3 ms·

I saw some blurb (I can't find it now) about the optimal time to launch an interstellar probe, and the punchline was that embarking too soon means you will be o

by abeppu 2y ago

I saw some blurb (I can't find it now) about the optimal time to launch an interstellar probe, and the punchline was that embarking too soon means you will be overtaken by a later vehicle which launched later but with better tech. When the rate of technological progress relative to the duration of a mission is above some threshold, it's rational to do more front-loaded research before actually committing to a particular plan to apply it. In the AI case, advancement has been so rapid that to a certain extent, it makes sense for large companies to delay some aspects of retooling, because better models that come out next month may have different characteristics or requirements. We haven't seen the big shifts in productivity because reorganizing around new tech is slow, and the tech itself is still changing so fast.

rickydroll 2y ago
I think another analogy that applies to AI is going to orbit. It's "not hard" for a sufficiently large rocket to be able to hit orbital heights. It's pretty hard for the rocket to enter orbit at the same height. In the AI world, we are at the suborbital stage of technology. For me, the 99% use case is accessibility tools (speech recognition, grammar correction) and coding. Image generation is like poor-quality Photoshop. Which, by the way, is the best target image generative AI could shoot for. Do what Photoshop can do, only better and easier.
avaldez_ 2y ago
>the optimal time to launch an interstellar probe, and the punchline was that embarking too soon means you will be overtaken by a later vehicle which launched later but with better tech. That's the wait calculation https://en.m.wikipedia.org/w/index.php?title=Interstellar_travel&diffonly=true#Wait_calculation https://en.m.wikipedia.org/w/index.php?title=Interstellar_tr...
TylerE 2y ago
Maybe, but all technology is built on what came before, so actually launching the thing and seeing what it does is a big part of enabling the next generation to be better.
- abeppu 2y ago
  I'm not sure that we disagree, but perhaps we're understanding "launching" in the metaphor somewhat differently. I think the key point about the interstellar probe example is it's assumed that once it's launched, it's not going to improve on the very long flight to Alpha Centauri. Maybe this isn't actually fair -- if you design a solar sail, pushed by a giant laser that stays in our system, you could likely upgrade the laser while the probe is in flight, but your probe can't build itself a new larger sail, or develop itself a fusion drive. I think research is rapidly building off the prior generation of techniques, but this improvement doesn't show up in the economic impact the article author is complaining about. Actual productivity changes are going to come when large organizations integrate AI models to replace or improve key systems or human-performed functions. This is slow, expensive, and probably a minority of transformations like this are even started, let alone showing fruit, and I think it's rational for a lot of companies to defer starting this work until the tech improves further. There's also a possibly prolonged intermediate phase where some stuff is being done both by an AI-centered system and either a human or a legacy system while the organization measures how well the AI version does (and how stable/reliable any associated vendors are), and during this period, productivity may even take a measurable hit.
gumby 2y ago
There was a funny (joke) academic paper circulating around 2000 that showed if you were plannig to do a computation that would take about a year that in 6 months you could pay 1/3 less and still get the computation done faster. What made it so funny is not only that it was written so well but that things were moving so fast back then it felt plausible.

by abeppu 2y ago

rickydroll 2y ago

I think another analogy that applies to AI is going to orbit. It's "not hard" for a sufficiently large rocket to be able to hit orbital heights. It's pretty hard for the rocket to enter orbit at the same height. In the AI world, we are at the suborbital stage of technology. For me, the 99% use case is accessibility tools (speech recognition, grammar correction) and coding. Image generation is like poor-quality Photoshop. Which, by the way, is the best target image generative AI could shoot for. Do what Photoshop can do, only better and easier.

avaldez_ 2y ago

>the optimal time to launch an interstellar probe, and the punchline was that embarking too soon means you will be overtaken by a later vehicle which launched later but with better tech. That's the wait calculation https://en.m.wikipedia.org/w/index.php?title=Interstellar_travel&diffonly=true#Wait_calculation https://en.m.wikipedia.org/w/index.php?title=Interstellar_tr...

TylerE 2y ago

Maybe, but all technology is built on what came before, so actually launching the thing and seeing what it does is a big part of enabling the next generation to be better.

abeppu 2y ago
I'm not sure that we disagree, but perhaps we're understanding "launching" in the metaphor somewhat differently. I think the key point about the interstellar probe example is it's assumed that once it's launched, it's not going to improve on the very long flight to Alpha Centauri. Maybe this isn't actually fair -- if you design a solar sail, pushed by a giant laser that stays in our system, you could likely upgrade the laser while the probe is in flight, but your probe can't build itself a new larger sail, or develop itself a fusion drive. I think research is rapidly building off the prior generation of techniques, but this improvement doesn't show up in the economic impact the article author is complaining about. Actual productivity changes are going to come when large organizations integrate AI models to replace or improve key systems or human-performed functions. This is slow, expensive, and probably a minority of transformations like this are even started, let alone showing fruit, and I think it's rational for a lot of companies to defer starting this work until the tech improves further. There's also a possibly prolonged intermediate phase where some stuff is being done both by an AI-centered system and either a human or a legacy system while the organization measures how well the AI version does (and how stable/reliable any associated vendors are), and during this period, productivity may even take a measurable hit.

gumby 2y ago

There was a funny (joke) academic paper circulating around 2000 that showed if you were plannig to do a computation that would take about a year that in 6 months you could pay 1/3 less and still get the computation done faster. What made it so funny is not only that it was written so well but that things were moving so fast back then it felt plausible.