Europe 2026
Stop Making Models Bigger, Make Them Behave — Kobie Crawford, Snorkel
Overview
This talk challenges the assumption that larger AI models are always superior, particularly for enterprise use cases requiring reliability and security. It presents research demonstrating that a smaller, 4-billion parameter model, when fine-tuned with Reinforcement Learning (RL) and high-quality data, can outperform a significantly larger 235-billion parameter model on a tool-use task for financial analysis. The core thesis is that focusing on improving model behavior and tool discipline through targeted data and training can yield substantial performance gains, often more effectively and efficiently than simply increasing model size.
Who should watch
- AI Engineers looking to improve model performance without solely relying on larger models.
- Product Managers evaluating cost-effective solutions for deploying AI in production.
- Builders seeking to enhance model reliability and tool-use capabilities.
- Researchers interested in data quality's impact on model behavior.
- Anyone facing challenges with model performance in constrained or security-sensitive enterprise environments.
Key takeaways
- Smaller models can outperform larger ones on specific tasks through targeted fine-tuning, especially when focusing on behavior and tool discipline.
- High-quality, expert-curated datasets are crucial for effective RL training, leading to significant performance improvements.
- Reinforcement Learning can be a cost-effective method for enhancing model capabilities, with training runs costing under $500.
- A 4-billion parameter model, after RL fine-tuning, successfully outperformed a 235-billion parameter model on a financial analysis tool-use task.
- The fine-tuned smaller model demonstrated superior tool usage, including discovering available tools, inspecting schemas, and self-correcting errors, unlike the larger model which hallucinated an answer.
- Focusing on tool discipline and error correction proved more impactful than general reasoning improvements for tool-use tasks.
- Training with single-table questions, even for a benchmark with multi-table questions, yielded the greatest performance uplift, suggesting that mastering fundamental tool interactions is key.
- Developing detailed rubrics for evaluation can help identify specific behavioral failure modes, guiding the creation of targeted datasets for improvement.
Notable quotes
*The point broadly speaking is that sometimes we find great wins to be had with the right data applied to the right problem statement.*
*Sometimes the idea is to find the specific behavior that's really the problem.*
Unofficial community note. Prefer the recording for nuance.