The Model Is the Product | Vintage Data

1. The Next Stage of AI Development: The Model Is the Product

There has been much speculation about the next cycle of AI development over the past few years. Agents? Reasoners? True multimodality? But now it seems time to state clearly: "The model is the product." All elements of current research and market development are heading in this direction.

2. Key Trends in Current AI Development

2.1 The Scaling Limits of General-Purpose Models

• The release of GPT-4.5 illustrates this well. "Performance increases linearly, but computing costs are growing exponentially." Even OpenAI cannot deploy this massive model at a reasonable price.

2.2 The Success of Opinionated Training

• The combination of reinforcement learning and reasoning has proven far more effective than expected at teaching models specific tasks. For example, even small models can solve math problems remarkably well, and coding models have moved beyond simply generating code to "managing entire codebases." Anthropic's Claude can "play Pokemon games with very limited context information and without dedicated training."

2.3 The Rapid Decline of Inference Costs

• Thanks to DeepSeek's optimizations, the cost has dropped so much that currently available GPUs could process 10,000 tokens per day for every person on earth. "The token-selling economy is no longer profitable for model providers." They must now move up the value chain.

3. The Uncomfortable Truth About Models Becoming Products

This direction reveals an uncomfortable truth for investors. "In the next phase of AI, the application layer is likely to be the first to be automated and collapse." This means investment strategies that relied solely on the application layer could be undermined.

4. New Model Forms: DeepResearch and Claude Sonnet 3.7

4.1 DeepResearch: A New Research Language Model

OpenAI's DeepResearch differs from conventional LLMs or chatbots. This model has learned browsing capabilities such as "searching, clicking, scrolling, and file interpretation," enabling it to "comprehensively analyze various websites and find specific information or write comprehensive reports."

"The model learned these capabilities through reinforcement learning on browsing tasks."

4.2 Claude Sonnet 3.7: A Code-Centric Model

Claude 3.7 was trained with complex code use cases in mind. This model replaces existing workflows and can "manage entire processes, not just code generation."

5. The Transition from Workflows to Agents

Anthropic defined agent models as follows:

"Agents perform tasks internally, dynamically directing their own processes and tool usage, and controlling how they perform tasks themselves." Many startups currently claim to be developing agents, but in reality they are only creating "workflows that orchestrate LLMs and tools through predefined code paths."

6. The Shift of Complexity: Training Is Key

Model training now anticipates more tasks and edge cases in advance, simplifying deployment. As a result, "value is created at the model training stage, and ultimately model trainers will capture most of it."

6.1 Automation of RAG Systems

Current RAG systems consist of multiple fragile workflows. However, new training techniques may consolidate them into two models, one handling data preparation and another handling retrieval and report generation. This process requires "sophisticated synthetic pipelines and new reinforcement learning reward functions."

7. Changes for Model Providers: From API to UI

Databricks' Naveen Rao said:

"All closed-source AI model providers will stop selling APIs within the next 2-3 years. Only open models will be offered through APIs." Closed-source model providers now aim to offer not just models but "applications with purpose-built UIs."

8. Problems in the Investment Landscape

The current AI investment landscape is based on the following assumptions:

• The application layer will create the most value independently of the model layer.
• Model providers will sell tokens at increasingly lower prices.
• Wrapping closed-source models will meet all demand.
• Building training capabilities is a waste of time.

But these assumptions fail to properly account for recent technological advances, particularly reinforcement learning (RL). "Model training is one of the areas with the greatest potential in the market, but the investment landscape is failing to properly evaluate it."

9. Future Outlook: The Importance of Model Training

Large research labs like OpenAI are now more likely to choose "partners to collaborate with at the training stage" rather than API customers. DeepSeek's founder Lian Wenfeng said:

"The current stage is not an explosion of applications but an explosion of technological innovation." "Once a complete industrial ecosystem is formed, we won't need to build applications ourselves."

10. Conclusion: The Start of a New War

Many Western companies are currently "preparing for the next war with strategies from the last war." In the new paradigm where the model is the product, small teams with training capabilities are likely to play an important role in the future AI market. "Model training will become the core competitive advantage in AI."