The AI stack expanded from model choice to application plumbing
As Copilot added another model, multi-turn training, GraphQL response generation and Kubeflow workflows also matured, moving competition beyond models and into the toolchain that puts them to work.
Artificial Intelligence··Midday
Grok 4.6 enters the Copilot picker
GitHub has made Grok 4.6, xAI's reasoning model, available in Copilot. It will appear in the model picker in Visual Studio Code, Visual Studio, Copilot CLI, the Copilot cloud agent, the Copilot app, JetBrains, Xcode and Eclipse. Access opens on the Copilot Pro, Pro+, Max, Business and Enterprise plans, with a gradual rollout. Business and Enterprise administrators must enable a separate Grok 4.6 policy in Copilot settings, off by default. Usage is billed per request at the provider's list price. GitHub says its internal testing found the model strong on terminal-based coding work in Visual Studio Code and Copilot CLI and suited to longer-horizon tasks that need sustained reasoning and tool use; that assessment is the company's own, not an independent measurement. Expanding the model menu shows that choice in the developer environment is tied as much to management and billing settings as to the product surface.[1]
Multi-turn rewards and GraphQL at request time
Amazon has made the serverless multi-turn reinforcement learning option in Nova Forge generally available, so teams can define multi-turn reward functions without running the training environment's infrastructure themselves. In the accompanying work, Amazon Nova Lite 2.0 is trained on 500 distinct programming tasks with a four-component weighted reward: correctness at 1.0, asking before coding at 0.6, a penalty for guessing immediately at 0.4, and a loop penalty at 0.2. Authors Maria Masood, Laurent Mombaerts and Nick Biso show that an aggregate reward curve can improve while individual components carry no learning signal at all: the asking behaviour rose from 34 per cent to 96 per cent while the correctness scorer returned the same value on every rollout, a problem caught by tracking each component's within-group standard deviation. The same day Expedia Group open-sourced mockql-rs, a Rust command-line tool that generates GraphQL mock responses from a language model at request time. It is the third public attempt at the same problem in six months, after Airbnb's @generateMock directive in April and a proposal opened at the GraphQL Foundation in February. According to Expedia's Samuel Vazquez, models are poor at inventing shapes and good at filling in a shape they are given; mockql-rs sits as a separate process between client and server, validates @mock fields against the schema with apollo-compiler, forwards real fields upstream, and merges generated and live data into one response.[2], [3]
From Kubeflow workflow to the production path
The Kubeflow project has announced updates for distributed AI and high-performance computing work on Kubernetes. Chief among them is Kale 2.0, which turns annotated Jupyter notebooks into production-ready pipelines; Kale 2.0 supports the second-generation Kubeflow Pipelines architecture and removes the need to author pipelines by hand, shortening the path from experiment to production for data scientists. The project is also nearing the second version of Kubeflow Notebooks, redesigned from the ground up around a declarative custom-resource architecture that gives platform teams templated control over JupyterLab and Visual Studio Code environments on Kubernetes; an alpha is available. Native Spark support in the software development kit lets users run Spark on Kubernetes without writing infrastructure configuration. The project is advancing toward graduation at the Cloud Native Computing Foundation. From the model picker to the reward function, from request-time GraphQL generation to the notebook-to-pipeline bridge, these four developments show competition concentrating not only on which model is chosen but on how it is trained, which API response it fills, and which production path runs it.[4], [1], [2], [3]
