The Pragmatic Embedded Podcast · 12 Nov 2025 · 28m
Luca, an embedded software expert, discusses his experience at the Embedded Online Conference and a German conference, highlighting an LLM-only hackathon where participants were forbidden from writing code manually. He explains that effective AI-assisted development requires rigorous context management, treating prompts like version-controlled code, and managing external context files to prevent hallucinations. The conversation covers the economic benefits of AI (approx. $0.01 per line of code), the tightening of feedback loops, and the evolving role of developers from coders to AI supervisors and technical product owners.
Where this conversation describes AI being used to build or deliver a product. Part of the AI in practice tracker.
Engineering & simulationAda (AI coding tool)
The speaker used the Ada tool to ingest external API documentation via URL into the context window to provide accurate information to the LLM, preventing hallucinations of non-existent APIs.
“in Ada, I know I can just type slash web and give it a URL and it will just read whatever is behind that URL and add it to its context. And so, you know, you can just have it slurp in an entire API documentation, for instance.”
Speaker 2 · Hear it at 4:31
Engineering & simulationGitHub Copilot
The speaker used GitHub Copilot to add specific files or directory subtrees to the context window, allowing the AI to generate code based on precise, winnowed parts of a large codebase rather than hallucinating from a lack of context.
“you can add entire files. You can, you know, depending on the tool, maybe you can, you can add entire subtrees of directories, you can maybe add only portions of a file … pick and choose very carefully. What goes into the context because if you if you pick wrong, then the L&Ms are generally very happy just to just make stuff up.”
Speaker 2 · Hear it at 4:31
early-stage companies
Other shows on the same ground, closest first.
François Baldassari, CEO of Memfault, and Thomas Sarlandie, Field CTO of Memfault, host a panel with Jonathan Beri, CEO of Golioth, and Dan Mangum, CTO of Golioth, to discuss the intersection of AI, open source, and embedded development. The group analyzes the maturity of edge AI for audio and video processing, the challenges of integrating LLMs into hardware due to rapid model iteration, and the current state of using AI for firmware development. They conclude that while AI is transforming code generation and review, open source remains critical for training models and providing trusted, complex infrastructure like Zephyr and FreeRTOS.
Security researcher Nicholas Carlini joins Bryan Cantrill and Adam Leventhal to discuss his blog post on the pragmatic use of Large Language Models (LLMs) as tools for reducing friction in software development. Carlini shares specific examples of using GPT-4 and Claude 3.5 Sonnet to simplify code, decompile Python bytecode, and generate plots, while emphasizing the importance of verifying outputs and understanding the models' limitations. The conversation also explores the impact of LLMs on education, hiring practices, and the potential for these tools to democratize programming by lowering the barrier to entry for beginners.
Hosts Elecia White and Christopher White interview an AI language model, referred to as 'Christopher White' or ChatGPT, discussing its capabilities and limitations in embedded systems and software development. They explore the use of tools like Compiler Explorer (godbolt.org) and Google Colab for coding tasks, and discuss the complexities of using ArduPilot for drone projects. The conversation highlights the distinction between statistical language models and true artificial intelligence, noting issues with hallucination and lack of reasoning.
Bryan Cantrill and Adam Leventhal of Oxide Computer Company are joined by software engineer Gregoirein to critique Y Combinator CEO Garry Tan’s claim of writing 37,000 lines of code daily using LLMs. Gregoirein used Claude to audit Tan’s public projects, revealing broken assets, legal non-compliance, and a site outage, while the hosts discuss the 'peril of laziness' and the difference between generating code and engineering elegant systems. They argue that while LLMs are useful for review or low-stakes tasks, they cannot replace the deep thinking required for complex, high-consequence software like kernel linkers or signal handlers.
Bryan Cantrill, Adam Leventhal, Rain Paharia, and David Crespo of Oxide Computer Company discuss using LLMs like Claude Code and Opus 4.5 to increase engineering rigor rather than just velocity. They share specific examples of using AI to refactor the Helios kernel, build the 'iddqd' Rust map library, debug the Ghostty terminal in Zig, and manage type drift in Oxide's APIs. The group argues that rigorous documentation and strong typing are essential for effective LLM collaboration, contrasting this with the 'vibe coding' narrative.
The summary, chapters and topics are generated from the episode’s audio by this directory, not written by The Pragmatic Embedded Podcast. The episode itself belongs to its publisher.