LLMs, Board Layout and Human Engineering

OpenAI's announcement of GPT-6 Astra, its latest model, included a demonstration of the LLM performing printed circuit board layout in KiCad, converting an electronic schematic into a manufacturable board by placing components and routing copper connections. The showcase drew frustration and apprehension from embedded engineers, who debated whether LLMs threaten their jobs and whether AI-designed electronics can be trusted. An engineer argues that board layout is only one downstream step in embedded development, and that human systems-level thinking around requirements, verification, and design intent remains essential.

Earlier this week, OpenAI announced their latest model Astra https://openai.com/index/gpt-6-astra/ - however with this announcement https://openai.com/index/gpt-6-astra/?video=1223245030 , came some frustration from the embedded engineering community. https://www.reddit.com/r/embedded/comments/1w7p2dq/openai is showcasing kicad integration with their/ This is a 15-second condensed playback of GPT‑6 Astra performing printed circuit board PCB layout in KiCad, turning an electronic schematic into a manufacturable PCB by placing components and routing copper connections. Integral to every electronic device today, PCB layout is a manual task and common source of latency in the electronics design process. Accelerating it means freeing engineers to invent, optimize, and test their next idea at a significantly higher cadence. The announcement showed the Astra model carrying out a board layout in KiCAD. LLMs have always been able to edit board layout files as well as schematic files , but OpenAI showcasing integration with KiCAD made a lot of engineers frustrated and apprehensive on the future. https://www.reddit.com/r/embedded/comments/1w7p2dq/openai is showcasing kicad integration with their/ The same arguments and counter-arguments are mentioned within the community; this threatens the job of what engineers have worked so hard to build a skill for vs can we really trust electronics designed by LLMs? Does it feel like more and more tasks by engineers will inevitably be learned by LLMs and AI? Systems Level Thinking From a personal standpoint, I’ve never felt like I’ve been brought into projects solely for my ability to carry out schematics design and board layout. Whether it’s been as an employee or consultant, it’s been to drive the engineering development from a broader standpoint , leaning on lived experience on different projects and giving insight into risks and opportunities with the project at hand. Carrying out design work such as schematic design, board layout or firmware has still been part of processes I have been involved in, but the execution comes downstream from design decisions. Depending on what stage I’ve been brought in on, an end-to-end embedded systems project can be simplified into the following stages: - Evaluating the problem of the user - Propose requirements - Verify the requirments - Development technical specifications based on requirements - Carry out schematic design - Carry out board layout - Liaise with manufacturers - Write firmware - System integration - Troubleshooting - Fix problems - Verify the technical system works - Validate the design The board layout is only one part of the process - however the worry amongst many engineers is that a KiCAD integration has crept up on the community, while LLMs are already doing a large part of the firmware process. There is a possible future where all engineering becomes Systems Engineering - here real lived experience by humans contributes to developing the system overall, even if tools like LLMs help expedite processes and timelines. Developing requirements and specifications and understanding how to verify and validate technology are key parts of the engineering process, and being good at these is has always been what makes good Systems Engineers including pre-LLMs . Risks and Rewards with Agentic Engineering The obvious risks with agentic workflows in engineering is that it could lead to more technical debt and critical issues. Embedded systems are already complex, and it can be argued the technical infrastructure for development has always been disjointed and disconnected - it’s easy for technical knowledge to fall through the gaps between subsystems e.g. electronics and firmware and not be preserved. A lot of this lost information is design intent, and without it LLMs lose the valuable context needed to carry out work. Having agents verify/validate designs however can be a more involved version of post-design checks. DRCs design rules checkers already catch several issues within the EDA, however can LLMs provide more involved checks that take all project requirements into account? For example, human error can easily lead to some track widths not being up to spec for boards handling high-current, and the rule not being added to the EDA’s rule. Could LLMs catching a human error like this be the difference between a board passing and failing emissions tests? Or overheating or not? Engineering in the Physical World When consulting for plant monitoring startup SmartyPlants https://www.smartyplants.co.uk/products/smartyplants-plant-monitor-green?tw source=google&tw adid=809273349549&tw campaign=23855786133&tw kwdid=kwd-301839407645&gad source=1&gad campaignid=23855786133&gbraid=0AAAAA nRJghmJJYI0O9XkG7lT V1IklTD&gclid=Cj0KCQjwzY7VBhDwARIsAFtPvBRiEwcnp-s-4sZE-y73dMXDv00--NZioAdznwkn8u0DCHRE0-9Br5UaAmB-EALw wcB , a significant amount of my time was spent on scientific testing. This included investigating soil mediums in different environments, light intensity and analysing the light sources to measure light quality. With hardware engineering the loop doesn’t start and end in a computer like it does with software. The processes are also not limited to just manoeuvring physical objects - for example when analysing the environment that the monitoring device was going to stay in, I had to video it because the deterministic characteristics did not tell the full story. How something looks and feels with human judgement is an important part of the validation process. If there’s a future where humanoid robots can recreate physical-world tests, it would be pessimistic to suggest it would be soon - and even then, I’d question how well it could compete against humans. Human Engineering Sticking with this plant monitor as an example, parts of the process I felt had “meaning” for a human to be working on it were: - The design process - Risk and failure mode evaluation analysis - Troubleshooting existing bugs - Scientific testing For tools to be helpful, they need to maximise the time on a project that involves meaningful human engineering - tasks that require data to be analysed and critiqued on lived experience on other projects. It’s not to say board layout is not meaningful to be done by humans, however the discussion needs to be less catastrophic - and put in context of what human engineering can look like in the future.