Using that, I can sort of get a prior for the likelihood of “big” discontinuities; which falls between 8% (4/50) and 24% (12/50). I can also get a rough probability of a discontinuity per year (~1% if the technology ever shows one). All of this has caveats, outlined in the post.
***
Your first point, that if I paper-push hard enough I can make anything look continuous doesn’t apply, because I’m not in fact doing that. For example, throughout WW2 there of were several iterations of the radar, each progressively less shitty, but progress was fast enough (due to the massive, parallel, state funding) that I’d still categorize it as a discontinuity (and note that it did get into the OODA loops of the Germans, the effectiveness of whose submarines greatly declined after the deployment of radar). Similarly, the Wright brothers also experimented with different designs, but overall their progress on heavier than air flight was rapid and anomalous enough that I categorized it as a discontinuity. Similarly, for transistors, there were of course many different transistors, but MOSFET transistors were substantially better on miniaturization than BJTs, even if MOSFETs were worse during their very first years. Similarly, transistors themselves were better than vacuum triodes, though I’m sure that if you squint you can also find something continuous somewhere.
Even if I were paper pushing, detecting 12⁄50 would still give you a lower bound for the probability of a “big” discontinuity (24% + however many I paper-pushed to the medium or small categories). Even if there wasn’t a clear line between “continuous” and “discontinuous”, I’d expect more continuous technologies to fall in the “medium”, “small” and “probably not” buckets, and more discontinuous technologies in the “big” and “medium” buckets.
Some of your gripes could conceivably apply to some parts of AI Impacts’ investigation (e..g, they don’t categorize printing as a “large and robust” discontinuity), but I am not them.
What? I feel like this comment doesn’t answer to the post above at all.
tl;dr of the post: If I look at 50 technologies which to a first approximation I expect to be roughly randomly chosen, I can broadly divide them into:
Probably with “big” discontinuities: Aviation, nuclear weapons, petroleum, printing, spaceflight, rockets, aluminium production, radar, radio, automobile, transistors, and PCR.
Probably with “medium” discontinuities: cryptography, glass, rail transport, water supply and sanitation, diesel car, automation, television, steam engine, timekeeping devices.
Probably with “small” discontinuities: cycling, furniture, robotics, candle making, sound recording, submarines, batteries, multitrack recording, paper, telescopes, wind power.
Probably not discontinuous: ceramics, film, oscilloscopes, photography, artificial life, calendars, chromatography, bladed weapons, condoms, hearing aids, telephones, internal combustion engine, manufactured fuel gases, perpetual motion machines, motorcylces, nanotech, portable gas stoves, roller coasters.
Using that, I can sort of get a prior for the likelihood of “big” discontinuities; which falls between 8% (4/50) and 24% (12/50). I can also get a rough probability of a discontinuity per year (~1% if the technology ever shows one). All of this has caveats, outlined in the post.
***
Your first point, that if I paper-push hard enough I can make anything look continuous doesn’t apply, because I’m not in fact doing that. For example, throughout WW2 there of were several iterations of the radar, each progressively less shitty, but progress was fast enough (due to the massive, parallel, state funding) that I’d still categorize it as a discontinuity (and note that it did get into the OODA loops of the Germans, the effectiveness of whose submarines greatly declined after the deployment of radar). Similarly, the Wright brothers also experimented with different designs, but overall their progress on heavier than air flight was rapid and anomalous enough that I categorized it as a discontinuity. Similarly, for transistors, there were of course many different transistors, but MOSFET transistors were substantially better on miniaturization than BJTs, even if MOSFETs were worse during their very first years. Similarly, transistors themselves were better than vacuum triodes, though I’m sure that if you squint you can also find something continuous somewhere.
Even if I were paper pushing, detecting 12⁄50 would still give you a lower bound for the probability of a “big” discontinuity (24% + however many I paper-pushed to the medium or small categories). Even if there wasn’t a clear line between “continuous” and “discontinuous”, I’d expect more continuous technologies to fall in the “medium”, “small” and “probably not” buckets, and more discontinuous technologies in the “big” and “medium” buckets.
Some of your gripes could conceivably apply to some parts of AI Impacts’ investigation (e..g, they don’t categorize printing as a “large and robust” discontinuity), but I am not them.
Big fan of your work, though.