GitHub Copilot: Coding Will Never Be The Same Again • Ryan J. Salva • YOW! 2023

Interesting. speech from Ryan J. SalvaVP of Product in Github, which talks about how Copilot works, and accompanies it all with several demos. The main theme of the talk is how Copilot helps engineers improve their productivity and change the way they work with code. The main thoughts of the speech are as follows: Copilot is able to make hints of code that may be useful to you, based on the context Contest is usually given by the prompt you write first as a natural language comment. In addition, the user has the opportunity to communicate with the assistant in the chat, asking questions on the code base or asking to generate some piece of code. The architecture of the system on the fingers consists of three parts: code editor> proxy <-> model Context is drawn from the entire editor. (The author shows the example of VS Code.). Other open tabs are also used for context. Different models are used for different purposes - conventionally low latency is needed for hints, so the model is simpler and faster. (GPT-3.5), and chat is acceptable for some waiting time, so you can use the model more powerful and slower. (GPT-4) In principle, in chat, you can ask Copilot to generate a response using the data of your entire project. (workspace) There is a possibility of using RAG. (retrieval augmented generation)when additional information from external sources is mixed with user questions to enhance copilot’s capabilities Example with documentation on an internal design system for creating interfaces Example of information from observability platforms (splunk, datadog, ...) for linking conditional failure, recent changes and authors' names (to connect them to the incident) Next, the author talks about fine tunning models, for example, to make a bias towards certain styles, API or SDK versions, programming languages. Here the author says that the model is trained on the feedback that is available to it. (accepted and rejected suggestions, code that passed the review code on the merge request and code that resulted in the review code being deleted, and so on) The guys at Github did a lot of experimentation on the effects of using LLMs, and the results showed significant effects. I particularly liked the study, where the quality of MRs was assessed blindly by engineers when they didn’t know whether or not they were using MRs. And this blind assessment showed significantly higher MR scores in those who used LLMs. It is interesting that LLMs are more likely to boost beginner developers, and the continuing effect is less Next, the author talks about the economic effect that they calculated as a result of a study with the McKinsey team and there was a number. 1.5 a trillion dollars over the next three years (Advisers like to draw big numbers.)

• Well, it ends with the fact that AI will not take away work and engineers, but rather relieve them of routine and in the future for engineers will be important creativity and system thinking, which will allow you to solve more complex problems, but learn the syntax of the language will have less:)

#AI #Software #Leadership #Future #ML #DataScience #Productivity #Engineering #Processes