Comments

It was either Hydrazine or YC. In either case, my point remains true: he's chosen to not dispose of his OA stake, whatever vehicle it is held in, even though it would be easy for someone of his financial acumen to do so by a sale or equivalent arrangement, forcing an embarrassing asterisk to his claims to have no direct financial conflict of interest in OA LLC - and one which comes up regularly in bad OA PR (particularly by people who believe it is less than candid to say you have no financial interest in OA when you totally do), and a stake which might be quite large at this point*, and so is particularly striking given his attitude towards much smaller conflicts supposedly risking bad OA PR. (This is in addition to the earlier conflicts of interest in Hydrazine while running YC or the interest of outsiders in investing in Hydrazine, apparently as a stepping stone towards OA.)

* if he invested a 'small' amount via some vehicle before he even went full-time at OA, when OA was valued at some very small amount like $50m or $100m, say, and OA's now valued at anywhere up to $90,000m or >900x more, and further, he strongly believes it's going to be worth far more than that in the near-future... Sure, it may be worth 'just' $500m or 'just' $1000m after dilution or whatever, but to most people that's pretty serious money!

Interesting interview. Metz seems extraordinarily incurious about anything Vassar says - like he mentions all sorts of things like Singularity University or Kurzweil or Leverage, which Metz clearly doesn't know much about and are relevant to his stated goals, but Metz is instead fixated on asking about a few things like 'how did X meet Thiel?' 'how did Y meet Thiel?' 'what did Z talk about with Thiel?' 'What did A say to Musk at Puerto Rico?' Like he's not listening to Vassar at all, just running a keyword filter over a few people's names and ignoring anything else. (Can you imagine, say, Caro doing an interview like this? Dwarkesh Patel? Or literally any Playboy interviewer? Even Lex Fridman asks better questions.)

I was wondering how, in his 2020 DL book The Genius Makers he could have so totally missed the scaling revolution when he was talking to so many of the right people, who surely would've told him how it was happening; and I guess seeing how he does interviews helps explain it: he doesn't hear even the things you tell him, just the things he expects to hear. Trying to tell him about the scaling hypothesis would be like trying to tell him about, well, things like Many Worlds... (He is also completely incurious about GPT-3 in this August 2020 interview too, which is especially striking given all the reporting he's done on people at OA since then, and the fact that he was presumably working on finishing Genius Makers for its March 2021 publication despite how obvious it should have been that GPT-3 may have rendered it obsolete almost a year before publication.)

And Metz does seem unable to explain at all what he considers 'facts' or what he does when reporting or how he picks the topics to fixate on that he does, giving bizarre responses like

Cade Metz: Well, you know, honestly, my, like, what I think of them doesn't matter what I'm trying to do is understand what's going on like, and so -

How do you 'understand' them without 'thinking of them'...? (Some advanced journalist Buddhism?) Or how about his blatant dodges and non-responses:

Michael Vassar: So you have read Scott's posts about Neo-reaction, right? They're very long.

Cade Metz: Yes.

Michael Vassar: So what did you think of those?

Cade Metz: Well, okay, maybe maybe I'll get even simpler here. So one thing I mentioned is just sort of the way all this stuff played out. So you had this relationship with Peter Thiel, Peter Thiel has, had, this relationship with, with Curtis Yarvin. Do you know much about that? Like, what's the overlap between sort of Yarvin's world and Silicon Valley?

We apparently have discovered the only human being to ever read all of Scott's criticisms of NRx and have no opinion or thought about them whatsoever. Somehow, it is 'simpler' to instead pivot to... 'how did X have a relationship with Thiel' etc. (Simpler in what respect, exactly?)

I was also struck by this passage at the end on the doxing:

Michael Vassar: ...So there are some important facts that need to be explained. There's there's this fact about why it would seem threatening to a highly influential psychologist and psychiatrist and author to have a New York Times article written about his blog with his real name, that seems like a very central piece of information that would need to be gathered, and which I imagine you've gathered to some degree, so I'd love to hear your take on that.

Cade Metz: Well, I mean... sigh Well, rest assured, you know, we we will think long and hard about that. And also -

Vassar: I'm not asking you do to anything, or to not do anything. I'm asking a question about what information you've gathered about the question. It's the opposite of a call to action: it's a request for facts.

Cade Metz: Yeah, I mean, so you know, I think what I don't know for sure, but I think when it comes time, you know, depending on what the what the decision is, we might even try to explain it in like a separate piece. You know, I think there's a lot of misinformation out there about this and and not all the not all the facts are out about this and so it is it is our job as trained journalists who have a lot of experience with this stuff. To to get this right and and we will.

Michael Vassar: What would getting it right mean?

Cade Metz: Well, I will send our - send you a link whenever, whenever the time comes,

Michael Vassar: No, I don't mean, "what will you do?" I'm saying what - what, okay. That that the link, whenever the time comes, would be a link to what you did. If getting it right means "whatever you end up doing", then it's a recursive definition and therefore provides no information about what you're going to do. The fact that you're going to get it right becomes a non-fact.

Cade Metz: Right. All right. Well... pause let me put it this way. We are journalists with a lot of experience with these things. And, and that is -

Michael Vassar: Who's "we"?

Cade Metz: Okay, all right. You know, I don't think we're gonna reach common ground on this. So I might just have to, to, to beg off on this. But honestly, I really appreciate all your help on this. I do appreciate it. And I'll send you a copy of this recording. As I said, and I really appreciate you taking all the time. It's, it's been helpful.

One notes that there was no separate piece, and even in OP's interview of Metz 4 years later about a topic that he promised Vassar he was going to have "thought long and hard about" and which caused Metz a good deal of trouble, Metz appears to struggle to provide any rationale beyond the implied political activism one. Here Metz struggles to even think of what the justification could be or even who exactly is the 'we' making the decision to dox Scott. This is not some dastardly gotcha but would seem to be a quite straightforward question and easy answer: "I and my editor at the NYT on this story" would not seem to be a hard response! Who else could be involved? The Pope? Pretty sure it's not, like, NYT shareholders like Carlos Slim who are gonna make the call on it... But Metz instead speaks ex cathedra in the royal we, and signs off in an awful hurry after he says "once I gather all the information that I need, I will write a story" and Vassar starts asking pointed questions about that narrative and why it seems to presuppose doxing Scott while unable to point to some specific newsworthy point of his true name like "his dayjob turns out to be Grand Wizard of the Ku Klux Klan".

(This interview is also a good example of the value of recordings. Think how useful this transcript is and how much less compelling some Vassar paraphrases of their conversation would be.)

I hear that they use GPT-4. If you are looking at timelines, recall that Cognition apparently was founded around January 2024. (The March Bloomberg article says "it didn’t even officially exist as a corporation until two months ago".) Since it requires many very expensive GPT-4 calls and RL, I doubt they could have done all that much prototyping or development in 2023.

'Social contagion' here being a metonym for all environmental factors, I think... For example, face-swapping apps: https://www.newyorker.com/books/page-turner/how-lucy-sante-became-the-person-she-feared This is an environmental effect which basically could not have existed until about 10 years ago at best. (What were you going to do in 2014, when face-swapping ML didn't exist? Hunt down a surviving oil portrait artist and pay them hundreds or thousands of dollars to paint you as if you were a woman, for no particular reason you're willing to admit to consciously?)

As far as journalists go, I'm not 'avoidant'.

I have talked to a number of journalists over the years. I actually agreed to interview with Cade Metz himself on the subject of SSC before this all came out; Metz just ghosted me at the arranged time (and never apologized).

What I have always done, and what I always advised the people who ask me about how to handle journalists, is maintained the unconditional requirement of recording it: preferably text, but audio recording if necessary.* (I also remind people that you can only say things like "off the record" or "on background" if you explicitly say that before you say anything spicy & the journalist assents in some way.)

I've noticed that when people are unjustly burned by talking to journalists, of the "I never said that quote" sort, it always seems to be in un-recorded contexts. At least so far, it has worked out for me and as far as I know, everyone I have given this advice to.

* oddly, trying to do it via text tends to kill a lot of interview requests outright. It doesn't seem to matter to journalists if it's email or Twitter DM or Signal or IRC or Discord, they just hate text, which is odd for a profession which mostly deals in, well, text. Nor do they seem to care that I'm hearing-impaired and so just casually phoning me up may be awesome for them but it's not so awesome for me... My general view is that if a journalist cares so little about interviewing me that wanting to do it in text is a dealbreaker for them, then that interview wasn't worth my time either; and so these days when I get an interview request, I insist on doing it via text (which is, of course, inherently recorded).

I don't think phrasing it as a classification problem is necessarily the best approach. It may be that there are many very similar images and which one out of a cluster you pick is fairly arbitrary, so any attempt at a classifying 'picked/'not picked' wouldn't tell you much, but clustering/sorting still makes it much easier/pleasant to do the curation.

(Speaking of classification and images and blogs, you might find it useful to know that we just launched the InvertOrNot.com API (HN) for dark-mode images.)

If you want to make it more efficient and spend less time fast-forwarding through redundant images, you could experiment with clustering and sorting images in a NN embedding space: https://github.com/MaartenGr/Concept https://every.to/napkin-math/6-new-theories-about-ai https://gwern.net/design#sort-by-magic

That's true, though, once the payment is a sunk cost. :) You don't have to eat it, any more than you would if someone handed you the identical snack for free. (And considering obesity rates, if your reaction isn't at least somewhat positive and happy when looking at this free snack, maybe you shouldn't eat it...)

I think they may have gotten it working for MDPs, but I see little evidence so far that they've gotten it for POMDPs, due to the general lack of explicit exploration. (Devin does ask questions but rarely, consistent with the base model capabilities and imitation learning.) The former would still be a big deal, however.

Load More