1. the model's responses are unreliable
2. to make them reliable, you have to tell the model the answer
3. even though you now have a process to figure out the answer, you still have to use the model because the VP wants you to
not getting into specifics (because i'm not an expert), but these conversations usually lack any discussion of individual vs. group. e.g., the 10th decile of wages may stay the same, but the people who were in the 10th decile 30 years ago are not the people in the 10th decile now. i think it's safe to say that many people (not all) start and end their careers at very different quantiles in the distribution.
lattner left google and was the primary reason they chose swift, so they lost interest.
if you're asking from an ML perspective, i believe the original motivation was to incorporate automatic differentiation in the swift compiler. enzyme would be the spiritual successor to that.
Revealing that associating anything AI with money inflates the bubble. This translates to "we need to hire a $335k employee to figure out how to use our product."
I'm not old enough to remember pre-2009 well enough to comment, but since then, yes, it's been scammy.
The SV/SFBay idea of success is what the rest of us would call a pump and dump: burn a bunch of money to drive existing enterprises in the rest out of the country out of business; while providing at best marginal, and usually no, product benefit to consumers; and dumping the money losing venture on institutional and retail investors.
That's success. "Exiting." Not building something new, valuable, or sustainable.
"Oh, but wait, there are sustainable businesses," you might say. With few exceptions, they are advertising companies. Advertising is a fixed pool drawn from the real economy; all they've done is divert it from existing businesses.
Just make sure to pretend you're changing the world.
you didn't provide any hints why it wouldn't work for you, so i recommend you look at scikit-learn's neighbors module (i.e., construct NearestNeighbors object and query it).
ads are essentially noise. some high-nines number of ads have no effect. to counteract this, demand-side providers came up with ways to fake an effect to look valuable. they then discovered clients couldn't tell the difference between real or fake, or didn't care because the clients also want to say their ads worked. then the race to the bottom began: buy the absolute minimum quality for as little as you possibly can. that put pressure on the supply side because their revenues fell.
examples of faking:
- the search "companyname" takes credit for "companyname.com"
- "we promise we didn't show you the one positive result from running 1000 tests"
correctness and verification are not close to being solved. it's feasible that AI will increase individual developer productivity and narrow the field. it's also possible, and probably more likely, that increased productivity and capability will just increase the complexity of software; that may be good (more interesting problems and broader real-world applications) or bad (overengineered monstrosities).