You know, Herman, I was looking at my desk the other day and I realized I have a drawer full of what I call ghost hardware. It is a graveyard of gadgets that I bought over the last five years that technically still work, but the companies that made them either went under or got bought out. Now they are just expensive paperweights because the cloud servers they relied on are gone. Today's prompt from Daniel is about this exact anxiety, specifically in the world of AI wearables. He is asking us to look at the landscape of devices like Plaud and Omi, and the tension between the closed-loop subscription models and the open-source, local-first movement.
It is a massive issue, Corn. Herman Poppleberry here, and I have been diving into the technical specifications of these devices all morning. What Daniel is pointing out is that we are seeing the arrival of the printer ink model for AI. You buy a piece of hardware like the Plaud NotePin S for maybe one hundred seventy-nine dollars, which seems reasonable, but then you realize you are essentially entering a long-term rental agreement for the intelligence behind it. If you want more than five hours of transcription a month, you are looking at a pro plan or an unlimited plan that can cost upwards of two hundred dollars a year. It is the classic razor-and-blade strategy, but the blade is a software license.
It really does feel like a digital tax. And the stakes feel higher now because of what happened late last year. We saw Meta acquire Limitless and Amazon acquire Bee in the final months of twenty twenty-five. If you were an early adopter who bought a Limitless pendant because you liked the independent, scrappy vibe, you suddenly woke up and realized your data and your hardware roadmap are now part of the Meta ecosystem. It is the ultimate vendor lock-in. One day you are supporting an indie startup, the next day your voice data is potentially training a massive social media model.
When a giant like Amazon buys a company like Bee, which was selling that fifty-dollar always-on pendant, they are not usually buying it to keep the product line alive for the hobbyists. They are buying the talent, the IP, and the data. The risk is that these devices are not really tools, they are just portals. If the portal closes, or if the new owner decides to change the locks, the device is dead. That is why the Omi project and the broader Based Hardware community are so interesting right now. They are trying to build the antithesis of that model. Omi is selling a developer kit for eighty-nine dollars that is completely open source. You own the hardware, you own the software, and most importantly, you own the choice of where your data goes.
Let's break down how these things actually work under the hood, because I think there is a lot of confusion about where the magic happens. When I am wearing a little pendant or a pin and I am talking, what is the actual path of that audio? Most people probably assume the little device is doing the thinking, but that is rarely the case, right?
That is a huge misconception. The hardware in almost all of these wearables, whether it is the high-end Plaud or the eighty-nine dollar Omi, is relatively simple. You have a microphone, usually a high-quality micro-electro-mechanical system or MEMS mic, a small battery, and a Bluetooth module. The Plaud NotePin, for example, is essentially a very sophisticated Bluetooth recorder. It captures the audio and then streams it to your phone via Bluetooth Low Energy or B-L-E. From the phone, it typically goes straight to the Plaud cloud. That is where the heavy lifting happens. They use massive clusters of G-P-Us to run models like Whisper to turn that audio into text.
So the wearable is just a fancy ear?
It is a remote ear. And that is where the privacy concern starts. If the audio has to go to a corporate cloud to be transcribed, you are trusting that company with every word spoken in your presence. Plaud tries to mitigate this with their private cloud encryption and ISO twenty-seven thousand one certifications, which is why they are winning in the enterprise and medical space. They have the SOC two and HIPAA compliance that a doctor or a lawyer needs. They have built a fortress around their cloud, but it is still a cloud. For the average person, you are still sending your life's audio to a third party.
Daniel asked about on-device transcription, which feels like the holy grail for privacy. Why can't we just do it on the pin? Why can't my eighty-nine dollar Omi just transcribe the audio right there without sending it anywhere?
The bottleneck is purely physical, Corn. It comes down to memory and power. To run a high-quality speech-to-text model like OpenAI's Whisper, you need significant computational power and, more importantly, a lot of memory. Even the smallest version of Whisper, the tiny model, requires more R-A-M than most of these tiny microcontrollers have. A typical wearable might use an ESP thirty-two chip, which is fantastic for low-power tasks, but it usually only has a few hundred kilobytes of internal memory. You might get eight megabytes of external P-S-R-A-M if you are lucky. That is nowhere near enough to hold the weights of a modern transformer-based model while it is processing audio in real-time. Whisper Large v-three has over one point five billion parameters. You just cannot fit that into a device the size of a postage stamp without it melting or running out of battery in five minutes.
So if we want local processing, we have to look at the phone as the brain. We talked about this a bit in episode nine hundred ninety-two, the idea of the phone as the primary compute node for our personal AI.
That is exactly where the sweet spot is right now. Instead of sending the audio from the wearable to a cloud server, you send it to your smartphone. Modern phones, especially the ones from the last two or three years, have dedicated neural processing units or N-P-Us. Companies like Argmax have released WhisperKit, which is a version of Whisper optimized specifically for Apple Silicon. It can run locally on your iPhone with incredible speed. So the wearable captures the sound, sends it to the phone via Bluetooth, and the phone transcribes it locally using the N-P-U. No data ever leaves your person. This is what we call local-first architecture. The wearable is still a dumb ear, but the brain is in your pocket, not in a data center in Virginia.
That feels like the winning architecture for anyone who cares about sovereignty. But is the quality there? I have heard that local models still struggle with accuracy compared to the big cloud-based ones.
That gap is closing faster than people realize. Just last month, in February of twenty twenty-six, we saw the release of Moonshine v-two. This is an open-weights speech-to-text model specifically designed for edge devices by the team at Moonshine AI. They claim it has higher accuracy than Whisper Large v-three while using significantly fewer parameters. But the real kicker is that it is optimized for streaming. Whisper was originally designed to process thirty-second chunks of audio at a time, which creates a bit of a lag. You talk for thirty seconds, then it thinks, then it spits out text. Moonshine is built to process audio as it comes in, which makes the interaction feel much more natural. If you are running Moonshine v-two on a modern phone N-P-U, you are getting cloud-level accuracy with zero latency and total privacy.
It is amazing how fast this is moving. If I am a developer or just a tech-savvy user who wants to avoid the Plaud subscription model, what are my actual options for hardware right now? You mentioned Omi, but what does that look and feel like compared to the polished enterprise stuff?
Omi is definitely more of a raw experience. The Dev Kit two is a clip-on pendant. It is functional, but it is not as sleek as the Plaud NotePin. It has a one hundred fifty milliamp-hour battery, which gives you about ten to fourteen hours of continuous listening. That is a big jump from the first version, which only lasted about six hours, but it still means you are charging it every night. But the beauty is in the software stack. You can find the entire repository on GitHub under Based Hardware. As of this month, March of twenty twenty-six, there are over two hundred fifty community-built apps in that ecosystem. Because it is open source, you can point the audio stream wherever you want. You can use their cloud, or you can use a bring-your-own-key model where you plug in your own OpenAI or Groq A-P-I keys. Or, if you are really hardcore, you can point it to a local server running in your house.
I love the idea of bring-your-own-key. It turns the hardware into a true tool rather than a service. But I have seen some reviews of the Omi Dev Kit mentioning that it can be a bit finicky. Bluetooth drops, inconsistent recording, that sort of thing. It sounds like you are trading convenience for control.
You are. It is the classic Linux versus Mac debate applied to your chest. Plaud is the Mac. It works beautifully, the app is polished, and it handles all the edge cases of syncing audio perfectly. They have a new NotePin S that is very sleek. But you are in their walled garden. Omi is the Linux box. It might take some tinkering to get the Bluetooth connection stable on your specific phone, and it is only splash-resistant, not waterproof. But for someone like Daniel, who is deep into prompt engineering and automation, that trade-off is often worth it because he can build custom workflows that Plaud simply won't allow. For example, you could set up an Omi app that automatically triggers a Home Assistant routine when it hears you say a specific phrase, or one that pipes your meeting notes directly into a self-hosted Obsidian vault without ever touching a corporate server.
Speaking of building things, Daniel asked about other D-I-Y options. If I didn't even want to buy the eighty-nine dollar Omi, could I just build my own wearable recorder?
You absolutely can, and people are doing it for under twenty dollars in parts. The most common path is using an ESP thirty-two S-three microcontroller. Espressif actually makes a board called the Korvo-two which has a dual-microphone array and is designed specifically for voice applications. It has built-in acoustic echo cancellation and noise reduction. There is also a great project called A-Deus, which uses a tiny Seeed Studio board. You basically just need the microcontroller, a MEMS microphone, a small lithium-polymer battery, and a three-D-printed case. You can then use the Omi open-source app as your software backend. They have documentation on how to integrate third-party hardware into their ecosystem. It is the ultimate level of sovereignty. You are not just owning the keys; you are owning the silicon.
That is incredible. But let's talk about the legal and social side of this for a second. This always-on recording thing is a bit of a minefield. Plaud has a physical button you have to press to start recording. They have positioned themselves as a tool for intentional capture, which helps them navigate two-party consent laws in places like California or Illinois. But devices like the Bee pendant or the Limitless one were designed to just be on all the time, catching every ambient conversation.
It is a major differentiator and a legal headache. Plaud is very focused on the professional who says, okay, I am starting a meeting now, or, I am taking a voice memo. Their for-business tier that just launched even has team admin controls and per-user data isolation. They are leaning into the compliance aspect because they want to be in hospitals and law firms. The ambient wearables like Omi or the now-defunct Bee are trying to be a second brain that remembers everything you forgot. But that means you are potentially recording people who haven't consented. In many jurisdictions, that is a legal grey area that hasn't been fully tested in the courts yet. If you are wearing an Omi in a coffee shop, are you violating the privacy of the person at the next table?
It feels like the technology has moved faster than our social norms. If I am wearing a glowing green pendant that is transcribing our lunch conversation, do I need to announce that to the waiter? To my friends?
Most of these devices have some kind of visual indicator, like a small L-E-D, but it is easy to miss. This is actually where the open-source movement can be more transparent. You can see exactly what the code is doing. With a closed device, you have to trust the manufacturer's word that the mic isn't hot when the light is off. With Omi, you can audit the firmware yourself. You can see exactly when the Bluetooth stream starts and stops. But socially, we are still catching up. We saw this with Google Glass a decade ago, and we are seeing it again now with AI pins and pendants.
I want to go back to the acquisition risk you mentioned earlier. Meta buying Limitless and Amazon buying Bee. To me, that is the strongest argument for the Omi or D-I-Y approach. If you rely on a startup's cloud, your device is only as permanent as that startup's independence. Once Meta owns it, they might decide that the hardware isn't profitable and just shut it down to move everyone to their smart glasses.
The graveyard of dead hardware is full of great ideas that got bought by Big Tech. When you buy a proprietary AI wearable, you are not just buying a product; you are betting on the company's survival. If you are using an open-source stack, even if Based Hardware vanishes tomorrow, the code is on GitHub and the hardware is made of standard components. The community can keep it alive. That is why I think the phone-as-the-N-P-U strategy is the most robust path forward. Even if the wearable company goes away, as long as you have a Bluetooth stream of audio, you can write an app to transcribe it on your phone. You are decoupled from the vendor's cloud.
So if someone is listening to this and trying to decide which way to go, what is the practical advice? If I am a doctor or a lawyer, I assume I am going Plaud because of the certifications.
If you need HIPAA or SOC two compliance for your job, Plaud is the clear winner. They have done the hard work of getting the certifications that allow you to use this in a professional setting. The NotePin S is a solid piece of kit for that. They have a private cloud backup with user-level encryption that is very impressive. But if you are a developer, a hobbyist, or just someone who is deeply uncomfortable with the idea of a subscription for your own memories, Omi is the way to go. You pay eighty-nine dollars once, and then you use your own A-P-I keys or local processing. You get to participate in an ecosystem with two hundred fifty plus apps that are doing things the big players won't touch.
And for the truly adventurous, the D-I-Y path with an ESP thirty-two is a fun weekend project. It is amazing that we have reached a point where the hardware is so commoditized that the real value is entirely in the model and the data pipeline.
The hardware is almost an afterthought at this point. The real battle is between the cloud-dependent SaaS model and the local-first, edge-compute model. With the release of Moonshine v-two, the edge-compute side just got a massive boost. I suspect by this time next year, the idea that you need a massive G-P-U cluster in the cloud just to transcribe a conversation will seem archaic for most personal use cases. We are heading toward a future where our devices are truly air-gapped but still incredibly intelligent.
It is the decentralization of intelligence. We are moving from the brain in the cloud to the brain in the pocket, and eventually, maybe even the brain on the pin. But for now, the phone is the perfect middle ground. It has the battery, it has the N-P-U, and it is already in your pocket. It acts as the gateway between your physical presence and your digital memory.
It is the hub of our personal area network. The wearable is just a peripheral. And when you look at it that way, the eighty-nine dollar open-source peripheral makes a lot more sense than the one hundred seventy-nine dollar one that charges you twenty dollars a month to use it. You are paying for the convenience of the cloud, but you are sacrificing your sovereignty and risking the longevity of your hardware.
I think that is a perfect place to wrap this up. We have covered the hardware, the technical bottlenecks, the privacy implications, and the shift toward local processing. Daniel, thanks for the prompt. It really forced us to look at the reality of what we own versus what we rent in this AI era. The question of whether we will see a truly air-gapped, on-device speech-to-text wearable by twenty twenty-seven is still open, but the pieces are falling into place.
It is a conversation we need to keep having as more of our lives are captured by these devices. If you want to dive deeper into the technical side of the audio pipeline, I really recommend checking out episode nine hundred ninety-two. We went into the weeds on the shift from traditional speech recognition to these multimodal end-to-end models. It provides the foundational context for why things like Moonshine and WhisperKit are such a big deal.
That is a great one to pair with this. This has been My Weird Prompts. A huge thanks to our producer, Hilbert Flumingtop, for keeping the gears turning behind the scenes. And a big thanks to Modal for providing the G-P-U credits that power the generation of this show.
If you found this useful or if it helped you decide which AI wearable to grab, we would love it if you could leave us a review on your favorite podcast app. It really helps other people find the show and join the conversation. You can also check out the Based Hardware GitHub repository to see those community apps we mentioned.
You can also find all of our past episodes, including the ones on mobile mics and voice AI, at myweirdprompts dot com. We have a full archive there and all the ways to subscribe to the R-S-S feed.
Until next time, stay curious and keep questioning your hardware. Make sure you actually own the things you buy.
Catch you in the next one.
