METR's David Rein says lab-government transparency is critical.
Transparency into every power player at the frontier of AI (labs, government, etc) is the only viable solution.
Figuring out the right transparency is hard, but it can't be he said she said between dario and the white house that determines the fate of the AI ecosystem.
Positive users support the researcher's call for institutionalized transparency from AI labs and government on security risks and evaluations, while negative users dismiss the approach as ineffective and mock the parties involved.
I think the White House should respond to national security threats quickly, including those posed by model deployments. I’m skeptical Fable is a significant threat, but I think a lot of AI safety people are too quick to condemn the admin here. If they’re just targeting Anthropic out of spite, that is bad. But if they’re genuinely worried about the national security threat posed by the model, then I think their actions are in the right direction given their understanding. I want them to be more transparent about their reasoning, but I think it’s possible they’re just freaked out. It’s also possible this is mostly politics! But I actually just don’t know. It’s also possible to be both. I don’t trust the government and I also don’t trust any AI company. There are great people in both trying to make this go well. I’ve worked with some of them and talked with many more. These people have hard jobs! It’s hard to know what to do about increasingly powerful AI agents! I hope everyone in labs and the government does their best to try to take the high road and focus on avoiding catastrophe. People will be retarded, I guarantee it. AI company CEOs will do stupid things. Admin people will do stupid things. But we got do our best anyway, cause the stakes are fucking high.
@Miles_Brundage I’m just actually not sure what the admin believes here! There are smart cyber people in OSTP and in the NSA and CAISI. I don’t know who is talking to who and how they’re assessing risk. Transparency here would be super helpful
@JeffLadish Talking about a manufactured crisis instead of otherwise preparing for inevitable crises, not using Fable for beneficial + research + security purposes, polarizing against one another for no reason
@JeffLadish I hope something like that happens but the last crisis did not seem to produce that outcome sadly
@Miles_Brundage Maybe this is naive but I’m hoping they work through this and come out with better working relationships, so next time when there’s more danger they can work together more effectively.
Like the labs and government are going to have to work together! The alternatives are bad
@Miles_Brundage Maybe this is naive but I’m hoping they work through this and come out with better working relationships, so next time when there’s more danger they can work together more effectively.
Like the labs and government are going to have to work together! The alternatives are bad
@Miles_Brundage In what way did we waste the days?
@Miles_Brundage True but I’m hoping this one had a better shot because it’s easier to evaluate claims about cyber risk than it is about contract language for military work
@JeffLadish I hope something like that happens but the last crisis did not seem to produce that outcome sadly
@natolambert “a country of geniuses playing telephone in a datacenter”
@natolambert imo the fix isnt making dario and the white house legible to each other its that the fate of the thing routing through 2 parties at all is the failure mode, legible or not so should be the one deciding, visibility doesnt fix that
a gud example 👇
@natolambert yes, and this moment will decide things for years to come. all this must be talked about.
it is time for demis' cern. i kind of don't see any alternative. when you play it out it all ends in the same nasty places, concentration, etc etc
@natolambert Would read the tea leaves... it's about trust.
"Transparency is hard" and it gets even harder when regulators also do not have a clear idea about safety and alignment, i think everyone is figuring out the right side of things, but no one actually knows cause the ai progress is fast and we are still behind in mech intrep research to know whats actually going on and then subsequently make decisions.
@natolambert transparency is easy to demand but hard to define
right now the conversation feels like it's missing a real structure beyond individual statements
@natolambert 5 “D”(s) of ai
Don’t Doubt Dario Dude Done.
As Dario was a president at openAI
@JeffLadish Hear, hear!
Admin and adjacent mouthpieces seem to be leaning heavily on random bizarre culture war-ish narratives and the cyber relevant info is being alleged to be from Amazon, which seems like a weird thing to claim when they could have asserted that CAISI/NSA found something. Based on this, I personally think it is unlikely this is being done of grounds of genuine security concerns.
@natolambert Agreed. We need institutionalized transparency, not personality-driven narratives. Clear safety evals, deployment reporting, and compute thresholds are a much better start.
