Nicolas Mace

Karma: 25

Nicolas Mace 23 Jun 2026 9:45 UTC
4 points
0 ∶ 0
on: Advocates Can Influence LLM Values By Editing Wikipedia
Hi Jonah, Jasmine and Miles,
Cool idea! Thanks for writing it up.
Why These Findings Already Apply to Frontier Model
I wonder if the causal effect of adding true stuff to wikipedia could be tested directly? For example:
- Treatment: claims that are on Wikipedia
- Control: claims that aren’t on Wikipedia
Design: Match by how likely the claim is to be true, and by how frequently the claim appears outside of Wikipedia.^[1] Compute contrasts.
1. ^
  A Wikipedia claim might be mirrored on various other parts of the training corpus. We might want to control for that.

Nicolas Mace 4 Jun 2026 9:30 UTC
3 points
0 ∶ 0
in reply to: Vasco Grilo🔸’s comment on: Saving human lives cheaply is the most cost-effective way of increasing animal welfare?
Which is the relevant table? I did not find “provider”
Ah, right, good question. My understanding is that provider variance is modeled as a random effect.^[1] I was looking at Supplementary Table 5: iiuc, provider effects explain R2_conditional—R2_marginal ~ 50% of the variance, whereas fixed effects explain 17% of the variance.
1. ^
  Paragraph 2.2.1: “Data provider was treated as a random effect to account for potential differences in sampling and analysis methods, and the selection of sampling sites”

Nicolas Mace 3 Jun 2026 16:30 UTC
3 points
0 ∶ 0
in reply to: Vasco Grilo🔸’s comment on: Saving human lives cheaply is the most cost-effective way of increasing animal welfare?
Super interesting and helpful! Decreasing uncertainty seems highest value, when it comes to assessing effects of interventions on nematode populations.
One thing that I find striking and that I think illustrates this point well: most of the explained variance in Li et al’s models comes from data provider^[1] systematic effects, rather than from environmental or land-use variables. (See their supplementary info).
1. ^
  “provider” being the person/group who supplied nematode data; ~50 providers contributed the dataset

Nicolas Mace 17 Feb 2026 17:42 UTC
1 point
0 ∶ 0
on: Quantifying the Global Burden of Extreme Pain from Cluster Headaches
that’s nearly 5 million person-days of extreme suffering (≥9/10 pain) annually
I notice you recently reported a slightly lower figure of 3 million DLES here, and in the accompanying research paper:
The Days Lived with Extreme Suffering (DLES) would then be 8,569 years x 365 days/year = 3,127,855
The difference doesn’t matter much in itself, I guess: It’s the same order of magnitude, and error bars are probably larger than 2 million DLES anyway. But I’m curious what caused this change in your estimate.

Nicolas Mace 26 Nov 2025 0:54 UTC
1 point
0 ∶ 0
in reply to: Jim Buhler’s comment on: 4. Why existing approaches to cause prioritization are not robust to unawareness
we could also assume that Emily only might have shoulder pain if she takes the shot
Yeah, if we’re clueless whether Emily will feel pain or not then the difference disappears. In this case I don’t have the pro-not-shooting bracketing intuition.
On B and C, we’re actually clueful on the out-bracket (the terrorist dwarfs Emily, so it’s better to shoot in expectation)
I was thinking on C we’re clueless on the out-bracket, because, conditional on shooting, we might (a) hit the child (bad for everyone except Emily), (b) nothing (neutral for everyone except Emily) or (c) the terrorist (good for everyone except Emily), and we’re clueless whether (a), (b) or (c) is the case. I might misunderstand something, tho.

Nicolas Mace 25 Nov 2025 0:31 UTC
3 points
0 ∶ 0
in reply to: Jim Buhler’s comment on: 4. Why existing approaches to cause prioritization are not robust to unawareness
1. I think there’s a difference between (A,B) and C: On the A’s or B’s in-bracket, we can’t say that one option is strictly better than the other.
  1. (Conditional on missing the terrorist / the child, A / B is indifferent between shooting and not shooting).
2. There’s also a difference between (B,C) and A: On B and C, we’re clueless on the out-bracket (conditional on hitting the terrorist, shooting is strictly better, and conditional on not hitting the terrorist it’s strictly worse). On A on the other hand we’re clueful on the out-bracket (it’s never strictly better for child+Emily to shoot).
I’m pretty unsure what to make of this. (I might also have misinterpreted the case). I think (1) is a point against A- and B-bracketings being action-guiding. (2) might be a reason to rule out A-bracketing. So considering A, B and C as candidate bracketings, I might go with C’s verdict.

Nicolas Mace 17 Nov 2025 19:18 UTC
1 point
0 ∶ 0
on: You should probably track your time (and it just got easier)
Hey Christoph, thanks for your work on this amazing product!
What’s the best way to reach out to you with questions regarding the app? It’d be amazing if there was e.g. a Slack or Discord channel where people could post questions, and you or other members of the community could answer them.
(For example, here’s one question I have: The app used to let you use your own API key. I don’t seem to be able to find the option anymore (on macOS). Is it just me, or was the functionality removed? Edit: I found the option again under “Permission” in the desktop app)

Nicolas Mace 19 Jun 2025 9:20 UTC
1 point
0 ∶ 0
in reply to: Anthony DiGiovanni 🔸’s comment on: antimonyanthony’s Shortform
vegan advocacy [...] has a lot of ambiguous effects on wild animal suffering (which don’t seem more “speculative”, given the fairly robust-seeming arguments Tomasik has written about)
I’d be curious to hear more about this. What are the robust-seeming arguments?

Nicolas Mace 19 Jun 2025 9:19 UTC
2 points
0 ∶ 0
in reply to: Anthony DiGiovanni 🔸’s comment on: antimonyanthony’s Shortform
Thanks for writing this, Anthony!
I find myself wondering what counts as “speculative” vs not. Here are some guesses at sufficient conditions for speculativeness:
An effect is speculative if it is highly sensitive to:
- Physical or metaphysical assumptions that are themselves speculative
  - E.g., theories of cosmology, solutions to the measurement problem in quantum mechanics, theories of consciousness
- Facts about the world / mechanisms we’re presently unaware of
  - E.g., improvements to SOTA AI architectures and training methods, entirely new mechanisms for extinction, facts challenging our current understanding of macroeconomics, biology, etc

Nicolas Mace 4 Jun 2025 11:50 UTC
16 points
2 ∶ 0
on: Hedging against deep and moral uncertainty
Thanks for writing this, that was an interesting read!
I will continue to illustrate with separate components, since that’s more general and can capture deeper uncertainty and worse moral uncertainty
Whether or not you think you can add separate components seems pretty important for the hedging approach.
Indeed, if a portfolio dominates the default on each individual component, then some interventions in the portfolio must dominate the default overall.^[1] So if you can compare interventions based on their total effects, the existence of such portfolios imply that some interventions dominate the default. Intuitively then, you would prefer investing in one of those interventions over hedging? (Although a complication I haven’t thought about is that you should compare interventions with one another too, unless you think the default has a privileged status.)
Given the above, a worry I have is that the hedging approach doesn’t save us from cluelessness, because we don’t have access to an overall-better-than-the-default intervention to begin with.
To put my two questions in more concrete terms:
1. In your example, how do we become confident that the wild animal intervention is worst-case net-positive compared to the default?
2. Given that we’re confident about (1), why prefer a portfolio to investing solely in the wild animal intervention?
1. ^
  Sketch of proof: Let $w_{i}$ be ressources allocated to intervention $A_{i}$ in your portfolio, and let $a_{i}^{j}$ be the worst-case effect of intervention $A_{i}$ on component $j$ . Then
  $\sum i, j w_{i} a_{i}^{j} = \sum i (\sum j a_{i}^{j}) w_{i} \geq 0$
  and there is an intervention for which worst-case effects are in aggregate non-negative.

Nicolas Mace

Why These Findings Already Apply to Frontier Model