Advertise on podcast: Embedded AI Podcast

We talk with Mark Deutel and Dr. Axel Plinge from Fraunhofer IIS about the art and science of running AI models on resource-constrained devices. Mark walks us through his PhD work on getting neural networks to run on microcontrollers with just 256KB of memory - a feat that requires clever optimization tricks, from neural architecture search to instruction-level tuning. Axel shares insights on why edge AI matters for robotics, where battery life and response time make local processing essential. The conversation ranges from their multilingual airport robot (complete with culturally appropriate gestures) to the practical challenges of teaching robots to grasp objects. We explore the shift from hand-engineered features to data-driven approaches, discuss why German privacy laws favor edge processing, and learn why the simplest microcontrollers sometimes give the most predictable performance. Mark and Axel make a compelling case that not every problem needs the biggest AI model - sometimes a specialized edge model is exactly the right tool. Key Topics: [03:30] Introduction to Fraunhofer IIS and its role bridging academia and industry[06:45] The multilingual airport robot: combining LLMs, speech processing, and local AI[15:20] Mark's PhD work: running neural networks on 256KB microcontrollers[19:40] The optimization pipeline: from neural architecture search to SIMD instructions[28:15] Why edge AI matters for robotics: battery life, latency, and the human analogy[35:50] Visual Language Action models and the challenge of world models in robotics[42:30] The complexity of grasping: what humans take for granted that robots must learn[48:20] German privacy laws (DSGVO) and why they favor edge processingNotable Quotes: "You're trying to run something that is way more compute-heavy on something that shouldn't actually be able to run it. It's kind of like the magic that guys in the 90s had when they wanted to run 3D games on old computers." — Mark Deutel "You cannot optimize without leaving stuff out. Mark says it's a magic trick, but it's not that much magic - yes, he got YOLO running on a system, but he also removed a lot of categories. Specialization is always part of the deal." — Dr. Axel Plinge "When you start to think about it, it's actually insane what considerations you make when you grasp something. You consider: is it fragile? Where's the best point to pick it? Should I pick it at the top or bottom? Are my hands sweaty? There's all kind of stuff you do without really thinking about it." — Mark Deutel Resources Mentioned: Fraunhofer IIS Efficient AI Team - Research group building AI solutions with fast MLOps workflow for embedded devicesHAI Store - Fraunhofer's platform for downloading and evaluating compressed AI modelsMark Deutel's Research Page - Personal homepage with details on PhD research about efficient neural networks on microcontrollersPIB Robot - 3D-printable educational humanoid robot from Incento used in the airport demonstrationNVIDIA Jetson Thor - Edge GPU system used for running local LLM and speech processing in the airport robotAgile Embedded Podcast Slack - Community discussion channel where you can reach Ryan and Luca, with a dedicated sub-channel for Embedded AI topics

We explore the resurgence of self-hosted LLMs and why they're worth considering despite the convenience of frontier models. From data privacy concerns to cost management and consistent performance, there are solid reasons to run your own models. We break down the hardware requirements (spoiler: it's all about memory bandwidth), compare different runtime environments from beginner-friendly LM Studio to production-ready VLLM, and discuss practical model choices. Luca shares his experience running dual RTX 3090s in his basement, while Ryan questions whether his 2004 Athlon is still viable (it might be). We also touch on the reality that local models lag frontier models by about six months—which means they're roughly where Claude Opus 4A was in February, and that's perfectly usable for many tasks. Key Topics: [02:30] Why self-host? Data privacy, cost control, and avoiding unpredictable changes from hosted providers[08:45] Memory bandwidth is the real bottleneck—why GPUs still win and what that means for your hardware choices[15:20] Hardware options: Mac Minis vs. discrete GPUs, and why second-hand RTX 3090s offer the best price-performance[22:10] Available models: Quen, Gemini, DeepSeek, and Mistral—what works well locally and the six-month lag behind frontier models[28:40] Runtime environments compared: llama.cpp for flexibility, LM Studio for ease of use, Olama for Docker-like experience, and VLLM for production[35:15] Practical considerations: context windows, model size vs. available VRAM, and why speed matters more than you think[40:30] Announcement: UddleSpec SDD framework now available, and the Agile Embedded Podcast is now Pragmatic EmbeddedNotable Quotes: "You need memory bandwidth. All of the memory. Every single token filters through all of those neural network layers—potentially 8 billion parameters per token." — Luca "The marketing behind these tools is very much geared toward 'don't worry what's happening behind the curtain.' But there are more efficient ways of solving the same problem without incurring the same cost." — Ryan "I really feel like the most rational choice is a second-hand 3090 still. They have 24 gigabytes of VRAM, they're nice and fast, and they have about half the bandwidth of a 5090—which is still pretty good and very much usable." — Luca Resources Mentioned: llama.cpp - Command-line LLM runtime environment with broad hardware support and REST APILM Studio - User-friendly GUI for running local LLMs with built-in model downloader and recommendationsOlama - Docker-like interface for managing and running local LLMs, good for experimentationVLLM - Production-grade LLM runtime with excellent parallel request handling and memory efficiencyHugging Face - Repository with hundreds of thousands of models, quantizations, and variants available for downloadUddleSpec - Luca's new SDD framework addressing real-world workflow needs—now available on GitHubPragmatic Embedded Podcast - Sister podcast (formerly Agile Embedded) covering embedded systems developmentAgile Embedded Podcast Slack - Community discussion channel where you can reach Ryan and Luca, with a dedicated sub-channel for Embedded AI topics

Ryan and Luca pull back the curtain on large language models, explaining what's really happening when you chat with an AI. They break down tokenization, context windows, and the stateless nature of LLMs—revealing why these tools aren't actually thinking or remembering, just generating the most likely next token based on massive matrices of weights. The conversation covers practical implications like why long sessions deteriorate, how caching affects costs, and why that apologetic "I won't do it again" from your LLM is meaningless. They also tackle the misleading anthropomorphization in AI marketing and share frustrations with features like Claude's opaque memory system. If you've ever wondered why your LLM seems to forget instructions or why it confidently states nonsense, this episode explains the mathematical reality behind the conversational illusion. Key Topics: [02:30] Tokenization: How LLMs break language into mathematical units[08:45] LLMs as stochastic algorithms: Predicting the next most likely token[12:20] Training models: Billions of parameters and matrix multiplication[15:10] Quantization: Trading precision for memory efficiency[18:00] Context windows and token limits: Why size matters and costs money[22:15] Stateless processing: LLMs don't remember, they reprocess everything[26:40] Caching and timeouts: The hidden costs of pausing your session[30:00] Hallucinations aren't bugs—they're the default mode of operation[35:20] Attention mechanisms and why large context windows cause degradation[40:15] Claude's memory system: Well-intentioned but problematic in practice[44:30] Mixture of experts: Sparse vs. dense models and routing tokensNotable Quotes: "The LLM doesn't even know what truth is, so it can't lie to you. It's not lying to the truth either. It's bullshitting—it doesn't care which way is true or false." — Ryan "All LLMs know how to do is hallucinate. They just generate chains of tokens. If you're lucky, those chains have some connection to the real world and are actually helpful." — Luca "I wish it wasn't programmed to just lie to me. If the LLM says 'I won't do it again,' yes it will, because it has no memory of this incident. It will behave the same way tomorrow." — Luca Resources Mentioned: Luca's Training and Consulting - Luca's website with links to AI and embedded systems training coursesTulip Tree Tech Emulator - Ryan's commercial emulator for embedded systems development (Raspberry Pi, STM, ESP, Microchip)Google's Attention Paper - The foundational 2013 paper introducing the attention mechanism that enabled modern LLMsOn B******t by Harry Frankfurt - Philosophical work distinguishing bullshitting from lying—relevant to understanding LLM outputsAgile Embedded Podcast Slack - Community discussion channel where you can reach Ryan and Luca, with a dedicated sub-channel for Embedded AI topics

We talk with Sebastian Boblest and Marco Mader Mendes from Bosch about the surprisingly challenging art of deploying AI models on resource-constrained embedded devices. They share insights from their work on ETAS Embedded AI Coder, a tool that generates optimized C code from neural network models for microcontrollers—sometimes with as little as 100 bytes of RAM available. The conversation covers practical strategies for model compression (often by factors of 100x or more), the counterintuitive benefits of float models over quantized ones on tiny devices, and why feature engineering still matters. Sebastian and Marco explain how they navigate the trade-offs between RAM, compute time, and accuracy, and why each project presents unique constraints—from confidential hardware specs to compiler quirks. They also discuss real-world applications, including Bosch's AI-powered wall scanner that uses radar and neural networks to detect cables in walls. Key Topics: [03:30] Why Bosch started exploring embedded AI on tiny hardware in 2020[06:45] Real-world application: AI-powered wall scanner using radar to detect cables[09:20] How ETAS Embedded AI Coder generates C code from neural network models[12:00] Deploying models with as few as 300 parameters and 100 bytes of RAM[16:30] Model compression strategies: squeezing networks down by 100x or more[21:15] When float models outperform quantized ones on tiny devices[28:00] Trading off RAM vs. compute time through code generation techniques[33:45] Benchmarking challenges with confidential hardware and compilers[40:20] AutoML and architecture search for constrained embedded targets[46:00] Free tool access for universities and opportunities for studentsNotable Quotes: "We go down to applications where we use 100 bytes of RAM and you can still do something useful with this on a Cortex-M0. It's very surprising how small you can get with neural networks." — Sebastian Boblest "Sometimes we have to squeeze it down not by one or two X, sometimes it's up to 100 X and more. This is really a regular task for us." — Marco Mader Mendes "One thing that's super counterintuitive for many people on these smaller devices is to go from a quantized model to a float model. It can actually help you. It's exactly the opposite of what people do on these larger systems." — Sebastian Boblest Resources Mentioned: ETAS Embedded AI Coder - Code generation tool for deploying neural networks on embedded devices; free for universitiesARM CMSIS-NN - Library containing functions for neural network layers on Cortex-M devicesMLPerf Tiny Benchmarks - Benchmarking suite for tiny ML systems that Sebastian's team participated inBosch AI-powered Wall Scanner - Radar-based power tool using neural networks to detect cables in wallsAgile Embedded Podcast Slack - Community discussion channel where you can reach Ryan and Luca, with a dedicated sub-channel for Embedded AI topics

Ryan and Luca tackle the hot topic of AI-driven security research, sparked by the release (and brief containment) of Anthropic's Mythos tool. Ryan, drawing on his 20+ years in cybersecurity, delivers a surprisingly reassuring message: if you've been doing proper engineering, AI-found vulnerabilities aren't the existential threat the hype suggests. We explore how Mythos and similar tools are flooding the CVE database with thousands of new vulnerabilities, but discuss why this doesn't automatically mean more successful attacks. Ryan explains the crucial difference between finding a bug and actually exploiting it, and why embedded systems developers shouldn't panic—but should definitely have their update processes sorted. The conversation covers everything from botnet refrigerators to Ukrainian security cameras, threat modeling with AI assistance, and why AI makes such a relentlessly effective hacker (spoiler: it doesn't get bored). Bottom line: the weapons haven't changed, you just need to install the bulletproof glass you should have had all along. Key Topics: [02:30] Introducing Mythos: AI tool for finding software vulnerabilities, its May release, US containment, and recent re-release[05:15] The flood of AI-generated bug reports: distinguishing real vulnerabilities from noise, and the burden on maintainers[08:45] Mythos by the numbers: 6,000+ critical vulnerabilities found, 90% true positive rate, but does it matter for your system?[12:00] The reachability problem: having a bug vs. being able to exploit it, and why solid engineering processes matter more than bug counts[16:30] Embedded systems challenges: BSP version conflicts, regulatory approval nightmares, and the CRA/FDA compliance push[21:00] No uptick in actual attacks: why more CVEs doesn't equal more breaches, and what motivates attackers (hint: money, not bugs)[28:45] Embedded systems' blessing and curse: air-gapped devices vs. internet-connected vulnerabilities, and the botnet refrigerator story[35:20] Using AI for defense: network analysis, threat modeling, traffic pattern recognition, and why AI is the best grep you've ever seen[42:00] AI as the relentless attacker: no social norms, no contracts, just pure problem-solving—and why that's both powerful and concerning[47:30] The offensive vs. defensive mindset: why AI bridges both motivational patterns and what that means for security teamsNotable Quotes: "Just because there's more CVEs doesn't really change the defensive posturing that you need to do. If you have a solid engineering process to provide updates to your system and test and verify those updates actually work, it doesn't matter how many bugs you find. You can find two. You can find 6,000." — Ryan Torvik "AI doesn't have social norms. AI will just go and do things. And yes, there are guardrails that they're trying to put on, but like, I don't know if you've seen AI just ignore parts of your prompt before." — Ryan Torvik "There's not a death star. This is not an existential crisis. This is not something new. It changes the game but not in a way that we can't handle. The weapons haven't changed, you just need to install the bulletproof glass." — Ryan Torvik Resources Mentioned: Anthropic's Mythos - AI tool for automated vulnerability discovery in software, released in May 2024, briefly restricted by US authorities, then re-releasedCVE Database - Common Vulnerabilities and Exposures database, currently experiencing significant growth due to AI-assisted vulnerability discoveryAgile Embedded Podcast Slack - Community Slack channel where Ryan participates, sister podcast to Embedded AITulip Tree Tech - Ryan's company focused on improving embedded development processes, including emulator-in-the-loop solutionsluca.engineer - Luca's website with links to training courses, LinkedIn, and other professional activities

We tackle the elephant in the room: will AI take our jobs? Spoiler alert - probably not, but your job will definitely change shape. Ryan and Luca dig into what software development actually is (hint: it's not just mashing keyboards), why embedded systems might be particularly safe from AI disruption, and what we can learn from the Luddites and steam engines. Drawing on Luca's experience with DevOps transformations and Ryan's work running a company, we explore why engineers keep solving the wrong problems, what Black & Decker actually sells (it's not drill bits), and why vibe-coding your email server is a terrible idea. The real question isn't whether AI will replace you - it's whether you understand what your actual job is in the first place. Plus: Pokemon evolution as career advice, and why AI is the new Agile fairy dust. Key Topics: [02:30] The real job vs. the mechanical task - why writing code isn't actually your job[08:45] Why embedded systems are particularly safe from AI disruption - complexity, hardware interaction, and undocumented quirks[15:20] The Black & Decker lesson: selling holes in walls, not drill bits - understanding what customers actually want[22:10] Historical parallels: steam engines, DevOps, and why demand grows faster than automation[28:40] The reality check: most developers aren't using AI systematically yet - you're not behind the curve[32:15] AI as the new Agile fairy dust - why magic solutions never work without process and understandingNotable Quotes: "Your job is not writing code. Your job is making product. If you think about the Luddites, their job was not operating a loom. Their job was making clothes." — Ryan Torvik "Writing code is really the smallest part of software development. Most of it is sitting in front of the screen, looking up and to the left, and figuring out what code to write." — Luca Ingianni "The CEO of Black & Decker was once quoted as saying: We are not in the business of selling drill bits. We are in the business of selling holes in the wall. If we had laser cannons that made holes in walls, people would buy the laser cannons instead." — Luca Ingianni Resources Mentioned: Agile Embedded Podcast Slack - Community discussion channel where you can reach Ryan and Luca, with a dedicated sub-channel for Embedded AI topicsLuca's Website - Contact Luca for consulting on using AI effectively in embedded systems contexts - multiple contact options available

Luca unveils WhittleSpec, his new open-source framework for AI-assisted development that challenges the waterfall assumptions baked into most coding tools. Born from frustration with frameworks like SpecKit that treat specifications as static documents, WhittleSpec emphasizes continuous learning through retros, refinement, and test-driven development. We explore why most AI coding frameworks make typing faster but don't help with the hard part—the thinking—and discuss how proper feedback loops and vertical slicing can lead to more trustworthy software. The conversation ranges from the philosophy of whittling away what doesn't fit (versus plowing ahead blindly) to practical implementation details: specification, planning, task breakdown, TDD cycles, and retrospectives. Luca argues that professional software engineering requires systematic processes that support learning at every step, not just tools that generate code quickly. We also touch on the missing pieces in current frameworks: support for safety-critical development, long-term roadmaps, and embedded systems considerations. Key Topics: [02:30] Introducing WhittleSpec and the problem with current AI coding frameworks[08:45] Why AI tools make the easy part easier but leave the hard part hard[15:20] The waterfall trap: static specifications vs. living documents[22:10] Mapping the landscape of AI development frameworks (SpecKit, BMAT, Kiro, etc.)[28:40] How WhittleSpec works: decide, specify, plan, tasks, and TDD cycles[38:15] The critical role of retrospectives and the 'refine' skill for course correction[45:30] Vertical slicing vs. layer-by-layer implementation: tracing bullets through the stack[51:00] Missing pieces: safety-critical development, long-term roadmaps, and embedded considerationsNotable Quotes: "No engineer ever said, 'I wish I could type curly brackets faster.' That was never quite the bottleneck. The hard part is sitting in front of your screen and going, 'hmm.'" — Luca Ingianni "SpecKit is just plain old horrendous waterfall. There are no provisions at all in it for learning. The idea is you specify something well enough and then you just walk away, sip a coffee, the machine does its thing. That approach has never ever worked." — Luca Ingianni "Your initial specification is not going to be right. As you implement the actual solution, you're going to learn things that's going to change what you need to accomplish with the spec. You might need to change completely what your expectations are." — Ryan Torvik Resources Mentioned: WhittleSpec - Luca's new open-source AI-assisted development framework emphasizing feedback loops, TDD, and iterative refinementSpecKit - GitHub's AI coding framework discussed as an example of waterfall-style specification-driven developmentAgile Embedded Slack - Community Slack channel now open to Embedded AI podcast listeners for discussion and questionsLuca.engineer - Luca's website with links to all his projects and ways to reach himTulipTreeTech - Ryan's company working on AI-generated models for pre-silicon firmware validation

Ryan and Luca sit down with Shawn Hymel, an educator and course creator focused on edge AI and embedded systems. We explore what's changed in the past few years—from basic keyword spotting to full object detection on microcontrollers, thanks to integrated NPUs. Shawn walks us through the messy reality of deploying ML models to embedded hardware: vendor-specific toolchains, dependency hell, and the ongoing challenge of making edge AI accessible. We discuss what students and experienced engineers need to learn (or unlearn) to work effectively in this space, and look ahead to exciting developments like reinforcement learning on tiny devices and neuromorphic computing. It's a candid, technical conversation about where edge AI stands today and where it's headed. Key Topics: [03:30] Why run ML on microcontrollers? Power, size, and application-specific advantages[06:00] Keyword spotting as the original killer app for edge AI[08:45] The game-changer: NPUs enabling full object detection on microcontrollers[13:20] Privacy benefits of on-device processing vs. cloud-based inference[16:00] How NPUs work under the hood and the vendor-specific deployment reality[24:30] The painful parts: dependency hell, graph compilers, and memory arena sizing[29:00] Using AI tools (LLMs) to navigate vendor documentation and generate code[33:45] What CS students and ECE students each need to learn for edge AI[40:15] Shifts in university enrollment: ECE rising, CS declining[44:00] When you don't need AI: PID loops and deterministic solutions still matter[47:30] Looking ahead: reinforcement learning on microcontrollers and neuromorphic computingNotable Quotes: "Five, six years ago, we didn't have full object detection on microcontrollers. Now with NPUs, we can do things like full YOLO on a 320x320 image—milliwatts of power, full object detection. That was not a thing five years ago." — Shawn Hymel "Expect to spend a day or two getting inference to actually run. The docs are still new, the graph compilers are fairly new. You're going to end up in dependency hell—both on the Python side and when you bring it over to the embedded side." — Shawn Hymel "If a PID loop solves your need, there is absolutely no reason to put AI on there. That is a solved problem. Don't do it—just use a PID loop." — Shawn Hymel Resources Mentioned: Shawn Hymel's Website - Shawn's main site with links to free and paid courses on edge AI and embedded systemsOpenMV - Computer vision platform for microcontrollers, including the new AE3 board with NPU supportEdge Impulse - Platform that simplifies ML model deployment to embedded devices, supporting various NPUsAndrew Ng's Coursera ML Course - Foundational machine learning course recommended for understanding the math behind MLTensorFlow Lite for Microcontrollers (LiteRT) - Framework for running ML models on microcontrollers across different platforms

Ryan and Luca tackle a challenge many AI users are facing: what happens when your AI provider starts acting up? Drawing from recent experiences with Anthropic's capacity issues, secret billing practices, and model degradation, we explore practical strategies for avoiding vendor lock-in. We discuss the three layers of complexity: the model itself, the harness (like Claude Code or GitHub Copilot), and your authored content (skills, MCP servers, prompts). Each layer presents different challenges when switching providers. Ryan shares his approach of stepping back to simpler, more granular prompting to stay provider-agnostic, while Luca experiments with maintaining escape hatches to other platforms. We also look at the realities of running local models and the tradeoffs between convenience and control. The bottom line? Pick one system, get proficient, but prepare your exit strategy - because in this volatile landscape, you'll likely need it sooner than you think. Key Topics: [02:30] Anthropic's recent troubles: capacity issues, model degradation, and gaslighting users[06:45] The Hermes.md billing scandal - secret charges for having a specific filename[10:20] Ryan's approach: stepping back to simpler, granular prompting for provider independence[15:00] The three layers of complexity: model, harness, and authored content[18:30] Why models aren't interchangeable - different flavors, tokenizers, and caching strategies[24:15] Luca's tone-of-voice challenge: getting consistent writing style across models[30:00] Running local models and private inference as alternatives to frontier models[35:45] Practical strategies: maintaining escape hatches without parallel systems[40:20] Luca's solution: versioning authored content separately with symlinksNotable Quotes: "The more you actually make use of AI in your work, the more you use it as a force multiplier, the more painful it becomes if that force multiplier goes away." — Luca "It's about total clock time. If you get a one shot and then have to redo it again, how much of that clock time is being used effectively?" — Ryan "Pick one, stick with it, be proficient in it. But prepare yourself to have to escape eventually, because the situation is so volatile." — Luca Resources Mentioned: Claude Code - Anthropic's AI coding assistant with hooks and skills supportOpenCode - Provider-agnostic AI coding harness that supports multiple modelsGitHub Copilot CLI - Multi-provider AI assistant with dropdown model selectionMCP (Model Context Protocol) - Protocol for extending AI capabilities across different harnessesOllama - Tool for running local AI models on your own hardware

We explore what it really means to run AI models locally instead of relying on cloud providers like OpenAI or Anthropic. From powerful desktop setups with dual NVIDIA RTX 3090s to tiny models running on embedded systems, we cover the full spectrum of local AI deployment. Luca shares hands-on experience running local models for client work, explaining the hardware requirements (spoiler: you need fast VRAM, not just lots of RAM), performance trade-offs, and practical tools like Ollama and LM Studio. We discuss how modern open-weight models from Meta, Google, and Chinese companies compare to hosted solutions - typically about a year behind state-of-the-art but surprisingly capable. We also look at edge AI applications, from elderly fall detection to traffic accident monitoring, where compact models shine. The conversation covers context window limitations, quantization techniques, and why getting started is easier than you might think - though you'll need to manage expectations about what local models can deliver compared to their cloud-based cousins. Key Topics: [00:00] Introduction: What are local models and why run them?[02:30] Hardware requirements: VRAM vs system RAM, and why graphics cards matter[05:45] Luca's setup: dual RTX 3090s and real-world client work with local models[08:20] Performance metrics: time to first token, tokens per second, and output quality[12:00] Tiny models for edge AI: Google's 270M parameter model and specific use cases[15:30] Tools and workflows: Ollama, LM Studio, and OpenAI-compatible APIs[18:45] Where models come from: Hugging Face, Meta's Llama, and the open-weight ecosystem[22:10] Context window limitations and quantization techniques[25:00] Getting started: realistic expectations and practical first stepsNotable Quotes: "What those models really need is tons of memory and as fast memory as you can get it. This is why people like Macs because they've got the unified memory architecture." — Luca Ingianni "Modern models are actually pretty good. They are smaller, so they will have less knowledge. They are a tad slower. They struggle with much smaller context windows. But if you can work within those bounds, they work pretty well." — Luca Ingianni "Using AI is a different mindset. It's a different way of thinking about solving a problem, but it's just another tool in your toolbox and it's just another way of getting work done." — Ryan Torvik Resources Mentioned: Ollama - Docker-like tool for running local LLMs with simple pull/run commands and OpenAI-compatible APILM Studio - Graphical user interface for downloading and running local models easilyOpen WebUI - Web interface for local models that mimics ChatGPT's chat interfaceHugging Face - Repository with hundreds of thousands of models in various sizes and configurationsMeta Llama - Open-weight model family from Meta that helped start the local LLM movementGoogle Gemma - Model family from Google including compact vision-capable models (270M parameters)

We talk with Julia Imlauer, Director for AI Strategy and Development, about the practical realities of introducing Gen AI tools into engineering organizations. Julia shares her experience training engineers and management across multiple companies, revealing what actually works—and what doesn't—when rolling out LLM-based coding assistants. The conversation covers the adoption curve from early enthusiasts to skeptics, the importance of giving engineers dedicated time to experiment, and why "word of mouth" matters more than top-down mandates. We discuss the messy reality of legacy codebases where LLMs shine, the challenge of managing AI-generated code in pull requests, and the emerging cost crisis as providers shift from subsidized to consumption-based pricing. Julia also addresses common fears about junior developer training and the misconception that LLMs can replace human expertise. Key Topics: [02:30] Julia's background in robotics and machine learning, and how she started introducing Gen AI tools into development workflows[05:15] Why LLMs help with messy, historically-grown codebases more than pristine template-based workflows[09:45] Different developer preferences: IDE integration vs. separate LLM interfaces[13:20] The adoption curve: early adopters, skeptics, and the crucial middle mass that needs convincing[16:00] Training sessions that work: giving engineers blocked time to experiment, not step-by-step tutorials[19:30] The power of word-of-mouth and why ambassador programs sometimes help spread adoption[24:10] Training management and executives: IP protection, managing AI-generated code in pull requests, and supporting skeptical engineers[29:45] The billing nightmare: token consumption, opacity in pricing models, and the shift away from all-you-can-eat plans[35:20] Common skeptic concerns: fear of replacement, adoption pain, and outdated first impressions from early LLM versions[38:50] Junior developer training with LLMs: why pair programming and code reviews still matter more than writing every line by hand[43:15] Communication breakdowns between departments: translating vocabulary between IT, legal, and engineering[46:30] Looking ahead: the need for agentic frameworks, standardization, and routing models to manage costsNotable Quotes: "What they do for you is really, they free your engineers to focus on things that matter. They write the boilerplate code. They write the code which you don't want to think about, which is a necessity, which you need to have to make everything work. But it's just like a doing concept." — Julia Imlauer "It sounds stupid when you say you need kind of a education training session for engineers to use this tool, but exactly this is what I experienced—they don't need you know a training in the ways of you need to click here and here and here. What they need is blocked two hours out of their daily coding hustle and really get kind of a free mind to take a look at the tool." — Julia Imlauer "There is this misconception of, okay, we don't need junior developers anymore. They just cost and then an LLM can do that. And let's just pile more on our senior developers. I mean, this is not working. And everyone who is in software development understands that." — Julia Imlauer Resources Mentioned: Caveman - A GitHub repo that compresses prompts to very basic words to reduce context window size and token consumptionOpen Claw / Nemo Claw - Open source agentic framework projects exploring standardized interfaces for LLM agentsMCP (Model Context Protocol) - An emerging standard for LLM tool integration mentioned as a potential path toward provider-agnostic frameworks

We sit down with Kwabena Ageyman, co-founder of OpenMV, to explore how microcontrollers have evolved from simple 8-bit chips to AI-capable systems that rival desktop computers. Kwabena walks us through OpenMV's journey from the CMU Cam days to their latest products—the OpenMV Cam AE3 and N6—which pack neural network accelerators, image signal processors, and H.264 encoders into single-chip packages. What makes these systems remarkable isn't just raw performance (250+ gigaops for AI inference), but what they enable: battery-powered computer vision deployments with no infrastructure requirements. Kwabena demonstrates running a complete web server with live video streaming—all in MicroPython on a microcontroller. We discuss the practical implications: doorbell cameras that don't phone home, parking lot monitors that run on solar panels, and industrial vision systems that don't require conduit runs. The conversation touches on hard technical choices (why debayer images even for AI?), the underappreciated value of MicroPython for complex applications, and the infrastructure costs that kill many promising AI deployments. Kwabena also previews what's coming: transformer support on microcontrollers and WiFi HaLow for long-range, high-bandwidth connectivity. For anyone working on edge AI or embedded vision, this episode offers both practical insights and a glimpse of what's possible when hardware acceleration meets thoughtful software design. Key Topics: [00:03] Introduction to OpenMV and the evolution from "impossible" computer vision on microcontrollers to AI-capable systems[00:05] Technical specs of the AE3 and N6: 250 gigaops performance, 5-64MB RAM, image signal processors, and H.264 encoding[00:12] Why image processing steps like debayering matter even for AI applications[00:18] The infrastructure cost problem: why power, connectivity, and deployment logistics kill many AI projects[00:28] Running MicroPython on microcontrollers: web servers, RTSP streaming, and complex applications without Linux[00:35] Live demo: complete web interface with video streaming running on a microcontroller[00:42] Real-world use cases: YOLO object detection, face tracking, drowsiness detection, and parking lot monitoring[00:52] The future: transformer support on microcontrollers and WiFi HaLow for long-range connectivityNotable Quotes: "Ten years ago when we started, if you Googled for computer vision on microcontrollers, you got a single Stack Overflow reply about how that was impossible. Since then, a lot has changed." — Kwabena Ageyman "The product dies when you have to tell people: I want you to put 10,000 of these in the field. They look at the infrastructure cost and say, what does that look like end to end? Is that actually going to be a net benefit, or is it just sexy and looks cool?" — Kwabena Ageyman "Having AI on the edge actually unlocks privacy. It is a decision to collect all the data and store it forever. If you have Edge AI locally on these devices, the device manufacturer can say: we're actually not going to go into a format where we have infinite data collection of everything." — Kwabena Ageyman Resources Mentioned: OpenMV - OpenMV's website with products, documentation, and community resourcesRoboflow - Cloud platform for training computer vision models, partnered with OpenMVEdge Impulse - Edge AI development platform, OpenMV partner for model trainingWiFi HaLow - Long-range, low-power WiFi technology (up to 10 miles) mentioned for future connectivityEmbedded Online Conference - Conference where Kwabena, Luca, and Ryan will be speaking

For E13 we recorded live in Luca's garden in Munich, with Ryan dropping by ahead of Embedded World week. Ryan and Luca talk about spec-driven development in the AI era: where the discipline came from, what changes when an LLM is doing the typing, and the failure modes that show up over and over again in trainings. The short version: vibe coding will get you something that demos beautifully, but the moment a stranger asks "what does this button do?", it tends to expose how little was actually thought through. The conversation circles around a few recurring themes — the iterative loop you cannot skip even when the AI lets you, the temptation to one-shot whole projects, and the awkward fact that the AI itself seems to actively prefer working in waterfall mode. We also get into why requirements engineering and product ownership matter more (not less) with AI in the picture, why TDD doubles as a way of describing the goal to your assistant, and why the engineer staying in the loop — with that loop running tighter and faster — is what actually makes this work in practice. Plus an honest digression about all the ditches Luca has fallen into building Claude Code skills around his daily workflow. Key Topics: [02:24] Three artifacts of spec-driven development — and what each one means in the AI era[05:45] Trainings, vibe-coded games, and the "what does this button do?" moment[11:27] Long-lived branches, three months of code in an hour, and why both fail for the same reason[22:25] Beningo's multiplier metaphor: what if the engineer's value is between -1 and 1?[24:25] Engineering as conversations; curly brackets as a side effect[27:17] Why requirements engineering and product ownership become more important with AI[30:59] The AI wants waterfall — and you need to fight it[33:54] One-shot prompts for whole projects: lying to yourself in five lines[37:27] Staying in control: one unit of work, AI in the loop, integrate, done[41:32] Filling ditches one at a time: Luca's Claude Code skills setup[43:06] TDD as the act of describing the goal — to yourself and to the AINotable Quotes: "If you've got a five-line prompt that generates 10,000 lines of code for you, then there's just going to be a lot of blank spots in there, a lot of ambiguity in there. That can't be good." — Luca Ingianni "I've had teammates create a long-lived branch and tell me 'I'll see you in a few months.' And I'm like — no. They don't understand how this is going to interact with the rest of the system. And you're basically doing that same thing — writing three months worth of code in an hour. Cool, now what?" — Ryan Torvik "Engineering is what happens when engineers talk to one another, and the differential equations and the C++ code are just side effects of those conversations." — Luca Ingianni Resources Mentioned: embeddedai.academy — Luca's AI trainings for embedded teamsAgile Embedded Podcast — sister show, more on agile in the embedded worldEmbedded World — annual embedded systems trade fair in Nuremberg, where Ryan was heading nextClaude Code — the AI coding tool Luca built his "skills" workflow aroundJacob Beningo's "multiplier" framing for AI in development teams (referenced from E12)

We sit down with Jacob Beningo, a real-time embedded systems consultant with 20 years of experience, to talk about what we've learned teaching engineers to use AI in their development workflows. Turns out, the hard part isn't getting AI to write code—it's all the systems engineering that comes before and after. We discuss common mistakes people make when starting out, like treating AI as a magic code generator instead of a pair programming partner, and why you absolutely cannot skip requirements, architecture, and critical thinking just because the AI can type faster than you. Jacob shares stories from his training sessions, including an AI that refused to follow test-driven development because "that would take too long." We explore why AI actually forces you to become a better engineer by taking away the dopamine hit of typing code yourself, and why IDE plugins might be leading people astray by keeping them at the wrong level of abstraction. The conversation gets real about costs—both in tokens and electricity bills—and why the "set it and forget it" YouTube hype doesn't match reality. If you're skeptical about AI in embedded systems, good—keep that skepticism. You're going to need it. Key Topics: [03:15] The real challenge: systems engineering, not code generation[08:45] Why requirements engineering skills matter more than ever with AI[14:20] The push-button module exercise: spending a full day on design before any code[16:30] When AI refuses to follow TDD: "That would take too long"[22:40] The temperature sensor exercise: when tests pass but the code isn't production-ready[28:15] Code is cheap, but experiments aren't free: finding the balance[35:50] The hidden costs of AI: token budgets and rising electricity bills[42:10] Why IDE plugins might be the wrong interface for AI-assisted development[48:30] Using AI as a pair programming partner, not a code completion tool[53:20] Keep your skepticism: why critical thinking is more important than everNotable Quotes: "The AI finished writing the code and all the tests in like four seconds. I'm like, how was that so fast? Well, I just wrote all the code. You didn't follow the TDD process? No, that would take too long." — Jacob Beningo "If you throw in such a vague requirement, the thing can't read their mind and neither can they read its mind. So it's really just a matter of luck what you're going to get." — Luca Ingianni "The AI is the best puppy you'll ever have. It'll go pick up the stick for you. And you're like, no, not that stick. What do you mean not that stick? You didn't tell me which stick. I grabbed you the stick." — Ryan Torvik Resources Mentioned: Jacob Beningo on LinkedIn - Daily posts about embedded systems development and modernizationBeningo.com - Jacob's consulting and training services for embedded systemsEmbedded Software Academy - Training courses including AI for embedded systems developmentEmbedded Online Conference - Annual May conference for embedded systems education and communityAgile Embedded Podcast Slack - Community discussion channel mentioned by the hosts

Ryan and Luca explore practical techniques for using AI to debug embedded systems -- from analyzing breadboard photos to parsing UART output and managing complex debugging workflows. LLMs work best as force multipliers rather than replacements for engineering expertise: they handle tedious tasks like adding printf statements, analyzing logs, and decoding resistor color codes, while the engineer guides the process and catches mistakes. A key theme is context management: balancing deterministic scripts for repeatable tasks with non-deterministic AI analysis, and using separate sessions to keep debugging focused. We share cautionary tales of LLMs getting stuck in loops or reverting to common patterns despite specific instructions -- human oversight remains essential. Experienced engineers benefit most because they can effectively steer the LLM and recognize when it goes off track. Key Topics: [02:30] Debugging hardware via photos -- having LLMs identify wiring errors on breadboards[06:45] The rubber duck effect -- LLMs as interactive debugging partners[11:20] Printf debugging with AI -- adding debug statements and analyzing UART output[15:40] Context management -- separate sessions, distilled datasheet summaries[22:15] LLM failure modes -- loops, pattern reversion, ignoring specific instructions[28:30] Force multiplier vs. replacement -- why experience matters more with AI tools[33:50] Deterministic scripts + non-deterministic AI analysis[38:20] Future: webcams and oscilloscope screenshots for real-time hardware debuggingNotable Quotes: "It's like the over-eager intern -- it's a little naive, but it can type like the devil." -- Luca Ingianni "You cannot use an LLM as a replacement for your brain or for your experience, but you can use it as a force multiplier." -- Luca Ingianni "The LLM does not have that emotional connection to the code. I don't think people understand how emotional that connection is." -- Ryan Torvik Resources Mentioned: Claude Code (Anthropic) -- AI coding assistant used for debugging, image analysis, and code generationArduino/Elegoo Development Boards -- hardware platforms discussed in context of voltage compatibility debuggingAgile Embedded Podcast -- Luca's podcast on agile practices in embedded development