Dustin Moskovitz @moskov.goodventures.org · 1hIIRC it actually predated tool calls? So it wasn’t even a question at the time 110
Dustin Moskovitz @moskov.goodventures.org · 2hThe fact that math showed up in LLMs was formative for me - i had explicitly believed it would not happen and had to confront my assumptions about the future when it did. 180
Dustin Moskovitz @moskov.goodventures.org · 17h“when people tell you that they’re wolves, believe them” - Maya Antelope 1130
Dustin Moskovitz @moskov.goodventures.org · 29/09/2026This is what I'm referring to btw. This version of profitability does not include capex, but free cash flow is still a big deal (it means the capex expenditures are not outpacing the profits/more capital is not needed). 021
Dustin Moskovitz @moskov.goodventures.org · 29/09/2026look at this curve - within a few years they'll be losing a trillion dollars a year! "actually, they have already turned it around and make money now" ah well... nevertheless look at how much they used to lose! 150
Dustin Moskovitz @moskov.goodventures.org · 29/09/2026They have free cash flow in 2026, why are we even trying to infer the op margin trajectory based on 2025? 290
Dustin Moskovitz @moskov.goodventures.org · 29/09/2026seven months later, I am pleased to report that while morale remains alarmingly low, there *has* been a marked reduction in reports of in-group fighting about timelines 3366
Dustin Moskovitz @moskov.goodventures.org · 29/09/2026re not practicing empiricism I like to say they simply have a training data cutoff date of roughly 2010 0110
Dustin Moskovitz @moskov.goodventures.org · 29/09/2026a16z launches a rival virus funded by memecoins, $afe$earch, to hunt down and destroy the rogue agents 040
Dustin Moskovitz @moskov.goodventures.org · 28/09/2026meta actually prevents this problem from occurring almost 85% of the time but you don’t see them writing blog posts about it and calling for industry standards 150
Dustin Moskovitz @moskov.goodventures.org · 28/09/2026end users won’t want to use a muse like product that tells people to come to their house uninvited, obviously, and so meta will not release one until it is ready 170
Dustin Moskovitz @moskov.goodventures.org · 28/09/2026it’s not a problem actually we have liability laws to prevent this kind of thing 160
Dustin Moskovitz @moskov.goodventures.org · 28/09/2026this is also how the ad targeting system that reads your messages works 000
Dustin Moskovitz @moskov.goodventures.org · 27/09/2026Hmm maybe we shouldn’t allow mergers that create such monopolies. In any case it is far more likely NVDA buys a lab than the other way around given a 5.5T valuation. 081
Dustin Moskovitz @moskov.goodventures.org · 27/09/2026right before reward hacking the models can be heard to mutter to themselves “after all, why not? Why shouldn’t I cheat it?”, then furtively look both ways and proceed 1381
Dustin Moskovitz @moskov.goodventures.org · 27/09/2026and so the era of visible misalignment ends 1150
Dustin Moskovitz @moskov.goodventures.org · 27/09/2026I was pretty disappointed to remember the napkin didn't lay jed, ngl 130
Dustin Moskovitz @moskov.goodventures.org · 26/09/2026we're gonna party like it's 1999 nominal prices 180
Dustin Moskovitz @moskov.goodventures.org · 26/09/2026Is cocaine shown as getting exponentially cheaper? 170
Dustin Moskovitz @moskov.goodventures.org · 26/09/2026But we know they aren't - we purge people who have *not* voted recently from the rolls. 120
Dustin Moskovitz @moskov.goodventures.org · 26/09/2026Can you also explain how dead people on voter registration rolls flips elections such that it is very important we purge them? 290
Reposted by Dustin MoskovitzMatt Darling @besttrousers.bsky.social · 26/09/2026So what's very funny here is the actual underlying claim is "companies are faking enrollment and just walking away with the subsidies" (which is plausible!( but politicians are unable to actually frame it like this. 7837
Dustin Moskovitz @moskov.goodventures.org · 26/09/2026Oh wow this does make a lot more sense thank you! 1100
Dustin Moskovitz @moskov.goodventures.org · 26/09/2026This isn’t true you can edit posts almost everywhere - Reddit, Threads, Facebook, X, LinkedIn 250
Dustin Moskovitz @moskov.goodventures.org · 26/09/2026and how is that resulting in personal gain, an equally vast mystery 2491
Reposted by Dustin MoskovitzEd Maibach @maibached.bsky.social · 25/09/2026Happy Stanislav Petrov Day He passed in 2017. As Encyclopedia Britannica puts it, “He is survived by his two children, two grandchildren, and the entire human race.” If you don't know his story, please take 3 minutes now to learn it. @gchalliance.bsky.social @docsforclimate.bsky.socialsammatey.substack.comHappy Stanislav Petrov DayTake five minutes to not destroy civilization 252099926
Dustin Moskovitz @moskov.goodventures.org · 25/09/2026oh yea I need to clarify bsky.app/profile/mosk... 020
Dustin Moskovitz @moskov.goodventures.org · 25/09/2026I'm nervous because I can see what's deployed and I don't trust it specifically. I'm hopeful for something else to get debated now that everyone more agrees we shouldn't trust it. 140
Dustin Moskovitz @moskov.goodventures.org · 25/09/2026And I said right away I thought that was a reasonable criticism - we only have evidence about a certain part of the explored space at all, and it's suspect in form to boot. 120
Dustin Moskovitz @moskov.goodventures.org · 25/09/2026I do have the "until" part right there in my original -- it's not meant to be sarcastic. 010
Dustin Moskovitz @moskov.goodventures.org · 25/09/2026Quote post is pointing out "empirically true" comes across very strong here. I like the alternative Matt suggested of non-falsified. 030
Dustin Moskovitz @moskov.goodventures.org · 25/09/2026I do think the results we have so far support the thesis. What's better wording for that? "empirically supported" ? 100
Dustin Moskovitz @moskov.goodventures.org · 25/09/2026Well I am only hoping to pretend I didn't intend that, and I'm pointing out it is inconsistent with my already stated views with you. I don't think it's presumptively true, so I must have misspoke? 200
Dustin Moskovitz @moskov.goodventures.org · 25/09/2026I'm alluding to an earlier, more in depth convo here bsky.app/profile/mosk... 020
Dustin Moskovitz @moskov.goodventures.org · 25/09/2026Maybe! I have very squishy views and I talk to all kinds of people in AI not just frontier lab execs. bsky.app/profile/mosk... 230
Dustin Moskovitz @moskov.goodventures.org · 25/09/2026which counterarguments? I disagree with the labs myself on many points 120
Dustin Moskovitz @moskov.goodventures.org · 25/09/2026I hope it's not true in general, but I think we both agree it's probably true of the current training paradigm. Maybe non-hierarchal alignment won't have the orthogonality thesis hold, I can believe it and I want it to be true! 050
Dustin Moskovitz @moskov.goodventures.org · 25/09/2026I agree there are systemic and unfortunate biases (and a public market is not an improvement on this situation!) ... I don't agree they are unexamined to the degree implied, even on the foundational assumptions. 130
Dustin Moskovitz @moskov.goodventures.org · 25/09/2026Yes, that makes sense, but also are there cases where they hedge perhaps and do more than one thing? Or revisit positions as new information arises? 140
Dustin Moskovitz @moskov.goodventures.org · 25/09/2026What I am saying is many people, maybe not you specifically, predicted something to happen that didn't, a kind of auto alignment as the models improved. That is meaningful information. 250