technicalities

Karma: 5,838

technicalities Jan 5, 2023, 10:13 PM
3 points
0 ∶ 0
in reply to: Noah Scales’s comment on: On being compromised
We’re not disagreeing.

technicalities Jan 5, 2023, 9:04 PM
15 points
5 ∶ 0
in reply to: Habryka [Deactivated]’s comment on: On being compromised
Good point thanks (though I am way less sure of the EU’s sign). That list of examples is serving two purposes, which were blended in my head til your comment:
1. examples of net-positive organisations with terrible mistakes (not a good list for this)
2. examples of very well-regarded things which are nonetheless extremely compromised (good list for this)
You seem to be using compromised to mean “good but flawed”, where I’m using it to mean “looks bad” without necessarily evaluating the EV.
Yet another lesson about me needing to write out my arguments explicitly.

technicalities Jan 5, 2023, 8:59 PM
10 points
2 ∶ 1
in reply to: gavento’s comment on: On being compromised
yes. The fire is in an entirely different room.

technicalities Jan 5, 2023, 3:47 PM
3 points
1 ∶ 1
on: On being compromised

technicalities Jan 3, 2023, 9:29 AM
4 points
1 ∶ 0
on: What are the most underrated posts & comments of 2022, according to you?
Title: The long reflection as the great stagnation
Author: Larks
URL: https://forum.effectivealtruism.org/posts/o5Q8dXfnHTozW9jkY/the-long-reflection-as-the-great-stagnation
Why it’s good: Powerful attack on a cherished institution. I don’t necessarily agree on the first order, but on the second order people will act up and ruin the Reflection.

technicalities Jan 3, 2023, 9:17 AM
2 points
0 ∶ 0
on: What are the most underrated posts & comments of 2022, according to you?
Title: Forecasting Newsletter: April 2222
Author: Nuno
URL: https://forum.effectivealtruism.org/posts/xnPhkLrfjSjooxnmM/forecasting-newsletter-april-2222
Why it’s good: Incredible density of gags. Some of the in-jokes are so clever that I had to think all day to get them; some are so niche that no one except Nuno and the target could possibly laugh.

technicalities Jan 3, 2023, 8:53 AM
2 points
0 ∶ 0
in reply to: Greg_Colbourn ⏸️ ’s comment on: What we tried (Covid response AMA)
Good question. Everyone feel free to have it in this thread

technicalities Dec 29, 2022, 12:17 PM
4 points
0 ∶ 0
on: What AI Safety Materials Do ML Researchers Find Compelling?
I take it the authors weren’t anonymised? Not actually that important though.

technicalities Dec 24, 2022, 12:06 AM
4 points
0 ∶ 0
in reply to: technicalities’s comment on: Understanding the diffusion of large language models: summary
https://twitter.com/sir_deenicus/status/1606360611524206592

technicalities Dec 21, 2022, 6:09 PM
2 points
0 ∶ 0
in reply to: Larks’s comment on: CEA’s work in 2022
Agree about the contest. Something was submitted but it wasn’t about blowup risk and didn’t rise to the top.

technicalities Dec 21, 2022, 5:04 PM
10 points
2 ∶ 0
on: Opinion: EA-jobs interviews are used by organizations/companies as free labor
I personally only offer paid work trials, and this is the norm in the orgs I’ve seen (e.g. OpenPhil). I hope the answer is that the ones you experienced actually can’t afford to do this (but I’m sure some could).

technicalities Dec 21, 2022, 3:56 PM
24 points
9 ∶ 0
on: Understanding the diffusion of large language models: summary
Your reasoning in footnote 4 is sound, but note that practitioners often complain that OPT is much worse than GPT-3 (or even GPT-NeoX) in qualitative / practical terms. Benchmark goodharting is real.
(Even so, this might be goalpost shifting, since GPT3!2022 is a very different thing from GPT3!2020.)
What links here?
- Understanding the diffusion of large language models: summary by Ben Cottier (Dec 21, 2022, 1:49 PM; 127 points)
- Understanding the diffusion of large language models: summary by Ben Cottier (LessWrong; Jan 16, 2023, 1:37 AM; 26 points)

technicalities Dec 14, 2022, 4:24 PM
18 points
0 ∶ 0
on: technicalities’s Shortform
Looks like we have a cost-saving way to prevent 7 billion male chick cullings a year.
I snipe at accelerationist anti-welfarists in the thread, but it’s an empirical question whether removing horrifying parts of the horrifying system ends up delaying abolition and being net-harmful. It seems extremely unlikely (and assumes that one-shot abolition is possible) but I haven’t modelled it.

technicalities Dec 13, 2022, 5:57 PM
7 points
0 ∶ 0
on: CEEALAR: 2022 Update
Greg and crew: Thanks for all your work, happy to see you made it through the pandemic.
Others: If you want a third-party opinion on what it’s like, DM me.

technicalities Dec 8, 2022, 4:43 PM
25 points
6 ∶ 0
on: Learning from non-EAs who seek to do good
I like all of your suggested actions. Two thoughts:

1) EA is a both a set of strong claims about causes + an intellectual framework which can be applied to any cause. One explanation for what’s happening is that we grew a lot recently, and new people find the precooked causes easier to engage with (and the all-important status gradient of the community points firmly towards them). It takes a lot of experience and boldness to investigate and intervene on a new cause.
I suspect you won’t agree with this framing but: one way of viewing the play between these two things is a classic explore/exploit tradeoff.^[1] On this view, exploration (new causes, new different people) is for discovering new causes.^[2] Once you find something huge, you stop searching until it is fixed.
IMO our search actually did find something so important, neglected, and maybe tractable (AI) that it’s right to somewhat de-emphasise cause exploration until that situation begins to look better. We found a combination gold mine / natural fission reactor. This cause is even pluralistic, since you can’t e.g. admire art if there’s no world.
2) But anyway I agree that we have narrowed too much. See this post which explains the significance of cause diversity on a maximising view, or my series of obituaries about people who did great things outside the community.
1. ^
  I suspect this because you say that we shouldn’t have a “singular best way to do good”, and the bandit framing usually assumes one objective.
2. ^
  Or new perspectives on causes / new ideas for causes / hidden costs of interventions / etc

technicalities Dec 1, 2022, 6:29 PM
4 points
0 ∶ 0
on: technicalities’s Shortform
On AI quietism. Distinguish four things:
1. Not believing in AGI takeover.
2. Not believing that AGI takeover is near. (Ng)
3. Believing in AGI takeover, but thinking it’ll be fine for humans. (Schmidhuber)
4. Believing that AGI will extinguish humanity, but this is fine.
  1. because the new thing is superior (maybe by definition, if it outcompetes us).
  2. because scientific discovery is the main thing
(4) is not a rational lack of concern about an uncertain or far-off risk: it’s lack of caring, conditional on the risk being real.
Can there really be anyone in category (4) ?
- Sutton: we could choose option (b) [acquiescence] and not have to worry about all that. What might happen then? We may still be of some value and live on. Or we may be useless and in the way, and go extinct. One big fear is that strong AIs will escape our control; this is likely, but not to be feared… ordinary humans will eventually be of little importance, perhaps extinct, if that is as it should be.
- Hinton: “the truth is that the prospect of discovery is too sweet.” He smiled awkwardly, the word hanging in the air—an echo of Oppenheimer, who famously said of the bomb, “When you see something that is technically sweet, you go ahead and do it, and you argue about what to do about it only after you have had your technical success.” As the scientists retreated to tables set up for refreshments, I asked Hinton if he believed an A.I. could be controlled. “That is like asking if a child can control his parents,” he said. “It can happen with a baby and a mother—there is biological hardwiring—but there is not a good track record of less intelligent things controlling things of greater intelligence.”
I expect this cope to become more common over the next few years.

technicalities Nov 30, 2022, 7:38 PM
5 points
0 ∶ 0
on: Estimating the marginal impact of outreach
Great work, and I was just about to ask for the code.
I think including personal fit (with say a 5 or 6 OOM range) will flip the sign on this though. Would also be good to show the intervals.

technicalities Nov 30, 2022, 2:54 PM
3 points
0 ∶ 0
in reply to: Denise_Melchin’s comment on: James Lovelock (1919 – 2022)
Seems like it would contribute to the profitability and feasibility of factory farming.

technicalities Oct 31, 2022, 10:48 AM
2 points
0 ∶ 0
in reply to: jh’s comment on: We can do better than argmax
Yep ta, even says so on page 1.

technicalities Oct 29, 2022, 10:56 AM
2 points
0 ∶ 0
on: We can do better than argmax
Ord’s undergrad thesis is a tight argument in favour of enlightened argmax: search over decision procedures and motivations and pick the best of those instead of acts or rules.