Around '05 or so, the house we were staying in started beeping. Smoke alarm sounds, as usual. For almost a day we couldn't find the source. Contacted the owners -- they didn't know it either. They bought the house few years prior, so not much of a help.
Standing in the hallways listening at different angles, as the article says, didn't help. No Fable in '05 either.
So I had to improvise. Got a ~1 meter stick and two headphones (every headphone is a microphone, even if it sucks as a microphone). Plugged them into the stereo mic port, and started recording in Audacity. The phase shift between L and R pointed directly to the source. Took about 4 or 5 beeps to get the sucker.
Turns out, it was hiding above the ceiling tiles in the garage. My wife still tells this story to all her friends as one of my biggest life achievements. Everything I shipped, the companies I founded -- pales in comparison to finding a dead smoke alarm
There's some beeping in my apartment, and we also cannot figure out where it's coming from. The problem is that it's really quiet - sometimes only my wife can hear it, but I can't. It also seems to occur randomly, we can only hear it every few days. I tried recording it with my phone, but it doesn't pick anything up. It drives me crazy
> Claude doesn’t have native audio input (yet), but it can run JavaScript and Python code that can analyze audio captured by a microphone.
LLMs can also do that with non-mic audio, btw. I had GLM (or was it qwen?) answer the question of "Why does the spotify version of the Metal Gear Rising" soundtrack sound so bad" - to great results.
Compared with the CD rips on archive.org, the noise floor was 27dB (!!!) higher, and, the whole thing was 10dB more compressed.
I cannot vouch for these specific numbers, but subtracting the good version from the bad one gave me a horrible delta wav file with lots to listen to, so there was definitely a _lot_ of difference.
Pff that sound is horrible, and sorta hard to locate (because of the pitch?). Anyway. None of my neighbours seem to care and in summer I used to hear an orchestra of beeps every 5 seconds. It was maddening, I was contemplating of buying 1000 batteries, print a manual and shoving it in peoples mailboxes, although somehow I think it wouldnt have made a difference. I moved out eventually (not because of the beeps)
It’s because it doesn’t have enough overtones, or a notable attack. You might have noticed that some trucks these days make a weird noise when they back up instead of a beep. They’re using white noise because it’s easier to locate; it has many overtones that allow your brain to pick the sound out of the background more easily, and thus more easily correlate the signals from both ears. You get a much more accurate location just from hearing that sound than you ever could with a pure tone.
Consider a pure sine wave, it's harder (in theory anyway) to match the signals from the left and right ears, and identify the time delay between the two signals: each cycle of the wave is just like the others, so there's no distinct signature to match and say "the sound reached the left ear 0.5 ms before it reached the right ear". This is especially the case for higher frequencies where there can be something like aliasing.
Animal species sometimes use this in different sounds they make. An alarm call has that kind of "smooth tone" characteristic, so a predator can't locate the source of the sound but it still works to raise the alarm. Meanwhile more social sounds are more "chirpy" and sharp, so they can locate each other.
Interesting! I also feel the pulse is just to short… starting to question yourself if you actually heard something or that you are imagining it. Here the cars and trucks play musicals tunes or a crappy voice recording hahah
That makes sense; we were searching for a family member’s blood sugar monitor just today and it barely seemed to come from any single direction. Ultimately it wasn’t the sound itself but the loudness of it that let us find it, because a blanket got shoved over it mid-search that we could deduce from there.
> I think the most interesting part of this project is that I demonstrated it’s possible to give Claude Code ears
Of course it can analyze audio. I made it detect Marijn Frank saying a high pitched "Hallo" in a much better way that I could have done myself (subtitle matching), namely by frequency analysis, in the process learning how that works with librosa: https://marijnfrankzegthallo.nl
It's still a bit surprising the first time. Especially if you've been playing with these things for years now almost exclusively text and perhaps checking for extra fingers now and then.
I didn't have birdnet to hand and wanted to identify something. It was hard to take a picture with my phone but easy to record. Turns out one of the models was able to identify it.
I guess the Americans might appreciate gunshot identification and localisation or someone might drive around in winter/summer recording the streets to listen for heat pumps that are on the way out so they can advertise repair services. I would like a local version for movement/presence detection based on footprints as an alternative to having pir sensors all over the place for my hall lights.
That was my proposed solution to finding something just today; it didn’t work, but one of the bags ended up containing an old contract from General Development Company. Apparently a late family member had been buying some land down there through them, but county appraisers showed that the land never changed hands until after GDC’s bankruptcy. Not even their kid had heard about this, so it really makes me wonder what the story was behind that considering they lived in the Midwest (but did have some relatives nearby in Florida).
Yes, but that's also O(log n) physical effort. This one's O(1) sitting on your butt.
(Out of curiosity: I think "physical binary search" is more properly O(n^{1/3} log(n)), or maybe O(n^{1/2} log(n)). Since physical objects have sizes, and moving things (or moving between things) takes time proportional to distance).
Good point about the moving effort required. But if moving an object takes constant time, then it's O(n) overall -- since <incorrect argument deleted>.
EDIT: My earlier analysis was wrong (it assumed you never move the beeping item after the initial halving), but the conclusion was right for a different reason: Each time you move items, you move half as many as you did last time. You initially move n/2 items, and the geometric series n/2 + n/4 + n/8 + ... + 1 sums to n-1. So O(n) moves are done, and this dominates the log n term for the number of steps.
You could also solve it by getting rid of the boxes of stuff you clearly don't need (if you needed it, it wouldn't still be in a box). The best way to solve a problem is to make the problem go away.
Why didn't you time the interval between the beeps? Then when the next beep was coming up, just listen to a particular part of the room or set of boxes until you found it.
This is the problem with technology. Some people just want to use it, so it's quite hard to tell whether this is really a quality of life improvement, or if it's just exciting for some weirdo (any programmer counts as weirdo here).
I've personally had to track down a beep before and used a similar technique as you suggest: time the beeps then be in right place to perform a kind of binary search to home in on it. I found a smoke alarm in the roof space above my flat, nailed to the top of the rafters that had been there since the building was built 20 years prior. It was annoying but I'm sure it took way less time and effort than "getting AI to do it".
This is also why it's been difficult for me to decide on portfolio projects to build. I just don't have anything I really need to build. I made a compound interest calculator as my first React app, but could have made an excel sheet or python script.
Ha! This reminded me of that old famous entry from bash org: "I can't find a server in my house. It responds to ping, but I can't physically locate it."
Several years ago, I had a bed that occasionally creaked, but I couldn't reproduce it, and by virtue of only happening when we were already in bed, it was not possible to just listen for where the creaking may be coming from. I had an idea to put a few esp32s with mics under the bed to do this exact thing, but figured it was not worth the effort. I guess I can just throw some tokens at it if it happens again
"The system worked as intended. It gave me the position of the beep to within about 40cm, then I dug through that specific pile of boxes and found an ancient carbon monoxide detector that desperately wanted its batteries replaced."
There was a post on HN recently about a suspended string robot, and how it could manage different loads (picking up dirty laundry) depending on on where it was in the room.
Simon’s microphone configuration is very similar except instead of payload kg it would be affecting detection accuracy.
I wonder why the accuracy was an OOM outside the expected result (40cm vs 7mm)? Also using claps for calibration of unknown devices feels like a different approach to measuring distances between devices with a tape and solving the polygon. How many devices would be needed to self calibrate — 1x beep and Nx listening — instead of relying on smudgy claps? Or maybe every sound is smudged in a room filled with boxes?
Not related to the beep but if you still have natural gas (guessing based on the old carbon monoxide detector) you likely also want to get a simple plugin alarm that also detects natural gas in its natural un burned state.
Carbon monoxide is only produced when it's burned and builds up. And is obviously very dangerous. But a small gas leak building up over time or an oven pilot out and left on is not great for you either.
Make sure you place it relatively near the ceiling as it pools up there apparently. Saved my life during Covid lockdowns
I'm only here with a friendly tip for the designers of beeping devices: make sure to gradually ramp high-pitched beeps up and down, so that there's no sharp edge that an adversary could time.
Not with any kind of precision if the slope is gentle and you're using random commodity devices with different frequency response, automatic gain control, etc.
This article is about a weekend project and not a serious product, so what I'm saying below is not criticizing the author's approach — I'm making a point about something tangential.
The post touches on my biggest pet peeves about how many software engineers currently use AI at work: they ask it to "write out a problems.md file describing the problem we want to solve" then lob the output over the fence for the rest of the engineering org to review, often times without even disclosing they used AI to generate it (it's, of course, obvious).
AI can not do the reasoning for us; thinking that because it is great at writing code that works, the same applies to documents whose primary purpose is to demonstrate critical thinking, explore trade-offs, and reason through a problem is absurd.
In fact (or IMO?), this "problems.md" document is entirely unnecessary. In many cases, simply describing the issue and providing the relevant context is enough for AI to come up with a working solution.
AI can not do critical thinking (yet), so don't ask it to (and stop sending your fellow software engineers AI-slop RFCs).
> In fact (or IMO?), this "problems.md" document is entirely unnecessary. In many cases, simply describing the issue and providing the relevant context is enough for AI to come up with a working solution.
I don't mean to respond to your 'professional work' scenario, but I find it quite useful to first let it draft a PLAN.md, even when "in many cases" it would one-shot some implementation without it. I mostly prompt it with a problem statement, "write a PLAN.md", "ask me any questions in prose and numbered", "don't start implementation until my say-so". The back-and-forth questions are often pretty significant to the final result, and I can (skim) review the plan and make other adjustments.
When you ask most models to write that plan, you see an extended thinking trace similar to implementation itself, but now all the decisions are made up front. If you one-shot it an implementation, those decisions are sprinkled all over the implementation phase, while half of it has been written and some architectural choices have already been locked in. I then frequently adjust that plan to end up (IMO) with a much better implementation, or much better suited to my situation, with less shortcuts.
Yeah, I wouldn't do this for code that I was expecting anyone else to review.
This is actually the first time I've told an agent to write out a "problems.md" document - it was an experiment to see if stating the problem provided decent results, as opposed to working directly on a spec.
It did seem to work well though. I was outsourcing as much judgement as possible to the model, and the model did a good job at making decisions that got me to my required end result.
In this case, the AI could do the reasoning for me. Opus 5.5 and Fable 5.2 are really good!
Oh, was gonna post a comment about this, I actually bought just last week their USBc version (Fotric Sonic 3), on eBay, for around $700. Works quite well, it has 64 mic array, I think it would be tough to replicate it to this level.
Lmk if anyone has any questions about it! It’s android-only, but I’ve gotten an agent to write a Mac version of it though haven’t tested yet.
For an intermittent beep, could you point it at a room and then leave it recording video for an hour, then come back later and scrub through that video looking for the source of the noise?
Are you able to gather the multichannel recordings off the device? I’ve often wondered if you could do that, then replay a video while focusing on specific parts of the scene. Kind of like a virtual stethoscope.
Does it work for infrequent sounds? The sound might for example come in the middle of the night for two seconds, loud enough to wake one up, then go away. If I leave it on, can it do continuous recording and max or threshold-based capture?
Yes it can, works like a thermal camera. No recording (apart from stills), but you can watch the video image and, depending on overlay options, new sounds will be localized on the screen. Use mine to track down air leaks in a compressed air system, thing is magic.
That should work. You could always have Claude Code run the server itself and explain to it the rough parameters of the sound (two seconds, late at night) - then let it listen for hours and occasionally tweak its own configuration to try to better capture what's happening.
In my case it identified that the likely beep was about every 62 seconds pretty quickly - then I confirmed that one of the recordings was the mystery sound and it locked in from there.
wonderful! Thanks for sharing. I thought about doing this to track the sounds moving around in my stomach... it is interesting how well the mic on your smartphone works at picking up stomach sounds
I had a beep that drove me crazy and the very long search turned up a surplus CO monitor that I didn’t recall buying and had gotten tucked into the laundry room for some reason.
Same brand as my smoke detectors.
Make sounds that can be detected directionally, assholes.
I spend $200/month on an Anthropic subscription which this all fit into easily, and I don't have a good estimate for the first part (in Claude Code Cloud with Opus 5.5), but the subsequent Fable 5.2 session on my laptop would have cost $31.04 at API prices (517.9k ctx / 331.4k out), according to `uvx agentsview serve`.
How did you "connect as many devices as you can spare to act as listeners"? I assume you didn't create iOS and Android versions of this and put them in app stores. Do you run Python on all of them, is there a direct way to connect devices as listeners I'm not aware of, or is this a PWA/web app? My guess is the last option but I'd be interested in this technical aspect.
Woah this is cool. Another interesting simonw project based around sound through browsers. Lots of simple real world problems that can be tackled with this.
What I'd like to know is: did it make you feel good? Did you get a sense of satisfaction at having solved a problem? Do you feel more stronger and more capable than you were before?
What I've been finding is just setting AI loose like this doesn't give me any of that. I've decided not to think, so I don't feel any struggle. No struggle means no relief, no feeling of accomplishment. I also don't learn anything because AI read the paper for me. Learning is what I live for.
There is no world in which I would have spent the days (more likely weeks) necessary to learn enough about audio analysis in both JavaScript and Python and the WebAudio APIs necessary to build this thing, just to help me identify an annoying beep.
Thanks to coding agents, I got to build a wildly ambitious system that still taught me a bunch of interesting things about audio analysis and network web application design, all in about an hour and a half of clock time.
The alternative wasn't learning those things at increased detail, it was not learning anything about them at all.
And I added more tools to my tool belt for the future. I now know that it's possible to measure distances to sounds using cheap devices over a network, and I've got code in a GitHub repository that shows exactly how to do that.
The trick where the server could inject additional JS modules into the clients without needing to refresh the page (and hence lose the microphone permission) was neat, too.
I don't believe you learnt anything to any appreciable depth in 1.5 hours. I guess you learnt what's possible, but not much more. I get it, learning is learning and you got 1.5 hours' worth. Fair enough.
The bit I struggle with, though, is celebrating things being easy. I'm not excited by easy things. I like hard things. It's a bit like reading an account of how easy it was doing the Alpe d'Huez in a car. I'm more excited by people cycling up it because I know how difficult that is, or building the car. But driving it? Meh.
I do wonder if I'm descended from the people who built Stonehenge. I get restless if things are too easy. I need something hard, like erecting huge stones. But damn it's been nice being able to get that kind of satisfaction just sitting at a desk.
> "I demonstrated it’s possible to give Claude Code ears"
You found out you can connect a microphone to a computer and have it record and process the audio? Well I'm glad you can be one of todays 10,000 https://xkcd.com/1053
The less obvious part is giving Claude five WebAudio ears spread around a room, wired together with WebSockets and using synchronized chirps to estimate each other's locations.
Around '05 or so, the house we were staying in started beeping. Smoke alarm sounds, as usual. For almost a day we couldn't find the source. Contacted the owners -- they didn't know it either. They bought the house few years prior, so not much of a help.
Standing in the hallways listening at different angles, as the article says, didn't help. No Fable in '05 either.
So I had to improvise. Got a ~1 meter stick and two headphones (every headphone is a microphone, even if it sucks as a microphone). Plugged them into the stereo mic port, and started recording in Audacity. The phase shift between L and R pointed directly to the source. Took about 4 or 5 beeps to get the sucker.
Turns out, it was hiding above the ceiling tiles in the garage. My wife still tells this story to all her friends as one of my biggest life achievements. Everything I shipped, the companies I founded -- pales in comparison to finding a dead smoke alarm
Hahaha amazing. Also cool to hear stories about headphones used as mics, I didn't think it would be sensitive enough for that.
There's some beeping in my apartment, and we also cannot figure out where it's coming from. The problem is that it's really quiet - sometimes only my wife can hear it, but I can't. It also seems to occur randomly, we can only hear it every few days. I tried recording it with my phone, but it doesn't pick anything up. It drives me crazy
I didn't mention this in the article, but this isn't the first time my family has been driven almost spare by a mysterious, hard-to-track beep.
Our dog goes outside and hides in the lavender bush if a smoke alarm starts making a "I need a battery" noise.
So I went into this project with the attitude that I would like to solve this class of problem forever.
> I didn't mention this in the article, but this isn't the first time my family has been driven almost spare by a mysterious, hard-to-track beep.
https://xkcd.com/3100/
This is very humanizing. I like that.
Nicely done!
> Claude doesn’t have native audio input (yet), but it can run JavaScript and Python code that can analyze audio captured by a microphone.
LLMs can also do that with non-mic audio, btw. I had GLM (or was it qwen?) answer the question of "Why does the spotify version of the Metal Gear Rising" soundtrack sound so bad" - to great results.
Compared with the CD rips on archive.org, the noise floor was 27dB (!!!) higher, and, the whole thing was 10dB more compressed.
I cannot vouch for these specific numbers, but subtracting the good version from the bad one gave me a horrible delta wav file with lots to listen to, so there was definitely a _lot_ of difference.
It got the job done is what I want to say.
Wait, can you share the results?
Pff that sound is horrible, and sorta hard to locate (because of the pitch?). Anyway. None of my neighbours seem to care and in summer I used to hear an orchestra of beeps every 5 seconds. It was maddening, I was contemplating of buying 1000 batteries, print a manual and shoving it in peoples mailboxes, although somehow I think it wouldnt have made a difference. I moved out eventually (not because of the beeps)
It’s because it doesn’t have enough overtones, or a notable attack. You might have noticed that some trucks these days make a weird noise when they back up instead of a beep. They’re using white noise because it’s easier to locate; it has many overtones that allow your brain to pick the sound out of the background more easily, and thus more easily correlate the signals from both ears. You get a much more accurate location just from hearing that sound than you ever could with a pure tone.
Consider a pure sine wave, it's harder (in theory anyway) to match the signals from the left and right ears, and identify the time delay between the two signals: each cycle of the wave is just like the others, so there's no distinct signature to match and say "the sound reached the left ear 0.5 ms before it reached the right ear". This is especially the case for higher frequencies where there can be something like aliasing.
Animal species sometimes use this in different sounds they make. An alarm call has that kind of "smooth tone" characteristic, so a predator can't locate the source of the sound but it still works to raise the alarm. Meanwhile more social sounds are more "chirpy" and sharp, so they can locate each other.
Interesting! I also feel the pulse is just to short… starting to question yourself if you actually heard something or that you are imagining it. Here the cars and trucks play musicals tunes or a crappy voice recording hahah
That makes sense; we were searching for a family member’s blood sugar monitor just today and it barely seemed to come from any single direction. Ultimately it wasn’t the sound itself but the loudness of it that let us find it, because a blanket got shoved over it mid-search that we could deduce from there.
> I think the most interesting part of this project is that I demonstrated it’s possible to give Claude Code ears
Of course it can analyze audio. I made it detect Marijn Frank saying a high pitched "Hallo" in a much better way that I could have done myself (subtitle matching), namely by frequency analysis, in the process learning how that works with librosa: https://marijnfrankzegthallo.nl
It's still a bit surprising the first time. Especially if you've been playing with these things for years now almost exclusively text and perhaps checking for extra fingers now and then.
I didn't have birdnet to hand and wanted to identify something. It was hard to take a picture with my phone but easy to record. Turns out one of the models was able to identify it.
I guess the Americans might appreciate gunshot identification and localisation or someone might drive around in winter/summer recording the streets to listen for heat pumps that are on the way out so they can advertise repair services. I would like a local version for movement/presence detection based on footprints as an alternative to having pir sensors all over the place for my hall lights.
You can also solve this in O(logn) using a physical binary search:
1) Move half the boxes to another room.
2) Listen for a beep and determine which room the beep comes from.
3) Take the boxes in that room and split them in half again.
4) Repeat until one box is left.
You are moving n/2 objects log(n) times, so its O(n.log(n))
It's O(logn) but you haven't accounted for how heavy each item in the array of n is.
If you randomly index them and there's enough of them then on average the partitions average out.
This is true, but it probably isn't a box of books that's beeping.
I also feel like ordering the moves by lightest box first is reasonable as a real-world optimization.
Lol classic “ignore the constant factor” gotcha when using Big-O.
Or you could use a very heavy blanked to just cover some boxes to supress the sound enough to simulate moving boxes.
That was my proposed solution to finding something just today; it didn’t work, but one of the bags ended up containing an old contract from General Development Company. Apparently a late family member had been buying some land down there through them, but county appraisers showed that the land never changed hands until after GDC’s bankruptcy. Not even their kid had heard about this, so it really makes me wonder what the story was behind that considering they lived in the Midwest (but did have some relatives nearby in Florida).
Yes, but that's also O(log n) physical effort. This one's O(1) sitting on your butt.
(Out of curiosity: I think "physical binary search" is more properly O(n^{1/3} log(n)), or maybe O(n^{1/2} log(n)). Since physical objects have sizes, and moving things (or moving between things) takes time proportional to distance).
Good point about the moving effort required. But if moving an object takes constant time, then it's O(n) overall -- since <incorrect argument deleted>.
EDIT: My earlier analysis was wrong (it assumed you never move the beeping item after the initial halving), but the conclusion was right for a different reason: Each time you move items, you move half as many as you did last time. You initially move n/2 items, and the geometric series n/2 + n/4 + n/8 + ... + 1 sums to n-1. So O(n) moves are done, and this dominates the log n term for the number of steps.
Yeah, that interpretation is more realistic!
O(log n) is if you can split the collection in half in an O(1) step, like, *shoving* the whole pile together as a unit
O(n) if you're moving items one at a time. So, splitting the pile in half once is O(n)
O(n^{3/2}) if the project is way out of control and walking distances start to matter
It’s not sitting on your butt if you have to code an entire system and set up notes throughout the house and monitor them.
That requires you to stand in the room and listen for which side. So it’s nlogn times a very long time constant
> Move half the boxes to another room.
Alternatively, no.
Alternatively, you can unpack
Came here to say this.
I am confident that inventing a new networked sound locating system is far easier than that.
More interesting for sure
Who's got that kind of time?
Yes, if you still didn't unpacked it when you don't need it.
You could also solve it by getting rid of the boxes of stuff you clearly don't need (if you needed it, it wouldn't still be in a box). The best way to solve a problem is to make the problem go away.
> if you needed it, it wouldn't still be in a box
I like to keep a lot of things in storage, and sometimes go through my boxes and “rediscover” things I haven’t used in a while.
Sure I have a lot of things I never use also. Some things I could sell, and some things I could give away, and some things I could throw away.
But just because a lot of things are in boxes, doesn’t mean that I might as well get rid of all of it for that reason alone.
Step one: burn down half the house. If the beeping stops, the device was in the first half. If not, burn down the other half of the house.
Why didn't you time the interval between the beeps? Then when the next beep was coming up, just listen to a particular part of the room or set of boxes until you found it.
This is the problem with technology. Some people just want to use it, so it's quite hard to tell whether this is really a quality of life improvement, or if it's just exciting for some weirdo (any programmer counts as weirdo here).
I've personally had to track down a beep before and used a similar technique as you suggest: time the beeps then be in right place to perform a kind of binary search to home in on it. I found a smoke alarm in the roof space above my flat, nailed to the top of the rafters that had been there since the building was built 20 years prior. It was annoying but I'm sure it took way less time and effort than "getting AI to do it".
This is also why it's been difficult for me to decide on portfolio projects to build. I just don't have anything I really need to build. I made a compound interest calculator as my first React app, but could have made an excel sheet or python script.
Ha! This reminded me of that old famous entry from bash org: "I can't find a server in my house. It responds to ping, but I can't physically locate it."
Several years ago, I had a bed that occasionally creaked, but I couldn't reproduce it, and by virtue of only happening when we were already in bed, it was not possible to just listen for where the creaking may be coming from. I had an idea to put a few esp32s with mics under the bed to do this exact thing, but figured it was not worth the effort. I guess I can just throw some tokens at it if it happens again
WHAT WAS MAKING THE BEEP!!?!?!
Hahahaha. You should slow down when you read
"The system worked as intended. It gave me the position of the beep to within about 40cm, then I dug through that specific pile of boxes and found an ancient carbon monoxide detector that desperately wanted its batteries replaced."
Oh, thanks. Not sure how I missed that in the very middle of the piece :D
Maybe I should have asked Claude to help triangulate it for me.
Finding the crux of the article buried somewhere in the middle... it's like poetry; it rhymes.
There was a post on HN recently about a suspended string robot, and how it could manage different loads (picking up dirty laundry) depending on on where it was in the room.
https://neufangled.com/payload
Simon’s microphone configuration is very similar except instead of payload kg it would be affecting detection accuracy.
I wonder why the accuracy was an OOM outside the expected result (40cm vs 7mm)? Also using claps for calibration of unknown devices feels like a different approach to measuring distances between devices with a tape and solving the polygon. How many devices would be needed to self calibrate — 1x beep and Nx listening — instead of relying on smudgy claps? Or maybe every sound is smudged in a room filled with boxes?
Not related to the beep but if you still have natural gas (guessing based on the old carbon monoxide detector) you likely also want to get a simple plugin alarm that also detects natural gas in its natural un burned state.
Carbon monoxide is only produced when it's burned and builds up. And is obviously very dangerous. But a small gas leak building up over time or an oven pilot out and left on is not great for you either.
Make sure you place it relatively near the ceiling as it pools up there apparently. Saved my life during Covid lockdowns
I'm only here with a friendly tip for the designers of beeping devices: make sure to gradually ramp high-pitched beeps up and down, so that there's no sharp edge that an adversary could time.
Thanks, Vetinaris clockmaker.
You can time the envelope of the ramp, though
Not with any kind of precision if the slope is gentle and you're using random commodity devices with different frequency response, automatic gain control, etc.
This article is about a weekend project and not a serious product, so what I'm saying below is not criticizing the author's approach — I'm making a point about something tangential.
The post touches on my biggest pet peeves about how many software engineers currently use AI at work: they ask it to "write out a problems.md file describing the problem we want to solve" then lob the output over the fence for the rest of the engineering org to review, often times without even disclosing they used AI to generate it (it's, of course, obvious).
AI can not do the reasoning for us; thinking that because it is great at writing code that works, the same applies to documents whose primary purpose is to demonstrate critical thinking, explore trade-offs, and reason through a problem is absurd.
In fact (or IMO?), this "problems.md" document is entirely unnecessary. In many cases, simply describing the issue and providing the relevant context is enough for AI to come up with a working solution.
AI can not do critical thinking (yet), so don't ask it to (and stop sending your fellow software engineers AI-slop RFCs).
> In fact (or IMO?), this "problems.md" document is entirely unnecessary. In many cases, simply describing the issue and providing the relevant context is enough for AI to come up with a working solution.
I don't mean to respond to your 'professional work' scenario, but I find it quite useful to first let it draft a PLAN.md, even when "in many cases" it would one-shot some implementation without it. I mostly prompt it with a problem statement, "write a PLAN.md", "ask me any questions in prose and numbered", "don't start implementation until my say-so". The back-and-forth questions are often pretty significant to the final result, and I can (skim) review the plan and make other adjustments.
When you ask most models to write that plan, you see an extended thinking trace similar to implementation itself, but now all the decisions are made up front. If you one-shot it an implementation, those decisions are sprinkled all over the implementation phase, while half of it has been written and some architectural choices have already been locked in. I then frequently adjust that plan to end up (IMO) with a much better implementation, or much better suited to my situation, with less shortcuts.
Yeah, I wouldn't do this for code that I was expecting anyone else to review.
This is actually the first time I've told an agent to write out a "problems.md" document - it was an experiment to see if stating the problem provided decent results, as opposed to working directly on a spec.
It did seem to work well though. I was outsourcing as much judgement as possible to the model, and the model did a good job at making decisions that got me to my required end result.
In this case, the AI could do the reasoning for me. Opus 5.5 and Fable 5.2 are really good!
MKBHD demos the the FOTRIC TD2e Acoustic Imaging Camera.
https://www.youtube.com/shorts/ai1gIWacf0I
It costs ~$1,000 so you'd probably rent it if you're only going to use it once, but it's a neat gadget.
You can also build one pretty cheaply and simply with the miniDSP UMA 16: https://www.minidsp.com/products/usb-audio-interface/uma-16-...
https://www.minidsp.com/applications/usb-mic-array/acoustic-...
Oh, was gonna post a comment about this, I actually bought just last week their USBc version (Fotric Sonic 3), on eBay, for around $700. Works quite well, it has 64 mic array, I think it would be tough to replicate it to this level.
Lmk if anyone has any questions about it! It’s android-only, but I’ve gotten an agent to write a Mac version of it though haven’t tested yet.
eBay link to what I bought: https://www.ebay.com/itm/960060893216
For an intermittent beep, could you point it at a room and then leave it recording video for an hour, then come back later and scrub through that video looking for the source of the noise?
Are you able to gather the multichannel recordings off the device? I’ve often wondered if you could do that, then replay a video while focusing on specific parts of the scene. Kind of like a virtual stethoscope.
It seems what happened with thermal cameras many years ago might be happening with acoustic ones now. They used to cost a lot more than that.
Great thing for tool checkout/rental at a makerspace!
Does it work for infrequent sounds? The sound might for example come in the middle of the night for two seconds, loud enough to wake one up, then go away. If I leave it on, can it do continuous recording and max or threshold-based capture?
Yes it can, works like a thermal camera. No recording (apart from stills), but you can watch the video image and, depending on overlay options, new sounds will be localized on the screen. Use mine to track down air leaks in a compressed air system, thing is magic.
That should work. You could always have Claude Code run the server itself and explain to it the rough parameters of the sound (two seconds, late at night) - then let it listen for hours and occasionally tweak its own configuration to try to better capture what's happening.
In my case it identified that the likely beep was about every 62 seconds pretty quickly - then I confirmed that one of the recordings was the mystery sound and it locked in from there.
Could you build this into a HUD for a cat, tuned to mouse squeeks? Asking for a friend.
wonderful! Thanks for sharing. I thought about doing this to track the sounds moving around in my stomach... it is interesting how well the mic on your smartphone works at picking up stomach sounds
I wonder if it would work to find an annoyatron?
I had a beep that drove me crazy and the very long search turned up a surplus CO monitor that I didn’t recall buying and had gotten tucked into the laundry room for some reason.
Same brand as my smoke detectors.
Make sounds that can be detected directionally, assholes.
Simon's LLM's beeping finder app vs https://annoyingpcb.com/
Who would win?? :-)
I lack the courage to plant such a PCB prank device... it really feels like you could drive people completely crazy haha
You know what might have been easier... Going through the boxes!
No, no. Instead of taking the time to unpack the boxes, it's better to spend time to play with AI, and leave the boxes unpacked forever.
Great post. Many levels of interesting information. Any sense of how much this cost?
I spend $200/month on an Anthropic subscription which this all fit into easily, and I don't have a good estimate for the first part (in Claude Code Cloud with Opus 5.5), but the subsequent Fable 5.2 session on my laptop would have cost $31.04 at API prices (517.9k ctx / 331.4k out), according to `uvx agentsview serve`.
How did you "connect as many devices as you can spare to act as listeners"? I assume you didn't create iOS and Android versions of this and put them in app stores. Do you run Python on all of them, is there a direct way to connect devices as listeners I'm not aware of, or is this a PWA/web app? My guess is the last option but I'd be interested in this technical aspect.
Woah this is cool. Another interesting simonw project based around sound through browsers. Lots of simple real world problems that can be tackled with this.
What I'd like to know is: did it make you feel good? Did you get a sense of satisfaction at having solved a problem? Do you feel more stronger and more capable than you were before?
What I've been finding is just setting AI loose like this doesn't give me any of that. I've decided not to think, so I don't feel any struggle. No struggle means no relief, no feeling of accomplishment. I also don't learn anything because AI read the paper for me. Learning is what I live for.
It makes me feel great!
There is no world in which I would have spent the days (more likely weeks) necessary to learn enough about audio analysis in both JavaScript and Python and the WebAudio APIs necessary to build this thing, just to help me identify an annoying beep.
Thanks to coding agents, I got to build a wildly ambitious system that still taught me a bunch of interesting things about audio analysis and network web application design, all in about an hour and a half of clock time.
The alternative wasn't learning those things at increased detail, it was not learning anything about them at all.
And I added more tools to my tool belt for the future. I now know that it's possible to measure distances to sounds using cheap devices over a network, and I've got code in a GitHub repository that shows exactly how to do that.
The trick where the server could inject additional JS modules into the clients without needing to refresh the page (and hence lose the microphone permission) was neat, too.
Well, I envy you, I guess.
I don't believe you learnt anything to any appreciable depth in 1.5 hours. I guess you learnt what's possible, but not much more. I get it, learning is learning and you got 1.5 hours' worth. Fair enough.
The bit I struggle with, though, is celebrating things being easy. I'm not excited by easy things. I like hard things. It's a bit like reading an account of how easy it was doing the Alpe d'Huez in a car. I'm more excited by people cycling up it because I know how difficult that is, or building the car. But driving it? Meh.
I do wonder if I'm descended from the people who built Stonehenge. I get restless if things are too easy. I need something hard, like erecting huge stones. But damn it's been nice being able to get that kind of satisfaction just sitting at a desk.
This is the way.
> "I demonstrated it’s possible to give Claude Code ears"
You found out you can connect a microphone to a computer and have it record and process the audio? Well I'm glad you can be one of todays 10,000 https://xkcd.com/1053
The less obvious part is giving Claude five WebAudio ears spread around a room, wired together with WebSockets and using synchronized chirps to estimate each other's locations.
https://www.metafilter.com/42331/ScoobyDoo-and-the-case-of-t...
Smoke alarm.