April 15, 2024
Pinterest Engineering
Pinterest Engineering Blog

Sam Wang | Sr. Technical Program Supervisor; Joe Gordon | Sr. Employees Software program Engineer

At Pinterest we’re repeatedly on the lookout for methods to enhance our developer expertise, and we’ve not too long ago shipped AI-assisted growth for everybody whereas balancing security, safety, and price. On this weblog put up, we share our journey of unlocking AI-assisted growth, from the preliminary concept to the Basic Availability (GA) stage. Be part of us as we delve into the alternatives, challenges, and successes we encountered alongside the best way.

Like many corporations, we initially disallowed the usage of Giant Language Fashions (LLMs) till we totally evaluated their authorized and safety implications. Throughout that point, many engineers expressed curiosity in adopting AI-assisted growth and started utilizing it for private tasks on the aspect, making them keen to make use of it at work as effectively.

To find out the true potential of AI-assisted growth, we wanted to judge the impression and advantages whereas additionally figuring out and addressing any vital dangers and considerations related to its implementation.

The primary determination we needed to make was construct versus purchase. Whereas Pinterest possesses intensive in-house AI experience and builds a lot of our developer instruments, we acknowledged growing every part from scratch was not important to our core enterprise. Opting to purchase a vendor answer allowed us to expedite this course of and supply our engineers with a elegant expertise with loads of nice Built-in Improvement Atmosphere (IDE) integration. After cautious consideration, we selected GitHub Copilot resulting from its characteristic set, sturdy LLM, and match with our present tooling ecosystem.

As with every new know-how, the adoption of AI-assisted growth comes with its fair proportion of dangers and considerations. Addressing our considerations and dangers required working cross functionally with quite a few groups all through the corporate. The agility of the Pinterest Engineering crew was actually on show as we had been in a position to scrappily pull collectively engineers from a number of groups exterior of a daily planning cycle to execute. Throughout each planning course of we at all times make certain to put aside a while for unplanned objects, as we’ve discovered issues can transfer shortly and we can’t plan for every part prematurely.

We performed a trial program to collect each qualitative and quantitative suggestions on the usefulness of Copilot. Whereas many corporations ran trials of fewer than 30 folks over just some weeks, we determined to run our trial with round 200 builders over an extended length. This was achieved to incorporate builders within the journey and provides people a chance to strive one thing innovative even when we ended up getting into a unique route. This bigger cohort additionally allowed us to make sure we had vital populations throughout varied developer personas. Operating the trial over an extended length helped us management for the novelty impact and different measurement points. Of the 200 or so individuals about 50% used vscode, with many crew members utilizing jetbrains IDEs as effectively. The breadth of supported IDEs accelerated Copilot adoption.

To guage the trial we leveraged all our prior work on how to consider and measure engineering productiveness, and utilized it right here. We checked out each qualitative and quantitative information — and frolicked sampling real-time consumer suggestions. Qualitative sentiment suggestions was collected weekly by way of a brief slack bot based mostly survey; beforehand we seen that slack based mostly surveys have increased completion charges than e-mail based mostly surveys, so we wished to fulfill builders the place they spend extra time and scale back friction for them to share suggestions. Getting good qualitative measurements was barely extra advanced. Our strategy was to match the relative change over time for the trial cohort vs a management from previous to the Copilot trial. Operating the trial for longer than just some weeks helped us isolate exterior temporal influences like holidays and so forth.

In shut collaboration with our authorized crew, we ensured that our utilization of AI-assisted growth adhered to all related licensing phrases and laws. Moreover, in partnership with our safety crew, we performed an intensive evaluation of the safety implications posed by AI-assisted growth. We aimed to make sure that the code produced by Copilot remained inside our management and was not employed for coaching future LLM fashions.

Moreover, we positioned excessive precedence on stopping vulnerabilities in our codebase. Our safety crew leveraged vulnerability scanning instruments to repeatedly audit all code launched by each Copilot individuals and non-participants. This complete strategy enabled us to successfully mitigate potential dangers to our sturdy safety posture arising from AI-assisted growth practices amongst our engineers.

Increasing In direction of Basic Availability:

Determine 1: Copilot Sentiment Week over Week

Qualitatively, we used a brief web promoter rating survey to collect suggestions. Early NPS outcomes had been actually constructive (NPS of 75), and we watched these improve because the trial continued. Our quantitative information was equally spectacular supporting the suggestions we heard that Copilot was serving to our groups be extra productive. This overwhelmingly constructive suggestions included feedback corresponding to “Over time, Copilot has been giving higher options based on the work I’m doing.” and ‘“Copilot was notably helpful after I needed to make a change in Scala, a language I’m not accustomed to. Being acquainted sufficient with normal language ideas, I might let Copilot handle the syntax and nonetheless really feel assured that I understood its options.” Primarily based on this constructive suggestions we made the choice to increase entry to Copilot to all of engineering prematurely of our annual Pinterest Makeathon, which after all was very AI targeted this yr. Since our shifting to Basic Availability, to extend Copilot adoption we ran coaching periods, streamlined the method to get entry to Copilot by way of integration into our entry management and provisioning programs, and partnered with our platform groups to assist people perceive tips on how to greatest reap the benefits of Copilot in numerous domains corresponding to internet, API and cellular.

The impression of our efforts spoke for itself. In the end, we unlocked AI Assisted growth safely from concept to scaled availability in lower than 6 months, elevated consumer adoption by 150% in 2 months — with 35% of our whole developer inhabitants utilizing Copilot frequently. This implies based on the Technology adoption lifecycle we’re effectively into the early majority section of adoption.

Shifting ahead, we’re devoted to additional bettering the standard of Copilot options by incorporating fine-tuning with our Pinterest supply code, and persevering with to make sure that as our groups leverage these applied sciences to go sooner — we additionally accomplish that safely by not introducing extra bugs or incidents. We additionally know that that is only the start, with the speedy growth of AI Assisted developer instruments, we’re continuously evaluating new alternatives to construct, purchase and incorporate new applied sciences to drive enhancements to our developer expertise and improve developer productiveness — to attain our aim of enabling each developer at Pinterest to do their greatest work.


This work wouldn’t have been attainable with out an enormous group of individuals working collectively over the previous few months. We’d prefer to thank Shriman Gurram, Scott Hebert, Mark Molinaro, Amine Kamel, Andre Ruegg, Nichelle Carr, Roger Lim, Brandon Black, Kalpesh Dharwadkar, Orna Toolan and Anthony Suarez

Moreover we’d prefer to thank all our trial individuals for his or her help and suggestions.

To study extra about engineering at Pinterest, try the remainder of our Engineering Weblog and go to our Pinterest Labs website. To discover and apply to open roles, go to our Careers web page.