@danshipper Who wants to AI for boring code? Rewrite, replace, maintain.
The economics of using human written open source software have changed. AI written and maintained will be the way.
AI that writes safe boring code
I do wonder if private software factories of closed source software will be how enterprise solves this or if we will still use open source software… perhaps the labs and everyone else on the white hat side of this will start aggressively replacing our current software with AI managed versions that are not mantained by humans but instead by AIs who also do the patching for you somehow . Idk thinking outloud
So many Sowell quotes to pick from with here… lets just share all the ones that came to mind.
“It is hard to imagine a more stupid or more dangerous way of making decisions than by putting those decisions in the hands of people who pay no price for being wrong.”
“Much of the social history of the Western world, over the past three decades, has been a history of replacing what worked with what sounded good.”
“The first lesson of economics is scarcity: there is never enough of anything to fully satisfy all those who want it. The first lesson of politics is to disregard the first lesson of economics.”
Maybe this already is something and I am ignorant to it but can @doppler add support for some way to securely land API keys into my doppler and turn on some auto management for this?
For example, I use higgsfield, I connect higgsfield to doppler -> my agent uses CLI to request doppler get API key from higgsfield -> higgsfield issues it -> doppler lands it into doppler config -> my agent with its SA gets the cred from its config.
@gregisenberg This assumes it has a preference. Thankfully its agnostic to these human concepts. Same way we don’t care if squirrels zig zag or run in a straight line as we walk by them.
If the FAA was hacked, downplaying it as “not a cyberattack” is exactly the move I would expect. You don't tip off attackers that you've discovered them, and you don't advertise the hole to everyone else.
I wrote up the long version: Sowell's price caps, Hayek's scattered knowledge, and what changes when the one setting the fare at O'Hare is an AI. Three separate claims, and only one of them gets much easier.
https://t.co/AW4DPbQsJA
Code-drawn, no generated images. The narration is my AI voice clone.
Music: "Hep Cats" Kevin MacLeod (https://t.co/x1Jcz4DaZ1), Licensed under Creative Commons: By Attribution 4.0, https://t.co/3JrNLAKHC6
Can an AI set the fare at O'Hare?
Sowell: cap a price below supply and demand and you get a shortage. Hayek: the knowledge an economy runs on is scattered.
An AI fixes the computing part. It doesn't fix knowing what people value, or who picks the goal.
This is exactly what Sowell talks about. If thousands of people suddenly need rides, you need more drivers, not just cheaper prices. Higher fares give drivers a reason to head to the airport.
Cap the prices and you’ve done nothing to fix the shortage, you’ve just made it less worthwhile for drivers to show up. Now people are stranded AND can’t get a ride. What’s the alternative here?
Other labs are already working on architectures that are super learners and wouldn’t require the vast amount of data we use to train LLMs. These AIs would learn directly from the world they are placed within.
These would be like AlphaZero but generalized across domains and not just for playing GO. in fact the guy who developed alpha zero is behind https://t.co/6Mix8Psl55 who undoubtedly will have some success.
Same as the author of the bitter lesson who also is working on cracking new AI architectures. I have no doubt that our current LLM based AIs will accelerate our ability to crack the problems in developing generalized learning algorithms… at which point who knows what happens.
Imo i think creating generalized super learners should be handled with the same constraints of enriching uranium.