I remember “ye olden days” when we (humans) designed electronic products by hand—conceiving circuits, selecting components, drawing schematics, laying out printed circuit boards (PCBs)… Can you imagine instead just telling an AI, “Make me a [your product here] for consumer use,” and it actually does it? Well, that day has arrived.
I’ve said it before, and I’ll doubtless say it again—things are moving fast in AI space (where no one can hear you scream). For example, GitHub Copilot is an AI-powered coding assistant developed in collaboration between GitHub and OpenAI. It’s designed to help software developers write code faster and more efficiently by suggesting code completions, generating functions, and even writing whole blocks of code directly inside the editor.
GitHub Copilot was announced in June 2021 as a technical preview. That original version was powered by OpenAI Codex, which was a modified production version of GPT3. Meanwhile, the ChatGPT that we know and love first began to impinge itself on the public consciousness when it was formally launched in November 2022. The initial ChatGPT release was based on GPT-3.5.
Did you ever wonder what the “GPT” part stands for? I’m happy to oblige. It stands for “Generative Pre-trained Transformer.”
“Generative” derives from the fact that the model generates new content (from text to code to images, to sounds in multimodal versions). In the case of text, for example, it doesn’t just retrieve facts; it produces original sentences or structures based on learned patterns.
“Pre-trained” refers to the fact that the model is initially trained on a massive dataset (including books, articles, code, and web pages) before being fine-tuned for specific tasks. This “pre-training” phase allows it to develop a broad understanding of language and reasoning. Later, it can be adapted or “fine-tuned” for more specialized purposes (e.g., answering questions, writing poetry, generating code).
“Transformer” refers to the Transformer Architecture, which was introduced in a 2017 research paper titled “Attention Is All You Need” by Vaswani et al. Transformers use self-attention mechanisms to analyze relationships between words (or tokens) in a sentence, enabling the model to capture context more effectively than older RNNs or LSTMs. This innovation is what made large language models (LLMs) like GPT practical and powerful.
Speaking of which, terms like large language model (LLM) and Generative AI (GenAI) didn’t start to appear in mainstream usage until late 2022 and early 2023.
Since those distant days, we’ve begun to see GenAI emerge in various electronic design automation (EDA) tools and disciplines. This process began with circuit design and schematic generation tools like Flux (see May the Flux (Copilot) Be with You!), Circuit Mind (see From Architecture to PCB Schematic in 60 Seconds! and From Power Supply Block Diagram to Completed Design in 60 Seconds!), and Celus (see From Thought to Circuit in Record Time with AI).
The folks at Flux introduced GenAI-powered PCB layout to their tool around December 2024. Since that time, we’ve seen a spate of AI-powered PCB layout tools, including JTIX (see Eeek Alors! AI-Powered PCB Layout Is Here and It’s Awesome) and Quilter (see You May Scoff, But AI-Powered PCB Layout Is Really Real).
Everything we’ve talked about thus far has featured Generative AI. While these tools do generate new output, they must be prompted to do so. The “latest and greatest” on the AI front is known as Agentic AI. This is generically defined as “AI systems that can act autonomously, make decisions, and plan and execute tasks toward goals with little human supervision.”
The first time I heard the term “Agentic AI” was earlier this summer (see The Wizard of Oz and Mozart Meet Generative and Agentic AI). Just a couple of months later, I was introduced to my first taste of Agentic AI in the wild (see Agentic AI Prevents Costly Machine Downtime in Factories and Semiconductor Fabs).
I really didn’t expect to see real-world deployments of this technology in the EDA arena for quite some time. I should have known better…
The reason for my meandering musings here is that I was just chatting with Matthias Wagner, who is the Founder and CEO at Flux. (See also Amelia Dalton’s Fish Fry podcast with Mathias: From Dreams to Reality: How Flux is Paving the Way for Ultimate Customization). Matthias was telling me about the latest offering from the folks at Flux, which involves… you’ve guessed it… Agentic AI. Matthias says that, with this latest release, you can say to the Flux CoPilot something like:
Design a temperature and humidity sensor node with Wi-Fi and Bluetooth capabilities, powered by a USB-C (5V) port, for consumer use. This should be a low-power environmental node featuring a digital T/RH sensor, an ultra-low-power MCU with Wi-Fi and Bluetooth (dual-radio), 2.4 GHz 802.11 b/g/n plus BLE 5.x, powered by a USB-C port (5 V). Include reverse/OVP/UVLO/OCP protection and plan for 0.5–3 A sources.
Then you press the “Go” button and let Flux perform its magic. Flux does this by creating and deploying AI agents to perform its bidding. Think of Flux as the conductor of an orchestra, and the AI agents as the musicians under the conductor’s control.
This new incarnation of Flux designs the circuit, generates the schematic, selects the components, and lays out the PCB. Along the way, Flux may ask questions of you, like “You haven’t told me what you want the design to do if […]” or “I originally selected this component because it’s the most cost-effective, but it’s on back-order from all the suppliers, so we have these options […], which would you like me to use?”
One well-founded point Matthias makes is that many PCB designers tend to stick with the same core selection of 50 to 60 integrated circuits (ICs) that they’ve grown to know and love. This reduces the learning curve, but it may not yield optimal designs in terms of capability and cost. By comparison, Flux has no such preconceptions and limitations (unless you impose them), which means it can explore innovative circuit solutions and part selections.
Matthias tells me that once the layout is complete and you’ve indicated your approval, Flux will ask questions like “How many boards do you wish to produce?” and “Do you want a conformal coating?” and so on. If you wish, Flux will then research PCB fabrication and assembly houses, reporting details such as prices and lead times. Matthias says that Flux won’t, of its own volition, upload the files to the target PCB/PCBA company and place the order… yet… but that the team is working towards this goal.
One comment on the Flux.ai website caught my eye. I’m thinking about the one that says:
The kids of today design their own iPhone apps.
The kids of tomorrow will design their own iPhones.
I’m a big science fiction fan. I’m thinking of replicators in the Star Trek universe—the machines that can create food, drinks, tools, clothing, or other objects on demand, seemingly out of thin air. Now I’m envisaging a future in which custom electronic products can be designed and fabricated on demand. Far-fetched? Maybe not as far-fetched as one might think.
