Requiring way too much resources will inevitably be solved by better, more efficient chips. It's a non-issue in the long run. Each generation requires less power than the previous one, and we don't even need Moore's law to make this happen. I don't see why people continue to float this talking point. If anything, what likely will happen will be data center build outs motivating investments in long neglected improvements in energy infrastructure, that people will gradually see the benefits of as better chips result in reduced power requirements. This is contingent on AI demand plateauing, which it inevitably will, and if it doesn't that means that the economic motivation for such build outs is justified.
replyThe risk to the global economic system is also oversold. AI enables the same work to be done for a lower cost, and more types of work to be done overall. No different than previous industrial automation. There are vast capabilities in medicine, energy, automated design, manufacturing, and science that are only now being unlocked. Does that sound like something that will crash the economy?
You will be right only if AI will be usable locally on commodity hardware, otherwise access to AI will become a weapon on various fronts - including your employer, your home country government, foreign countries, and international orgs.
replyThere's nothing that indicates LLMs running in data centers is more evil than LLMs running on your own rig. The issue is centralization of providers into a duopoly or triopoly that can use their power to absorb more training data and operate a pricing cartel, as well as absorb knowledge from other individuals and corporations.
replyRight now the models actually run most optimally and economically in a data center b/c multiple requests can be batched on a single GPU. This may change of course...
Providers cannot impose daily quota on local AI usage
replyI don't think ai can annihilate us from a misalignment issue but I think there exists a real risk of real world disaster caused by one of these poorly executed benchmarks. They don't seem to be too aware of what they're up to when they let them run for days and days and I could see a scenario like huggingface but for something that matters. Dumb ai is more dangerous than anything right now.
replythe topic literally started with an unrealistic approach: "Lets assume I'm spiderman and have spider-powers [...]" ... I would prefer being Phoenix, or maybe Thor from the Avengers, the cartoon version, possibly not the cropped down, dumbed down lowly-God Avengers endgame version...
replyDo you have an argument? Actually, could you just give me a workable definition of alignment?
replyI'm weird and worry about AI x-risk though, so feel free to dismiss me for that if you like.