I couldn’t attend the interpretability hackathon and was hoping to get acquainted with LLM interpretability research as a sofware dev with no experience in interpretability or transformers. So here’s a starting point following in the footsteps of this submission (see their writeup here):
Neel Nanda live coding/testing transformer hypotheses learning (1.5hrs) https://www.youtube.com/watch?v=yo4QvDn-vsU&t=3403s
Go through Transformer theory:
Attention is all you need paper: https://arxiv.org/abs/1706.03762
Good summary: https://nostalgebraist.tumblr.com/post/185326092369/the-transformer-explained
Positional embeddings explained (9mins): https://www.youtube.com/watch?v=1biZfFLPRSY
Basically I am thinking we can use the hackathon as a collaborative study session to become more familiar with transformers and interpretability, ultimately culminating in replicating the results in the linked submission (it took them 3 days but since we have a starting point, possibly we can replicate their project and grok what they did much quicker).
Not shoehorned to this idea though. If you think there is a better avenue to using the hackathon to upskill in LLM interpretability and transformers, do share.
Count me in as someone interested in joining this project! Something I miss from the old GWWC before the website redesign was different donations had “badges” and it felt like a collectible game to donate to each EA charity at least once. Maybe something like this could be added back.
A bit of donation gamification can go a long way.
How do you pick between FWI and SWP? I’m looking between those myself.
This post’s philosophical implications—which has remained foundational to informing my donations—is explored further in OP’s new charity research that he subsequently conducted:
So just FYI:
I’ve had members in my community point out concern that the post being taken down is evidence of censorship in EA.
The message “Sorry, you don’t have access to this page” should probably read something like “Sorry, this post has been removed by the author.” This is not even “only about optics”. This is just updating a message so that it says something true rather than false.
Just want to let you know there are Georgist EAs out there. In my head I put something like 80% odds that Georgism (or maybe just LVT) will be a cause area (i.e. similar to how at first dismissed areas like mental health, wild animal suffering or great power conflict are now considered major cause areas in EA) within the next ten years.A second reason I’m writing this comment is so I can look back on it in ten years.
I’m trying to more succinctly understand what you’re saying since your second last paragraph has confusing wording. You’re saying that Nonlinear can scale as EA scales (as opposed to scaling by their ability) and thereby attract competent clout like Emerson (since EA has become more famous as a whole it attracts big-shots), but that as an organization they don’t yet produce enough value/output for someone like Emerson to be a good fit at their organization? And that this plausibly has a causal relationship to why there has been conflict? e.g. Emerson being a bad fit leads to him more easily getting frustrated with other employees?(PS: just a note that this doesn’t excuse Emerson mistreating employees if he was indeed mistreating employees. My comment here is just trying to understand what the comment above is saying since it confused me, but I think it might be valuable to clarify)
Threw me off as well at first. I’d second that it’s probably best to reword it in the future for better clarity
Now I feel dumb, but at least I’m smarter. Thanx.
Time-poor long-time deep EA with imposter syndrome here with a forum post draft now probably over a year old.
One point you didn’t hit on that I think strongly applies to me (and probably others like me) is just when I think I’ve struck upon some sort of insight or found a topic I want to dive into to write about… I find a lengthy well-written EA Forum post has already been written on the topic and not only that, that I agree with it and that there is nothing new I would add (or if there is it is better suited as a quick comment rather than a followup post).
In other online communities I have at times found myself writing profusely because I find areas of disagreement , areas where I can add value, influence discourse etc. In EA… I’m just “another EA” that has roughly the same views and values as so many other EAs. I’m weird and interesting and insightful outside of EA. Within EA… Not so much. And what I find myself wanting to say so often has already been said.
It actually seems to be a paradox of sorts. Due to being an EA “insider” I’m less likely to generate any unique valuable insight compared to someone who is more EA-adjacent or an EA “outsider.”
The lethal autonomous weapons (LAW) case seems fundamentally different, unfortunately, for two reasons:
In the case of CBW we had the work of a dedicated group of scientists persuading Congress, the people, and the government that biological weapons would not be useful. It will be a lot harder to argue that LAWs cant be useful. As devices that are wholly man-made I’d think there will always be a significant group of scientists and engineers arguing they just need to be better built to provide huge value. Unlike CBWs you don’t have complicated issues of containment to worry about. Unlike CBWs, with LAWs the potential for creative fine-grained designs seems boundless since you’re not restricted by what a biological agent is capable of. LAWs are only restricted by physics and how well your AI is programmed.
Secondly, the public and Congressional pressure on Nixon I see as largely stemming from the “creep” factor of CBWs. These are weapons that bring to the public’s mind horrendous diseases gone rampant—something we as a species have deep evolutionary and culturally ingrained reasons to be creeped out about, both aesthetically and morally.
LAWs don’t give that same creep factor. If anything, they give a “cool” factor, probably as a result of being wholly man-made. So I have a hard time seeing concerned scientists having anything but a very long upward battle trying to convince the public and Congress that LAWs should be banned analogously to the biological and chemical weapons conventions. This is just making me wonder if it might be intrumentally better to argue for LAWs being banned without arguing they are analogous to CBWs. I don’t have expertise in this area, but this is what is coming to my mind.
sidenote: I don’t believe you explicitly defined “CBW.” I am inferring BW = Biological Weapons. What does the C stand for? “Conventional”?
I think his accent might actually have the effect of being a draw due to the novelty for Americans. Plus it then additionally forces a listener to listen more closely if they cant follow as easily—and it is hardly a bad thing for your audience to be paying more close attention to you if you’re making nuanced arguments.
But I’m just projecting my own experience here, probably.
This is a good comment and is a perfect example of why the new system that seperates “upvote/downvote” and “agreement/disagreement” is a good idea.
I notice you haven’t been active on the EA Forum since posting this comment. I hope you’ll return one day! The downvotes here are likely more because people disagreed with you, not because they didn’t appreciate your input. On the comments in newer posts you’ll see that it is possible to both appreciate a comment but disagree with its contents (or vice versa think a comment is bad, but agree with its contents).
Hi Michael, any chance you could fix the post so the table displays properly again? I continue to be surprised how I keep coming back to your table. Feel free to point me towards some other posts or sources you think is better than this post if your views have changed since I realize this post is relatively old.
Another thing is just I wonder: if a strong enough pronatalist argument was presented to me maybe that in itself would make me excited enough to have kids. I do adapt enthusiastically to EA arguments telling me to donate here vs there, to change my career etc. Though naturally, sometimes I adapt with resistance and begrudgingly. I wonder if there is some pronatalist argument I haven’t heard that will firmly slot me into the former group where I adapt enthusiastically.
But as you point out, maybe I don’t even need to stress too much about merely being ‘not excited’ if a good enough pronatalist argument convinces me I should have kids. This is something I’d love to get feedback on from EAs who have kids (and I can think of zero EAs in my social circle that have kids). Jeff Kaufman, do you have kids of your own that makes you more confident in your statement?(meta-note: I don’t know if it is possible to tag someone in a comment to notify them they have been mentioned)
Hi Prof. Miller,
Do you think I should spend effort and research and care about the particulars of having kids ahead of time even though I largely perceive myself as possibly/probably never wanting to have kids? It certainly seems reasonable to strongly expect I might eventually fully change my mind about having kids just because having kids is—well—genetically mandated and such. Preparing and planning now seems like a wise hedge against my lack of desire to have kids right now?
I am a 30 year old male now and am still sorta mostly leaning towards not having kids. Is that normally a sign that this is how I should expect my psychology to be throughout my thirties and forties? And I should just not sweat it and continue to focus on studying AI Alignment?
Secondly, do you think there is a good pronatalist argument to be made that an EA that doesn’t feel like they want kids should still regardless have kids? I’ve heard the “in expectation if EAs all have kids we’ll get more geniuses—low prob. in each case, but such high potential impact that we ought to have kids” argument. I take this line of thinking seriously, but I still find it hard to think that this argument could ever be strong enough to—on consequentialist grounds—mean someone who doesn’t want kids should have kids regardless.
Thirdly, I find the idea of adopting pulls at my heartstrings, but have lost track of how many times I’ve heard the argument that I should have my own kids to pass on my genes. Naturally sperm bank donations are always an option if I am rationally committed to the latter. But the question still remains which is, as an EA and a psychology professor, how do you think about adoption vs having your own kids?
I have the hypothesis that the two of you are just different types of people personality-wise with different communication styles that led to a far larger inferential distance between the two of you than either of you could have anticipated. Possibly also since each of you saw the other as “a fellow EA” this made it even easier for both of you to underestimate how large the inferential distance could be between the two of you.My understanding is you and Constance only talked for an hour. That doesn’t sound like that long and if anything sounds like the perfect amount of time for both parties to come away confidently and incorrectly believing that the object level information had been properly transmitted and understood.