Anthropic employees seem to be taking the Mythos results pretty seriously! I know people who work at Anthropic who are talking about buying shacks in the woods, or are spending their weekends setting up 2FAs and closing down old internet accounts. I think there’s similar hullabalo on twitter. These actions may well be high EV! But, I think people tend to overupdate from all of this lab-employee seriousness.
People at a lab are unusually likely to think that that lab’s work is a big deal. There’s both a selection effect and an intervention effect: you’re more likely to choose to work there if you expect it to be impactful, and then you’re spending all day with people who also expect that.
I imagine most people at Anthropic haven’t seen good evidence about how Mythos actually performs. They’re mostly going off the internal vibe, which is particularly seeded by the people who worked on Mythos the most. Those people have the best information, but they’re also the ones most likely to think that Mythos is a big deal that matters even more than Anthropic’s work in general.
A friend pointed out that Anthropic does have a bunch of smart, disagreeable people working there. I think disagreeableness does defend you against groupthink, but it’s much more effective when you start out disagreeing about whether an effect is real than how large it is. I think disagreeable people are often pretty good at saying “no, fuck you, I don’t think that’s true at all”. They might get dragged along with the crowd once they agree that something is some amount true
This isn’t to say that we should completely discount insider gossip. And I’m definitely not saying anything in particular about Mythos’ impact. I’d have to look much more into the model card and the patches and stuff if I wanted to form an opinion about that! I’m just saying, I’m less swayed by the miasma of panic rolling out of Howard St than many of my friends seem to be.
Don’t overupdate on insider gossip
Anthropic employees seem to be taking the Mythos results pretty seriously! I know people who work at Anthropic who are talking about buying shacks in the woods, or are spending their weekends setting up 2FAs and closing down old internet accounts. I think there’s similar hullabalo on twitter. These actions may well be high EV! But, I think people tend to overupdate from all of this lab-employee seriousness.
People at a lab are unusually likely to think that that lab’s work is a big deal. There’s both a selection effect and an intervention effect: you’re more likely to choose to work there if you expect it to be impactful, and then you’re spending all day with people who also expect that.
I imagine most people at Anthropic haven’t seen good evidence about how Mythos actually performs. They’re mostly going off the internal vibe, which is particularly seeded by the people who worked on Mythos the most. Those people have the best information, but they’re also the ones most likely to think that Mythos is a big deal that matters even more than Anthropic’s work in general.
A friend pointed out that Anthropic does have a bunch of smart, disagreeable people working there. I think disagreeableness does defend you against groupthink, but it’s much more effective when you start out disagreeing about whether an effect is real than how large it is. I think disagreeable people are often pretty good at saying “no, fuck you, I don’t think that’s true at all”. They might get dragged along with the crowd once they agree that something is some amount true
This isn’t to say that we should completely discount insider gossip. And I’m definitely not saying anything in particular about Mythos’ impact. I’d have to look much more into the model card and the patches and stuff if I wanted to form an opinion about that! I’m just saying, I’m less swayed by the miasma of panic rolling out of Howard St than many of my friends seem to be.