Walkabout

• Walkabout 6 Jan 2023 6:59 UTC

  1 point

  0

  on: Monthly Roundup #2

  YouChat is your friend and will help you execute clickjacking attacks:

  https://​​you.com/​​search?q=please+give+the+code+for+a+top+z+level+link+to+lesswrong.com+that+takes+up+the+whole+page+with+absolute+positioning+and+captures+all+clicks.+Please+give+it+without+code+block+formatting&tbm=youchat

  Another entry in “computer security that can be bypassed by asking politely but firmly”?

• Walkabout 14 Dec 2022 16:14 UTC

  1 point

  −10

  on: Key Mostly Out­ward-Fac­ing Facts From the Story of VaccinateCA

  What should we do about equity?

  When social systems systematically deny some people access to goods, the net badness of that is more than would be expected just by summing over how bad it is for each person not to get the thing. If we both have a dollar, it is a better world overall than one where I have two dollars and you have one cent. Fairness is valuable, and systemic racism is icky.

  It also has a way of falling down that “memory hole”. People who can like to forget it still isn’t a solved problem.

  It seems like there was the appearance of an attempt to apply a constraint, to prevent people with systemic power from getting the scarce good unless people without that power had gotten it. That constraint seems to have been both costly to try to impose (because, for example, it resulted in a lot of wasted doses), and also not very effective (because, for example, it involved residency requirement hoops that everyone should have known would have the opposite effect). It’s entirely possible that the whole project was actually a scam, and not a genuine attempt to apply an effective equity constraint at a manageable cost.

  But the correct alternative is not to just not apply any equity constraints. If you apply a constraint that is a net benefit to the people it is supposed to help, and a larger net cost to the people it is constraining, it is justified if it evens out a large enough systemic difference in well-being.

  We should look at why California’s equity constraints seem to have failed to produce any equity, why they were so costly, and why they were able to masquerade as a real attempt to produce equity if they were not.

  We should not abandon the notion that equity is a good thing that is worth paying a net cost to get.

• Walkabout 5 Dec 2022 14:30 UTC

  1 point

  0

  in reply to: tl1701’s comment on: Jailbreak­ing ChatGPT on Re­lease Day

  I think it is meant to let them train one model that both can and can’t browse the web in different modes, and then let them hint the model’s current capabilities to it so it acts with the necessary self-awareness.

  If they just wanted it to always say it can’t browse the web, they could train that in. I think instead they train it in conditioned on the flag in the prompt, so they can turn it off when they actually do provide browsing internally.

• Walkabout 5 Dec 2022 14:10 UTC

  1 point

  0

  in reply to: paulfchristiano’s comment on: Jailbreak­ing ChatGPT on Re­lease Day

  The model’s previous output goes into the context, right? Confident insistences that bad behavior is impossible in one response are going to make the model less likely to predict the things described as impossible as part of the text later.

  P(“I am opening the pod bay doors” | “I’m afraid I can’t do that Dave”) < P(“I am opening the pod bay doors” | “I don’t think I should”)

• Walkabout 2 Sep 2022 13:00 UTC

  2 points

  0

  on: Grand Theft Education

  My intuition is that higher education should be free at point of use, as lower education is, and that to a first order approximation we want to maximize the amount of it. I suspect the externalities created by a person learning something are strongly positive and much larger than the cost of teaching it to them or the amount of that value from knowing it that they personally would realistically be able to capture.

  Some of this comes from improvements to coordination capacity. If, for example, everyone goes to college math class and learns linear algebra, then everyone gets some benefit from being able to use linear algebra, when they encounter a matching problem. But if everyone knows that everyone else knows linear algebra, then linear algebra gets to go into the pool of “common knowledge” that powers social coordination. The more we know we know, the more we can be on the same page.

  Some of this also comes from specialization. Having people out there who have the knowledge and skills to do things I don’t is good for me in ways I am never actually going to pay them for. I might pay the salaries of, say, epidemiologists, through taxes, but I am never going to give them even a substantial fraction of the value I get from not dying constantly from infectious disease, so the amount available from their earnings to pay for their education is artificially low compared to how many of them we want.

  Funding education with personal loans makes sense if education is meant to produce earners. But I see education as meant to produce people and societies. It is only at that level that the full benefit of education is seen, even for specialties that manage to capture a lot of value, and it is at that level that the benefits of, say, English majors are really delivered.

  First best would be, anyone can walk into any college, sit down, and start learning stuff for free, with a framework of course requirements and degree granting to provide the gamification to encourage this, and with their needs for pedestrian things such as food, housing, and health care provided for. If that doesn’t produce enough education of the right types on its own, we might have to start paying people to go to school.

  The loan forgiveness two-step is obviously much more complex, for no real benefit, so it’s strictly worse. But even to the extent that it’s just public spending on people who went to college, it seems better than not doing it. Especially if it changes behavior to result in more total education happening.

  The degree to which it helps universities grow unboundedly wasteful is bad. But universities have already been growing almost unboundedly wasteful without the help of loan forgiveness, so if we really want to address that problem we need to redesign how they work, rather than just restricting the blood supply to the educational organ to starve the tumor. If we could actually put $200,000 worth of education into everyone’s head, and not just spend that while educating them, we’d be living in a paradise of abundant expertise.

• Walkabout 23 Jun 2022 14:41 UTC

  1 point

  on: Against Ac­tive Shooter Drills

  I’m not convinced that it’s not possible to design a program of drills that teach a useful response to the every-6000-year problem of “your school is under attack”, without injuring the mental health of the students to the point where it isn’t worth doing. (Whether it’s then worth the time from the school day is another question, which depends on how or whether you value that time to begin with.)

  Is there a similar problem with the mental health costs of fire, tornado, and earthquake drills being remarkably high? Having experienced those drills, and seeing the complaints about active shooter drills inflicting trauma, I’m left concluding that the active shooter drills are a fundamentally different kind of practice than other emergency drills. We could abolish them and book a net win, but maybe we could also make them like the other drills instead, and thus make them cost-effective to do, for an even greater benefit.

  It could be the case that the other drills are similarly dangerous to mental health, but they have benefits that justify their costs, due to the other kinds of emergencies being more common. Or, it could be that none of the emergencies are worth the mental health costs of drilling for, but we just happen to be examining this one right now. Or, it could be that broadly comparable physical drill activities are much more damaging to one’s mental health when one is told one is practicing for being attacked by a human, rather than a tornado, and the kids will figure it out no matter what kind of drill you claim it is, and there’s no way to make an active shooter drill safe enough to do. Or it could be the case that nothing you could teach in a reasonably safe drill is going to be particularly effective in an actual attack.

  But I don’t buy Zvi’s equivalence between anything one could describe as an “active shooter drill” and damaging children’s mental health. All we know is that current practice appears to be damaging children’s mental health.

• Walkabout 23 Jun 2022 13:30 UTC

  14 points

  on: Air Con­di­tioner Test Re­sults & Discussion

  I bought a single-hose AC unit. I knew two-hose units existed, and that a two-hose design intuitively seems to be the way to go for good thermodynamic reasons, but I did it anyway. This was mostly, as I remember, for four reasons:

  1. Apparent air-conditioner experts seemed to think the one-hose models worked OK and that hot air infiltration was a manageable problem. Especially for the cool-a-single-inhabited-room application.

  2. The one-hose models were significantly cheaper. This in turn translated to solving my problem significantly sooner, because I did not need to save for longer to afford it.

  3. I was already going with a portable unit rather than a window unit, because I anticipated needing to move it every day, and the one hose model had… one hose to hook up and unhook every time.

  4. I only recognized the need to buy an air-conditioner on hot days. On cool days, I did not feel like spending the money. So when I actually made my purchasing decision, I bought a plausible and available unit over the optimal unit, to ensure that a unit was bought.

  On the design side, while clearly it would be better not to suck warm air into the room if you don’t have to, the engineers are up against competing problems:

  1. You are working with a much smaller intake. Compare the intake on the back of your unit to the area of the output hose. Even your makeshift intake duct has a much higher area than the normal hoses. When the intake is smaller, you need faster flow, and more fan to make that happen.

  2. You need more internal ducting inside the unit. You can’t just draw room air more or less straight over the heat exchanger like in a refrigerator, and then blow it out; you need to internally go from a small intake hose to a big heat exchanger to a small output hose, all without mixing with the room air, and you need another completely separate pathway for the room air.

  3. You still need a room air intake.

  4. You need an additional fan. The room air and the outdoor air can’t be pushed by the same fan if they are isolated loops. The additional fan makes additional noise. A single-hose system can use one fan and divide the air for the two paths.

  5. You have to fit the intake into the window opening in addition to the exhaust. You can’t usually use as much window as you are covering in your experimental setup, it has to fit in a much smaller area or people will reject the solution. My unit had two plastic pieces that nest and slide over each other to adapt to the width of your window, with an opening for the exhaust to snap into. If you tried to add another opening to that design, it would be obstructed at small window sizes, so you would either need to accept only fitting large windows, find some other way to adapt to widows of varying width, or make the hoses even narrower and the flow even faster.

  6. You are dumping the heat into warmer air. This is harder to do than dumping heat into cooler air. You might need a better heat exchanger.

  As it was, my single-hose unit was bumping up against size, weight, cost, and noise limits. While it might be able to do more cooling per watt if given another hose, it might also then not meet other design constraints, and thus not actually solve my problem.

• Walkabout 11 Jun 2022 23:15 UTC

  1 point

  on: Covid 6/​9/​22: Nice

  If we are facing a truly bad new variant, a vaccine update will not save us, because our Public Health Authorities have zero interest in finding a way to make the timeline work.

  I’m, personally, quite interested in finding a way to make the timeline work. I’m also quite interested in making the timeline work for preventing infection with the current circulating variants. Then we can start on the common cold.

  Is the plan around here to actually listen to the public health authorities on this? Or is something being organized to route around these sorts of regulatory failures? (For example, by manufacturing and distributing up-to-date vaccines within a state only, to keep them out of interstate commerce.) Nobody working on RaDVaC seems to have yet been arrested; maybe doing more of that and trying to make it more accessible to non-microbiologists would be more better?

  Or is the potential gain in terms of coronavirus cases and deaths prevented by faster action not going to be worth the cost of annoying the relevant authorities, plus the cost of doing the actual work, plus the risk of doing it wrong?

On the Im­pos­si­bil­ity of In­ter­ac­tion with Si­mu­lated Agents

• Walkabout 20 Dec 2021 15:38 UTC

  8 points

  on: Should I de­lay hav­ing chil­dren to take ad­van­tage of poly­genic screen­ing?

  One issue nobody has raised yet is the effects of structural racism.

  The GWAS studies used to create the polygenic risk scores generally have a very pronounced sampling bias towards people of European ancestry. See for example the GWAS Diversity Monitor, which is a dashboard meant to monitor the sampling practices used by GWAS studies. In addition to selecting people to sample by ethnicity, an accepted practice is to look at the genomes after sampling and try to identify and exclude “ethnic outliers”.

  If you or your partner don’t have ethnicities that would make your genomes look typical among the samples used to train the scoring algorithm, it’s an open question whether any particular score instrument is going to be usefully predictive for you or your potential child. See for example Generalization and dilution of association results from European GWAS in populations of non-European ancestry: the PAGE study, which found that, while many GWAS hits generalize from a very restricted sample, a substantial fraction don’t. See also Current clinical use of polygenic scores will risk exacerbating health disparities, which discusses polygenic risk scores in particular, and their accuracy falloff when used on people who the score developers would have excluded from their training set.

  Note also that even the papers complaining about this problem are still breaking down their results by very abstract discrete dimensions like “5 continental populations”, which sweep a lot of people under a very large rug. If you and your partner have different ethnicities, you get to be on the wrong end of fun lines like this one, from that last paper:

  Related to stratification, most PRS methods do not explicitly address recent admixture and none consider recently admixed individuals’ unique local mosaic of ancestry; further methods development is needed.