The Era of Vertical AI Models

By AI Daily Brief

Categories: AI

Summary

Intercom's vertical AI model Finn outperformed GPT-4 and Claude at customer service, challenging the 70-year "bitter lesson" that general models always dominate. The shift to post-training on proprietary data may finally enable specialized models to beat frontier models in specific domains.

Key Takeaways

1. Intercom's Finn model achieved highest performance, speed, and cost efficiency for customer service, beating GPT-4 and Claude Opus—suggesting vertical models can win in narrow domains where domain-specific data matters more than raw scale.
2. The bitter lesson (Rich Sutton, 2019) shows general methods with massive compute beat specialized models across chess, Go, speech, vision, and language—but post-training on last-mile user interaction data may change this equation for vertical AI.
3. Post-training on proprietary user interaction data ("last mile usage data") rather than pre-training from scratch is the new battleground—companies like Cursor can fine-tune open-weight models to exceed frontier models in specific domains without building from scratch.
4. Bloomberg GPT, a 50-billion parameter finance-specific model, was outperformed by general models, but 2026's vertical models succeed by taking state-of-the-art open models and adding domain-specific post-training rather than training from scratch.
5. The agent labs thesis predicts that if pre-training data limits are hit, post-training with proprietary data becomes the differentiator—shifting competitive advantage from model scale to data quality at the point of user interaction.

Topics

• Vertical AI Models
• Post-Training vs Pre-Training
• Last-Mile Usage Data
• The Bitter Lesson
• Domain-Specific Model Fine-Tuning

Transcript Excerpt

Today we are revisiting the bitter lesson and exploring whether we are now in the era of vertical AI models. Welcome back to the AI daily brief. I noticed this really interesting story yesterday where intercom announced that their new dedicated customer service focused model Finn had achieved something very significant. CEO Euan McCabe called it objectively the highest performing, fastest, and cheapest model for customer service, beating the very best models in the industry, including GPT 5.4 an...