The setting
The setting
- Hardware: GPUs became the mainstay of AI compute, and cloud vendors released dedicated AI chips (such as Google's TPU, announced in 2016); phones grew powerful enough to run models locally.
- Networks: 4G reached full coverage, and 5G licenses were issued in 2019; short video and live streaming became the bulk of traffic.
- Capital: early on, capital chased AI and short video; later platform regulation tightened and the industry turned to cost control.
- Users: growth peaked, and user time became what every platform fought over; the 2020 pandemic moved work, education and government services online at scale (health codes, remote meetings).
- Policy: the GDPR (2018), the Digital Services Act (2022) and the AI Act (2024) turned compliance from optional into a precondition for launch.
The pressure
What needed to change
- User acquisition was expensive, so each user had to see more, stay longer and convert better.
- Once a monolithic app was split into hundreds of microservices, deploying, scaling and troubleshooting all became extremely complex.
- Every business line reinvented the wheel, and data was scattered across their separate systems ("data silos").
- Cloud systems run 24/7, and a single failure affects tens of millions of users, so staffing it by hand is unrealistic.
- With iOS, Android, Web and mini programs running in parallel, the same feature had to be built several times over.
- Large-scale data abuse and breaches (the 2016 Xu Yuyu case, the 2018 Facebook-Cambridge Analytica scandal) left users no longer trusting.
- Ransomware, IoT botnets and open-source component vulnerabilities broke out on a massive scale (2016 Mirai, 2017 WannaCry, 2021 Log4Shell).
- Deep learning models needed separately labeled data and separate training for every task, which was costly and generalized poorly.
The building blocks
The building blocks
- Recommendation algorithms and deep learning (CTR prediction, short-video recommendation feeds) and A/B testing -> pain point 1: content "finds" the user, and user time became the core metric.
- Cloud native: Kubernetes, service mesh, serverless and observability (logs, metrics, distributed tracing) -> pain point 2.
- The middle-platform strategy, data warehouses/data lakes and unified data platforms -> pain point 3.
- SRE (site reliability engineering; Google published the SRE book in 2016), error budgets and chaos engineering -> pain point 4: managing reliability with an engineering approach.
- Cross-platform frameworks (mini programs in 2017, Flutter in 2018), design systems, component libraries and TypeScript -> pain point 5: one design and one set of components, reused across platforms.
- Privacy compliance engineering: data grading and classification, data minimization, privacy policies and consent dialogs, and personal information impact assessments; plus "usable but invisible" data techniques such as federated learning and differential privacy -> pain point 6. At the legal level, three data laws responded.
- Zero-trust architecture, software supply chain security (dependency scanning, SBOM), and MLPS 2.0 covering cloud computing, IoT and mobile internet -> pain point 7.
- Pretrained models: the Transformer (2017), BERT (2018) and GPT-3 (2020), together with scaling laws -> pain point 8: pretrain one general model on huge amounts of data, then adapt it to various tasks. This led directly into the next era.
The people
The people
Ops → SRE & platform engineeringDevOps Engineer
Setting SLOs, managing Kubernetes clusters, building monitoring and alerting, and running on-call and incident reviews. The ops role page at `/roles/devops/` marks the "Ops → SRE" evolution.
FrontendFrontend Engineer
Engineering tooling (Webpack, TypeScript), "big frontend" (mini programs, cross-platform), the BFF layer, and the development of component libraries and design systems.
BackendBackend Engineer
Microservices and cloud-native development, increasingly wired into data compliance requirements (masking, permissions, auditing).
ProductProduct Manager
Data-driven product managers looking at funnels, running experiments and driving growth; the "strategy product manager" emerged.
DesignInteraction Designer
Design systems, design tokens, co-building component libraries with engineering, and measuring experience.
SecuritySecurity Engineer
Expanding from offense and defense into compliance. New data security and privacy compliance roles appeared (the DPO under GDPR, the person in charge of personal information protection domestically), working on zero trust and supply chain security.
Full-stackFull-stack Engineer
Product teams owned the interface, the API and the deploy pipeline, with cloud services replacing much of the server room.
Trust & safety
Trust & safety
Selected moments
Selected moments
AlphaGo beat Lee Sedol 4:1
Sources: en.wikipedia.org
The Cybersecurity Law was passed (in effect 2017-06-01)
Sources: flk.npc.gov.cn
The WannaCry ransomware outbreak
Sources: en.wikipedia.org
"Attention Is All You Need" was published, introducing the Transformer
Sources: arxiv.org
The Spectre/Meltdown vulnerabilities were disclosed
Sources: arxiv.org
The EU's GDPR began to apply
Sources: eur-lex.europa.eu
The MLPS 2.0 national standards took effect
Sources: openstd.samr.gov.cn
The GPT-3 paper was published
Sources: arxiv.org
The Data Security Law and the Personal Information Protection Law were passed
Sources: flk.npc.gov.cn · flk.npc.gov.cn
The Log4Shell vulnerability was disclosed
Sources: en.wikipedia.org
The Provisions on the Administration of Algorithmic Recommendation of Internet Information Services took effect
Sources: www.cac.gov.cn
OpenAI released ChatGPT
Sources: openai.com
The EU Digital Services Act is adopted
Sources: eur-lex.europa.eu
Looking forward
Looking forward
Every business had to train its own model, and algorithms served only recommendation and recognition; at the same time, pretraining plus scaling laws proved that "one large enough model can do many things". At the end of 2022, ChatGPT let ordinary people use natural language to call on that general capability directly for the first time, and the industry's question shifted from "how do we train a specialized model" to "how do we use large models reliably, cheaply and compliantly".
Dates and scope
Era boundaries
- Boundary start
- 2016 AlphaGo, cloud-native platforms and the GDPR debate; the traffic dividend peaked
- Boundary end
- 2022-11-30 ChatGPT is released; from 2023 large models reshape the industry