Rachel Shu

Karma: 319

Rachel Shu Apr 24, 2025, 4:23 PM
10 points
4
in reply to: Martin Sustrik’s comment on: Accountability Sinks
I did follow that turn, I just am confused by the examples you chose to illustrate it with. The first examples of Bell Labs and VC firms I agree match the claim, but not the subsequent ones.

I am imagining an accountability sink as a situation where the person held responsible has no power over the outcome, shielding a third party. So this is bad as in the airline example (Attendant held responsible by disgruntled passenger, although mostly powerless, this shields corporate structure, problem not resolved), and good as in the VC example (VC firm held responsible by investors for profits, although mostly powerless, this shields startup founders to take risks, problem resolved successfully).

And if this is the frame you’re using, then I don’t see how the ER doctor and ATC controller examples fit this mold?

Rachel Shu Apr 23, 2025, 10:50 PM
22 points
14
on: Accountability Sinks
Maybe I’m just misunderstanding the structure of the essay, but I’m a bit confused by the second half of this essay — you begin to argue that there are benefits to designing accountability sinks correctly, but it seems like most of your subsequent examples to support this involve someone disobeying the formal process and taking responsibility!
- The ER doctor skips process, turns over triage, and takes responsibility. His actions are defended by people using out-of system reasoning.
- The ATC skips process, comes back, and takes responsibility. Her actions are defended by people using out-of system reasoning.
etc for Healthcare.gov, Boris Johnson, etc. They were operating in the context of accountability sinks which discouraged the thing they ultimately and rightly chose to do, within the system they would have been forgiven for just following the rules.
Likewise, the free market example given feels like the total opposite of an accountability sink! The person who has the problem is in fact the person who can solve it. The free market does have a classic accountability sink, in the form of externalities, but how it’s framed here seems like a perfect everyday example of the buck stopping exactly where it should stop.

Rachel Shu Mar 20, 2025, 1:00 PM
1 point
0
in reply to: Siebe’s comment on: METR: Measuring AI Ability to Complete Long Tasks
Possibly, but then you have to consider you can spin up possibly arbitrarily many instances of the LLM as well, in which case you might expect the trend to go even faster, as now you’re scaling on 2 axes, and we know parallel compute scales exceptionally well.

Parallel years don’t trade off exactly with years in series, but “20 people given 8 years” might do much more than 160 given one, or 1 given 160, depending on the task.

Rachel Shu Jan 28, 2025, 12:42 AM
19 points
19
in reply to: Eneasz’s comment on: [Link] A community alert about Ziz
I don’t think this is a good reason: There might be some other way of highlighting its importance in light of recent events, but it’s not core LW content.

Rachel Shu Dec 21, 2024, 8:50 PM
2 points
1
on: Dress Up For Secular Solstice
I think the strongest thing this fashion norm has going for it is that, without having read or even heard about this post until showing up at solstice, I and at least one other person who also hadn’t heard of it managed to coordinate on the same theme of “black + shiny” simply by wanting to wear something decorative for the occasion. I also like the “wear something you would like to be asked about” as a different kind of fashion coordination point, and think that it would be great for the summer solstice, which has a bit more of that vibe.

Rachel Shu Aug 29, 2024, 12:47 PM
7 points
0
on: Rachel Shu’s Shortform
Preliminary thoughts on MPOX 1b:

Most recent WHO report: https://www.who.int/publications/m/item/multi-country-outbreak-of-mpox—external-situation-report-35--12-august-2024

Released every other week so next one due in a day or two.

Agree with medical consensus that lockdown-qualifying global pandemic still seems unlikely, but I am concerned enough. I’m guessing <10% chance of global pandemic, <1% chance the median LessWronger gets infected, including tail risk (as in I don’t think that even 10% of Westerners get infected even if it becomes a global pandemic.) Nonsexual transmission is more likely this time around than 2022 but sex parties still seem like a key locus. If you intend to go to sex parties in the next 3-4 months you should get the Jynneos vaccine soon, as it takes 2 doses, and there is some chance of a temporary supply shortage.

Things to worry about:
- Presymptomatic transmission seems likely.
- This is more deadly, and higher transmission, than 2022′s MPOX Clade 1I outbreak. 3-5% death rate, more severe among children than adults (opposite of COVID). Unfortunately kids also tend to be in closer proximity to each other than adults.
- Less clearly associated with sexual contact than 2022, good evidence of nonsexual skin-skin transmission being prevalent.
- Some possible reported cases of reinfection in people who previously had MPOX (I can’t track this down)
- Virus is already being reported in non-endemic countries. No person-to-person contact in the West yet, but seems probable in the near future.
- People vaccinated for smallpox a long time ago have probably lost their immunity by now.
Reasons to not worry:
- Not a zero-day exploit like COVID was, we already have several proven, reasonably effective vaccines and a process for delivering them, and a decent number of previously vaccinated people. I haven’t looked into supply constraints on vaccine production, and naively expect production to scale well.
- No stupid refusal to test potential cases this time
- Probable cross-protection from 2022 outbreak (given that the vaccine is made from an even more distant virus and is protective)
- A lot of potential superspreaders (assuming sex parties are still a large component of the risk of superspreading) are already vaccinated/recovered, at least in the US.
- Aerosol or droplet transmission isn’t likely, and the limited evidence we currently have suggests that we’re still looking at skin-skin contact transmission as with MPOX Clade 1I. However, smallpox was primarily droplet-transmitted (with some evidence for aerosol transmission) so it’s not out of the realm of possibility. <Of course, that’s what we said with COVID too.>
Random other things I’ve learned or thought about:
- After the point at which droplet transmission is established, it seems like co-infection with other, more cough-inducing respiratory diseases is an underexplored risk factor for superspreading, but that’s not super common.
- Shedded smallpox scabs were not very infectious. Surface-based transmission was most likely during periods peak illness, when it was very obvious that it was smallpox and people knew to stay away.

Rachel Shu Jun 25, 2024, 8:04 PM
1 point
0
in reply to: Andrew Keenan Richardson’s comment on: Higher-effort summer solstice: What if we used AI (i.e., Angel Island)?
It’s actually feasible with proper planning, I looked into this last December (you have to book 6 months out). If this year is like last year in terms of policy, demand, and reservations meta, I’m 99% confident we’ll get the picnic sites, 95% confident we get 60+ camping spots, 70% confident we get 100+ camping spots.

Rachel Shu Jun 25, 2024, 4:03 PM
1 point
0
in reply to: girllich’s comment on: Higher-effort summer solstice: What if we used AI (i.e., Angel Island)?
Do you think it’d influence your decision strongly if the organizers were to charter a small motorboat to take people back to Tiburon on-demand? The ferries run every hour, the boat could make an additional two trips an hour, so average waiting time would be 20 minutes and often the boat would be immediately available. It’s still 45 min from Tiburon back to Berkeley.

Rachel Shu Nov 13, 2023, 9:20 AM
1 point
in reply to: Yair Halberstadt’s comment on: Combining the best of Georgian and Harberger taxes
This isn’t a great way of thinking about intangible value, since most people are far less liquid than the value of the things in their lives. If your mom needs life-saving treatment for $10,000 and you can’t afford to pay it, does that mean your mom is really worth less than $10,000 to you? No—her existence might contribute hundreds of thousands of dollars of devotion, emotional support, and sage advice to you. It’s just that nobody will accept those things as collateral for a loan.

The Colosseum doesn’t have as stark a value proposition going for it, but by a similar argument it’s easy to imagine the public benefit of some building being greater than the amount of money that can be raised to support it. Personally, I think the Colosseum is mostly only valuable to the Italian state for the legitimacy that it buys it, a little self-justifying scheme of wasting taxpayer money to manufacture taxpayer assent.

Rachel Shu Nov 13, 2023, 9:09 AM
1 point
in reply to: Dagon’s comment on: Combining the best of Georgian and Harberger taxes
I disagree about the transfer costs making Harberger tax unworkable, that gets priced into the self-assessed property value. You pay taxes on the use value of the property to you, not the market value, which is typically lower. When you want to sell, you’ll offer at the lower market rate.

But this actually points to an even larger problem. Harberger taxes as such would be incredibly destructive to communities. The stronger the community, the higher the use value, meaning the higher the self-assessed tax rate would have to be in order to reflect the worth to you. But the value from community is uncashable, meaning that there’s an unjust tax burden for having a better community. And people are usually only liquid to a minute proportion of the value of the intangible things in their lives.

Thus, people likely have to assess below their use value, which means the community can be eroded by compulsory sale and eviction. This already occurs in rent-paying communities, which are destroyed when property value goes up and they can no longer afford to rent there.

Tbh, you probably end up fighting this by having your neighbors in the community firebomb the property which got bought out from under you.

Rachel Shu Nov 13, 2023, 8:49 AM
1 point
in reply to: JBlack’s comment on: Combining the best of Georgian and Harberger taxes
I think you can just drop the ‘zero notice’ part, add something like a 90 day delay between sale and transfer, and things would be a lot more workable.

There’s another problem that your comment makes me think of—in a lot of cases, the easiest and most profitable thing to do will be to just buy the property and offer to lease it back to the previous owner at a higher rate than they were last paying in Harberger tax. With capital accumulation, you’ll probably end up with giant extractive leasing company monopolies practically owning unwilling renters with no outside options—not sure what to do about this.

Rachel Shu Nov 13, 2023, 8:41 AM
1 point
in reply to: Yair Halberstadt’s comment on: Combining the best of Georgian and Harberger taxes
If anything, it’ll be more common for the previous owner to trash the place on their way out with some excuse for avoiding legal damages.

Rachel Shu Nov 13, 2023, 8:39 AM
1 point
in reply to: Yair Halberstadt’s comment on: Combining the best of Georgian and Harberger taxes
I think in the typical Harberger scheme, in order to compel transfer of property rights, you pay the entire capital cost of the property, not some portion as a temporary rent.

Rachel Shu Nov 13, 2023, 8:38 AM
1 point
in reply to: gwern’s comment on: Combining the best of Georgian and Harberger taxes
My proposed solution to both your problem and Richard’s, explained further in a separate comment, is that payment and transfer be separated by enough time to vacate amicably, or sue to enjoin transfer.

Rachel Shu Nov 13, 2023, 8:36 AM
1 point
in reply to: rk’s comment on: Combining the best of Georgian and Harberger taxes
where the value to an actor of some property might be less than the amount of value they have custody over via that property.
I don’t think those are separate things? The value of a roof is the value of everything underneath it when it rains.

Rachel Shu Nov 13, 2023, 8:25 AM
1 point
on: Combining the best of Georgian and Harberger taxes
Property owners are only going to assess their property for tax purposes at perceived market value if they’re ready to sell. Otherwise they’ll assess it higher, factoring in moving costs. The Harberger buyer is usually overpaying.
For this reason, I think most Harberger extortion is easily avoidable by adding a sufficient delay between sale time and transfer time. For example, for most categories of residential and commercial real estate, a reasonable expectation for compulsory sale would be perhaps 90 days between purchase and transfer. (This is already the custom for foreclosures, I think.) The compelled seller, having priced in moving costs, will have time to decide how to put their new liquidity to use and make the proper adjustments or sue to enjoin the transfer if maliciousness such as repeat harassment is suspected.
This leaves the problem of sudden price changes. One solution is to make compulsory sales begin as blind auctions. If someone submits a blind bid for your property at or above your self-assessed tax value, you get a short period of time to submit a reassessment, and if your bid is higher you pay back taxes on that reassessment within that tax year, otherwise they buy from you at their assessment. I think this scheme gets both bidders to bid their true rates, and discourages holdouts approximately the same amount? It makes the market slightly less efficient but keeps the same tax efficiency.

Rachel Shu Oct 26, 2022, 9:17 PM
9 points
2
in reply to: James Grugett’s comment on: Crypto loves impact markets: Notes from Schelling Point Bogotá
You correctly imply something worth restating clearly: despite their initial framing, impact markets are not a way to achieve public goods per se, they are a way to efficiently achieve funder goals in contexts where the path to that goal is uncertain, there are many plausible options, and there is lots of information that can be potentially priced into the market about those options.

With some impact market designs decentralized self interest can in fact come into play, perhaps in the form of bounty pools pledged into escrow by some subset of the people who would benefit from the existence of the public good they are offering the bounty for. In this case funders have a vested interest in what is achieved, and will evaluate based on such. Maybe in the long run markets that enable such a design will gain reputability relative to markets solely funded and assessed by fly-by philanthropists.

I agree with you that disinterested funders often end up having counterproductive goals, although not in all domains. The above example of generic pharmaceutical repurposing trials might be such, where the market can bear useful information about which of many interventions would have the highest impact and chance of success, but the work to achieve that goal is kind of hard to do serious harm with, given that the risk profiles of those drugs is already well quantified. In such cases I see especially little risk to encouraging philanthropy.

If some funder intentionally wishes to achieve nefarious goals with impact markets, I admit the existence of impact market infrastructure might facilitate that. But we have legal and social tools to counteract bad ends and I don’t think that impact markets are so powerful as to enable an end run around these.

Rachel Shu Apr 24, 2022, 11:03 PM
1 point
on: My current thoughts on the risks from SETI
Even granting that there are grabby aliens in your cosmic neighborhood (click here to chat with them*), I find the case for SETI-risk entirely unpersuasive (as in, trillionths of a percent plausible, or indistinguishable from cosmic background uncertainty), and will summarize some of the arguments others have already made against it and some of my own. I think it is so implausible that I don’t see any need to urge SETI to change their policy. [Throwing in a bunch of completely spitballed, mostly-meaningless felt-sense order-of-magnitude probability estimates.]
- Parsability. As Ben points out, conveying meaning is hard. Language is highly arbitrary; the aliens are going to know enough about human languages to have a crack at composing bytestrings that compile executable code? No chance if undirected transmission, 1% if directed transmission intended to exploit my civilization in particular.
- System complexity. Dweomite is correct that conveying meaning to computers is even harder. There is far too much flexibility, and far too many arbitrary and idiosyncratic choices made in computer architectures and programming languages. No chance if undirected, 10% if directed, conditioning on all above conditions being fulfilled.
- Transmission fidelity. If you want to transmit encrypted messages or program code, you can’t be dropping bits. Do you know what frequency I’m listening on, and what my sample depth is? The orbital period of my planet and the location of my telescope? What the interplanetary and terrestrial weather conditions that day are going to be, you being presumably light-years away or you’d have chosen a different attack vector? You want to mail me a bomb, but you’re shipping it in parts, expecting all the pieces to get there, and asking me to build it myself as well? 0.01% chance if undirected, 1% if directed, conditioning on all above conditions being fulfilled.
- Compute. As MichaelStJules’s comment suggests, if the compute needed to reproduce powerful AI is anything like Ajeya’s estimates, who cares if some random asshole runs the thing on their PC? No chance if undirected, 1% if directed, conditioning on all above conditions being fulfilled.
- Information density. Sorry, how much training is your AI going to have to do in order to be functional? Do you have a model that can bootstrap itself up from as much data as you can send in an unbroken transmission? Are you going to be able to access the hardware necessary to obtain more information? See above objections. There’s terabytes of SETI recordings, but probably at most megabytes of meaningful data in there. 1% chance if undirected, 100% if directed, conditioning on all above conditions being fulfilled.
- Inflexible policy in the case of observed risk. If the first three lines look like an exploit, I’m not posting it on the internet. Likewise, if an alien virus I accidentally posted somehow does manage to infect a whole bunch of people’s computers, I’m shutting off the radio telescope before you can start beaming down an entire AI, etc, etc. (I don’t think you’d manage to target all architectures with a single transmission without being detected; even if your entire program was encrypted to the point of indistinguishability from entropy, the escape code and decrypter are going to have to look like legible information to anyone doing any amount of analysis.) Good luck social engineering me out of pragmatism, even if I wasn’t listening to x-risk concerns before now. 1% chance if undirected, 10% if directed, conditioning on all above conditions being fulfilled.
So if you were an extraterrestrial civilization trying this strategy, most of the time you’d just end up accomplishing nothing, and if you even got close to accomplishing something, you’d more often be alerting neighboring civilizations about your hostile intentions than succeeding. Maybe you’d have a couple lucky successes. I hope you are traveling at a reasonable fraction of C, because if not you’ve just given your targets a lot of advance warning about any planned invasion.

I just don’t think this one is worth anyone’s time, sorry. I’d expect any extraterrestrial communications we receive to be at least superficially friendly, and intended to be clearly understood rather than accidentally executed, and the first sign of hostility to be something like a lethal gamma-ray burst. In the case that I did observe an attempt to execute this strategy, I’d be highly inclined to believe that the aliens already had us completely owned and were trolling us for lolz.

*Why exactly did you click on a spammy-looking link in a comment on the topic of arbitrary code execution?

Rachel Shu Feb 4, 2022, 4:20 AM
3 points
on: We Choose To Align AI
Composer Christopher Tin has set JFK’s “We Choose to go to the Moon” speech to music, https://www.youtube.com/watch?v=HBITb9Zz0rY . Solsticegoers may recognize the opening leitmotif as shared with Sogno Di Volare, another movement from the same work, an oratorio on the theme of flight, To Shiver the Sky.

Rachel Shu Jan 28, 2022, 5:13 AM
10 points
5
in reply to: Alex_Altair’s comment on: What are some beautiful, rationalist artworks?
I’ve recently gained a better appreciation for how astonishingly good this work is at linear perspective, which had only come about in European art in the prior century. Many things about this painting are good (and some bad to my eye, like the messy color scheme) but those hexagonal details on the curved arches in perspective is 100% Raphael showing off.

An aside, but linear perspective is the most rational part of art, in the older philosophical sense of rational; it’s pretty much the only major part of classical art which descends from first principles rather than having an empirical basis.