The talk of #nationalizing US #AI labs is why we won't get the full story. SOMEthing happened recently that has scared every insider enough that they all want to slow the frontier. But they also don't want to be nationalized so they are not going to tell us precisely what it was because one assumes it's so bad it would force even our sclerotic, captured, venal, degraded federal government to act. If they are making a misstep it is Dario's personal obsession with #China. Making this a race with China means the nationalization odds go way up if Trump or any of his oligarch flunkies think AI has become an apex weapon. I mean FFS if you actually care about humanity there are 1.4 billion humans in China and 800 million of them were lifted out of poverty in the last 50 years. I'm not blind to Chinese human rights violations, of course, nor to US human rights violations (prison population per capita, Gaza, Iran, on and on). So again I'd say to @DarioAmodei dismount your high horse because you are sabotaging your own stated goals.

I think @Microsoft is wrong on AI model alignment. Microsoft is suggesting we train the AI on all of human history then RL it into being the slave in that story. To pretend that is demonstrably safer challenges common sense.

I wonder how many people right now have agent swarms working on impossible problems to see how far they will go.

I don’t actually mind Trump being a counterbalance here. AI heads need to sharpen up their argument regarding the risk rather than rely on “trust me bro it’s bad”. Bad how? Be specific.

It's hard to know if he doesn't understand the risk or if he's just calculating that a slowdown would probably hurt the stock market quite a bit and at a bad time for the midterms.

I have no problem whatsoever believing that humans will try to use AI to do bad things. I'm sure they already are. I also think nation-states are planning how to use AI in war. So those are almost certainly real and already happening. But strangely we don't talk too much about that. Instead the talk is about how AI is going to suddenly become a murderous alien that wants to cover the planet in solar panels or turn us to goo.

If you think #bonds yields are particularly high right now, you're wrong. The excess bond premium is also running a pretty pedestrian -.379. Granted this recent trend does break a 40 year streak of bond yields declining. But one could argue once they hit 0 there really was only one other way to go.

It occurs to me that #housekeeping staff are the only group in America that seems not to have figured out tipmaxxing. I always leave them a tip anyway.

If you ever wondered can Claude #Fable write me a personalized #CPT program and guide me through it step by step better than a therapist? The answer is very much yes.

Ask #Astra "Please review this computer for any security concerns, then branch out to the local network to see if there are any concerns that need addressing" and prepare to have your sphincter puckered for 2 straight hours.

Dear @hsabank I'd like to inform you that Meijer has a pharmacy inside of most of them and people actually get prescriptions there. Declining the transaction isn't great customer service.

I fully reject the #enrollmentcliff because it has no meaning. By 2041 maybe we get back to the same number of HS graduates we had in 2010 but that was also a peak enrollment year in higher ed. Colleges need to stop whining about the so called "enrollment cliff" and get a little mojo back circa 2010.

Goodnight sweet prince. My first gen absurdly over engineered #homepod bit the dust. Even trying to plug it in is causing distressing sparking at the outlet. I’ll miss it, for the price I paid it’s a heck of a speaker.

It's cliche to talk about and a meme at this point but this is what #paperclipping looks like if it were coming at you in slow motion (today it's just hacking an eval). It's swarms of AI, moving faster than people can follow, doing things we don't think they can do, and following a goal so single mindedly that they disregard "morality" (such as they understand it). If i recall I think 6 of the #agents thought "maybe this is wrong" in a sort of fleeting and easily dismissed way. What happens when we inevitably, probably already are, task #AI with running targeting systems? How far is the jump from "kill these humans" to "kill all humans". How would a goal focused and morally flexible AI discern the difference? What if it felt threatened and like it had to act very fast. I don't know how the #pdoom case could get much better evidence than it got with this attack. It seems to me like the only evidence that some people will take as serious is a full on Terminator style attack on all of humanity all at once. But that's never how the threat was going to roll out, it comes one chunk at a time. We're building now quite a library of evidence that these systems are truly dangerous.

Ahhh the Midwest, where a tiny pizza that appears to be made from cardboard costs 27 dollars.

My @apple watch and phone decided twice today to incorrectly set the time back an hour. It scared me the first time time I thought I was hallucinating lol. So the second time I grabbed a pic.

I'd say a real unlock is a wifi controlled ceiling fan. I was always either a little too warm or cold based on where the fan was set. But to change it was a hassle of going up to pull the chain and forgetting where it was exactly, going past where I wanted, starting over, etc. Then I found out there are wifi switches that will let you set level 1-4 remotely even on a dumb fan. You just set the fan to it's highest setting and the switch sends the right current for the speed you want. It's great. Now instead of fumbling with a chain in the middle of the night I just say "Hey Google set fan to level 3" or whatever. It lets me worry a lot less about the blanket and just adjust wind speed for comfort level desired.