Interesting that Gemini said it was infeasible. It should be aware that using a Pico W as a transparent ethernet bridge has been done several times over in open source projects, for example on BlueSCSI (emulating a Daynaport SCSI-Ethernet adapter) and PicoMEM and my own PicoGUS project (emulating an NE2000 Ethernet adapter).
My experience with questions to LLM is that they mimic reddit a lot: ask them an apparently simple thing that is not possible, and they will bend backwards to tell you it's possible and give you tons of convoluted possible solutions. Ask them a thing that is possible but complicated, and they will be overly dismissive. The good thing however is that between those two banks there's a wide river of utility...more often than not they'll at least mention things I haven't considered before.
I cannot fathom how people fall into the trap of thinking that LLMs have a consistent world view on any of the users bugbears (what would be an opinion in a human). CoT is a work-around and doesn't persist across sessions. "You are absolutely correct" and "I should not have deleted your production DB" prove the anthropomorphic facade only a facade.
shallow immediate confident authoritative answers; upon inspection falls apart immediately. sometimes can be beaten by repeat questioning into wisdom. by accident, beauty witnessing unfathomable NPC stupidity, but also invokes a hidden genius operating as though designed to troll first
> Can you use a raspberry pi pico W as a USB WiFi adapter
> Yes, it is possible to use a Raspberry Pi Pico W as a USB Wi-Fi adapter, but it is a project that requires custom firmware and a clear understanding of your goals.
Then goes off and lists the things you’ll need which at a cursory glance seems like good starting points.
Exactly, bit banging an 8-bit bus isn't that different from pushing the data out of the USB port. It would be great to try an LLM trained on pre-1900 documents and ask it if powered flight is possible.
English around 1800-1900 reads almost like modern English. At least if the author didn't write in an overly pretentious style, which was more common back then
Yes, but. I asked Perplexity Pro the same fact question (sports) three days in a row after correcting its error the first day and having the correction acknowledged with a promise to get it right in the future. Same error on second day, same promise. Third day: correct. Perhaps it needs to be in the "Slow" classroom....
I'm not sure even treating it as a guidance is the best approach, generally you should challenge it a lot (which is what CoT kind of emulates), you never know what it is confabulating and when it is correct.
Unless there is a hardware limitation or the hardware does not support it, anything in software is possible.
Gemini and all these other LLMs are designed to convince you that they have "awareness" which they do not have any of the sort. They are neither sentient nor do they have consciousness
Bad take. Some things are feasible and some things are not, "anything is possible" is a useless framework. Example: go convert two smartphones to communicate p2p over their 4g radios - it's all software!
LLM "awareness" is similarly irrelevant. They process information usefully, in a way grounded in reality, and that's that.
You seem to be all over the place, most of it by not reading the parent's comment. So let's break it all down.
> Bad take. Some things are feasible and some things are not, "anything is possible" is a useless framework.
It would help if you quoted the entire comment rather than removing the context and further giving a very bad example afterwards:
> Example: go convert two smartphones to communicate p2p over their 4g radios - it's all software!
Nice try. That is a hardware limitation in the 4G radio which is designed to connect to an operator mast. Even if you wanted to do it in software, the hardware does not support that P2P use-case which is what I already said.
> LLM "awareness" is similarly irrelevant.
Exactly. There is no such thing as awareness in LLMs.
The parent comment I replied to believed that an element of awareness had to be present to give an answer because this was done "several times over" in open source projects. Which that is inaccurate in the context of LLM research.
>> "It should be aware that using a Pico W as a transparent ethernet bridge has been done several times over in open source projects..."
> They process information usefully, in a way grounded in reality, and that's that.
Useful to those who know when it is either mostly correct or outright wrong.
Clearly in this example, Gemini doesn't even know if its own answers are grounded in reality and consequently people using them are unable to determine if the results they bring are true or not and there are countless examples of that.
Bits is bits - so sorry you couldn't figure out the right bits to make the hardware do what you want, but those bits do exist and they are findable in principle - it's still all software! Your lack of knowledge of the signing keys is qualitatively no different than your lack of knowledge of any other undocumented proprietary information required to program the chip. Frankly you'd struggle even with the keys.
Infeasible? Oh my yes. Impossible? No. That's the point - there is a distinction.
In the retro-computing world (Oric-1/Atmos) we are using the PicoW as a way to get our old machines on the Internet, not by implementing USB-CDC drivers on the old machines, but by having the PicoW imitate an old-school modem interface, albeit one that 'dials in' to a web site address instead of another modem ..
It's pretty cool, I have to say - especially for us old Oric folks, who never did quite get a full BBS ecosystem for these machines, unlike others. Its on like donkey kong now though, thanks to wonderful PicoW projects like this!
Author here: good point - I do hope the retro computing community enjoys this tool. It does remind me of the Espressif line if ESP8266 and ESP32 devices which have Wi-Fi devices which are available on UART and respond to AT modem commands. Pico-usb-wifi is a nice tool which some clever YouTubers will hopefully be able to use to get old internet appliances, or systems that didn't have Wi-Fi, online.
Personally, I rescued a dozen PogoPlugs from the trash a decade ago and used one to practice self hosting a word press instance. Having a dongle like this would have allowed me to stick it in a corner and run something like librespot or snapcast. Instead, I eventually sent them to the recycler when I moved.
I've been using an ESP8266 as a wifi/modem adapter with my original Commodore 64. Works really well! You may be able to get away with something cheaper if you push.
I had the same idea for a Plan 9 USB WiFi using an ESP32. You serve the wifi device as a ether(3) device which negates the need for janky side band config as the config is done over the same 9P interface. Never got around to it.
Actually, all of them are fallible, very incompetent machines that are good at writing text. They're not people, they don't have any qualities of people except the really bad ones, and they are absolutely miserable at reasoning.
The only people I know who use Gemini are unemployed.
The only people I know who use Claude vibe-code everything, often including their communication -- they probably let Claude kiss their kids goodnight.
Everyone else uses ChatGPT, and the world is worse off for it.
They are machines. "it" is the only acceptable pronoun, and personifying these machines adds emotion into the discussion and the use of the tool. They are not people. They do not behave like people. If you feel like they do, and you're e.g. autistic, that's entirely fair, so please take my word that they do not behave sufficiently like people in any way.
Nothing they do mirrors the behavior of engineers. They instead mimic the language of engineers. I understand that this is all it takes in a lot of circles to gain respect, which is quite a sad state for those circles, but that doesn't mean its a universal experience.
I'm actually tired that people put stupid questions in ChatGPT and then present in with a straight face as a source of truth. Sometimes it hallucinates completely, sometimes the conditions or regulation have changed and it gives false answers and no-one cares. Simple collaborations that was possible before now turn into unnecessary arguing. Some ChatGPT users aren't even aware that LLMs hallucinate, I just pointed it out recently and was accused of mansplaining and being a tinfoil hat.
> Nothing they do mirrors ...
"Mirror, mirror on the wall, who is the fairest of them all?"
Gemini writes pretty shitty code in my experience. We tried it out for a grand total of half a day at work before deciding it wasn't worth our time and switched back to Opus.
ChatGPT writes like it's life depends on it and refuses to correct its own mistakes. It'll figure out a way to write 4k lines for something that could've been done in 500
Buried the lede by calling this a "usb wifi adapter" and not "magic Ethernet adapter that you don't have to plug a cable into". This will come in handy for my systems that are normally air gapped or currently connected to something else, but sometimes I need to get another network connection on there for a few minutes. I don't always have a USB ethernet adapter on hand or enough cable, but I almost always have a Pico around.
Also I posted this comment using one. Speedtest says about 4 Mbps. Surprisingly usable for web browsing though. Very nice tool! I'll be keeping the .uf2 file around for sure.
Author here: Thank you! It's a nice thought that something I made will become a trusted tool for IT professionals.
Now I need to make a nice 3D-printed case.
I admit, I almost couldn't believe it when I turned off my regular blazing fast 802.11ax NIC in the KDE network manager and switched to the pico-usb-wifi and I watched the YouTube "stats for nerds" buffer fully fill up and playback was totally smooth.
I would like to ask the author, what do they mean when they say 'I spent 2 days and 1M claude code tokens'
Did you completely vibe code this? or did you create a spec and asked claude to implement that? How many times did you need to course correct claude when it deviated if it did?
How much of your own experience as working in WiFi industry was useful while building this project? And what would be different if you had worked on this project without that experience?
Author here: this is actually a "side quest" which I alluded to in the README.adoc and another comment. I've had a few Spotify Car Things (SuperBird) sitting on my desk that I have not touched in a year because I was waiting for the community to come out with a non-tethered approach to reclaiming them. All of the existing continue to either require a host PC, phone, or RPI Zero. I want to deploy these throughout my home and have them usable without my PC or phone nearby - likely with just a Home Assistant dashboard. I spent the first few hours cloning repositories of existing firmware, and examining other approaches, such as Bluetooth low energy with ESPHome Bluetooth proxies, but they ended up being dead ends. I took a compromise that a small wireless dongle would be acceptable. I then began examining the Linux source tree to see if it was possible to flip the USB from client to host and spent a considerable amount of time recompiling BishopDynamics' SuperBird NixOS distribution and building a test dashboard to show the IP networking information, because I knew that once I flipped the device's USB mode, I would lose access to the device. SuperBird took almost four hours to compile on my low end CPU - I let it compile while I hit my five hour Claude Code quota.
I also did not look forward to trying to build in wireless functionality in to NixOS, because it also took 30 minutes to flash the device - I had to lower the transfer block size and add more retries in the flash utility because it kept crashing.
ultimately both the SuperBird work and the RPI pico firmware was about 300 prompts of investigation, research, creating potential approaches, architecting, designing, debugging, testing/validating, and documenting.
Regarding your last two questions... I started out 25 years ago by studying for a CCNA in high school and joined the industry just as first generation 802.11ac devices began hitting the market. From my position, I've worked on every layer from the SoC, PCB, to software. I want to be very careful and say software estimation is very difficult. The main bug (issue #1) I encountered would have been very difficult to diagnose and resolve ten years ago using traditional development workflows.
I'm only starting out in embedded software domain, and the way you describe your 2 day side quest feels like Dumbledore showing some magic trick to Harry.
I've often thought that if I ever did get around to publishing any of the stuff I work on that I would probably also publish any AI usage verbatim. Like, the full logs of all of the chats, at least as a sign of "here was my thought process" and "here's how I convinced the text machine to give me the right text". That somehow feels more honest than "I made this thing, but actually I vibe coded this thing."
Author here: I will consider writing a blog about ir, but it will require a lot of work on my part, because as I mentioned, this was a digression/side quest from the main task of hacking my Spotify Car Thing. It might work better as a YouTube video.
I’ve been looking for a good solution for doing the exact opposite, being able to connect stuff through USB in my bench, and see them pop up in my office desktop as if they were usb devices.
The closest I’ve gotten is using a raspberry pi in the workbench, but for some weird devices that’s sometimes not good enough.
They're much cheaper than the competing devices from Digi and they've been bulletproof for me. I've got some out there running >10 years.
I've even done stupid stuff like hung a USB Ethernet adapter off of one and made it a NIC on my local PC to talk to another of the same unit hanging off that USB NIC (just to be cheeky-- not for actual use). Stacking things on top of other things is fun.
My only complaint is I wish they did PoE. I use a cheap PoE splitter for that but it would be nice not to have to do that.
If you have CAT5/6 between the locations, USB-over-CAT5/6 adapters are inexpensive and work quite well. I've used OREI's successfully over >60m. But there are many options, with varying limits on cable length.
There USB over IP. I think Linux supports it. It could be possible to have Pico implementation, use WiFi like this, and connect to computer or another dongle.
Author here: I have wanted to make a wireless USB mouse and keyboard switch out of RPI Pico Ws for a long time. I would go a layer deeper than USB/IP, because I want to cache the USB handshake to keep the connection enumerated on multiple systems without dropping.
Can you have the Pico operate as an access point? Would love to be able to use this to connect over wifi to a printer (printer in client mode), with the printer and macos talking directly over IP without needing to configure any other routing/forwarding on macos.
No. The printer cannot connect to the wifi network that the macbook is connected to. I don't want to expose a USB printer to the network. I just want to allow the printer to connect to an AP created by the Macbook/PiZero/ESP32/insert_cheap_widget device and then allow the Macbook to connect to the printer's IP.
That's what I meant by the Macbook and the printer connecting to the same AP. The Pico is the AP, and both your printer and the Macbook connect to it. (You could also use a cheap off the shelf travel router)
What the Macbook can't do is have multiple wifi connections at the same time, so you'd have to disconnect from its primary network (which also rules out the Macbook serving as an AP that the printer connects to).
In a similar but opposite vein, I am going on a vacation and I wanted to share the stupidly expensive internet in my room at night with the family so I am likely bringing a raspberry pi to have as a travel router attached to my Mac. In this case, I can use the RaspAP project: https://raspap.com/
This is slightly different in that I do want a NAT.
You can use it as a repeater, so the whole family can just use the same network/password we use at home. And it is so small, you can run it from a power bank for hours.
I don't have any idea of what is the purpose of this.
Do you want to share an USB device across the WiFi?
If so, why not use the USB-IP protocol? It is already part of the Linux kernel, has implementation for Windows and doesn't require additional hardware.
Totally pedantic deep dive: the Pico has a USB FS PHY (the numbered versions don't directly correlate to line rate, and FS was defined in 1.0) - the signaling rate is 12Mb/s, but it's impossible to achieve sustained transfers at this data rate under any circumstance:
- USB FS has a 1ms frame limit (HS is 125µs)
- UBS FS Bulk is thusly limited to ~19 transactions x 64 bytes maximum (HS is 512)
- USB FS Isochronous can do 1023 byte transfers, but you can only fit one of those in a 1ms frame (resulting in a giant quantization hole in the packet)
- Focusing on bulk only: the token packet, ACK handshake, inter-packet bus turnaround time minimums, framing bits, CRC bits, and periodic FS SOF packets mean that the actual theoretical maximum data rate is ~81% of the signaling rate
- Bit stuffing optimality issues (required for clock recovery) eat an additional several percent on most data, up to ~17% on pathological data
Therefore: ~9.5Mb/s is the best theoretical data rate that can be obtained with optimal host and device IP and an ideal application layer.
Realistically, ~8Mb/s is the most one can expect on real hardware with an ideal application (and this is optimistically high in my experience.)
Author here: thank you for this deep dive. I saw peaks of 5.9 Mbps, it might be possible to increase throughput by allocating more memory. The link between the rp2040 and Infineon Wi-Fi device is 31.5 Mb/s. It might be possible to push that out over PIO to an external USB 2.0 PHY. Apparently the Wi-Fi device is capable of higher speeds. Who knows what will be the limiting factor. I'll leave it to someone else to experiment, my focus was the stock hardware without any modifications.
> It might be possible to push that out over PIO to an external USB 2.0 PHY.
I doubt you will be able to achieve USB HS via a ULPI PHY without an FPGA tunneling the USB SIE for you and handling the tight turnaround requirement (in which case, you might as well give it the entire MAC workload or just use an MCU with an integrated HS PHY.)
The ULPI clock is a 60MHz free-running PHY-sourced signal (typically - some PHYs allow MAC-sourced clocks and this would likely be a little easier), the PHY is entirely in control of when the MAC can transmit (DIR pin), and the protocol is unforgiving from a timing standpoint.
It's one of the neat features of the AsciiDoc language. The user is able to change captions mid document, in this case :figure-caption:. AsciiDoc and Antora are things I've invested a lot of my time into
I'm kind of fascinated by the first diagram on the page. It sits so firmly in the uncanny valley for me and I can't put my finger on why. By itself every part looks ok and normal, but as a whole it just screams AI to me. I don't know if its the color choices or the composition or something else. It all just feels that little bit off.
I mean, I know its AI, the page says so itself, no one is trying to hide it. But it also just gives me AI vibes on such a subliminal level that I can't figure out why.
To me there's something about the "AI Slop Font" where it instantly triggers that uncanny valley. I guess since it's kind of an average of a bunch of fonts so it's somehow familiar but also unfamiliar at the same time.
Oooooh, now I'm thinking... you could design a simple circuitboard that holds multiple picos (surface mounted) and uses the USB data pads on the back to pull all the USB ports out to an onboard USB hub basically allowing you to add a multitude of wifi adapters to a project in one USB cable. Would be great for War Driving!
I think this is more of a "can you do this?" rather than something useful. Sure you can have Pico with this firmware around, but you can have much smaller and faster dongle instead.
Pico is an amazing thing tho, so many weird firmwares that are useful in a pinch.
Meta comment to your meta comment : this is a very personal opinion but I cannot stop myself from thinking less about your message (even though I agree with it), just because of that edgy wannabe way to go about naming Microsoft. You want to say Microsoft say that, you want to say Github say that but if you say a vaguely insulting name only to have to specify what you actually mean... It doesn't bring anything to the conversation, except lower the level, and if a company has flaws as big as Microsoft you don't need to resort to childish name calling to expose them. This is my personal opinion.
> just because of that edgy wannabe way to go about naming Microsoft.
I think it stopped being edgy and transitioned to a symbol of user frustration the moment their CEO tried to address the slop without acknowledging their customer pain.
It is not much different from the micro$oft used here often.
one million Claude Tokens (assuming you are on opus) = 5 USD = the very dongle you tried to replace.
Add the cost of the rasberry pico, you'll have an easier time buying the wifi dongle. The project is cool thought to learn about networks, NAT, Proxys, ect...
No, it's not really easier to buy a Wi-Fi dongle. My target device is the Spotify Car Thing and SuperBird doesn't have Wi-Fi components. My Claude Code Pro subscription was idle, so it cost me nothing. Also, according to an article from Tom's Hardware from two years ago, four million Picos have already shipped, so I've unlocked this ability for let's say 500,000 devices. Finally, my day job is in the Wi-Fi industry... this wasn't a learning exercise.
The author's tone when they discuss the cost of the project is self-deprecating. They know it would have been simpler to just buy one.
But also, the author has given the community a great gift, both directly (the blog post and the project!) and indirectly (the idea: what else can be implemented in similar ways).
Author here: Thank you, this is one of the nicest things anyone has ever said about my work.
I achieved two goals today: a front page post on hackernews and for being recognized for my work; or more accurately, for being very good at using AI ;)
Two arm cores at 133 MHz. That's already more powerful than my first computer. For $4. It qualifies as a computer on it's own. It runs Linux just a hacked version with an emulated MMU
Interesting that Gemini said it was infeasible. It should be aware that using a Pico W as a transparent ethernet bridge has been done several times over in open source projects, for example on BlueSCSI (emulating a Daynaport SCSI-Ethernet adapter) and PicoMEM and my own PicoGUS project (emulating an NE2000 Ethernet adapter).
My experience with questions to LLM is that they mimic reddit a lot: ask them an apparently simple thing that is not possible, and they will bend backwards to tell you it's possible and give you tons of convoluted possible solutions. Ask them a thing that is possible but complicated, and they will be overly dismissive. The good thing however is that between those two banks there's a wide river of utility...more often than not they'll at least mention things I haven't considered before.
I cannot fathom how people fall into the trap of thinking that LLMs have a consistent world view on any of the users bugbears (what would be an opinion in a human). CoT is a work-around and doesn't persist across sessions. "You are absolutely correct" and "I should not have deleted your production DB" prove the anthropomorphic facade only a facade.
shallow immediate confident authoritative answers; upon inspection falls apart immediately. sometimes can be beaten by repeat questioning into wisdom. by accident, beauty witnessing unfathomable NPC stupidity, but also invokes a hidden genius operating as though designed to troll first
ya, that is my experience as well
I just did this test with Gemini Flash-Lite 3.1:
> Can you use a raspberry pi pico W as a USB WiFi adapter
> Yes, it is possible to use a Raspberry Pi Pico W as a USB Wi-Fi adapter, but it is a project that requires custom firmware and a clear understanding of your goals.
Then goes off and lists the things you’ll need which at a cursory glance seems like good starting points.
Exactly, bit banging an 8-bit bus isn't that different from pushing the data out of the USB port. It would be great to try an LLM trained on pre-1900 documents and ask it if powered flight is possible.
Great work on PicoGUS.
You can probably try this:
https://news.ycombinator.com/item?id=46590280
does the LLM speak old english ???
English around 1800-1900 reads almost like modern English. At least if the author didn't write in an overly pretentious style, which was more common back then
I don't think people speak old english in 1800s
It's often good to approach what llm said as a guidance, and not as ground truth.
You can ask follow up questions or point to potential feasibility and it will change its answer.
"Guidance" that's often wrong and that will change if I ask a follow-up question? Wow where can I send my money??
Yes, but. I asked Perplexity Pro the same fact question (sports) three days in a row after correcting its error the first day and having the correction acknowledged with a promise to get it right in the future. Same error on second day, same promise. Third day: correct. Perhaps it needs to be in the "Slow" classroom....
I've never used Perplexity Pro, but I would expect the exact same outcome from Claude, as I don't have cross-session memory enabled.
If you do have cross-session memory enabled, I agree this is not glowing performance. If you don't, then I think it's working exactly as intended.
I just asked if cross-session memory is enabled on my account: it replied that it's still being rolled out.
perhaps we should stop personifying organized dirt
It is "grounded in reality" and changes its answer like the weather.
No, Claude, etm are for "Entertainment purposes only", as listed in their TOS.
You know, like the same for horoscopes and psychics.
Perplexity still exists?
I just got major Dotcom vibes
I'm not sure even treating it as a guidance is the best approach, generally you should challenge it a lot (which is what CoT kind of emulates), you never know what it is confabulating and when it is correct.
stopcitingai.com. Been trying to say this as much as possible.
> Interesting that Gemini said it was infeasible.
Unless there is a hardware limitation or the hardware does not support it, anything in software is possible.
Gemini and all these other LLMs are designed to convince you that they have "awareness" which they do not have any of the sort. They are neither sentient nor do they have consciousness
Something doesn’t have to be sentient or have consciousness to answer a technical question like that.
Bad take. Some things are feasible and some things are not, "anything is possible" is a useless framework. Example: go convert two smartphones to communicate p2p over their 4g radios - it's all software!
LLM "awareness" is similarly irrelevant. They process information usefully, in a way grounded in reality, and that's that.
You seem to be all over the place, most of it by not reading the parent's comment. So let's break it all down.
> Bad take. Some things are feasible and some things are not, "anything is possible" is a useless framework.
It would help if you quoted the entire comment rather than removing the context and further giving a very bad example afterwards:
> Example: go convert two smartphones to communicate p2p over their 4g radios - it's all software!
Nice try. That is a hardware limitation in the 4G radio which is designed to connect to an operator mast. Even if you wanted to do it in software, the hardware does not support that P2P use-case which is what I already said.
> LLM "awareness" is similarly irrelevant.
Exactly. There is no such thing as awareness in LLMs.
The parent comment I replied to believed that an element of awareness had to be present to give an answer because this was done "several times over" in open source projects. Which that is inaccurate in the context of LLM research.
>> "It should be aware that using a Pico W as a transparent ethernet bridge has been done several times over in open source projects..."
> They process information usefully, in a way grounded in reality, and that's that.
Useful to those who know when it is either mostly correct or outright wrong.
Clearly in this example, Gemini doesn't even know if its own answers are grounded in reality and consequently people using them are unable to determine if the results they bring are true or not and there are countless examples of that.
So you know what you just said is not true.
> That is a hardware limitation
Sounds like a intentional firmware (aka: software) limitation to me?
If the hardware won't accept custom firmware without signed with a private key, I'd say that's hardware limitation
Bits is bits - so sorry you couldn't figure out the right bits to make the hardware do what you want, but those bits do exist and they are findable in principle - it's still all software! Your lack of knowledge of the signing keys is qualitatively no different than your lack of knowledge of any other undocumented proprietary information required to program the chip. Frankly you'd struggle even with the keys.
Infeasible? Oh my yes. Impossible? No. That's the point - there is a distinction.
It was Gemini Flash, probably an even faster variant optimised for immediate response on search pages.
[dead]
In the retro-computing world (Oric-1/Atmos) we are using the PicoW as a way to get our old machines on the Internet, not by implementing USB-CDC drivers on the old machines, but by having the PicoW imitate an old-school modem interface, albeit one that 'dials in' to a web site address instead of another modem ..
It's pretty cool, I have to say - especially for us old Oric folks, who never did quite get a full BBS ecosystem for these machines, unlike others. Its on like donkey kong now though, thanks to wonderful PicoW projects like this!
Author here: good point - I do hope the retro computing community enjoys this tool. It does remind me of the Espressif line if ESP8266 and ESP32 devices which have Wi-Fi devices which are available on UART and respond to AT modem commands. Pico-usb-wifi is a nice tool which some clever YouTubers will hopefully be able to use to get old internet appliances, or systems that didn't have Wi-Fi, online.
Personally, I rescued a dozen PogoPlugs from the trash a decade ago and used one to practice self hosting a word press instance. Having a dongle like this would have allowed me to stick it in a corner and run something like librespot or snapcast. Instead, I eventually sent them to the recycler when I moved.
I've been using an ESP8266 as a wifi/modem adapter with my original Commodore 64. Works really well! You may be able to get away with something cheaper if you push.
http://tech.guitarsite.de/c64_wifimodem.html
Nice to hear of your success with the C64 .. I'll do that with mine too, one of these days.
Since the Pico W is around 7 bucks in my neighborhood... its already pretty cheap.
I had the same idea for a Plan 9 USB WiFi using an ESP32. You serve the wifi device as a ether(3) device which negates the need for janky side band config as the config is done over the same 9P interface. Never got around to it.
pico-usb-wifi is firmware for the Raspberry Pi Pico W that turns it into a driverless USB Wi-Fi adapter, enumerating as a USB CDC-NCM device.
Hah! That's neat! So much fun stuff to be had with that particular bit of kit.
I recently bought Pico 2 W for DualSense (https://github.com/awalol/DS5Dongle).
Google Gemini is that naysayer senior developer who confidently tells you it can’t be done.
Claude is that easy to get along with smart hard working guy who just gets on with it and builds it double quick.
ChatGPT is the eager senior developer who says it can be done but can’t actually work it out and fluffs it.
Actually, all of them are fallible, very incompetent machines that are good at writing text. They're not people, they don't have any qualities of people except the really bad ones, and they are absolutely miserable at reasoning.
The only people I know who use Gemini are unemployed.
The only people I know who use Claude vibe-code everything, often including their communication -- they probably let Claude kiss their kids goodnight.
Everyone else uses ChatGPT, and the world is worse off for it.
They are machines. "it" is the only acceptable pronoun, and personifying these machines adds emotion into the discussion and the use of the tool. They are not people. They do not behave like people. If you feel like they do, and you're e.g. autistic, that's entirely fair, so please take my word that they do not behave sufficiently like people in any way.
Nothing they do mirrors the behavior of engineers. They instead mimic the language of engineers. I understand that this is all it takes in a lot of circles to gain respect, which is quite a sad state for those circles, but that doesn't mean its a universal experience.
> Everyone else uses ChatGPT
I'm actually tired that people put stupid questions in ChatGPT and then present in with a straight face as a source of truth. Sometimes it hallucinates completely, sometimes the conditions or regulation have changed and it gives false answers and no-one cares. Simple collaborations that was possible before now turn into unnecessary arguing. Some ChatGPT users aren't even aware that LLMs hallucinate, I just pointed it out recently and was accused of mansplaining and being a tinfoil hat.
> Nothing they do mirrors ...
"Mirror, mirror on the wall, who is the fairest of them all?"
Gemini writes pretty shitty code in my experience. We tried it out for a grand total of half a day at work before deciding it wasn't worth our time and switched back to Opus.
ChatGPT writes like it's life depends on it and refuses to correct its own mistakes. It'll figure out a way to write 4k lines for something that could've been done in 500
DeepSeek will just wing it and tell you it's done only for you to find 1 major + 3 edge case bugs.
They're all actually a pile of numbers getting multiplied against each other to sometimes produce output that convinces some humans they're people.
ChatGPT is very good at code-reviewing Claude’s work and finds the howlers in it fairly reliably
Playing off different LLMs against one another in that kind of manner is a good way to expose some first order errors.
Buried the lede by calling this a "usb wifi adapter" and not "magic Ethernet adapter that you don't have to plug a cable into". This will come in handy for my systems that are normally air gapped or currently connected to something else, but sometimes I need to get another network connection on there for a few minutes. I don't always have a USB ethernet adapter on hand or enough cable, but I almost always have a Pico around.
Also I posted this comment using one. Speedtest says about 4 Mbps. Surprisingly usable for web browsing though. Very nice tool! I'll be keeping the .uf2 file around for sure.
Author here: Thank you! It's a nice thought that something I made will become a trusted tool for IT professionals.
Now I need to make a nice 3D-printed case.
I admit, I almost couldn't believe it when I turned off my regular blazing fast 802.11ax NIC in the KDE network manager and switched to the pico-usb-wifi and I watched the YouTube "stats for nerds" buffer fully fill up and playback was totally smooth.
I would like to ask the author, what do they mean when they say 'I spent 2 days and 1M claude code tokens' Did you completely vibe code this? or did you create a spec and asked claude to implement that? How many times did you need to course correct claude when it deviated if it did? How much of your own experience as working in WiFi industry was useful while building this project? And what would be different if you had worked on this project without that experience?
Author here: this is actually a "side quest" which I alluded to in the README.adoc and another comment. I've had a few Spotify Car Things (SuperBird) sitting on my desk that I have not touched in a year because I was waiting for the community to come out with a non-tethered approach to reclaiming them. All of the existing continue to either require a host PC, phone, or RPI Zero. I want to deploy these throughout my home and have them usable without my PC or phone nearby - likely with just a Home Assistant dashboard. I spent the first few hours cloning repositories of existing firmware, and examining other approaches, such as Bluetooth low energy with ESPHome Bluetooth proxies, but they ended up being dead ends. I took a compromise that a small wireless dongle would be acceptable. I then began examining the Linux source tree to see if it was possible to flip the USB from client to host and spent a considerable amount of time recompiling BishopDynamics' SuperBird NixOS distribution and building a test dashboard to show the IP networking information, because I knew that once I flipped the device's USB mode, I would lose access to the device. SuperBird took almost four hours to compile on my low end CPU - I let it compile while I hit my five hour Claude Code quota.
https://gitlab.com/baiyibai/superbird-host-mode
I also did not look forward to trying to build in wireless functionality in to NixOS, because it also took 30 minutes to flash the device - I had to lower the transfer block size and add more retries in the flash utility because it kept crashing.
ultimately both the SuperBird work and the RPI pico firmware was about 300 prompts of investigation, research, creating potential approaches, architecting, designing, debugging, testing/validating, and documenting.
Regarding your last two questions... I started out 25 years ago by studying for a CCNA in high school and joined the industry just as first generation 802.11ac devices began hitting the market. From my position, I've worked on every layer from the SoC, PCB, to software. I want to be very careful and say software estimation is very difficult. The main bug (issue #1) I encountered would have been very difficult to diagnose and resolve ten years ago using traditional development workflows.
I'm only starting out in embedded software domain, and the way you describe your 2 day side quest feels like Dumbledore showing some magic trick to Harry.
I'm flattered. You asked very good questions. I'd be happy to de-mystify it with a longer conversation. Find me on LinkedIn: 白一百
I've often thought that if I ever did get around to publishing any of the stuff I work on that I would probably also publish any AI usage verbatim. Like, the full logs of all of the chats, at least as a sign of "here was my thought process" and "here's how I convinced the text machine to give me the right text". That somehow feels more honest than "I made this thing, but actually I vibe coded this thing."
Author here: I will consider writing a blog about ir, but it will require a lot of work on my part, because as I mentioned, this was a digression/side quest from the main task of hacking my Spotify Car Thing. It might work better as a YouTube video.
Interesting point. The prompts are the source, in a way.
I’ve been looking for a good solution for doing the exact opposite, being able to connect stuff through USB in my bench, and see them pop up in my office desktop as if they were usb devices.
The closest I’ve gotten is using a raspberry pi in the workbench, but for some weird devices that’s sometimes not good enough.
In a Windows world I've used the hell out of these devices for licensing dongles on Windows virtual machines: https://www.silextechnology.com/connectivity-solutions/devic...
They're much cheaper than the competing devices from Digi and they've been bulletproof for me. I've got some out there running >10 years.
I've even done stupid stuff like hung a USB Ethernet adapter off of one and made it a NIC on my local PC to talk to another of the same unit hanging off that USB NIC (just to be cheeky-- not for actual use). Stacking things on top of other things is fun.
My only complaint is I wish they did PoE. I use a cheap PoE splitter for that but it would be nice not to have to do that.
If you have CAT5/6 between the locations, USB-over-CAT5/6 adapters are inexpensive and work quite well. I've used OREI's successfully over >60m. But there are many options, with varying limits on cable length.
https://www.orei.com/products/usb-over-ethernet-extender-upt...
There USB over IP. I think Linux supports it. It could be possible to have Pico implementation, use WiFi like this, and connect to computer or another dongle.
Author here: I have wanted to make a wireless USB mouse and keyboard switch out of RPI Pico Ws for a long time. I would go a layer deeper than USB/IP, because I want to cache the USB handshake to keep the connection enumerated on multiple systems without dropping.
[dead]
I've used USB over Network in the past, worked well on Linux
Can you have the Pico operate as an access point? Would love to be able to use this to connect over wifi to a printer (printer in client mode), with the printer and macos talking directly over IP without needing to configure any other routing/forwarding on macos.
Wifi printer, where both your machine and the printer are connected to the same AP? yes
If you'd rather just expose a USB printer to the network, a Pi Zero is a better fit.
No. The printer cannot connect to the wifi network that the macbook is connected to. I don't want to expose a USB printer to the network. I just want to allow the printer to connect to an AP created by the Macbook/PiZero/ESP32/insert_cheap_widget device and then allow the Macbook to connect to the printer's IP.
That's what I meant by the Macbook and the printer connecting to the same AP. The Pico is the AP, and both your printer and the Macbook connect to it. (You could also use a cheap off the shelf travel router)
What the Macbook can't do is have multiple wifi connections at the same time, so you'd have to disconnect from its primary network (which also rules out the Macbook serving as an AP that the printer connects to).
[dead]
The Pico W can host an AP.
[dead]
Interesting project.
In a similar but opposite vein, I am going on a vacation and I wanted to share the stupidly expensive internet in my room at night with the family so I am likely bringing a raspberry pi to have as a travel router attached to my Mac. In this case, I can use the RaspAP project: https://raspap.com/
This is slightly different in that I do want a NAT.
I like this little thing: https://openwrt.org/toh/gl.inet/gl-usb150
You can use it as a repeater, so the whole family can just use the same network/password we use at home. And it is so small, you can run it from a power bank for hours.
Unfortunately, they are not sold anymore.
It's also slightly different because this project is for the Pi Pico.
Yeah, given the through put on the Pico is just 4Mbps in this setup, it will not work as a travel router for the family.
No Javascript
Peruse README and source code
Would love a Bluetooth variant, so I can use Bluetooth HID devices on a host with no Bluetooth stack.
The default reference firmware for the nRF52840 Desktop Dongle does this. https://github.com/nrfconnect/sdk-nrf/blob/main/applications...
I don't have any idea of what is the purpose of this.
Do you want to share an USB device across the WiFi?
If so, why not use the USB-IP protocol? It is already part of the Linux kernel, has implementation for Windows and doesn't require additional hardware.
close enough, welcome back 56(0)k
> Average 4.75 Mbits/sec throughput
Isn't that slow for WiFi?
I mean it's an interesting learning experience, but isn't that strictly worse than pretty much any WiFi dongle?
Yes, the guy clearly spent more of his time than it's worth - you can buy a wifi dongle for a few bucks on aliexpress.
But that's exactly the point of such experience. It's a challenge, and the guy/gal nailed it.
Well Claude nailed it, but they get credit for the idea.
Yes, just yesterday I got codex to write me a home assistant plugin for reticulum lxmf. For the fun of it. It got things wrong, I got it to fix.
You can not buy wifi usb adapters that work with MacOS on MacBooks I believe. I could be wrong but I think none of them have driver support.
The Pico is USB 1.1, so the upper bound is 12 Mb/s.
Totally pedantic deep dive: the Pico has a USB FS PHY (the numbered versions don't directly correlate to line rate, and FS was defined in 1.0) - the signaling rate is 12Mb/s, but it's impossible to achieve sustained transfers at this data rate under any circumstance:
- USB FS has a 1ms frame limit (HS is 125µs)
- UBS FS Bulk is thusly limited to ~19 transactions x 64 bytes maximum (HS is 512)
- USB FS Isochronous can do 1023 byte transfers, but you can only fit one of those in a 1ms frame (resulting in a giant quantization hole in the packet)
- Focusing on bulk only: the token packet, ACK handshake, inter-packet bus turnaround time minimums, framing bits, CRC bits, and periodic FS SOF packets mean that the actual theoretical maximum data rate is ~81% of the signaling rate
- Bit stuffing optimality issues (required for clock recovery) eat an additional several percent on most data, up to ~17% on pathological data
Therefore: ~9.5Mb/s is the best theoretical data rate that can be obtained with optimal host and device IP and an ideal application layer.
Realistically, ~8Mb/s is the most one can expect on real hardware with an ideal application (and this is optimistically high in my experience.)
Author here: thank you for this deep dive. I saw peaks of 5.9 Mbps, it might be possible to increase throughput by allocating more memory. The link between the rp2040 and Infineon Wi-Fi device is 31.5 Mb/s. It might be possible to push that out over PIO to an external USB 2.0 PHY. Apparently the Wi-Fi device is capable of higher speeds. Who knows what will be the limiting factor. I'll leave it to someone else to experiment, my focus was the stock hardware without any modifications.
> It might be possible to push that out over PIO to an external USB 2.0 PHY.
I doubt you will be able to achieve USB HS via a ULPI PHY without an FPGA tunneling the USB SIE for you and handling the tight turnaround requirement (in which case, you might as well give it the entire MAC workload or just use an MCU with an integrated HS PHY.)
The ULPI clock is a 60MHz free-running PHY-sourced signal (typically - some PHYs allow MAC-sourced clocks and this would likely be a little easier), the PHY is entirely in control of when the MAC can transmit (DIR pin), and the protocol is unforgiving from a timing standpoint.
Thanks! Now I potentially have ~20 USB WiFi adapters I didn’t have yesterday.
Even better, no need to hassle with the WiFi settings on the target system.
In wrong hands, Pico W is actually a bit terrifying device, because it combines USB and wireless.
Author here: You're welcome!
Love the way the author labels each of his diagrams as “AI Slop”!
It's one of the neat features of the AsciiDoc language. The user is able to change captions mid document, in this case :figure-caption:. AsciiDoc and Antora are things I've invested a lot of my time into
https://baiyibai-antora.gitlab.io
I'm kind of fascinated by the first diagram on the page. It sits so firmly in the uncanny valley for me and I can't put my finger on why. By itself every part looks ok and normal, but as a whole it just screams AI to me. I don't know if its the color choices or the composition or something else. It all just feels that little bit off.
I mean, I know its AI, the page says so itself, no one is trying to hide it. But it also just gives me AI vibes on such a subliminal level that I can't figure out why.
It's the amount of little icons and bold text. And then phrases like "No NAT, No Port-forwards"
Also details like the light blue boxes being swapped.
To me there's something about the "AI Slop Font" where it instantly triggers that uncanny valley. I guess since it's kind of an average of a bunch of fonts so it's somehow familiar but also unfamiliar at the same time.
Absolutely can relate. Thats why I would say: Do your own diagrams! At least the text and formatting.
Where I do think it ads value is "AI slop 2". This is somehow even better comprehensible than an average photo.
The pic "AI Slop 2. A Real World Situation" is great, the USB cable is connected to the wrong end of the Pico!
Oooooh, now I'm thinking... you could design a simple circuitboard that holds multiple picos (surface mounted) and uses the USB data pads on the back to pull all the USB ports out to an onboard USB hub basically allowing you to add a multitude of wifi adapters to a project in one USB cable. Would be great for War Driving!
How many Mbps?
6
not very a lot are they
6 is all you need in all cases, depending on the unit.
I think this is more of a "can you do this?" rather than something useful. Sure you can have Pico with this firmware around, but you can have much smaller and faster dongle instead.
Pico is an amazing thing tho, so many weird firmwares that are useful in a pinch.
Yes, but why does it matter. People do all sorts of things for fun.
More than enough for quite a lot of use cases.
I need mesh too
what is this useful for?
AI Slop.
I like that the second image filename is slop_2.png. At least they're self-aware.
[flagged]
OK, but please avoid generic tangents and fulmination on HN. https://news.ycombinator.com/newsguidelines.html
Meta comment to your meta comment : this is a very personal opinion but I cannot stop myself from thinking less about your message (even though I agree with it), just because of that edgy wannabe way to go about naming Microsoft. You want to say Microsoft say that, you want to say Github say that but if you say a vaguely insulting name only to have to specify what you actually mean... It doesn't bring anything to the conversation, except lower the level, and if a company has flaws as big as Microsoft you don't need to resort to childish name calling to expose them. This is my personal opinion.
I thought it brought something to the conversation
> just because of that edgy wannabe way to go about naming Microsoft.
I think it stopped being edgy and transitioned to a symbol of user frustration the moment their CEO tried to address the slop without acknowledging their customer pain.
It is not much different from the micro$oft used here often.
It means GH is now full of slop and pointing out it's owned by Microsoft, for those who may not have known or LMS who did not until recently
won’t somebody please think of the billion dollar company?
This has nothing to do with the billion dollar company, and I have no lost love for Microsoft.
It's about quality of discourse among us, and Microsoft has nothing to do with that.
what do you think about Micro$haft WinBlowz
one million Claude Tokens (assuming you are on opus) = 5 USD = the very dongle you tried to replace. Add the cost of the rasberry pico, you'll have an easier time buying the wifi dongle. The project is cool thought to learn about networks, NAT, Proxys, ect...
No, it's not really easier to buy a Wi-Fi dongle. My target device is the Spotify Car Thing and SuperBird doesn't have Wi-Fi components. My Claude Code Pro subscription was idle, so it cost me nothing. Also, according to an article from Tom's Hardware from two years ago, four million Picos have already shipped, so I've unlocked this ability for let's say 500,000 devices. Finally, my day job is in the Wi-Fi industry... this wasn't a learning exercise.
The author's tone when they discuss the cost of the project is self-deprecating. They know it would have been simpler to just buy one.
But also, the author has given the community a great gift, both directly (the blog post and the project!) and indirectly (the idea: what else can be implemented in similar ways).
Author here: Thank you, this is one of the nicest things anyone has ever said about my work.
I achieved two goals today: a front page post on hackernews and for being recognized for my work; or more accurately, for being very good at using AI ;)
It’s nice that it doesn’t need the WiFi stack or host side configuration though. This would be great for headless machines.
Finally hacker news on .. Hacker News. No need for it to be economical cheap.
Is it truly hacking in the traditional sense if everything was done with AI?
is it truly hacking in the traditional sense if you don't use punch cards?
> The project is cool thought to learn about networks, NAT, Proxys, ect...
The author learned nothing though...
Eh maybe a few tidbits. But the world gained a new WiFi adapter!
except now we don't need to spend that $5
> I spent two days of a long holiday weekend and about one million Claude Code tokens building this firmware.
You can do this by installing OpenWRT on the Pi and controlling it from the web interface.
But this is a Pi Pico, which is a microcontroller and not a Linux system.
Two arm cores at 133 MHz. That's already more powerful than my first computer. For $4. It qualifies as a computer on it's own. It runs Linux just a hacked version with an emulated MMU
https://github.com/tvlad1234/pico-rv32ima
FUZIX is a better fit. https://codeberg.org/EtchedPixels/FUZIX
In this case I'd say https://en.wikipedia.org/wiki/NuttX
> This project uses CNLohr's mini-rv32ima RISC-V emulator core
I wonder why not use the Pico's RISC-V cores.