I observe the software engineering landscape shift as Poolside approaches a $3 billion valuation. Jason Warner, the former chief technology officer of GitHub, leads the Paris-based startup alongside co-founder Eiso Kant. Eiso Kant previously launched Athenian, a developer tooling startup that the Linux Foundation acquired in 2023. The company moved its headquarters to Paris last August and raised $126 million in initial funding. The company engages in discussions for a $400 million round with Bain Capital Ventures and DST. Redpoint and Felicis also participated in previous funding. I find the malibu and point foundation models to be the center of their development. The malibu model handles complex engineering challenges such as complex code generation and documentation. The point model provides rapid code completion and suggests design patterns through context awareness. The company intends to give its models an edge over other AI systems that target programming use cases through a training approach it calls reinforcement learning from code execution feedback. This technique allows models to train on more data by exposing them to real-world code bases and providing feedback based on code execution. This process currently powers one million container images and allows 10,000 code executions per minute on AWS. The company aims to eventually train models that allow business users to create software.
AWS infrastructure and security
The partnership between AWS and Poolside makes the models a first-party offering in Amazon Bedrock and EC2. This collaboration allows enterprise customers to integrate AI software development solutions into their AWS environments. Amazon Bedrock provides developers access to high-performing models through a single API as a fully-managed service. Enterprise customers can customize the generative AI Assistant for software development with their own data to meet security and privacy concerns. AWS Trainium chips power the inference for malibu and point to improve price performance. I find the security benefits of this integration to be the most important factor for regulated industries. AWS ensures that customers deploy models where their data stays within their firewalls. No data shares back with Poolside.
| Feature | Capability |
|---|---|
| malibu | Complex engineering and documentation generation |
| point | Rapid code completion and algorithm optimization |
| RLCEF | Synthetic data generation via code execution |
| AWS Trainium | High-performance inference for foundation models |
The company also rents AI training infrastructure, such as a cluster of 504 H100 graphics processing units from Iren. These tools allow the models to improve by completing millions of tasks in tens of thousands of real-world software projects. I would skip the hype around the scale of training if the models cannot effectively handle the proprietary knowledge of large enterprises.
GitHub Copilot model retirement
GitHub instructed users on October 2, 2026, to move to supported models after it retired certain AI models in Copilot. This change affects Copilot Chat, inline edits, ask and agent modes, and code completions. Enterprise admins must allow other models through the model policies in Copilot settings to ensure developers see alternatives in the model picker. You should check if your team’s specific integrations rely on a model that is no longer supported.
The mid-September updates to GitHub Copilot aim to make the tool more enterprise-ready. One update adds three tiers to Auto model selection: Efficiency, Balance, and Intelligence. Efficiency favors speed for quick edits, while Intelligence prioritizes reasoning for complex debugging. Copilot also suggests allowed values for custom properties, such as environment, data sensitivity, team ownership, or compliance tier, to maintain consistent repository metadata. I find the removal of the CodeQL prerequisite for AI Scan to be a significant improvement for security teams. This change allows wider coverage across repositories that were previously ineligible for vulnerability detection on pull requests. The Copilot usage metrics API now reports agentic CLI customizations, such as custom agents, skills, MCP servers, slash commands, and plugins.
| Copilot Update | Function |
|---|---|
| Auto Model Selection | Efficiency, Balance, and Intelligence tiers |
| AI Scan | Vulnerability detection without CodeQL |
| Usage Metrics API | Telemetry for custom agents and MCP servers |
| Impact Dashboard | Feature engagement over a 28-day window |
The Impact Dashboard shows feature engagement over a rolling 28-day window to help leaders find adoption patterns or enablement gaps. Code review improvements also help by providing clearer findings and smarter auto-resolution for addressed issues. Does the move to these new models resolve the long-standing issue of developer fatigue during code reviews? Copilot now writes the commit message when eligible suggestions are applied in batches.
