Every now and then I read articles by a couple software dev bloggers I trust who are, I think, looking carefully at *how* to use LLMs in a way that might (just maybe) set devs up to be "centaurs", ie the drivers of technology, versus Doctorow's "reverse-centaurs", ie, meat-puppets directed by the system a la Amazon drivers. I don't have any particular faith that those can be achieved in any broad sense in the current context of the industry, though...

My site is honestly barely half finished, I wasn't really expecting this to go anywhere, and this was the first thing I published in years. I'll be sure to keep writing ️

I started work as a full-time software dev in 1988 (or a few years earlier if you count summer jobs), and I was lucky enough to enjoy it pretty much through to retirement a few years ago, right around the first chatgpt release. One of the parts I found most personally fulfilling was mentoring younger devs (which definitely included me learning from them and them challenging my old-guy reflexes) and to see what's happened in the very few years since is profoundly discouraging...

And thanks for @tante for forwarding this...

️

@tante

Take care ️ reach out for help if it becomes too much please

And thanks ️

In truth, I was never a coder, so let's hope I can find other ways to feed myself and my family.. it's rough out here for everyone.

@tante

Having ridden out the implosion of the dot-com crash, I’m staring at all this buildout and wondering what this time smells like. Capital spend crushed a few big names. Example: Lucent went to a nickel on the dollar.

Your story resonates (also bc I'm a career changer so old enough to have other experiences that are similar) - it's both hilarious and stark.

Well done

@tante

Also I can help nerding out a bit, but running on the assumption that Nvidia is talking about technical reality and not some marketing ploy, the raw compute power does not necessarily equate to better performance. Compute power is more subtle than gigaflops. This means that more compute power may not be found necessarily more gigaflops but in the surrounding architecture and that can be good or less good.

https://www.thundercompute.com/blog/cuda-cores-explained-ai-training

I think the thing to be pointed out is that cost of the systems versus their operational costs are very different things and this page gives you an idea of what the power draw is. Primarily the cost is in increased energy, demand and waste.

This is what I meant. Is that Moore’s law, is only been kept alive by stacking components on the silicon. For GPU this is intrinsic and ideal. You can’t contravene thermodynamics though.

https://northflank.com/blog/best-gpu-for-ai