[1/2] Achieving Productivity Gains with AI-Based IDE Features: A Journey at Google (Category Whitepaper)
Last night I read Google’s new paper, dated 27 January 2026, on adding two AI features to its internal IDE: conventional code completion and code transformation, where you select code and describe the change you want. To make developers faster, the engineers had to solve latency, UX and suggestion-quality problems through experiments and data. They approached this as a product funnel, measuring losses at each step from model uncertainty, delays, irrelevant suggestions or low engagement. That helped them prioritise the factors worth improving.
Here are the two use cases.
1️⃣ AI code completion A transformer predicts code after the cursor using fill-in-the-middle, fine-tuned on Google developers’ actual edit histories. Adaptive caching reuses recent suggestions to reduce lag during fast typing. Speculative decoding uses a small predictor model alongside other latency optimisations. The model also receives more context: relevant nearby code from open files and their surroundings.
The reported effects were:
- Adaptive caching produced roughly 35% cache hits and reduced median latency by 9%, with p90 down 2%. Suggestion acceptance increased around 17%, and the fraction of ML-generated code rose 41%.
- Additional context increased acceptance by another 5% and FCML by 11%, although median latency rose 46%.
Overall acceptance reached roughly 45%. The model now generates around 28.7% of all code, or up to 70.6% when copy-paste is excluded. An accepted suggestion averages about 62 characters.
2️⃣ Transform Code: edits from a description Developers select code and give a textual instruction, such as “replace X with Y.” A Gemini model fine-tuned on internal code receives the file context, selection and request, then returns a diff. Key challenges were discoverability, displaying large changes clearly and producing good suggestions even for unfinished code.
Google made several improvements:
- A button shown when code is selected increased requests by 40% and new users by 64%.
- A simplified diff made review 7% faster and improved acceptance by 2.2%, or up to 4.5% for large changes.
- Fine-tuning on edits raised acceptance from roughly 55% to 63%.
Transform Code already delivers value: developers accept around 68% of proposed edits, with an average response time < 1 second.
In the next part, I will explain how the authors measured productivity changes from these tools. That part is particularly interesting.
#Engineering #AI #ML #Software #Bigtech #Productivity #Management #Leadership #Processes #Architecture