竞争与市场管理局：2023年人工智能基础模型初步报告（英文版）（127页）.pdf

1、AI Foundation Models:Initial Report18 September 2023 2 1.Introduction.5 How we conducted the review.6 The structure of this report.6 2.Background.8 What are FMs?.8 How are FMs developed?.10 Data preprocessing and architecture design.10 Pre-training.11 Fine-tuning.11 Computing power.12 Deployment,rou

2、tes to market and monetisation strategies.14 AI supply chains and vertical relationships.16 How FMs are evaluated.18 The FM landscape.20 Potential application of FMs.25 3.Competition and barriers to entry in the development of FMs.27 Introduction.27 Data requirements.28 Pre-training.28 Fine-tuning.3

3、0 Alignment.30 Domain-or task-specific fine-tuning.32 Synthetic data.32 Computational resources.33 Pre-training.34 Fine-tuning.36 Inference.37 Technical expertise.38 Access to funding.39 Open-source models.40 Pre-training.40 Fine-tuning.41 Uncertainties.41 Will access to proprietary data become nece

4、ssary to compete?.42 Will models become larger?.44 Will FMs be highly generalised?.46 Is cutting edge performance required to compete?.47 Will large technology companies and first movers have an advantage over others?.48 Will open-source models remain a key part of the market?.50 Conclusion.52 4.The

5、 impact of FMs on competition in other markets.54 Introduction.54 Deploying FMs in downstream markets.55 3 FMs could become an important input in a wide range of markets.55 Firms can access FMs in a number of ways.55 Firms are monetising FM services in different ways.57 Potential impact of FMs on co

6、mpetition in downstream markets,including potential risks from vertical integration.58 How FMs could drive competition and disrupt incumbent firms.58 Different types of vertical integration and partnerships.66 Features of downstream markets that could affect competition in upstream FM development.67