in 2024) -> pain point 3: models can call APIs, operate browsers, and write and run code.

Alignment training (RLHF), red-teaming, guardrails (input/output filtering) and least-privilege tool authorization -> pain point 4.

A filing system, explicit and implicit labeling of generated content (in effect 2025-09-01), watermarking, and compliance review of training data -> pain point 5.

AI coding assistants and code agents, together with benchmark sets, automated testing and code review standards -> pain point 6.

Software supply chain security (SBOM, dependency signing), multi-model redundancy and canary releases -> pain point 7.

The people

FrontendFrontend Engineer

Using AI to generate and rewrite code, building conversational interfaces (streaming output, generative UI), and building operating interfaces for agents.

BackendBackend Engineer

Wrapping business capabilities into tools the model can call (API/MCP), and handling asynchronous long-running tasks, cost and rate limiting.

ProductProduct Manager

AI product managers who define "the boundary of model capability plus a fallback experience", using evaluations to replace part of the requirements document.

DesignInteraction Designer

AI-native interaction (conversation, multimodality, presenting uncertain results), with AI-generated assets entering the design workflow.

OpsDevOps Engineer

LLMOps and GPU cluster operations, responsible for model versions and for the stability and cost of inference services.

SecuritySecurity Engineer

AI security (prompt injection, jailbreaks, the model supply chain), content security and filing compliance, and labeling of generated content.

Full-stackFull-stack Engineer

Builds the interface, retrieval and tool-calling layer around a model, and owns the evaluation loop.

Selected moments

2023-01-10

The Provisions on the Administration of Deep Synthesis of Internet Information Services took effect

Sources: www.cac.gov.cn