The Era of Vertical AI Models
Summary
Intercom recently announced its new customer service-focused AI model, Finn Apex, which CEO Eoin Mac Caba claims outperforms GPT-4 and Opus 4.5 in performance, speed, and cost for customer service tasks. This development challenges the "bitter lesson" in AI, which posits that general methods leveraging computation consistently outperform specialized, human-knowledge-encoded systems. While past efforts like Bloomberg GPT failed to beat general models, Intercom's success, alongside Cursor's Composer 2 model for coding, suggests that "last-mile usage data" and post-training on open-source base models can create domain-specific models that rival or exceed frontier general models. This shift implies significant business model implications, potentially reducing reliance on API-based general models and fostering a "full-stack" approach where companies develop their own specialized AI layers.
Key takeaway
For CTOs and entrepreneurs evaluating AI strategy, Intercom's Finn Apex and Cursor's Composer 2 demonstrate that specialized vertical models, built on open-source foundations with proprietary post-training, can now surpass general frontier models in specific domains. You should assess your organization's unique "last-mile usage data" as a potential asset for developing cost-effective, high-performing custom AI solutions, rather than solely relying on API-based general models. This shift could redefine competitive advantage in AI-driven products.
Key insights
Domain-specific AI models, enhanced by post-training on last-mile usage data, can now outperform general-purpose frontier models.
Principles
- General methods leveraging computation are ultimately most effective.
- Post-training on experiential data can close performance gaps.
- Durable differentiation will move to the model layer.
Method
Take a strong open-source base model and apply extensive reinforcement learning and post-training using proprietary, domain-specific interaction data to achieve superior specialized performance.
In practice
- Explore post-training open models with proprietary interaction data.
- Evaluate in-house model development for domain-specific tasks.
- Consider acquiring companies with strong domain-specific evaluation data.
Topics
- Vertical AI Models
- Intercom Apex
- Bitter Lesson
- Last-Mile Usage Data
- Post-Training
Best for: Investor, Entrepreneur, CTO, Director of AI/ML, AI Architect, AI Product Manager
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by The AI Daily Brief: Artificial Intelligence News.