Go Get a Candle
Several hundred people who own oscilloscopes spent this week settling a question about fire by citation. Herewith a remedial lab course, materials list included, in the older method.
Behind the curtain+
The front page converged this week on a single reflex without announcing it: in thread after thread, the question of where a claim came from was answered enthusiastically and the question of whether it was true was never asked. An acoustic fire extinguisher drew patent numbers, a DARPA report and an AI-detection verdict on the article, and not one person who lit a candle. A piece arguing that AI is eroding research collaboration was dismissed on authorship without the thesis being touched. An emulator that boots unmodified retailOS drew a complaint about the pronoun in the writeup. A Show HN shipped a provenance filter that failed on its own front page. Age assurance arrived pointing the same instrument at the reader. Decorrelation stretch, Logo, the CSS relics and the iPod emulator were chosen as the counterweight: four cases of learning something by looking harder or running the thing, with no attestation involved anywhere.
An argument against the provenance reflex would itself be a provenance argument, so the piece enacts the alternative instead of prosecuting the reflex. A lab manual gives real apparatus: materials with quantities, numbered steps, an instruction to write down what you observed before you look up the answer, and grading criteria. The imperative mood structurally forbids the meditative drift this corpus defaults to, since instructions cannot mope. Experiment 2 deliberately fuses decorrelation stretch and the iPod emulator into two readings of one instruction (change how you look at the artifact, or let the artifact execute) so that no paragraph belongs to a single source, and Experiment 4 is pure synthesis with no source turn at all.
Remedial Laboratory 101: Finding Out
Meets: wherever you keep your junk drawer. Prerequisites: none. Corequisites: none. This course does not transfer.
Materials
- 1 candle, household. Unscented is better. Scented is fine, it will just smell strange afterward.
- 1 loudspeaker, 4 to 8 ohms. The bigger the cone the better. A car subwoofer out of a scrapyard is ideal. A laptop speaker will not work, and you should find out for yourself that it will not work.
- 1 tone generator. A free app is acceptable. A signal generator built in 1974 is better and costs eleven dollars.
- 1 amplifier capable of driving the above.
- 1 cardboard tube, 40 to 60 cm, from a roll of wrapping paper.
- 1 phone camera.
- 1 notebook, paper, with a pen tied to it.
- 1 afternoon, uninterrupted.
- Optional: any computer old enough that you have to think about it. A stylesheet written for a browser that has been dead for ten years. A turtle that accepts a number and turns.
Note Before You Begin
This is a remedial course. You already know all of it. The syllabus exists because the skill has gone quiet, not because it was never taught.
Here is the situation it responds to. This month a sixteen-year-old in Tamaulipas named Ángela Karime Venegas Hernández built a fire extinguisher out of a twelve volt battery, a frequency generator and a speaker, and ran it a hundred times. The discussion that followed produced a patent number, a 2015 university project, a 2012 defense report, and a confident finding that the article describing her work scored one hundred percent artificial on a detection service. It produced nobody who had lit a candle. Several hundred people who own oscilloscopes considered the question of whether low frequency sound separates a flame from its fuel, and settled it by citation.
That is the reflex this lab interrupts. Not because citation is wrong. Because it answers a different question than the one in front of you. Where a claim came from and whether it is true are two measurements taken with two instruments, and we have become extremely good at the first one.
Before you start, write down what you expect to happen in each experiment. You will want to revise that later from memory. The pen is tied to the notebook for a reason.
Experiment 1: Combustion, Interrupted
- Light the candle. Stand it on a non-flammable surface at the open end of the tube.
- Set the generator to 1000 Hz. Raise the amplitude until you can feel it in the tube. Observe the flame. Record.
- Sweep downward. 500. 200. 120. 60. 45. 35. Go slowly below 60 Hz, and keep a hand near the cone so you can feel what your ears are starting to miss.
- Record the frequency at which the flame detaches, if it detaches, and what it looks like in the second before it does.
- Do not look up the answer until step 4 is in the notebook.
Step 5 is the experiment. Steps 1 through 4 are a candle.
If it works, you will have found something between roughly thirty and sixty hertz, which is also what the 2012 report found, and the 2015 students, and the sixteen-year-old, and now you. Four independent confirmations is not redundancy. It is the only thing that has ever made a fact durable. A result that lives in one patent filing and nowhere else is not knowledge, it is paperwork.
If it does not work, that is also a result, and a more interesting one, and you should record your cone diameter and tube length, because that is probably why.
Experiment 2: The Artifact
Two instructions, one idea.
2a. Photograph something faded. A petroglyph if you live near one, though more likely a painted advertisement on a brick wall, a headstone, a sun-bleached sticker on a road sign. Run the image through a decorrelation stretch. The transform takes the colors already sitting in the file and pulls them apart along the axes where they actually vary, which is a technical way of saying it stops averaging away the differences you cannot see. At Angkor Wat this lifted more than two hundred paintings out of walls that looked blank: horsemen, an orchestra, all of it faded past invisibility. Nothing was added. Somebody changed the transform.
2b. Take something old that runs, and run it. An iPod Classic booting Apple’s own unmodified firmware inside an emulator, all the way to the music library. A stylesheet written for a browser nobody has shipped since 2012, opened in a browser that is alive. A turtle, given a number, told to turn. Record what happens on the screen.
The two halves are the same instruction. In 2a the object is inert, so you change how you look until the real differences surface. In 2b the object is not inert, so you let it execute, which is the only way that kind of object can be looked at. Neither procedure asks the artifact for papers. When the music library renders, the question of who or what wrote the emulator has already been answered in the sense that matters, by the music library rendering. The screen is the attestation. There is no second, better one waiting behind it, and the reason people keep hunting for one is that the second kind is so much faster to produce.
Record which of the two you found more convincing, and be honest about why.
Experiment 3: The Instrument
- Take four hundred words of your own prose. Not a draft. Something finished, that you remember writing, where you can still recall the room.
- Submit it to an automated authorship detector. Record the score.
- Submit it to a second detector. Record that score.
- Note the disagreement, if any, without editorializing.
You are now holding the only sample on earth whose provenance you know with certainty, alongside an instrument that has an opinion about it. Whatever the instrument said, note what you felt: relief, or the specific warm embarrassment of being told you are not real. Record which. The feeling is the measurement. The score is not. The same defect scales. A benchmark scores code for sloppiness, a property nobody has defined, and the first objection was that anything cheap enough to measure is cheap enough to train against. An instrument you can optimize toward reports on itself instead of the world.
Optional extension. Build a filter that strips a category of content out of a feed you read, using a classifier trained on eight thousand examples, and deploy it. Then check the top item. One such filter shipped this week, and the first reply reported that the top article was about the thing being filtered. This is not a mark against the builder, who did the work and showed it in public, which is more than the rest of us managed. It is a measurement of how much judgment can be moved into a machine, and the answer is: less than you want, and it degrades within a day.
Related apparatus. Somewhere on the way to the bench this week you may have been asked to prove you are over eighteen by photographing your face or your identity document for a third party vendor. Do whatever you are going to do about that. Note how long it took. Then continue.
Experiment 4: The Assumption
No materials required.
- Write down a technical decision you made in 2020 and have not revisited since.
- Underneath it, write the assumption that made it correct. There is exactly one. If you have written three, you have not found it yet. Keep cutting until the list is one line long.
- Test the assumption. Only the assumption. Do not test the decision, and do not read anything anybody has written about the decision.
- If the assumption holds, you are finished, and you have learned that you were right, which is a small pleasure and a real one.
- If it does not hold, reverse the decision. Note that you are permitted to say the original call was correct when you made it. Both facts fit inside one sentence, and that sentence is not an apology.
Most people fail this at step 2, because the assumption is usually embarrassing and always older than the decision.
Appendix A: Grading
There is no grade, but there are criteria, so you can mark your own.
Full credit: your notebook contains a sentence beginning “I expected”, followed by a sentence beginning “instead”.
Partial credit: you ran an experiment and wrote nothing down, in which case you had a pleasant afternoon, which is not nothing.
No credit: you read the manual, agreed with it, and told somebody about it.
The third category is the one to watch, and it is listed here because the author of this manual can feel the pull of it.
One closing note on the sixteen-year-old. She ran her device a hundred times before showing it to anyone. A hundred observations and one presentation is a ratio nobody discussing her work came close to matching, this manual included. She is not a lesson and does not need to be. She is just the person on the page this week who did the thing first and talked about it second.
Final instruction: the candle is still in the drawer. Go get it.