There’s a formatting issue with the link, should be: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2634591/
XFrequentist
Karma: 2,322
Preventing neural network weight exfiltration (by third parties or an AI itself)
This is really really interesting; a fairly “normal” infosec concern to prevent IP/PII theft, plus a (necessary?) step in many AGI risk scenarios. Is the claim that one could become a “world expert” specifically in this (ie without becoming an expert in information security more generally)?
Indeed, as Vladmir gleaned, I just wanted to clarify that the historical roots of LW & AGI risk are deeper than might be immediately apparent, which could offer a better explanation for the prevalence of Doomerism than, like, EY enchanting us with his eyes or whatever.
I am saddened that this doomerism has gained so much track in a community as great as LW
You’re aware that Less Wrong (and the project of applied rationality) literally began as EY’s effort to produce a cohort of humans capable of clearly recognizing the AGI problem?
It’s probably based on GPT-4.
Bing literally says it’s powered by “GPT 4.0 technology” in this chat, is that synonymous with GPT-4 (genuinely unsure)?
I’ve actually wondered if some kind of stripped-down sign language could be a useful adjunct to verbal communication, and specifically if a rationalist version could be used to convey epistemic status (or other non-obvious conversational metadata).
In the (outstanding) show The Expanse, a branch of humanity called “Belters” have been mining the asteroid belt for enough generations that they have begun to diverge (culturally, politically, and even physically) from <humanity-main>. They have such an adjunct sign language, originally developed to communicate in the void of space, fully integrated into their standard communication.
This seems so useful! I’m so frequently frustrated in conversations, trying to align on the same meta-level as my conversational partner, or convey my epistemic status effectively without derailing the object-level conversation.
An unrelated anecdote, on the general awesomeness of signing. Years ago, I was heading home on the NYC subway late at night, and the usual A train din precluded conversation. Most passengers were mindlessly scrolling on their phones or sullenly staring out windows, but four young men were carrying on a silent, boisterous conversation via signing, with full-body laughter and obvious joy.
In that environment, their (presumed) general disability translated to local advantage. Still makes me happy to think about.
<sarcasm>
And obviously, the entire public health community is up in arms about this…
</sarcasm>
[Narrator: They were not, in fact, up in arms.]
There might be another strain in the future. I don’t know how likely this is, but that’s the most likely way that things ‘don’t mostly end’ after this wave
I agree, and I also don’t really have great mental handles to model this, but this seems like the most consequential question to predict post-Omicron life. My two biggest surprises of the pandemic have been Delta and Omicron, so sorting this out feels like a high VOI investment.
Here’s a messy brain dump on this, mostly I’m just looking for a better framework for thinking about this.
The amount of transmission obviously matters, since more generations provide more opportunities for mutation. All else equal, VOCs are more likely to arise where cases are high.
Is a partially vaccinated population more likely to generate VOCs? Either in the sense of a large number of single dosed people, or a large proportion unvaccinated, or some complex interaction of the two?
If mutation and selection is happening within an immunocompromised individual (as opposed to gradually accumulating in a population), does this imply regions with high HIV-AIDS rates are most likely to produce VOCs? Are there other clusters of immunocompromised people, or are they fairly evenly distributed?
The two VOC (Beta and Omicron) with the most immune evasion arose (or were at least first detected) in South Africa. Is that a coincidence? Is the presence of a BSL-4 (the only such institution south of Gabon)? Origin vs detection is confounded by the relative abundance of sequencing resources in SA vs the rest of Africa, which makes detection in SA quite likely even if VOCs emerge elsewhere, but the pattern of spread of both Beta and Omicron are more consistent with SA origin.
Is any of this modifiable through policy decisions? The WHO is urging wealthy countries to forgo boosters in favour of distributing doses to the developing world, which is on-brand lunatic messaging given current events, but post-Omicron wave could this actually be good policy? Maybe this depends on sorting out some of the above?
The lightcone is such a great symbol. It also kind of looks like an hourglass, evoking (to me) the image of time (and galaxies) slipping away. Kudos!
you really could have been the first mover on a few of these new enterprises back in 2021 if you had brainstormed a bit. Describe one of them.
Fun!
Curated data services for forecasting.
High-trust paid newsletter/research service oriented around interesting markets. Maintain a publicly-verifiable scoreboard linked to market positions to demonstrate reliability.
Insurance markets—take an extreme position as a catastrophe hedge, use this to subsidize market making.
Social network—rationales and discussion for various markets. Content discovery algorithms could include historic accuracy so truth propagates more easily.
Infrastructure to annotate opinion pieces with concrete predictions, link to market positions.
… [will update if more occur to me]
Lots of things, but the biggest win is probably snow removal services.
For $200 a year I save several dozen hours of drudgery, there’s no management/coordination overhead to speak of, and my plow guy does a better job than I would have.
The commentary below has focused on child care—a more salient pain point for our demographic, surely—but the “elder care” angle actually seems much more promising. Still labor-intensive, but fewer regulatory nightmares (?).
Note there are some very large regional players in this game, but there don’t appear to be any Starbucks-size winners (so says my wife, who often works with the elderly).
Thanks! LW was malfunctioning when I posted this, otherwise I would have.
This.
Also, schlep alert: this might be the densest regulatory thicket outside of healthcare, with huge variation in standards at (at least?) the state/province level. In my little environment of 13 million Ontarians, a recent arbitrary change of the teacher/child ratio allegedly drove a good many daycares out of business.
Also, parents are insane (source: am parent).
Program good ethics into artificial intelligence
Assemble a group of scientists who on their own could eradicate mosquitoes and just do it. Don’t wait for official approval.
The appeal of this route is obvious, but I don’t think it should be discussed on a public forum.
Agreed! What would be the best approach (I’m a PhD student and vector-borne disease epidemiologist)?
Writing one or more popular/lay articles
Writing one or more technical/scholarly articles
Writing a popular/lay book
Writing a technical/scholarly book
Starting an advocacy non-profit
Performing an explicit cost-benefit analysis
Modelling to determine the necessary conditions for eradication
Something else… ?
Yar, have taken the scurvy survey, says I!
Your definition what counts as “AI related” seems to be narrower than mine, but fine. I trust readers can judge whether the linked resources are of interest.
Possibly you’re thinking about this: https://www.quantifiedintuitions.org/pastcasting