Everyone is vibe coding. Is vibe prototyping next?

It was only on February 2, 2025, that the term "vibe coding" came up in a post on X by Andrej Karpathy, even though the concept feels as if it has been around for years, but now there is new "vibing" emerging: vibe prototyping. Currently, there is no established terminology for the application of AI in EDA (Electronic Design Automation) or within the hardtech sector in general, so I will use the term "vibe prototyping".

In the last 12 months we have seen the same developments in EDA tools that we saw in regular AI tools a little bit earlier. EDA no longer relies only on task-specific machine learning, but rather on multi-agent orchestration. Just like in coding tools such as Claude Code and Cursor.

Platforms like Synopsys are advancing rapidly, and there will come a time when you can vibe engineer chip design, just as you vibe code software today.

Atoms vs. bits.

One might assume that, if software is going to be solved soon, hardtech will follow closely behind, but that overlooks the fundamental differences between the two.

LLMs (Large Language Models) benefit from a massive, universally accessible backlog of human digital history. better said: they operate in a closed loop, where they consume digital bits and also output digital bits. They do not have to cross boundaries of the physical reality.

However, AI that designs physical hardware is outside of this loop. If we ask algorithms to generate (or model) physical solutions without the necessary data from the physical world, it's like asking AI to learn how to swim without ever touching water. To achieve the same level of success in hardtech applications, we need to give the AI access to this water.

One option is to use HIL (Hardware-in-the-Loop), which essentially means having a physical prototype connected directly to the digital system. This way, you get a constant flow of new data. You can also leverage devices already in the field, much like Tesla uses its car sensors. It's actually one of the smartest ways to gather physical world data, because you have already sold the product for a profit and are now getting a constant flow of valuable data as a bonus.

At the end of the day, there is no method for acquiring data from the physical world that comes even close to the speed we have in the digital world.

Hardtech is, quite literally, hard-and acquiring data is not the only hard problem new hardtech startups face. You cannot revert your mistakes in the real world as easily as you can in software. If you burn a microcontroller, it stays burned; you cannot simply 'git revert' it. Consequently, developing a physical product requires significantly more resources.

So we are not getting vibe prototyping after all?

Actually we are, just a little bit slower. We cannot skip the implications of the physical world, but eventually we will have tools to fully model chip designs and generate synthetic data through more advanced simulations such as digital twins.

When that time comes, the world will be in shock-an even greater shock than AI already is today. As technology progresses, we need to be where next big disruption happens. The physical world is the next disruption and when you think about it, its scale will be vastly larger than the digital world.