Hey Martin very cool to see you on Substack, I just "recommended" your Substack and am looking forward to following your writing. I will probably be a regular in your comments section haha.
If you get a minute, checkout my Substack, I mainly investigate corruption/fraud in academia but also blog about prediction markets, economics, culture wars etc.
The DNA computing part? I have to do some real research in that space. Probably better for storage than compute? Silicon transistors are pretty damn small but I wonder if there are thermodynamic benefits to DNA.
I meant the whole section, starting with the web3, and especially the predictions about AGI. There's some progress in AI, that in 20 years will give us more sophisticated tools, but I don't see a capable AGI on the horizon in 20 years, unless we have some kind of black swan style breakthroughs, which is very very unlikely. My bet: at least 100 years.
We probably have some time before true AGI, but I think we're not far from there in some incomplete sense. LLMs are certainly causing a stir, and I think when added to Nilsson's triple-tower model, you have a real case for something that is close to at least faking it. The REPL style LMs/GPT-3 gives us now is obviously far from human-like, it feels more like calling an API.
If you feel like LLMs are doing something that resembles human reasoning (which cannot be *much* different than some undefined vector optimization going on in our wetware), then I'd say we're not far off. But if you feel like LLMs are a parlor trick, then I would guess the onus is on you to explain what we're missing. Penrose/Rumsfeld explanations that we don't know what we don't know seem lacking, to me. Computers have checked a lot of boxes in things humans can't do well, and obviously need to check more boxes in things humans do easily a la McCarthy. Yet, the new text-to-image models are quite impressive, and check off something a 5-year-old can do that a computer previously couldn't.
The God of Mischief Returns
I say this with Love my good man.
Yes, indeed! Although I think I will make a little less trouble this time around.
Hey Martin very cool to see you on Substack, I just "recommended" your Substack and am looking forward to following your writing. I will probably be a regular in your comments section haha.
If you get a minute, checkout my Substack, I mainly investigate corruption/fraud in academia but also blog about prediction markets, economics, culture wars etc.
https://karlstack.substack.com/
Is the crypto section satire? Also, welcome back home Shkeli.
not at all! just fantasizing about the potential future outcomes. they don't have to have high probabilities.
Glad you are back. Good read
Good to see you're back man. Interacted with you briefly on the blog when you were in prison, and on reddit years earlier when wsb was in its infancy
If you ever feel like doing YouTube streams in the future we'd love to hear from you again
The last prediction for "Computing" is all wrong lmao
The DNA computing part? I have to do some real research in that space. Probably better for storage than compute? Silicon transistors are pretty damn small but I wonder if there are thermodynamic benefits to DNA.
I meant the whole section, starting with the web3, and especially the predictions about AGI. There's some progress in AI, that in 20 years will give us more sophisticated tools, but I don't see a capable AGI on the horizon in 20 years, unless we have some kind of black swan style breakthroughs, which is very very unlikely. My bet: at least 100 years.
We probably have some time before true AGI, but I think we're not far from there in some incomplete sense. LLMs are certainly causing a stir, and I think when added to Nilsson's triple-tower model, you have a real case for something that is close to at least faking it. The REPL style LMs/GPT-3 gives us now is obviously far from human-like, it feels more like calling an API.
It also depends on your opinion of LLMs (I can guess where you stand!). This article on "attention heads" was quite interesting: https://www.quantamagazine.org/researchers-glimpse-how-ai-gets-so-good-at-language-processing-20220414/
If you feel like LLMs are doing something that resembles human reasoning (which cannot be *much* different than some undefined vector optimization going on in our wetware), then I'd say we're not far off. But if you feel like LLMs are a parlor trick, then I would guess the onus is on you to explain what we're missing. Penrose/Rumsfeld explanations that we don't know what we don't know seem lacking, to me. Computers have checked a lot of boxes in things humans can't do well, and obviously need to check more boxes in things humans do easily a la McCarthy. Yet, the new text-to-image models are quite impressive, and check off something a 5-year-old can do that a computer previously couldn't.
Let me know what you think!