Posted by sshah03 1 hour ago
They're also far from caring about anyone's privacy and were caught spying on their users so..
I'm not sure what you mean here.
https://www.opensecrets.org/federal-lobbying/clients/nvidia-...
https://www.opensecrets.org/federal-lobbying/clients/openai?...
https://www.opensecrets.org/federal-lobbying/clients/meta?cl...
https://www.opensecrets.org/federal-lobbying/clients/alphabe...
https://www.opensecrets.org/federal-lobbying/clients/spacex?...
https://www.opensecrets.org/federal-lobbying/clients/anthrop...
Off by a factor of ~1000:
"The Financial Times reported that Anthropic nearly tripled its lobbying expenditure to $3.53 million."
https://finance.yahoo.com/technology/ai/articles/anthropic-n...
You can effectively lobby for something for the fraction of the cost of a house in many urban areas.
What do you mean by this?
I don't have a strong opinion on Anthropic but the situation is essentially a prisoners dilemma with no good option; either dominate and risk messing it up or relinquish dominance and hot potato that responsibility to someone else.
I'm interested in seeing which way that behavior & honesty skews post-IPO.
Hacker News has plenty of the 'Legalize recreational plutonium' type, that doesn't mean said people have put a second of thought beyond their own selfish interests.
This is social media all over again where "it's a platform, we're not responsible for the content" is combined with "but we still want to ban people we don't like from using it".
They did say the system was inspired by Sapling etc... & That discussion is really around 1) What responsibility do companies have to make sure Agents haven't gone off and reverse engineered existing things, and 2) If there were ex-meta, ex-google engineers that produced code, or had AI produce code based on their memory of architecture at Google/Meta, at what point does that become an issue.
As a whole it's a different conversation than honesty around AI safety though.
Low if it's only an internal tool. Very high if you're going to make a blog (marketing) post bragging about it.
1) RSI is very dangerous, and such self improving models should not fall into the wrong hands.
2) Releasing an open model is like opening Pandora's box, you can't recall it once it's released. If that model is capable of RSI, there's nothing you can do to stop bad actors from getting superintelligence.
3) Therefore, the safest thing to do is to ban open weights above a certain capability level. The problem is we won't know what that capability level is until it happens.
I don't know what the "weakening safety pledges" part means, would be curious to learn what they're referring to.
If not, and the US law enforcement can't touch the rogue actors, then your US regulation only disadvantages good actors in the US (and hugely advantages anthropic et al by killing the competition ofc), and the rogue actors get their open models from China or whatever anyway.
Again, you can disagree with the premise of RSI being real and near, all I'm saying is their logic is consistent.
That’s Anthropic: they really do believe in catastrophic AI risk, and they also really do believe - sincerely, but mistakenly - that getting huge piles of cash and winning the recursive self-improvement race is the way to avert catastrophe.
There is little point in (for example) evaluating Dario’s sincerity, when he sincerely believes that slowing down a race without changing its direction turns “race to the bottom” into “race to the top” and not a “leisurely stroll to the bottom”.
The whole world can see the writing on the wall, Anthropic has said a million times that they are on a quest to get rich or kill everyone including themselves. And what did the world do? Gave them a trillion bucks. So what does your word do vs that? Help you sleep at night? Pointless twaddle.
I can only hope you don’t escalate to posting a kick me sign on Dario’s back!