Poolside: the French alternative to GitHub Copilot
Based in Paris, the start-up Poolside presents itself as an innovative response to GitHub Copilot. Its launch in May 2023 in the United States was marked by a relocation to France, thereby strengthening its national anchorage. Founded by Jason Warner, former CTO of GitHub, and Eiso Kant, an expert in artificial intelligence for software engineering, Poolside develops specialized AI models for code, intended to be deployed on-premise.
The start-up adopts a unique model by offering customized AIs for each client while ensuring data security. As a sovereign alternative, Poolside is committed to respecting confidentiality, thus circumventing common practices that involve sending data to the cloud.
A solution focused on confidentiality
Poolside stands out with its radically different approach to data privacy. Unlike competitors such as GitHub Copilot, Tabnine, or CodeWhisperer, which require sending information to remote servers, Poolside installs directly in the clients’ environment. Jason Warner states: “Asking clients to give up their data in exchange for an AI solution is making a pact with the devil.”
This strategy ensures that the AI system operates exclusively on the companies’ data. Poolside’s models also exclude any code derived from GPL v2/v3 licenses. These licenses could compel users to share derived code in open source, thereby posing a legal risk for proprietary software development.
Training Poolside models
Jason Warner reveals that all players in the AI industry have access to the same data on the Internet. Poolside adopts a distinct approach by using “reinforcement learning via code execution feedback” to enrich its models. Rather than seeking new data, the company generates its own datasets.
This functional process relies on an iterative system: when faced with a programming problem, the model proposes several solutions which are then tested in real conditions. Eighty to ninety percent of the solutions may be incorrect, but each attempt enriches the training database.
This method should allow Poolside to produce between 1.5 and 2 billion new “tokens” of code in the next 18 months, adding to the 3 billion already available online.
Customized models for each client
Poolside also distinguishes itself by its ability to refine its models to suit the specific needs of each company. This customization occurs through a multi-layer fine-tuning process. The model first analyzes existing codebases, integrates technical documentation, and then adjusts to the daily practices of developers.
For now, Poolside focuses its efforts on complex environments suited to sectors such as finance, defense, and technology. A partnership with AWS enables companies to access Poolside services via EC2 and integrate the service into their infrastructures while benefiting from AWS security and governance standards.
Clients have the choice to run their models on Trainium or Nvidia chips. Poolside’s AI compatibility extends to most IDEs, such as VS Code and IntelliJ, with future integration prospects for Visual Studio and Eclipse.
Frequently asked questions about Poolside: the French solution to GitHub Copilot, emphasizing confidentiality
What is Poolside and how does it differ from other AI tools for code?
Poolside is a French start-up that develops artificial intelligence models specifically for code generation. Unlike tools like GitHub Copilot, Poolside installs locally in the user’s environment, ensuring that data does not leave the company and respects the confidentiality of sensitive information.
How does Poolside ensure the confidentiality of user data?
Poolside does not require sending data to the cloud to operate. Instead, the AI models run on the clients’ local servers, meaning that the company has full control over its data and can keep it private, with no risk of information leakage.
What types of licenses does Poolside avoid and why?
Poolside refuses to use code derived from GPL v2/v3 licenses, which require that any derived code must also be published as open source. This protects the legal interests of its clients, especially those in highly regulated sectors where proprietary code must remain confidential.
What training process does Poolside use for its AI models?
Poolside uses a reinforcement learning method through code execution feedback. This iterative process generates multiple solutions within a model that are then tested to evaluate their effectiveness, continuously enriching the training database with realistic results.
How can Poolside models adapt to the specific needs of a company?
Poolside’s AI models are designed to gradually adjust to the specifics of each company. This is done through a multi-layer fine-tuning process, incorporating elements like the existing codebase and the company’s technical documentation to create a unique and tailored model.
Which sectors benefit the most from using Poolside?
Poolside primarily targets demanding sectors like finance, defense, technology, retail, and system integrators. These fields require robust solutions compliant with security and regulatory standards, making Poolside an ideal solution.
Is Poolside compatible with common development environments?
Yes, Poolside is designed to be compatible with the majority of integrated development environments (IDEs) on the market, including Visual Studio Code, IntelliJ, and soon Visual Studio and Eclipse, thus facilitating its integration into developers’ workflows.
What types of solutions does Poolside offer developers?
Poolside offers a variety of solutions, ranging from code completion models to a conversational assistant, allowing developers to interact with the AI intuitively and effectively according to their specific needs.
