Black swans: When exploring something new, large chunks of hypothesis space are hypotheses you haven’t even thought of yet. Often these unknown unknowns add up to more than 50% probability. I don’t think we tend to appreciate this fact enough on LW, so I thought I’d share a story. I had a professor in college who would give the class actual engineering problems that cost real companies many tens of thousands of dollars, and give us an hour discuss it before trying to guess the answer. I always enjoyed those games, so I’ll present my little mishap in a similar format.
I was getting my 3D printer up and running again after over a year of inactivity (due to computer problems, then the Windows 10 update). The leading software had changed, so I installed it and did my best to put in print settings that were reasonable for my printer. I noticed that test prints would start out fine, and then get stringier and stringier as it printed subsequent layers. The first couple layers were 100% dense, but it seemed like there was more empty space than plastic in the upper layers. PLA filament slowly absorbs water from the air, so one thing I was worried about was that it had absorbed moisture over the past year, and that it would boil off inside the printer. However, rather than jumping to conclusions and spending several days dehumidifying the plastic filament, I performed several tests. I wanted to rule out alternatives, like the extruder motor skipping steps by trying to push stiff filament through the extruder nozzle before it had a chance to melt. I turned up the nozzle heat to make this less likely, but it is possible to get similar effects if the driver chip has to turn itself off for an moment to avoid overheating.
When there are many competing hypotheses, the most efficient experiments are ones that rule out entire classes of hypotheses at once. After a print, I had to physically push about a centimeter of filament into the nozzle before anything would be extruded out the other end. Otherwise, all I got was the slowly oozing drips, jut like an inactive hot glue gun. I marked the filament every centimeter, and watched it closely over a half hour of printing a tiny tower at low speed. The filament was indeed fed into the extruder nozzle at a constant rate, with no skipping. So, whatever the problem was, I reasoned that it had to be downstream of that. I noticed that I was confused, since the filament being fed in had to be going somewhere after all. I started to wrack my brain for things that could be going on inside the nozzle to extrude plenty of filament at first, and then not enough.
Maybe the molten filament just before the print was baked dry, but as the print continued and new filament was melted, the steam would force a bit of extra plastic out of the nozzle. Eventually the steam would have to come out of course, but if it reciprocated between extruding plastic and gas this might explain the observed behavior. The only other thing I could think of was that the filament itself was doing something strange somewhere, but I couldn’t form a coherent concept of exactly what it might be doing, so I assigned a 75% probability to the moisture hypothesis.
Solution, encrypted with a Caesar cipher shifting all letters 13 characters: Orsber guebjvat gur erry bs svynzrag va n ohpxrg jvgu n qruhzvqvsvre naq tbvat gb orq, V qrpvqrq gb ybbx bire nyy gur frggvatf bar zber gvzr. Nyzbfg vzzrqvngryl, V qvq n snprcnyz. Gur cevagre gubhtug V jnf hfvat 3 zz svynzrag, juvyr gur npghny svynzrag qvnzrgre jnf 1.75 zz. Nf n erfhyg, vg jnf srrqvat svynzrag vagb gur abmmyr zhpu zber fybjyl guna vg arrqrq gb or qbvat. Boivbhfyl sbe lbh gb cerqvpg guvf fcrpvsvpnyyl lbh jbhyq arrq n yvggyr xabjyrqtr bs 3Q cevagvat, ohg gur zber trareny ulcbgurfvf gung gur srrq engr jnf gbb ybj vf boivbhf va ergebfcrpg. Znlor V jnf whfg biregverq ng gur gvzr, ohg gur jubyr vqrn bs fgrnz pnhfvat plpyrf bs bire-rkgehqvat naq haqre-rkgehqvat abj frrzf pbzcyrgryl vafnar. Jul jbhyq vg tb va plpyrf naljnl, va bccbfrq gb whfg pbafgnag ohooyrf va gur rkgehqrq svynzrag? Jul jbhyq V unir gb rkgehqr n pragvzrgre bs svynzrag nsgre gur raq bs gur cevag orsber V fgnegrq trggvat nalguvat ohg bbmvat svynzrag qevccvat bhg bs gur raq bs gur abmmyr? Znlor nyy bs lbh vzzrqvngryl jbaqrerq guvf, naq znlor vg jnf boivbhf jura V jnf cevivyrtvat gur ulcbgurfvf, ohg V bayl whfg oneryl pnhtug zl zvfgnxr.
Black swans: When exploring something new, large chunks of hypothesis space are hypotheses you haven’t even thought of yet. Often these unknown unknowns add up to more than 50% probability. I don’t think we tend to appreciate this fact enough on LW, so I thought I’d share a story. I had a professor in college who would give the class actual engineering problems that cost real companies many tens of thousands of dollars, and give us an hour discuss it before trying to guess the answer. I always enjoyed those games, so I’ll present my little mishap in a similar format.
I was getting my 3D printer up and running again after over a year of inactivity (due to computer problems, then the Windows 10 update). The leading software had changed, so I installed it and did my best to put in print settings that were reasonable for my printer. I noticed that test prints would start out fine, and then get stringier and stringier as it printed subsequent layers. The first couple layers were 100% dense, but it seemed like there was more empty space than plastic in the upper layers. PLA filament slowly absorbs water from the air, so one thing I was worried about was that it had absorbed moisture over the past year, and that it would boil off inside the printer. However, rather than jumping to conclusions and spending several days dehumidifying the plastic filament, I performed several tests. I wanted to rule out alternatives, like the extruder motor skipping steps by trying to push stiff filament through the extruder nozzle before it had a chance to melt. I turned up the nozzle heat to make this less likely, but it is possible to get similar effects if the driver chip has to turn itself off for an moment to avoid overheating.
When there are many competing hypotheses, the most efficient experiments are ones that rule out entire classes of hypotheses at once. After a print, I had to physically push about a centimeter of filament into the nozzle before anything would be extruded out the other end. Otherwise, all I got was the slowly oozing drips, jut like an inactive hot glue gun. I marked the filament every centimeter, and watched it closely over a half hour of printing a tiny tower at low speed. The filament was indeed fed into the extruder nozzle at a constant rate, with no skipping. So, whatever the problem was, I reasoned that it had to be downstream of that. I noticed that I was confused, since the filament being fed in had to be going somewhere after all. I started to wrack my brain for things that could be going on inside the nozzle to extrude plenty of filament at first, and then not enough.
Maybe the molten filament just before the print was baked dry, but as the print continued and new filament was melted, the steam would force a bit of extra plastic out of the nozzle. Eventually the steam would have to come out of course, but if it reciprocated between extruding plastic and gas this might explain the observed behavior. The only other thing I could think of was that the filament itself was doing something strange somewhere, but I couldn’t form a coherent concept of exactly what it might be doing, so I assigned a 75% probability to the moisture hypothesis.
Solution, encrypted with a Caesar cipher shifting all letters 13 characters: Orsber guebjvat gur erry bs svynzrag va n ohpxrg jvgu n qruhzvqvsvre naq tbvat gb orq, V qrpvqrq gb ybbx bire nyy gur frggvatf bar zber gvzr. Nyzbfg vzzrqvngryl, V qvq n snprcnyz. Gur cevagre gubhtug V jnf hfvat 3 zz svynzrag, juvyr gur npghny svynzrag qvnzrgre jnf 1.75 zz. Nf n erfhyg, vg jnf srrqvat svynzrag vagb gur abmmyr zhpu zber fybjyl guna vg arrqrq gb or qbvat. Boivbhfyl sbe lbh gb cerqvpg guvf fcrpvsvpnyyl lbh jbhyq arrq n yvggyr xabjyrqtr bs 3Q cevagvat, ohg gur zber trareny ulcbgurfvf gung gur srrq engr jnf gbb ybj vf boivbhf va ergebfcrpg. Znlor V jnf whfg biregverq ng gur gvzr, ohg gur jubyr vqrn bs fgrnz pnhfvat plpyrf bs bire-rkgehqvat naq haqre-rkgehqvat abj frrzf pbzcyrgryl vafnar. Jul jbhyq vg tb va plpyrf naljnl, va bccbfrq gb whfg pbafgnag ohooyrf va gur rkgehqrq svynzrag? Jul jbhyq V unir gb rkgehqr n pragvzrgre bs svynzrag nsgre gur raq bs gur cevag orsber V fgnegrq trggvat nalguvat ohg bbmvat svynzrag qevccvat bhg bs gur raq bs gur abmmyr? Znlor nyy bs lbh vzzrqvngryl jbaqrerq guvf, naq znlor vg jnf boivbhf jura V jnf cevivyrtvat gur ulcbgurfvf, ohg V bayl whfg oneryl pnhtug zl zvfgnxr.