You have your orders placed for the Anthropic IPO, right /biz/? Do you realize how important this company will be if they solve AGI? It will make Nvidia look like peanuts.
>singularity annonced episode 45037
Hey Avital! What do you mean: If?
>>62884970I recently just re-read "superintelligence" by Nick Bostrom...and one of the things i found really strange was his timeline. The trajectory from human-level intelligence to god-like intelligence, and his fast one is hours to a few days, medium is weeks to a month, and slow is several months to years.It's hard to get an intuitive understanding of exponential growth. Like image a water droplet fall in your room, next second 2 droplets fall, the second after that 4 droplets, etc.How long untill you are drowning? an hour? a day? how long untill the blue planet becomes entirely blue, the sea level finally reach Himalaya? A day? week? month?It seems like he, Nick Bostrom, believes the singularity will happen almost instantaneously, hours or weeks are basically the same, you have no ability to react. you can't do shit.
>>62885015I love that book. It's great.I've always wondered why there's a bit of red on the cover, it looks like a printing error. Mine has it too.
>>62885015Interesting but most of it is speculation pulled right out of someone's ass. In reality no one knows how the singularity could happen or how because it never happened before, and we have no way of predicting or modelling it
>>62884970It can't happen because we don't fully understand the human brain yet.
>>62885068Yeah and guess what? We don't fully understand how AI even works either
Not a luddite and I'm extremely bullish on AI but most likely, it will do something similar to what SPACEX did. Dump for weeks until it finds its bottom and up and to the right from there.
>>62885073uhh I think we understand it pretty well. "concepts" are mapped to coordinates in space, these coordinates have weighted relationships to one another, and with simple arithmetic repeated billions of times, a computer can take an input, digest them into "concepts", analyze the relationships between the concepts, and produce the most likely output that directly follows those concepts. And you can greatly finetune that output by adding in very specific context pertaining to the input and desired output
>>62885073We built it.
2 more weeks bro
>>62884970https://youtu.be/atzuvm7vy-w?si=6Q1fp2I9590P5Tue
>>62884970>Two more weeks
>>62885068https://ai-2027.com/
>>62884970Translation: Buy the IPO :)We are literally creating a genie and instead of keeping the upside for ourselves we graciously decided to share it with you for 3 trillion USD :) you're welcome :). Also your pension fund will be buying our shares if we end up in the top 500 :)Lmao
>>62884970I remain skeptical about the potential of RSI. No matter how sophisticated an LLM and its tooling get, there are limits to domain knowledge that an LLM cannot expand with information it does not have. We've seen in math for instance the explosion of discoveries when the LLM was trained on the highly specific domain knowledge of mathematicians and drew reasonable conclusions from that information. But now that information has been exhausted; what further exponential progress can be made, once it has "used up" that knowledge? It needs more guidance to continue forward. So certainly the speed of research will be enhanced, but the idea that we will roll into a self-sustaining dynamo of discovery seems thermodynamically impossible. There are also physical limitations. Suppose with the domain knowledge of quantum physicists and fusion engineers, an LLM is able to solve two of the hardest applied physics problems: the mass qubit alignment problem, and fusion reactors that reliably produce net energy. There will nonetheless be lag time before a sophisticated quantum computer and fusion reactor will be built, and mass production will take years. This is assuming these problems can be solved at all, or if we will discover unfortunately that we are limited by the laws of physics and there are simply no scalable solutions for these technologies. We just don't know.
>>62885056This. When AI replaces all the researchers (is in progress), experiments and training are still bound on compute. Computers can't rewire themselves to have more FLOPS, and every next iteration has to be shipped over the Pacific.
