Insights into AI R&D Management, Cost Efficiency, and Education Solutions at New Oriental AI Research Institute

Preface

New Oriental AI Open Platform (https://ai.xdf.cn/) aims to combine education and AI innovation, offering high‑quality AI capabilities such as speech, OCR, face/body detection, and natural language understanding at the lowest industry price to help small and medium companies innovate cost‑effectively, and invites feedback.

Since August 2020, with minimal manpower and R&D investment, New Oriental built a leading full‑category AI capability for education within a year and delivered many AI education solutions; this article condenses the author’s reflections on that passionate entrepreneurial journey, providing broad, thought‑provoking insights for AI practitioners.

AI Development and Challenges

There is a saying that technology is over‑estimated in the short term and under‑estimated in the long term.

The history of AI dates back to 1943; breakthroughs in the 1980s (Hopfield) and later work by Hinton led to AlexNet in 2012, followed by many models (ResNet 2015, BERT 2018) that accelerated industrial adoption and sparked optimism about AI’s future.

While AI performs well in perception, it still lags far behind human cognition; competitive advantage now lies more in deep business‑rule knowledge and massive, domain‑specific data than in the public algorithms themselves.

Humans can infer correct conclusions from sparse clues, whereas neural networks remain black‑box approximations that rely heavily on big data and compute, making 100% accuracy unattainable for many NP‑complete problems.

Beyond algorithms, data, and compute, success requires thorough business‑rule understanding, robust data engineering, strong organization, and efficient infrastructure to achieve lower cost, faster speed, and better algorithmic outcomes; after 2018 AI hype surged, many enterprises now view AI as over‑hyped in the short term, with real‑world deployment still facing significant hurdles.

Problem Analysis

Technological value typically passes three stages: theoretical breakthrough in academia, lab‑scale engineering, and finally industrial application that reduces cost and lowers entry barriers.

Software engineering’s 40‑year evolution—from elite‑only programming to mass‑trained developers and sophisticated management methods—parallels AI engineering, creating many ordinary roles that can be filled with simple training.

Historical examples (WPS, Foxmail) show that individual developers once built flagship products, whereas today similar products require large, coordinated teams.

Given this, AI engineering is likely to follow a similar 20‑year cycle, where early adopters gain advantage but later entrants must keep pace or fall behind.

Education + AI

New Oriental’s culture stresses “doing the right thing and doing it correctly,” emphasizing careful planning, risk anticipation, efficiency, cost reduction, and execution to make ideas happen.

Regardless of AI hype, the combination of education and AI is a correct long‑term direction; the article shares management practices, cost‑reduction methods, and platform ideas to help other AI teams achieve similar benefits.

Performance Management

What should KPIs or OKRs be for an AI R&D department?

The primary goal of any company is profit; AI R&D must serve business by either increasing revenue or reducing cost.

Key KPI considerations: ability of the algorithm team to support business needs, effectiveness of the solutions, and cost‑effectiveness compared with third‑party services.

AI teams should leverage unique company data and business rules; long‑term maintenance and iterative improvement are essential.

Case study: New Oriental’s self‑developed face‑attendance system costs ~850 CNY per unit versus ~4,000 CNY for a third‑party solution, yielding a cost advantage of tens of millions of RMB while maintaining higher accuracy and privacy.

R&D Innovation

AI is still a relatively new technology; many patents arise from applying AI to business. The institute focuses on applied research and engineering, leaving fundamental theory to non‑profit research institutions.

AI teams must proactively discover business opportunities, innovate at the product level, and collect massive high‑quality data (e.g., 1 million+ English speaking evaluations) to improve models while reducing manual labeling costs.

Cost Reduction and Efficiency

AI department costs fall into three categories: GPU training/inference servers, data, and personnel. While personnel costs are hard to cut, productivity can be boosted through system support, automated labeling, and DevOps‑driven GPU resource pooling.

Core KPI/OKR summary: Algorithms must be business‑driven. Drive product innovation to increase AI value creation. Generate business‑related patents. Continuously reduce cost and improve efficiency through supporting software systems.

After a year of comprehensive governance, New Oriental achieved >10× AI output per person and reduced server costs to one‑third of the original while increasing patent output.

AI Institute Organizational Management

The institute’s structure combines specialized technical teams with dedicated software systems, forming an AI middle‑platform that enables low‑coupling, high‑cohesion, and strong execution.

(1) Algorithm Development Team

Core team focuses on three AI domains—NLP, computer vision, and speech—conducting extensive experiments, data preparation, code development, paper reading, and model acceleration, while also handling API packaging, deployment, and external communication.

(2) Data Annotation Team

To lower manual labeling costs and protect data assets, the institute built its own AI data annotation system that leverages workflow decomposition and crowdsourcing.

(3) Software Development Team

Supports all algorithm‑related software, builds CI/CD pipelines, enables GPU resource pooling and elastic scaling via Docker, and ensures stable online services.

(4) Hardware Development Team

Develops AIOT management systems that provide standardized services such as device control, remote updates, and monitoring, serving as a foundation for all smart hardware projects.

(5) Research Management Team

Handles invention patent documentation and liaison with patent agencies, freeing engineers to focus on algorithm work.

(6) Testing Development Team

Performs cross‑validation of algorithms, ensuring models that excel in specific business scenarios also meet broader performance standards.

Team Collaboration

Collaboration diagram illustrating interactions among algorithm, data, software, hardware, and testing teams (image).

Education AI Solutions

Readers are invited to visit the New Oriental AI Open Platform for solution overviews and to try the AI services.

(1) Smart Classroom

Low‑cost (<3,000 CNY per room) smart classroom features include invisible face check‑in, AI supervision, and AI tutoring, supported by patented dual‑teacher AI devices.

(2) AI Dual‑Teacher

The institute’s dual‑teacher AI course delivers AI‑driven interactive recorded lessons for thousands of classes.

(3) Computer Vision Empowerment

Face recognition, pose detection, and other CV technologies are applied to attendance, behavior monitoring, handwriting recognition, and more.

(4) Intelligent Grading

AI‑based elementary formula grading outperforms mainstream domestic AI platforms.

(5) English Essay Correction

Handwritten English recognition meets the needs of essay grading and dictation, achieving top results in international competitions.

(6) TOEFL Speaking Evaluation

The algorithm quantifies speaking abilities according to ETS standards, pinpointing pronunciation issues across fine‑grained dimensions.

(7) AI Data Annotation System

The self‑built annotation system improves labeling quality and speed, safeguards data, and integrates with the AI open platform and model factory to automate the full training‑to‑deployment pipeline.

Additional promotional sections about community benefits, QR code for data‑science collection, and calls for likes and shares are included at the end of the original source.