simonshen patreon

【國際關係深度評：經典復刻】第四次工業革命：人工智能要多久，才會顛覆全球?

🕑 Added 2025-02-17 07:01:01 +0000 UTC

Comments

Carl Yang

只想得到資本回報…

堅離地書院 College

最上游的會壟斷，其他的韭菜，未來世界的 polarization 可能更誇張

堅離地書院 College

Yes unfortunately polarization is the trend. A few extremely well-off elites, Vs "the others", in the coming new world.

堅離地書院 College

佢係刻意想傳遞呢類訊息

Good Year

AI做唔做到1件事, 行業有無引入另一回事, 華人社會係無創新風氣, 又睇重監管, 監管方無進步, 整體都無人會做

Good Year

資訊科技界的員工都唔一定有前途, 因為AI 搞掂哂, coding 第1個受影響最慘, 因為AI 集translation ( 人係唔洗同AI講專業用語佢可以自己理解, 配對) 寫code, debug 於一身, 仲更快理解D客要咩只會更死。AI暫時都無創意, 人類就只有創意呢個最後一關。

lyk

「要多久，才會顛覆全球」 — we can only talk about the things we know: 1/ most guesses about technology diffusion are wrong, and sometimes wrong by orders of magnitude 2/ constraints of getting there — e.g. how many mega data centres do we need, and where are they? How much electric power do we need, and how many nuclear power plants can we build in the US? How long would it take to rebuild a half century old electricity grid? How many chips do we need, and where do we make them? 3/ Algorithms, narrow AI, AGI, agents on one hand; automation, robots, humanoids on the other, all have different constraints. Who makes them and where, and how many? 4/ Learning from human experience from the internet is one thing, how do they create new knowledge, collect data, research and experiment in the real world? 5/ There will be economic cycles. How much capital is required and who pays for all this? What are people willing to pay for — just to do something, do it cheaper or do it better? 6/ Humans are general purpose. Job specifications generally has more than one item. A lawyer offers legal advice, is also a sales person and negotiator. When will we reach AGI? 7/ Who will help deploy them in specific industries and globally? Suppose a humanoid does surgery in an emergency room, who trains it? who wants to be operated on by a humanoid? who is responsible when it goes wrong? is it more likely a human doctor will be assisted by a machine or replaced by a machine? 8/ Who will be managing events and projects? Who will deploy the vaccines for the next Covid, who will put out the next fire in LA, who will build the next building, who will experiment, test, build and fix the next plane? 9/ Institutional, physical infrastructures are built around humans. Humans will resist some of these changes. Who will be organising and managing humans, and how long would it take for these infrastructures to evolve? 10/ Who is going to serve the Digital Overlord, supposing there is one? I know if I try to answer these questions I will be wrong. Coming back to the timeline question, we do know: 1/ software advances has been accelerating exponentially enabled by breakthrough in compute (GPUs), but there will be constraints building them 2/ some technology adoption like smartphone happens very fast, while energy transitions (charcoal to coal to fossil fuel) can take decades or a century. AI is much broader in scope than either of them 3/ the more dependencies there are, the longer it takes. There are fewer dependencies in things like a general chatbot, more in others applications 4/ At this point, it looks like it will be more of an assist than a replacement for most scenarios, which isn’t offering enough value to justify the current cost for wider adoption. Despite the exponential technical advancements, the deployment so far looks narrow and seems more like an evolution more than a revolution. But that could change any time. From the above, it seems like if you are good at what you are doing, no matter what field you are in, we and our kids have a good chance of surviving, insofar as this continues to look like an evolution and a human assist in the foreseeable future. The divergence between who is good at something and who is not will increase. Life long learning is the key. Technologists have unlimited imagination and tend to get lost in talking about what AI can potentially do. Playing Go on a chessboard is neat, the rules are so well defined, but the real world is messy. Elon prophesied about FSD in 2 years in 2015, which is narrow AI, we are still nowhere close at all. The real world has constraints, just tabling a few of them here.