Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
In 2003, Clayton Christensen, a professor at Harvard Business School, proposed in The Innovator’s Solution: Disruptive innovation often makes complex products or services simpler, cheaper, and more accessible, allowing more people to participate… Over the next two decades, Internet pioneers have embraced this view, and it still holds true in the new AI wave.
For most people, technology evolution is “invisible” (like the iceberg underwater); they care about the “visible” part: what specific problems AI solves. Taobao is a typical player—integrating AI into every scenario to address user needs. In March this year, Taobao fully upgraded its AIGX technology system (covering AIGI, AIGR, AIGB, AIGA, AIGC, AIGD), making AI the “gene” of Taobao’s algorithms.
AI Era: Generative AI Drives Productivity Leap
Looking back, AI breakthroughs (ChatGPT, Deepseek, Sora 2) have accelerated rapidly. Unlike previous tech revolutions (PC/ mobile Internet) that focused on connection efficiency, the AI era’s biggest surprise is generative AI’s intergenerational productivity boost. Multimodal intelligence—enabling AI to understand/generate text, images, audio, video—has become the core, allowing natural human-AI collaboration.
Taobao has invested heavily in AI since 2023, developing multimodal, search/recommendation, and video generation models. Its AIGX system powers products like AI Universal Search, AI Try-On, and Advertising LMA Model.
Multimodal Intelligence: Taobao’s Core AI Domain
In July, Taobao launched RecGPT (a 10B-parameter generative recommendation model), the first in the industry to use LLM to “systematically” transform recommendation algorithms. Here’s how it differs from traditional recommendations:
Traditional vs. Generative Recommendation
Traditional Recommendation: Uses collaborative filtering/deep learning to predict clicks based on historical data. Efficient but limited by cold starts and “information cocoons.”
Generative Recommendation (RecGPT): Uses multimodal models + world knowledge to generate personalized recommendations. Creative, avoids data limits, and understands user needs better.
For example: If RecGPT detects a user group bought baby cribs, Stage 1 formula, and soothing toys since March 2024, it infers a newborn. When the baby turns 1, it recommends walkers/Stage 3 formula. During Double 11, it combines brand preferences to suggest winter kids’ clothing bundles. Another case: It recommended lunar eclipse observation gear to astronomy fans before the September 8 eclipse was widely known.
Taobao Star Video Generation Model 3.0
Taobao upgraded its video generation model to V3.0, using a compact 16x16x4 spatio-temporal VAE. It maintains fast inference while increasing DIT parameters, with better semantics, natural movements, and realistic visuals. The model will soon launch on Taobao’s e-commerce platforms.
TStars-Omni: Full Multimodal Model
Taobao’s TStars-Omni supports text/image/video/audio input and text/audio output. It enables deep product understanding—e.g., if a user asks, “Can I put this fridge in my kitchen?” (with fridge + kitchen floor plan images), TStars-Omni analyzes and answers with suggestions.
AI Agent: iFlow CLI
iFlow CLI follows “one kernel, multiple applications”: Programmers use it as a terminal/IDE plugin; business developers integrate via SDK. It’s open, secure, and free for personal users, supporting scenarios like ad creativity, academic writing, and flowchart design.
Technology Openness & Ecosystem Resonance
Taobao open-sourced two frameworks in 2025:
1. ROLL: A reinforcement learning framework for small-to-600B+ models, enhancing LLM performance in preference alignment and complex reasoning.
2. RecIS: A unified deep learning framework for large-scale sparse-dense computing, used in Alibaba’s advertising, recommendation, and search.
By sharing internal capabilities, Taobao aims to accelerate the industry’s move toward “Artificial Super Intelligence (ASI).” Based on current trends (AI complexity growing 5-10x yearly, error rates dropping 50% yearly), narrow AGI (outperforming 95% of humans in most tasks) may arrive in 5-10 years.