Baidu Commercial AI Competition Explained

Baidu Commercial AI Competition Explained | Generated by AI

Home 2026.07

Question: What’s CTI (百度商业智能大赛)?

Answer:

CTI here means 百度商业AI技术创新大赛 (Baidu Commercial AI Technology Innovation Competition), not the common meanings like “Computer Telephony Integration” or “Cyber Threat Intelligence”. (cti.baidu.com)

It is an AI competition organized by Baidu’s commercial AI team. The goal is to explore industrial AI + business scenarios, especially recommendation, advertising, large-model optimization, and AI engineering talent. It started in 2023 and is open to universities, research institutes, and industry participants. (cti.baidu.com)

The interesting part is that it is not a pure Kaggle-style accuracy competition. It is closer to real-world AI engineering:

Typical tracks

1. Commercial prediction / recommendation

User behavior prediction
CTR (click-through rate) prediction
Conversion prediction
Advertising ranking models

Example:

user history
     |
     v
embedding + features
     |
     v
Transformer / DeepFM / DIN / DLRM
     |
     v
P(click | user, ad)

The goal is maximizing business metrics like CTR/CVR. (cti.baidu.com)

2. Large model inference optimization

This is closer to AI infra:

LLM model
   |
   +-- quantization
   +-- KV cache optimization
   +-- CUDA kernels
   +-- batching
   +-- memory optimization
   |
   v
lower latency / higher throughput

For example, the 2023 competition included AIGC inference optimization, focusing on improving large-model serving efficiency. (cti.baidu.com)

3. Generative recommendation / advertising

Recent competitions moved toward:

Transformer-based recommendation
Generative ranking
Long user behavior sequence modeling
High-performance inference

The 2026 competition focuses on “生成式推荐广告排序推理性能优化” (generative recommendation advertising ranking inference optimization). (cti.baidu.com)

For someone with your background (nanoGPT, GPT-2 training, GPU/ROCm experiments), the interesting part is probably not the traditional recommendation model, but:

AI Infra
    |
    +-- vLLM style serving
    +-- CUDA/ROCm optimization
    +-- quantization
    +-- distributed inference
    +-- model compression

This is actually close to what companies need:

A model that scores 1% better but costs 10x more GPU → bad
A model that is 95% accuracy but serves 100x cheaper → valuable

Commercial AI is often an optimization problem:

[ \text{Business Value} = \frac{\text{Model Quality} \times \text{Traffic}}{\text{Compute Cost} \times \text{Latency}} ]

The competition is basically training people to solve that equation. (cti.baidu.com)

References:

百度商业AI技术创新大赛官网 (cti.baidu.com)
2023 百度商业AI技术创新大赛介绍 (cti.baidu.com)

Back Donate