We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
VS Code Blog·5 min read
How Prompt Tuning Improved GPT-5.5 in VS Code
TL;DR
A two-week experiment in VS Code tested prompt tuning for GPT-5.5, aiming to reduce exploration time and increase validation speed. By adjusting the system prompt, the team sought to enhance token efficiency, which directly impacts user costs.
✦ Why It Matters
Engineers can implement prompt tuning strategies to optimize AI model performance and reduce operational costs today.
Key Takeaways
- Experiment with prompt adjustments to improve AI model responses.
- Monitor token usage to identify areas for efficiency gains.
- Implement a structured approach to evidence, action, and validation.
- Collaborate with AI experts to refine model performance.
Full Summary
In a recent experiment, VS Code collaborated with OpenAI to optimize the GPT-5.5 model's performance through prompt tuning. The goal was to encourage the model to explore less and validate its actions sooner, thereby improving token efficiency—an important metric given usage-based billing.
The team identified that the model often spent excessive tokens on unnecessary exploration before making edits. By implementing two specific prompt adjustments, they measured the outcomes against a control group using live traffic data.
The results showed that the tuned model could deliver faster responses while reducing costs for users. This approach not only enhances user experience but also minimizes expenses associated with token usage in coding tasks.