I wouldn't label this a "Revolution", not yet, it's a strong word. I think the real next tech revolution will be along the lines of room-temperature superconductor, or finding a solution for stable qubits. Former enabling the latter would be exciting time to witness.
This is precisely what so many do not get when they are arguing against independent, skeptical academics or scientists. The academics/scientists get basically nothing from their position. Sure, they get papers and maybe tenure and so on, but they can get those holding either position. Meanwhile, on the other side we have people with literally millions riding on the position. I wonder who will provide a better unbiased take?
This is further muddied in the LLM debate, as we have a generation (of mostly nerdy middle class white men) who grew up on sci-fi desperately want their boyhood fantasies to come true. I was one of them, but I got over it - we're not colonizing Mars to save humanity and an LLM is not anything resembling AGI. People not invested in the narrative can see that LLMs, while amazing at generating sensible sounding text, are not even remotely close to reasoning AGI.
The article argues that AI won't make companies a huge amount of money, which isn't exactly the same as what the title implies. For example the article discusses how open source models have significantly improved. That's bad for OpenAI, but good for open AI.
It's funny because this article will be forgotten but it will be so wrong soon that the author will not want to be reminded. Classic case of Kruggerman thinking.
I don't think it's slowing down yet. Every day new cool models appear. I'm having a field day experimenting.
It's now possible to run a gpt-3.5 quality model at home which is pretty amazing imo. Further dedicated hardware will bring the cost down like the ASICs did with bitcoin mining.
I do think the usefulness is overrated. Microsoft is heavily pushing CoPilot for office as an assistant that can do anything you want just by asking, but it is really disappointing. It's great at making quick demo presentations but at real work it either gets it totally wrong or it just throws its hands up and refuses. Microsoft seems to be aware of its limitations and is now pushing more towards training on more complex prompting. But I don't really see the point: that effort could be spent teaching people to do the work too.
But for some use cases like rewriting, summarising and translating it shines and is reliable.
If this article was posted on HN as commentary, it would be shredded for lack of substantiation. In the WSJ, it’s particularly pathetic. First, there needs to be a revolution, then it can lose steam. Second, the LLM revolution may just be the sign that a real AI revolution could be in the cards. Revolutions take patience, ask any historian
17 comments
[ 3.1 ms ] story [ 54.0 ms ] threadicloud private relay?
[0] https://www.msn.com/en-us/money/other/the-ai-revolution-is-a...
for context, Marcus has a quite bearish view on AI and lots of others, most prominently Yann LeCun, disagree with Marcus.
"It is difficult to get a man to understand something, when his salary depends on his not understanding it." -Upton Sinclair
This is further muddied in the LLM debate, as we have a generation (of mostly nerdy middle class white men) who grew up on sci-fi desperately want their boyhood fantasies to come true. I was one of them, but I got over it - we're not colonizing Mars to save humanity and an LLM is not anything resembling AGI. People not invested in the narrative can see that LLMs, while amazing at generating sensible sounding text, are not even remotely close to reasoning AGI.
There is a big hype around AI.
Some people mention AGI but have no metrics or definition about cognition, conscience, what is intelligence in terms of evolution.
Almost nobody talk about intelligence in small brains like insects, while it's probably much easier to simulate.
Generally computer scientists will probably dismiss psychology because it's not hard science. It says a lot.
All the sales bros will find something new and we’ll go back to solving MDPs and life will be fine
It's now possible to run a gpt-3.5 quality model at home which is pretty amazing imo. Further dedicated hardware will bring the cost down like the ASICs did with bitcoin mining.
I do think the usefulness is overrated. Microsoft is heavily pushing CoPilot for office as an assistant that can do anything you want just by asking, but it is really disappointing. It's great at making quick demo presentations but at real work it either gets it totally wrong or it just throws its hands up and refuses. Microsoft seems to be aware of its limitations and is now pushing more towards training on more complex prompting. But I don't really see the point: that effort could be spent teaching people to do the work too.
But for some use cases like rewriting, summarising and translating it shines and is reliable.
Doesn’t to me indicate a slowdown in the wider trend. Stuff like figure, sora, chat bots etc seem here to stay