The watch has ended for me but new shall come with more dignity and fire. The wind of change will one-day whirl over the Mecca of Nepalese Engineers. I couldn't be but there would be some and their songs will be remembered, if not written.
Your Favorite NFT Marketplace on @Aptos is now a one-stop aggregator
Explore Now - https://t.co/Jtumhl7lCX
To celebrate the launch, we’re organizing a 100M GUI @guiinuonaptos, trading competition.
Only Users Maintaining 100% Loyalty will qualify for prizes.
More details - https://t.co/4hMkpxAKVC
I have mad respect for Karpathy. But RL agents will not find exploits in physics that give us infinite energy, or anything like that.
As somebody who knows a thing or two about both AI and physics, I am quite certain of this. The so-called standard model of particle physics (which includes QM, which my grandfather played a founding role in) is _extremely_ good at predicting reality. Its biggest problem is that it is too good at predicting experimental outcomes, leaving very little room to look for improvements. We know it doesn’t mesh well with gravity / general relativity. But the answers here are unlikely to come from direct experimentation because the relative strengths of the effects are many orders of magnitude off from what can be measured - unless you get black holes involved, which thankfully nobody has a clue how to do in a lab.
Finding bugs in simulations is something RL can do - and it is both hilarious and inspiring when it happens. Maybe our reality is itself a simulation - this line of thought occurs to most people who ponder physics for long enough, and might never be disproven. But we have already pushed the bounds of our maybe-simulation many many orders of magnitude away from anything that can be experienced in daily life, and never seen even the slimmest hint of any sort of “bug”.
AI for fundamental science research - absolutely! Throwing out crazy ideas during live interviews - yes, please! Conservation of energy, thermodynamics, basic laws of physics - completely unassailable, no matter how smart the algorithms are.
Yes we know our understanding of physics is incomplete - but we know what those gaps look like well enough to know that there are no infinite sources of energy that can be discovered by AI-guided-random search through experiment space.
AI-guided random search to get a robot to do house chores - absolutely!
AI-guided scientist to figure out dark matter - maybe???
Perpetual motion - never gonna happen.
Has innovation in AI boiled down to quality training data and tricks in scaling. The small language model with 2.7 B params Phi-2 beating 25 times larger models in different benchmarks is amazing.
Just used it's hf implementation loved it