This feels a bit like a reframing of the rather absurd "Eternal Sloptember" nonsense, desperately trying to pivot from Luddism to visionary (and yes, I know this is the "hacked the iPhone" guy. I'm not a cult of personality person and I positively do not care). Also incredibly weird how it repeatedly talks about people moving to San Francisco, which ... what in the world is that nonsense about? People talking about the concerns of AI and automation are in no universe considering moving to SF as the protection...

"Where’s all this new magical software that the productivity improvements should imply?"

This is a recurring gotcha in the anti-AI marketplace of denialism. It's a bit like saying "I saw a fat guy, so why do people keep telling me that GLP-1s change everything?"

It takes time. Like already I would say just about every programmer has replaced a number of tools with random shit they spit out from LLMs. It percolates out from there as some things become products, etc.

And to anyone actually paying attention, and not just feeding their delusions, the impact is utterly enormous. Incontestable. The "Where's the software? / Where's the change?" people are absolutely going to find themselves in the dustbin of history.

It's also fascinating how often people do the stochastic parrot horseshit.

The other night I had to do a large scale compression test with libjxl, which notably is software that has seen an enormous amount of optimization interest and you would assume has little extra to be eked out. I've traced through this software before and the compression path is insanely complex. It's the sort of software that is headache inducing. Anyways, curious what the state of the platform was I grabbed the latest source and asked Fable to look for low-hanging fruit in the lossy and lossless compression paths. It suggested a few, created a test harness to A:B bitwise compare with the original, and implemented its optimizations. It achieved a 14% performance increase in a single pass, using just the remaining quota I had on a subscription as my week drew to a close. And all it did was some high level logical optimizations, some more efficient memory allocations, and so on. All of its code was completely in the style of the project, was no more significant than necessary, and so on. Anyone that isn't utterly blown away by that -- who gets the hype -- is lying to themselves.

I love LLMs, I hate hype