Thinking Machines与桥水的金融AI实验：当「专家判断」变成一种可训练资产

This article introduces Thinking Machines, a new AI startup founded by OpenAI's former CTO, and its ongoing financial AI experiment in partnership with Bridgewater Associates. The company focuses on converting corporate expert judgment into trainable AI assets. Key takeaways are as follows:

1. The company's flagship product Tinker is an enterprise-focused model training API. It eliminates the need for users to build complex GPU clusters or handle the engineering details of distributed training, enabling fine-tuning for a wide range of open-source large models and drastically lowering the barrier to entry for custom model development.

2. Experiment results show that even cutting-edge general-purpose large models, after prompt engineering optimization, achieve a top accuracy of less than 80% on financial information screening tasks. In comparison, the custom model fine-tuned via Tinker reached 84.7% accuracy, cut error rates by nearly 30%, and reduced inference costs to 1/14 of the original level.

3. A clear industry direction has emerged: the next wave of AI business value will come from building customized, dedicated AI for enterprises, converting accumulated tacit expert experience into scalable organizational capabilities.

This article outlines an emerging trend of AI adoption in enterprise services, offering clear insights for brands looking to integrate AI and build differentiated competitive advantages. Key takeaways are as follows:

1. The consumer and marketing space is currently facing massive information overload, and brand operations rely heavily on tacit expert judgment for tasks such as content value screening, user demand identification, and marketing asset rating. Brands can adapt the experiment's framework to convert this expert judgment into scalable AI capabilities, reduce labor spent on low-value work, and improve operational efficiency.

2. Brands do not need to blindly pursue top-tier general-purpose large models for AI adoption. Fine-tuning open-source models aligned with your brand's specific judgment standards and user preferences can deliver a more business-adapted, lower-cost solution. It also converts your brand's years of accumulated operational experience into exclusive AI assets to build differentiated competitive moats.

3. Tools like Tinker lower the barrier to custom model deployment, allowing brands of all sizes to test AI transformation without building an in-house specialized AI engineering team.

This article maps out new opportunity directions for sellers adopting AI-driven operations, alongside risk warnings and practical guidance. Key takeaways are as follows:

1. Opportunity: The AI industry is shifting from broad general-purpose model adoption to customized dedicated AI development. Sellers hold large volumes of tacit operational experience for tasks like product selection judgment, traffic content screening, customer inquiry triage, and supply chain information filtering. All of this can be converted into AI capabilities using the experiment's framework, drastically improving operational efficiency and freeing up human labor.

2. Practical implementation: Training APIs like Tinker drastically cut the technical and cost barriers to custom model development. Sellers do not need to build complex in-house GPU clusters or hire specialized AI engineering teams; they can control training logic and fine-tune business-specific models entirely via API, greatly simplifying deployment.

3. Risk warning: Do not blindly overestimate the capabilities of general-purpose large models. Their accuracy in vertical seller use cases often falls short of operational requirements. Custom fine-tuning informed by your own business's expert experience is required to deliver production-ready results.

This article offers clear insights for factories advancing digital and AI transformation, and outlines new business opportunity directions. Key takeaways are as follows:

1. Factory processes including production, design, and quality control have accumulated large volumes of tacit experience from senior engineers and skilled technicians. Historically, this experience was hard to replicate and relied on specific personnel. Factories can now adapt Bridgewater's experimental framework to convert this expert judgment into trainable, scalable AI model assets, accessible capability across the entire factory, reduce reliance on senior talent, and improve the stability and efficiency of production and design.

2. Factories do not need to invest heavily in building underlying AI infrastructure and in-house engineering teams to adopt AI. Using training APIs like Tinker, factories only need to focus on organizing their own experience data and designing training logic. Underlying engineering work such as GPU scheduling and distributed training is handled entirely by the API, drastically lowering the capital and technical barriers to AI transformation.

3. Future AI competition in manufacturing will center on differentiated dedicated AI. The earlier factories convert their accumulated experience and data into custom AI capabilities, the easier it will be to build new competitive moats.

This article clearly lays out the development trend of the enterprise AI service industry, identifies core customer pain points, and outlines a feasible solution direction. Key takeaways are as follows:

1. Industry trend: The core growth track for enterprise AI services in the next phase will no longer be general-purpose large model inference services. Instead, it will be helping enterprises convert internal proprietary data and expert experience into customized dedicated AI to build differentiated competitive capabilities, a market segment that will see rapid demand growth.

2. Core enterprise customer pain points: On one hand, building in-house AI training infrastructure carries very high technical and capital barriers. On the other, general-purpose large models fail to meet business accuracy requirements for vertical professional use cases, and full expert data annotation is prohibitively expensive, blocking widespread deployment.

3. Feasible solution reference: Providers can follow Tinker's model to launch training API services that open up custom training logic. The platform handles underlying engineering work including GPU scheduling, resource management, and distributed training, while leaving full training autonomy to clients. For data annotation, a workflow of non-expert pre-annotation + machine screening of disputed samples + expert review controls costs while improving data quality. This framework has already been validated effective through experiment.

This article reveals shifting demand from enterprise customers for AI platforms, and clarifies directions for AI platform operation, merchant acquisition, and risk mitigation. Key takeaways are as follows:

1. Shifting demand: Enterprise customer demand for AI platforms has evolved. Clients are no longer satisfied with generic large model services provided by platforms. A growing number of vertical industry enterprises need the ability to perform custom model fine-tuning, and want to autonomously control training logic without bearing the cost of developing and maintaining underlying infrastructure. This is a new demand opportunity platforms need to capture.

2. New direction for platform operation: Platforms can launch model training API services similar to Tinker, opening up custom permissions for training logic, loss functions, and data workflows, while the platform handles all underlying engineering work. This meets enterprise customization demands while opening new revenue streams. Platforms can also onboarding vertical enterprises and AI service teams around custom AI services to enrich platform ecosystem.

3. Risk mitigation: Platforms should not blindly allocate all resources to building ever-larger general-purpose large models. They need to recognize the growth of enterprise customization demand,布局差异化服务提前, and avoid getting trapped in the red ocean competition of general-purpose large models, missing out on the new growth opportunity of enterprise AI services.

This article presents the latest developments in AI industry deployment, raises new questions, summarizes new business models and practical methods, and offers high research value. Key takeaways are as follows:

1. New industry development: The AI industry has entered a new stage of development. Competition has shifted from parameter wars between general-purpose large models to building enterprise-exclusive differentiated AI. The core path is converting hard-to-articulate expert judgment within enterprises into trainable, productizable, scalable model assets, representing a new direction for AI deployment in high-skill vertical scenarios such as finance and manufacturing.

2. Core new finding: This research explicitly validates the structural limitations of general-purpose large models in professional use cases. Prompt engineering can only encode knowledge that experts can explicitly articulate, but the most valuable professional judgment often takes the form of tacit experience that experts cannot fully describe. This is the core reason for the poor performance of general-purpose large models in these scenarios.

3. New business model and research reference: This trend has spawned a new business model of enterprise-focused training APIs, and the experiment provides a validated set of effective training methods. Improvements including low-cost data cleaning mechanisms, interleaved batch processing, CISPO loss with asymmetric clipping, and strong teacher online policy distillation all provide reusable practical cases for research in related fields.