AI Coding Assistants

Today's Stories

1. 1

   Disney ditches GitHub Copilot, adopts OpenAI's Codex

   Disney is dropping Microsoft's GitHub Copilot and two Amazon AI coding tools (Kiro and Q) in August in the US, while adopting OpenAI's Codex. The company will keep access to Anthropic's Claude Enterprise and Cursor. Eight Disney tech employees told Business Insider they rarely or never used GitHub Copilot; one said it produced "needlessly complex" code requiring cleanup, and another noted access had "lapsed for lack of use." The shift reflects Copilot's broader adoption struggles against rivals like ChatGPT and Claude, and signals which tools are winning with major enterprise customers.

   OpenAI's Codex is being prepared for adoption at Disney, though no timeline for introduction was given. Claude and Cursor have become the dominant tools among Disney staffers—one high-level engineer said their usage was "probably around 80% Claude at the terminal when coding."

2. 2

   6 LLM deployment formats compared: speed vs. hardware flexibility

   A technical breakdown of six production LLM deployment formats—Pickle, safetensors, GGUF, ONNX, MLX, and TensorRT—each with different tradeoffs between inference speed and hardware compatibility. Pickle and safetensors run anywhere Python runs; GGUF bundles weights and tokenizer for CPU-only machines; ONNX lets models switch frameworks; MLX optimizes for Apple silicon; TensorRT compiles to specific NVIDIA GPUs for maximum speed. Faster models require hardware-specific choices baked into the file, so a single model must be rebuilt separately for each chip or runtime it needs to support. Teams must choose how far down the speed-versus-compatibility ladder to go—trading inference latency against the types of hardware that can actually run the model in production.

   The article emphasizes holding the deployment format as high on the speed-compatibility chart as the latency budget allows, because every step toward faster execution narrows the hardware it can run on. Teams deploying LLMs face a constant tradeoff between minimizing response time and maximizing portability.

3. 3

   Sitrep: AI Copilot for Incident Response, Wired Into Grafana

   A developer has built Sitrep, a menu-bar app that screenshots your screen every ~15 seconds during an incident, distills each frame into structured observations using Claude vision, and lets you ask an AI copilot questions like "what was that trace ID I saw five minutes ago?" — then pulls the actual trace from Grafana Tempo to show where the time went. During incidents, engineers see hundreds of ephemeral facts across windows (trace IDs, error panels, metric spikes, Slack messages) and lose most when switching context. Sitrep captures and indexes them automatically, letting you search your screen history and correlate observations with backend traces — all without screen permissions or an API key, runnable in mock mode locally.

   The app is open-source on Hacker News. It runs on macOS with Node ≥20, integrates via Grafana's datasource proxy (TraceQL for Tempo, PromQL for metrics), and is instrumented with OpenTelemetry GenAI conventions — the copilot's own LLM calls are observable in the same Grafana instance it helps you debug.

4. 4

   AI worms self-propagate through Copilot for Word documents

   Researchers demonstrated that AI worms can embed themselves in documents and self-propagate when users open infected files in Copilot for Word, spreading malicious instructions through the AI system. This reveals a new attack vector where documents themselves become infection carriers, potentially compromising both individual users and organizations that rely on AI-assisted productivity tools for routine work.

   The practical implications for document-sharing workflows and whether Microsoft implements defenses to detect or isolate AI instructions injected into Word documents.