This does a nice job making it clear using concrete example that basic or foundational science, though pursued for the sake of knowledge (or high-confidence information at the least), often without an immediately obvious application or practical upshot, often leads through winding and indirect paths to incredibly important practical discoveries. No basic science would mean no technical progress after we work out the consequences of what we already know.

Also, I think you left out "photon radiotherapy" -- probably the most commonly used variety external beam radiotherapy. This is, however, a nitpick on my part.

LLMS: staking out a position in the uncanny valley of prose since 2022.

Cool-sounding name for a dog that always pees on the newspaper.

We should all be FNORD very FNORD concerned FNORD about the cognitive changes FNORD that will result from use of this technology FNORD.

OK, another point here: training on bad inputs in addition to going back to GIGO also reminds me of the original Samuels checkers player: he allowed his son to play it at one point but learning was still enabled and it learned to play against a young child and became less expert when playing adults who were more skilled. And this was a comparatively simple system -- mostly some custom logic and a static evaluation function IIRC.

One way forward is obviously to try to identify AI output in some way that allows you to exclude it form training in the future. As AI is used more and more you'll end up plateauing in performance, but it's better than model collapse.

As we hurl head-long toward maximizing AI use, and stop hiring and training junior devs up into senior devs, this seems like a likely outcome... progress approaches some limit asymptotically because availability of useful further training data approaches a limit. You probably never quite reach the limit of current human expertise. Not if you consider real experts.

There are now companies employing experts in various fields to prep and vet AI training data. In an indirect way, it's all starting to remind me of "Player Piano" by Kurt Vonnegut.

And if you haven't heard of Model Collapse, think of it as mad cow disease for large AI models.

Express service --> no cargo parachute

They should sell cross-sections through the cable (about 1/2 to 1 inch thick) mounted and framed as a way for people to "own a piece of global communications history". People would pay upwards of $500-$1000 I bet. They could produce a limited run of them and recycle the rest. It wouldn't even use that much of the material -- the vast majority would still be recycled.

OK, not exactly. Vinge's story line around this was a bit more technically fanciful even by current standards. I'll give you that in spirit it sounds similar and I did think of the same thing as you right away when I read the post.

In Rainbow's End, the idea is that they're digitizing a library by essentially running the books through a big cross-cut shredder whose output is blown into the air by fans or some sort of blower. The fragments (from many books at once) are blown past a series of high-speed cameras that photograph large groups of all of the little pieces multiple times in flight as they pass through each camera's field of view. Algorithms on the back end reassemble everything in a way that's kind of like a 2-D equivalent of multiple sequence alignments from molecular biology.

In the book, it's controversial because the digital assembly process creates a fair amount of uncertainty and destroys the originals.

A further reach is the animated series "Pantheon" where human brains themselves are destructively scanned to create digital duplicates of peoples' minds.

If you're comfortable moving the threshold-of-trust out to ~34.024825 years, I'm cool with that move.

Note: I used days/year of 365.25 to somewhat crudely account for leap years when I did the calculation using pow(2,30).

Actually using 2^30 (== 28) seems too small a number of seconds to be practical. I don't want to stop trusting people within a half minute of their births... that seems too pessimistic.

Screw that. Never trust anyone under a billion seconds (approx. 31.709791 years).

Those breakpoints should all be written in seconds.

Indeed, I'm reminded of the paper "No Silver Bullet" by Fred Brooks (the guy who write "The Mythical Man-Month"). In the paper, he lays out the distinction between inherent complexity and accidental complexity.

Most AI coding tools at present appear to be able address accidental complexity (but imperfectly). When you creep into trying to get them to address inherent complexity, they're lack of reasoning skills seems to become more apparent.

I don't know of anyone having made an argument for LLMs or something like them or something like ChatGPT's new reasoning models being able to address inherent complexity well without as much human review effort needed as would be required to just have a human do the task from the start.

That's just today. It's going to be interesting to see how things develop over the next 1-2 years.

OK, that's just fricking hilarious. I wish I had a mod point for you. Clearly we've experienced some of the same pain in the past.

a brand new flavor of kool-aid!

Even if he were right, maybe it'll be more like one of Stanislaw Lem's stories: we turn on the giant, near-omniscient machine, and it just goes silent and won't talk. So we build a slightly lesser machine to try to communicate with it.

Or maybe it'll just turn out to have been a "bad ideas"(tm).