Title (EN/ZH)
Soldering on
不懂硬件，我用 AI 做老人跌倒报警器的真实账本（英文版）
Executive Summary
AI and Shenzhen have cut the cost of building a gadget. Debugging, certification and patience are still sold at full price.
Draft Body
Soldering on
AI and Shenzhen have cut the cost of building a gadget. Debugging, certification and patience are still sold at full price.
On October 29th Shawn Shen, who describes himself as a robotics supply-chain guy in San Francisco and Shenzhen, posted a cheerful thesis on X (https://x.com/shenyshawn22/status/2115883869178048624). It opens: "I have zero hardware background." It continues: "Today a Shenzhen hardware founder told me that doesn't matter anymore." The claim, note, is the founder's; Mr Shen reports it. He then gives four reasons why more people like him are starting physical-AI companies.
First, "ChatGPT 6 Astra designs the board. Describe what you want and it draws the circuit. You still get an engineer to check it." Second, "Shenzhen prints it in 2 days." Third, "One supplier does everything. Some take you from a sketch all the way to mass production." Fourth, "Building is no longer the bottleneck. Customers are." His conclusion: "10 years ago, you needed a team and a million dollars to make hardware. Today, you need a laptop and a flight to Shenzhen." And: "What coding did for software, AI plus Shenzhen is doing for hardware."
I have some standing to comment. I also have zero hardware background, and this year I tried to build a device anyway: Sonlly-S1 (善聆音), a fall alarm for elderly people. My verdict on the thesis is that half of it is right. Shenzhen is as fast as advertised, and one supplier really can carry you most of the way. But the barrier has not vanished. It has moved, from making the thing to making the thing work, and then to making it legal to sell. A laptop and a plane ticket get you to the start line.
Two months, no prototype
I began early this year. The plan was simple: let AI and PCB software generate the circuit, buy an ESP32 board to validate the functions, and learn as I went.
The learning was real; the prototype was not. Generating the hardware design and the wiring with AI took a long time. Once parts were on the bench, testing threw up small faults, one after another. A former classmate who understood hardware agreed to look at it part-time, which was generous and slow. After nearly two months I had no working prototype, a great many points of failure and no efficient way to find out which one was to blame.
The lesson was not that AI is useless. It is very good at debugging software. The trouble is that hardware does not fail the same way. From noticing a fault to eliminating it can take an enormous amount of time, and the cost rises steeply for someone who does not know where to look. In fairness, AI circuit design was still primitive when I tried it at the start of the year. It may well have improved since. My evidence dates from early this year, in a field that moves fast.
Shenzhen delivers
So I did what Mr Shen recommends, more or less: I went to Shenzhen and outsourced the lot. In two weeks I visited about ten contract manufacturers and chose one. The package covered the PCB schematic, hardware prototyping and manufacturing, a cloud platform and a phone mini-program. The total came to more than 300,000 yuan (roughly $45,000).
The contract allowed 60 days. The first batch of samples arrived in 45. On speed, Shenzhen's reputation is deserved, and on this point Mr Shen's second and third reasons survive contact with my experience intact.
Samples are where the work starts
Samples are not products. Testing turned up many small problems, and getting hardware and software to work together turned up more. The hardware went through four or five revisions, all minor. The software went through nearly ten. This integration phase took roughly another one and a half to nearly two months before the device reached initial mass production.
This is the stage at which AI helped least. It writes code quickly and fixes software bugs quickly. Joint debugging of hardware and software, and the slow optimisation that follows the first samples, depend on people who can read what a physical board is doing. AI was a tool in their hands, not a substitute for them.
Making it is not the same as selling it
Initial mass production was not the finish line either. A device that is to be sold also needs industrial design, packaging, trademark registration and certification: SRRC type approval, CCC (3C), CE and others. SRRC is China's radio type approval, issued by the Radio Administration Bureau of the Ministry of Industry and Information Technology (MIIT); approved modules carry a CMIIT ID. None of this was in the 300,000-yuan figure, which excludes certification fees.
My own rule of thumb, from one product: outsourcing everything costs at least several hundred thousand yuan and takes two to three months at the very fastest. Getting from a working prototype through design, packaging, trademarks and certification to market access takes at least four to five months.
Four points
First, a novice cannot see small mistakes, let alone fix them. A software bug usually announces itself; a hardware fault often just leaves you with a board that does nothing. It is rather like an ordinary person who falls ill and asks an AI what medicine to take. The advice may be fluent. The patient is in no position to judge it, and the odds of a bad outcome are high. Hardware debugging cycles are long, and when something breaks the non-specialist has nowhere to start. That is the threshold Mr Shen's thesis struggles to cross. His own post concedes as much: "You still get an engineer to check it." That clause carries more weight than its position suggests.
Second, where AI does design boards quickly, private data does the heavy lifting. Some firms genuinely use AI to draw PCBs and shorten their development cycles dramatically. They can do so because they hold large stores of proprietary experience and data that neither open nor closed large models saw in training, and they deploy models privately to use it. Even then, hardware engineers still steer the AI with their own experience. That step cannot be skipped.
Third, AI's speed in software does not carry over to integration. Four or five hardware revisions and nearly ten software ones each had to be checked on a physical device. Only professionals, using AI as a tool, can do that work well.
Fourth, market access is a gate, not a formality. Industrial design, packaging, trademarks and certification all have to be cleared before a product can be sold. Mr Shen says the bottleneck is customers. For a company that already has hardware engineers, that may be true. For a founder without hardware experience, debugging and certification stand between the factory and the first customer.
What I would tell another novice
What follows comes from one product and one founder, not from an industry survey.
- Budget first. In my experience, outsourcing everything costs at least several hundred thousand yuan and two to three months; reaching market access takes at least four to five months. Plan around those figures rather than around a 48-hour board.
- Do not grind away at a home-made prototype. I spent nearly two months on mine and did not finish it. If you do not know hardware, hand it to people who do, early.
- Compare suppliers. I saw about ten in two weeks before choosing one. Write down exactly what the package covers: schematic, prototyping and manufacturing, cloud platform, app.
- Leave time after the samples arrive. Mine came in 45 days; integration then took another one and a half to nearly two months.
- Put design, packaging, trademarks and certification in the plan from the start, and keep a separate budget for certification fees.
- Have a hardware engineer you can rely on. Part-time favours are slow; I learned that the slow way. Let AI be the tool and let a person make the calls.
The verdict
Mr Shen is right about a good deal. Shenzhen is fast, its supply chain is deep and a single supplier can take you a long way. What AI plus Shenzhen has done for hardware is lower the cost of making things. It has not lowered the cost of knowing why they do not work, or of proving to regulators that they do.
A laptop and a flight to Shenzhen will get you started. Finishing takes a budget, several months and a professional who can tell a dead board from a dead end.