Original Vibe Coding
In early 2025, vibe coding became a brand new term in the computer world, coined by none other than the legendary OpenAI co-founder Andrej Karpathy. With LLMs gaining more power and intelligence, vibe coders became masters of their own software stack almost overnight, coding in plain English. Traditionally, model capabilities improved in areas such as reasoning with text, image recognition, and computer use. However, models becoming more advanced at using computer software tools enabled an entirely new vertical.
OpenAI's recent launch of GPT-6 “Astra” showed that model capabilities are entering a new era of computer use, with the next frontier being CAD tool use. Vibe CAD engineering is the next frontier AI labs are enabling, and we are getting it not tomorrow, but today. At the launch of the new “Astra” model, OpenAI emphasized how capable the model is at using computer tools for research, spreadsheets, and more. However, the company also enabled new capabilities that align with design work that underpins modern electronics – Printed Circuit Boards (PCBs).
The biggest difference between vibe coding and what we are now seeing with hardware is that the model is no longer just generating text-based code. It is interacting directly with the tools engineers use to build physical products. Instead of describing a piece of software and receiving a block of code, an engineer can describe a board, its components, constraints, and intended purpose, and let an AI agent work directly inside the CAD environment. This is what I would call Vibe CAD Engineering.
The community on X exploded with people trying the model and seeing whether the new Astra could design PCBs from a simple description. Interestingly, the results seemed to be rather good, and traditional PCB design may be becoming a thing of the past.
How the Community Ended Up Using New Capabilities
There are countless examples of how the community on X ended up using “Astra” to power their projects. One of the posts that gained the most traction was from user @i2cjak, who created a Bluetooth AI wearable with two mics for beamforming, an IMU, an LED, and a button.
In about 30 minutes, Astra-Medium did this, a Bluetooth AI wearable with two mics for beam forming, IMU, LED, button. With KiStack and T3CAD, my little fork of T3Code. This level of performance is beyond what OpenAI demo’d and is the EASIEST way to steer the model as it works.
@i2cjakView post on XTo let “Astra” interact with the open-source KiCad PCB design software, it uses a specific set of agentic skills called KiStack. The model uses these capabilities and skills to steer its behavior and tool usage, allowing it to know exactly how to design, place, route, and potentially test the PCB design all on its own.
Another good example is the PCB design from X user @Peter05704721, who designed a PCB for a first-person-view drone. The PCB was designed entirely by “Astra” end to end by describing what needed to be done and what the design should resemble. After giving “Astra” a reference board, the model took it and redid the component placement and routing, creating a much cleaner design. Interestingly, this user actually took the design to a factory and had it manufactured, with pictures of the real design below, made by “Astra.”
fab sent photos. real FPV board is done. GPT-6 Astra + EasyEDA. PCB + SMT finished. package still in transit. next when it arrives: power on, then PX4.
@Peter05704721View post on XThe final example includes an unknown six-layer PCB design, entirely designed and routed by “Astra.” This one has also been sent to the manufacturing contractor, so we can expect updates in the near future.
OpenAI’s Example
Today, on OpenAI’s official GPT-6 “Astra” product page, there is a sample where the model is performing PCB layout in KiCad, showing just how useful the design is. Seemingly, “Astra” picked up the componentry, read the layout specifications, and laid out all of the components where they needed to be. Then the model optimized the layout and routed everything to create interconnections, producing a fully working electronics component, just like an engineer would.
After this, an engineer would run checks to ensure that the design is working properly, and we are talking more about that next.
PCB Design Rules
Typical PCB design tools, alongside most CAD design tools, have features that automate component placement and routing. These features are pre-defined mathematical pathfinders that treat the PCB as a canvas with multiple layers used to interconnect components like chips, SMTs, and other plugs, adapters, and more. Using these pre-defined mathematical rules, traditional CAD software for PCB design can automatically route and even place components. However, since those algorithms have no real sense of the 3D world, the math falls apart as they produce less-than-ideal routing and designs that are not really optimal for manufacturing. This is where the magic of modern AI models steps in and dictates the future.
LLMs like GPT-6 “Astra” are extremely good at using computer tools and software, while using their internal intelligence engine to reason about components and placement. When connected to a CAD tool, a model that has been specialized in computer use can provide a workflow similar to what a human can. As we have seen from the examples above, “Astra” is capable of understanding what the desired specification and outcome look like, then optimizing its work against design rules.
In any modern PCB CAD software, there are built-in design rule check (DRC) engines that test the layout, component placement, and electrical and physical properties of a PCB. For example, for a certain component, the DRC checks whether it makes electrical sense to be there, as well as whether layers might overlap and create electrical issues within the design. These DRCs highlight issues at the end of the PCB design workflow, which are later fixed manually by an engineer to remove any problems before the design is sent to manufacturing.
Design rule checks are only one part of the engineering process, however. A PCB can pass its basic rules and still require extensive validation for signal integrity, thermal behavior, power delivery, component availability, electromagnetic compatibility, and real-world operation. This is where the role of the engineer starts to change rather than simply disappear. The AI can increasingly handle the execution inside the CAD environment, while the human remains responsible for defining what the product needs to achieve and determining whether the final design actually meets those requirements.
For an agentic AI, this is an amazing baseline. When AI can use a tool like KiCad, it will automatically place and route PCB components, run design rule checks, get the results, and fix whatever the problem is. Essentially, an infinite feedback loop can run on the selected CAD tool until the design is perfect. That works in theory, but compute and agents are not running indefinitely, only until the end goal from the user’s prompt is achieved. Using tool calls, internal reasoning, and many other aces up its sleeve, OpenAI has optimized “Astra” for computer use in a way that allows it to thrive when there are lots of rules to follow.
RL Environments
How OpenAI achieved this is now a different topic altogether. Inside its labs, the company set up dedicated reinforcement learning environments, where the model’s agentic use is being optimized by getting rewards for correctly completed tasks and being “punished” for going off course. These RL environments have become so important that modern post-training efforts now dedicate massive compute capability to this task alone. After the model finishes its pre-training and fine-tuning, RL environments step in to steer the model toward correct ways of applying its intelligence across a broad range of capabilities.
Model performance and improvements can be achieved through RL environments, which are scaled until reaching an AI lab’s compute limits. Most of the latest model releases show that this is the case as the model continues to learn more about its task and RL environments. Xiaomi’s MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL model launch showed that scaling RL environments in real time costs a lot of compute, but provides significant performance increases when running popular benchmarks. As more RL steps are applied, models continue to improve through various techniques, which are described in detail here.
If Xiaomi is doing this on its models, we can assume that OpenAI has done much larger-scale RL optimization runs on its models with more compute capabilities. The split and relationship between pre-, mid-, and post-training optimizations at large labs is unknown and can be left to guesswork, but seeing how good these models are becoming at specialized tasks, we can assume that RL is playing an increasingly important role in this sector.
Future
Models will continue to scale their RL environments and work to get computer use to a perfect level. As they become more aware, tasks like PCB design will become almost as trivial as vibe coding an application today. However, that will also carry its own set of prerequisites, as you first need to know a thing or two about hardware engineering, product specification gathering, and how that would need to happen in the first place. Creating a simple PCB for a Bluetooth accessory, an FPV drone, or something similar is only one half of the story. The real test will be how these systems scale from relatively simple boards to increasingly complex products, from motherboards and GPU boards to networking hardware and server accelerator boards. The second half would need to be actual, production-grade PCB designs as we see in modern server rack-scale solutions. These are notoriously complex designs that carry a huge number of PCB layers, are equipped with tens of thousands of components, and carry massive power to the destination, aka AI accelerators.
Once AI models become viable in those scenarios, we could see more commercial adoption there as well, but for now it remains a tool to transform a vibe-coder into a vibe hardware engineer, with everything needed being knowledge of what to build, in plain English.
