Will Aldred

Karma: 4,111

Will Aldred 15 Sep 2021 21:36 UTC
3 points
0 ∶ 0
in reply to: MichaelStJules’s comment on: It takes 5 layers and 1000 artificial neurons to simulate a single biological neuron [Link]
If interested, here’s some further evidence that it’s just really hard to map:
An adult [Drosophila fly] whole brain dataset is already publicly available and focussed reconstruction is under way, but its 40× greater volume [than a larva brain] would require ∼500–5000 person-years of manual labour.
—Learning from connectomics on the fly

Will Aldred 30 Mar 2022 1:47 UTC
16 points
0 ∶ 0
on: Impactful Forecasting Prize for forecast writeups on curated Metaculus questions
“To give examples of our target audience: [...] 3. Aspiring generalist researchers at any stage in their career.”

I agree that writing up forecasting reasoning is one way for aspiring generalist researchers to build generalist-type research skill, but also want to highlight some other options:
- Summarize/Collect previous posts/articles/papers (I think this is the probably the best skill-building activity for an aspiring generalist researcher)
- Read, then write book reviews (see posts tagged under ‘books,’ and also suggestions from Michael Aird and from Buck Shlegeris; also related is Holden Karnofsky’s ‘Reading books vs. engaging with them’)
- Build inside views (see Holden Karnofsky’s ‘Learning by writing’ and Neel Nanda’s ‘How I formed my own views about AI safety’
- From Linch Zhang’s shortform: “Deep dive into seminal papers/blog posts and attempt to identify all the empirical and conceptual errors in past work, especially writings by either a) other respected EAs or b) other stuff that we otherwise think of as especially important.”
- Apply for jobs/internships/research training programs (and view the process of writing written responses in your applications as skill-building)
- Possibly other things suggested in Aird’s ‘Notes on EA-related research, writing, testing fit, learning, and the Forum’

Will Aldred 30 Mar 2022 2:40 UTC
3 points
0 ∶ 0
in reply to: MichaelA’s comment on: Nuclear Risk Overview: CERI Summer Research Fellowship
Many thanks for this comment, especially the part below, which I was aware of but forgot to include, and which I’ve now incorporated into the main text of my post.

Regarding “Who else is working on the problem?”, people might also find useful the “nuclear risk” “view” of my Database of orgs relevant to longtermist/x-risk work

Will Aldred 2 Apr 2022 14:21 UTC
14 points
0 ∶ 0
on: Where’s WALY?
Where are the rest of the upvotes for this post?

My guess is that the wordplay here flew over the heads of our US friends, for whom the franchise is not Where’s Wally, but Where’s Waldo.

And while we’re on this topic, here’s a (non-exhaustive) list of some other silly-sounding things Americans say =D

Will Aldred 7 Apr 2022 18:17 UTC
11 points
0 ∶ 0
on: Call For Distillers
See also: “Research Debt” (Olah & Carter, 2017).

Will Aldred 29 Jun 2022 14:32 UTC
20 points
0 ∶ 0
on: Let’s not have a separate “community building” track
I appreciate this post’s high info to length ratio, and I’d be excited about more EAs taking actions along the lines of your corollaries at the end.
On another note, I find it interesting that for the amount of time community builders should be spending skilling up for direct work / doing direct work / skilling up in general:
- You say roughly 20%
- Emma Williamson says >50% (in her post from yesterday)

Will Aldred 29 Jun 2022 21:32 UTC
13 points
0 ∶ 0
on: Limits to Legibility
I enjoyed reading this post, and I think I agree with your assessment:
It is pretty clear that some of the main cruxes of current disagreements about AI alignment are beyond the limits of legible reasoning. (The current limits, anyway.)
(In addition to the Christiano-Yudkowsky example you give, one could also point to the Hanson-Yudkowsky AI-Foom Debate of 2008.)
In addition to “Epistemic Legibility” and “A Sketch of Good Communication,” which you mention, I’d recommend “Public beliefs vs. Private beliefs” (Tyre, 2022) to others who enjoyed this post – Tyre explores a somewhat related theme.

Will Aldred 29 Jun 2022 22:30 UTC
11 points
0 ∶ 0
on: Four reasons I find AI safety emotionally compelling
I think I mainly agree with the other comments (from Devin Kalish and Miranda Zhang), but on net I’m still glad that this post exists. Many thanks for writing it :)
Specifically, I think positive/x-hope (as opposed to negative/x-risk) styled framings can be valuable. Because:
1. There’s some literature in behavioral psychology and in risk studies which says, roughly, that a significant fraction of people will kind of shut down and ignore the message when told about a threat. See, e.g., normalcy bias and ostrich effect.
2. Mental contrasting has been shown to work (and the study replicates, as far as I can tell). Essentially, the research behind mental contrasting tells us that people tend to be strongly motivated by working toward a desired future.^[1]
3. Anecdotally, I’ve found that some people do just resonate strongly with “inspirational visions” and the idea of helping build something really cool.
1. ^
  Technically, mental constrasting only works if the perceived chance of success is high. I’m not sure what percentage “high” in this context maps to, but there is an issue here for folks who have a non-low P(doom) via misaligned AI… To counter this, there is the option of resorting to the Dark Arts, self-deception in particular, but that discussion is beyond the scope of this comment. See the Dark Arts tag on LessWrong if interested.

Will Aldred 30 Jun 2022 15:14 UTC
10 points
1 ∶ 0
on: Fanatical EAs should support very weird projects
There’s a bit of a history of estimating how low the probabilities are that we can ignore.
See also “Exceeding Expectations: Stochastic Dominance as a General Decision Theory” (Tarsney, 2020); West (2021) summarises this Tarsney paper, which is pretty technical. A key sentence from West:
Tarsney argues that we should use an alternative decision criterion called stochastic dominance which agrees with EV in non-Pascallian situations, but, when combined with the above argument about uncertainty, disagrees with EV in Pascallian ones.

Will Aldred 11 Jul 2022 17:39 UTC
7 points
0 ∶ 0
on: Three Ways to Advance Science (Bostrom, 2008)
Notable exception(s) to the premise, “accelerated scientific and technological progress would be great”: AI capabilities research. Also, synthetic biology and nanotechnology.
Nevertheless, though these are important exceptions, I feel that if one refutes cognitive enhancement based on these exceptions, then one has missed the point of Bostrom’s core argument.

Will Aldred 14 Jul 2022 12:24 UTC
9 points
0 ∶ 0
on: [2] Hail Mary, Value Porosity, and Utility Diversification (Bostrom, 2014)
P.S. I’m personally confused by the “Hail Mary” terminology...
… Wikipedia: “A Hail Mary pass is a very long forward pass in American football, typically made in desperation, with an exceptionally small chance of achieving a completion. Due to the difficulty of a completion with this pass, it makes reference to the Catholic ‘Hail Mary’ prayer for supernatural help.”
However, Bostrom’s Hail Mary approach was intended to have a high probability of success for achieving a small win (where small win = short pass, in football speak), i.e., the opposite of what Wikipedia describes.

Will Aldred 21 Jul 2022 8:29 UTC
20 points
0 ∶ 0
on: Will Aldred’s Shortform
TAI makes everything else less important.
One of my CERI fellows asked me to elaborate on a claim I made that was along the lines of,* “If AI timelines are shorter, then this makes (direct) nuclear risk work less important because the time during which nuclear weapons can wipe us out is shorter.”
There’s a general point here, I think, which isn’t limited to nuclear risk. Namely, AI timelines being shorter not only makes AI risk more important, but makes everything else less important. Because the time during which the other thing (whether that be an asteroid, engineered pandemic, nuclear war, nanotech-caused grey goo scenario, etc.) matters as a human-triggered x-risk is shortened.
To give the nuclear risk example:^
If TAI is 50 years away, and per-year risk of nuclear conflict is 0.5%, then risk of nuclear conflict before TAI is 1-(0.995^50) = 22%
If TAI is 15 years away, and per-year risk of nuclear conflict is 0.5%, then risk of nuclear conflict before TAI is 1-(0.995^15) = 7%
This does rely on the assumption that we’ll be playing a different ball game after TAI/AGI/HLMI arrives (if not, then there’s no particular reason to view TAI or similar as a cut-off point), but to me this different ball game assumption seems fair (see, e.g., Muehlhauser, 2019).
*My background thinking behind my claim here has been inspired by conversations with Michael Aird, though I’m not certain he’d agree with everything I’ve written in this shortform.
^A couple of not-that-important caveats:
- “Before TAI” refers to the default arrival time of TAI if nuclear conflict does not happen.
- The simple calculations I’ve performed assume mutual independence between nuclear-risk-in-given-year-x and nuclear-risk-in-given-year-y.
What links here?
- Will Aldred's comment on Will Aldred’s Quick takes by Will Aldred (22 Jul 2022 15:28 UTC; 6 points)

Will Aldred 21 Jul 2022 15:37 UTC
2 points
0 ∶ 0
on: Paper summary: Staking our future: deontic long-termism and the non-identity problem (Andreas Mogensen)
Hi GPI, love these paper summaries. Just an FYI that “Mogensen” appears to be spelled incorrectly (“Morgensen”) on your website alongside this “Staking our future” summary.
(Edited to add: I was hoping I’d be able to delete this comment once GPI acted on it, but it seems like the Forum settings do not allow me to do that.)

Will Aldred 22 Jul 2022 15:28 UTC
6 points
0 ∶ 0
on: Will Aldred’s Shortform
We could spend all longtermist EA money, now.
(This is a sort-of sequel to my previous shortform.)
The EA funding pool is large, but not infinite. This statement is nothing to write home about, but I’ve noticed quite a few EAs I talk to view longtermist/x-risk EA funding as effectively infinite, the notion being that we’re severely bottlenecked by good funding opportunities.
I think this might be erroneous.
Here are some areas that could plausibly absorb all EA funding, right now:
- Biorisk
  - Better sequencing
  - Better surveillance
  - Developing and deploying PPE
  - Large-scale philanthropic response to a pandemic
- AI risk
  - Policy spending (especially in the US)
  - AI chips
    either scaling up chip production, or buying up top-of-the-range chips
  - Backing the lab(s) that we might want to get to TAI/AGI/HLMI/PASTA first
(Note: I’m definitely not saying we should fund these things, but I am pointing out that there are large funding opportunities out there which potentially meet the funding bar. For what it’s worth, my true thinking is something closer to: “We should reserve most of our funding for shaping TAI come crunch time, and/or once we have better strategic clarity.”
Note also: Perhaps some, or all, of these don’t actually work, and perhaps there are many more examples I’m missing—I only spent ~3 mins brainstorming the above. I’m also pretty sure this wasn’t a totally original brainstorm, and that I was remembering these examples having read something on a similar topic somewhere, probably here on the Forum, though I can’t recall which post it was.)

Will Aldred 24 Jul 2022 18:21 UTC
4 points
0 ∶ 0
on: Odds of recovering values after collapse?
On Loki’s Wagers: for an amusing example, see Yann LeCun’s objection to AGI.

Will Aldred 7 Aug 2022 20:34 UTC
12 points
0 ∶ 0
on: What’s the likelihood of irrecoverable civilizational collapse if 90% of the population dies?
This question feels similar in spirit to the one I asked a couple weeks ago, “Odds of recovering values after collapse?”, so OP you might be interested in checking out that question and the responses to it.

Will Aldred 9 Aug 2022 14:14 UTC
9 points
0 ∶ 0
on: Risks from atomically precise manufacturing—Problem profile
Widespread access to atomically precise manufacturing could lead to widespread ability to unilaterally produce things as destructive as nuclear weapons.
See below an excerpt from the atomically precise manufacturing (APM) section of Michael Aird’s shallow review of tech developments that could increase risks from nuclear weapons (link):
The concern related to nuclear risk is that APM could potentially make proliferation and/or huge nuclear arsenal sizes much more likely. Specifically:
- APM could make it much harder to monitor/control who has nuclear weapons, including even non-state groups or perhaps individuals.
- APM could make it much more likely that non-state groups or individuals would not only attain one or a few but rather many nuclear weapons.
  This could land us in the very risky-seeming “easy nukes” scenario described in Bostrom’s (2019) The Vulnerable World Hypothesis paper.
- APM could make it so that nuclear-armed states can more cheaply, easily, and quickly build hundreds, thousands, tens of thousands, or even millions of nuclear weapons.
  The enhanced ease and speed of doing this could undermine arms race stability.
  This might also mean that even just the perception of APM being on the horizon could undermine strategic stability, even before APM has arrived.
  And nuclear conflicts involving huge numbers of warheads are much more likely to cause an existential catastrophe, or otherwise increase existential risk, than nuclear conflicts involving the tens, hundreds, or thousands of nuclear weapons each nuclear-armed state currently has.
  A counterpoint is that at least some nuclear-armed states already have the physical capacity to make many more nuclear warheads than they do (given enough time to build up nuclear manufacturing infrastructure), but having many more warheads doesn’t appear to be what these states are aiming for. This is some evidence that nuclear weapon production being easier to make might not result in huge arsenal sizes. But:
  This is only some evidence.
  And some nuclear-armed states (e.g., North Korea, Pakistan, and maybe India) may be developing nuclear weapons about as fast as they reasonably can, given their economies. As such, perhaps having access to APM would be likely to substantially affect at least their arsenal sizes.
(Note that we haven’t spent much time thinking about these points, nor seen much prior discussion of them, so this should all be taken as tentative and speculative.)

Will Aldred 9 Aug 2022 14:51 UTC
3 points
0 ∶ 0
in reply to: simeon_c’s comment on: What’s the likelihood of irrecoverable civilizational collapse if 90% of the population dies?
(DM’ed you)

Will Aldred 10 Aug 2022 13:46 UTC
7 points
0 ∶ 0
on: Radio Bostrom: Audio narrations of papers by Nick Bostrom
Tell us which papers should feature in the “Introduction to Nick Bostrom” series, and in what order. We’re especially keen for suggestions on what the first three papers should be.
If the series is to be 3 papers long, I’d suggest:
If the series is to be a dozen papers long, I’d suggest:

Will Aldred 16 Aug 2022 14:57 UTC
12 points
0 ∶ 0
on: A concern about the “evolutionary anchor” of Ajeya Cotra’s report on AI timelines.
Thanks for writing this post, have added it to my collection “AI timelines via bioanchors: the debate in one place”.