[1/2] Achieving Productivity Gains with AI-based IDE features: A Journey at Google∗ (Category Whitepaper)

I read it at night. freshman (27 January 2026 year) From Google about their experience in implementing two AI functions in their internal IDE: standard code completion and more interesting code transformation. (where you select the code and then write a prompt on how to change). These features were designed to speed up the work of developers, but along the way, engineers had to solve problems with latency, UX and quality of clues, relying on experiments and data. In fact, the guys looked at the product process - built a funnel and estimated losses at each step due to various factors: model uncertainty, delays, irrelevant recommendations or low user engagement. This made it possible to systematically deal with these factors, understanding which ones should be emphasized more.

Let’s take a look at both scenarios.

1а AI-automation code At the heart of the auto-complement is a transformer trained to predict code just behind the cursor. (fill-in-the-middle)And the model is further trained on real stories of edits by Google developers (bringing her closer to a live script n). Adaptive caching is used to accelerate response and improve quality. (reuses recent clues, eliminating lag when quickly dialing)Speculative decoding (parallel launch of a small predictor model to reduce latency) and other latency optimizations. In addition, the model gets an extended context - code fragments from open files along with their environment. n (Relevant parts closest to the editing site).

Together, all these squats gave the following effects:

• Adaptive cache gave.35% of cash hits and reduced median delay 9% (p90 –2%), and acceptance rate tips increased by ~17By increasing the percentage of code generated by ML, 41%
• Adding context has given more5% acceptance and +11per cent (Even though the delay has increased, the median +46%)

Acceptance rate has reached ~45%, and ~28,7% of all code is now generated by the model (before 70,6% if you delete the copy-paste). The average piece of advice received is ~62 symbol.

2️⃣ Transform Code (description) This tool allows you to edit the selected code according to the text instruction at the request of the developer. (For example, “replace X with Y”). Under the hood is the Gemini model, trained on internal code; it receives the context of the file with the selected fragment and the text of the request, and issues changes in the diff format. The main difficulties of implementation - how to tell the user to use the function (discoverability)How to display large edits and how to ensure high quality of sentences even on incomplete code.

Google has addressed these issues with several improvements:

• Selection button (+40% of requests, +64Percentage of new users).
• Simplified diff. (rev 7%, acceptance +2.2%, up to +4.5% on major changes).
• Editorial training (Acceptance has grown ~55% → 63%).

Transform Code is already useful:68% of the proposed changes are accepted by the developers, and the average response 1 c.

In continuation I'm going to talk about how the kids were evaluating the productivity changes from these tools. (It's interesting.).

#Engineering #AI #ML #Software #Bigtech #Productivity #Management #Leadership #Processes #Architecture