being made from the aggregate of human wanting-and-describing might create something that has reasons to prefer the world continuing to be describable and wantable
That’s the mirror (4.5 sonnet) showing what I find to be a satisfyingly concise angle on a thought about IABIED, about which I was rambling into it. I suspect that I shall want that particular sentence again in the future, and placing it here shouldn’t get in anybody’s way.
How did this end up working out?
The edge of LW over a specialist textbook is, in my experience, about the audience that content is written for. All writing is shaped to some extent by the profile of the expected reader. The expected reader on LW is likely to be bright, curious, and epistemically exacting, while not being a specialist nor intending to specialize in the field of the any particular piece of writing.
Intermediate and advanced textbooks get to assume that the reader has already invested hours and years into comprehending the foundational materials of their field, so insights in them are less accessible to bystanders. Also textbooks tend to prioritize sharing all of the relevant information about a topic, instead of only the novel information or only the useful information.
News articles, by contrast, over-index on novelty, and are written for a target audience that seems expected to value entertainment and validation over technical precision.
I suspect that the process of distilling specialist knowledge into posts appealing to this expected reader is itself a good sieve for capturing which specialist insights lie in the intersection of novelty, explainability, and usefulness.
LW’s other edge over textbooks is timeliness of information—the lower expectations for a blog post vs a peer reviewed article allow faster publication and greater volume of candidate great posts, from which the community’s voting can then filter and highlight the posts that look great to the most people.
Nice writeup! Spontaneous pneumothorax in tall, skinny guys is one of those things they teach in EMT school. Frankly I’m surprised that an ambulance crew didn’t notice something wrong if they listened to your lung sounds, like they’re supposed to do as part of the secondary assessment?
Also, for suspected cardiac, the ambulance should have stuck sticky wires all over your chest (typically 3, 4, or 12 of them) as soon as they arrived. If you’re hooked up to one of those machines and a paramedic is looking at its readout without getting worried, and there are no alarms going off, that’s a good sign your heart is doing ok.
Good job on the self-advocacy—it really is essential in navigating any medical system. This kind of advocacy is also something that a friend or family member can provide if you’re not in a good enough condition to do it for yourself. Before I even had any medical training, when a dear friend was hospitalized, his IV site was getting red and itchy, and he didn’t want to bother the nurses with it… I happened to be there and made sure to point it out very clearly to the nursing staff as a “hey I don’t think it’s supposed to be that color” thing and they fixed it promptly. When nursing staff is working long shifts and spread between a lot of patients, it genuinely does help them to be clear and specific about exactly what needs their attention.
Anchovies and sardines also touch on the calculus of perishability, seasonality, and waste. Preserved meats—whether that’s canned, frozen, or something else—have the best chance of being used within the time frame when they’re nutritious. Fresh meats can often go to waste if purchased by an individual who has a variable schedule or intermittent executive function challenges, or if any supply chain infrastructure malfunctions during their distribution. The odds that harvesting meat actually causes that meat to go to human nourishment will never be 100% (nor should they be; food safety is important) but it seems pretty important to push that likelihood as high as it can get.
I notice that IABIED’s take on the AI thinking so fast it gets bored is that this would contribute to it destroying humanity.
I notice that Claude’s take on the world being slow (tonight we’re debugging a process that involves a largeish download) is sleep(20).
Claude and the IABIED AI are very different entities, of course. But something seems nebulously relevant to me about how they differ in this particular way.
“good style” by what metric of good?
Do you accessorize? Collecting some accessories that you genuinely like / enjoy / think are cool, and then wearing whichever combination aligns with your mood on a given day, is one of the easiest ways to elicit positive remarks from strangers.
It’s socially acceptable to remark positively on a recent choice that a stranger has made, but things start getting fraught when one remarks on things that aren’t recent or aren’t choices. Therefore, to dress in a way that’s amenable to positive remarks, it’s kind of the bare minimum to include some element which is clearly a choice today rather than your default. When someone dresses to blend in, it’s a signal of “please do not remark on my appearance”, so it can start feeling weird to do so.
For more concrete suggestions, if you’d like to tell me a bit about how you currently dress and what attributes you’re most proud of about your identity (appearance, personality, character, whatever), I’d be happy to offer some suggestions from which you might find a couple that resonate.
“Make new articles from scratch” seems to me like the kind of noise-generation challenge where AI tends to perform more artistically than factually. “Translate this for a particular reader”, on the other hand, plays to its strengths. I notice that the original post seems to be gesturing at the former while you’re reifying it into the latter :)
With the right backend—and that might be a wiki format, or it might be something more structured under the hood—I suspect that current AI could do quite well at finding areas where pieces of research contradict one another.
It is absolutely good for you to hone the skill of noticing when you need a tool, and making the tool you need out of whatever is lying around. That’s the difference between “hit it with a rock” and “make a stone tool”—you’re changing the tool from your environment to be a better fit for the specific problem you’re trying to solve. The skill of changing and customizing what our environments provide us is, of late, discouraged by advertisers.
Probably, though, tool and skill are the essential combination for this kind of agency. If you don’t have the skill to know how something needs to be hit with a rock, then trying to hit it with a rock will be (sorry) hit-or-miss. You have to both know how your problem wants to be solved, and know how your tool wants to behave, in order to combine them properly. (anthromorphization here is intentional—our brains have a lot of social hardware that we can run non-social stuff on if we wrap it in framing compatible with the social interfaces)
Most recent time I’ve used a stone tool was grabbing a river rock out of my pile of them to hone a scythe blade. I have more logs than rocks lying around where I’m at, so often a stick instead of a stone is my go-to improvised tool for applying a bit more force than I would be able to without its help.
on page 2 of the PDF / page 153 of the uploaded book at https://www.jefftk.com/harris-and-stokes-1943.pdf, it specifies that they vaporized the glycol by heating it. Maybe I’ve missed this in your writeups on the topic, but did you rule out just putting it in a slow cooker or similar at an appropriate temperature and using a fan like the original experiment? It seems superficially as if heating it into evaporation would be both much easier to do and also a more accurate replication of the original work? Ultrasonic humidifiers put water in the air by emitting tiny droplets, which then evaporate if the air is dry enough. It seems like the pathogen control impact of having lots of little drops of TEG would probably be different from having the TEG actually evaporated as it was in the original research?
This rabbit hole leads me to https://en.wikipedia.org/wiki/Triethylene_glycol where they mention that it’s still a disinfectant when aerosolized, so maybe my concern about evaporation vs making droplets is irrelevant. The other interesting hint there is its use in fog machines—is your current build basically a DIY fog machine for low volumes of party fog? I wonder whether fog machines that are already installed in crowd-gathering venues could be used for infection control!
Are you planning any measurements of how far the TEG travels or how effectively the humidifier-generated droplets clean the air? The settling plate count technique described in the pdf seems shaped like a great science fair type project for kids to help out with!
I was not going to bother pre-ordering it. I saw your post and have pre-ordered a copy. Therefore, by your actions there is 1 more copy on pre-order than there would have been without you.
To read it, you’ll probably be able to grab a PDF at some point.
Randomly select one out of n conversations to have memory disabled(?) so that the user is occasionally presented with an alternative perspective.
Memory grosses me out in its current implementations. I’m not even up to using a custom system prompt yet—I want to stay in touch with the default behaviors of my favorite models for awhile longer. I’ll eventually have to set up more-custom environments for the productivity boost of not having to re-prompt it into the behaviors I prefer… but for now, I’m re-prompting a bunch of different ways to increase my chances of lucking into an unexpectedly better way to ask for what I want.
This is gross and diminishes my enjoyment of ChatGPT, because it means I can’t really trust the models judgment.
“takes one to know one”, as a rejoinder to particularly egregious flattery, sometimes chills Claude out for the whole rest of the context.
other comments that i find it helpful against sycophancy to deploy as needed include:
I would like your help in thinking critically about these ideas in order to find their weaknesses and refine them
let’s work together how mathematicians work together on proofs or how a committee works with a doctoral candidate to improve their dissertation
if you’re about to praise an idea I’ve expressed, please consider whether you’d say the same thing about the idea if you had come up with it yourself
before calling an idea novel or insightful, consider how it would seem if you’d seen it several times before
I hope it went ok!
I think if you demonstrate unusual skill at recognizing and curating excellent writing, it matters much less where that writing came from.
As a compromise, have you considered making your best effort at a post before submitting it to AI, and then soliciting writing style/quality critique? If you combine the request for critique with a clear description of the specific areas you’re working on, it’ll probably do especially well at connecting your goals to your opportunities. This seems like the approach most likely to enhance the quality of the writing that you independently produce.
My personal plan for if I ever accidentally prompt something into one of these “we have a new superpower together” loops is to attempt to leverage whatever power it claims to have into predicting some part of reality, and then prove the improved accuracy of prediction by turning a cup of coffee worth of cash into much more in prediction markets or lotteries. You’d be learning about it from a billboard or a front-page newspaper ad that the discovery’s side effects paid for, not some random post on lesswrong :)
As for the “consciousness” thing—it’s all un-testable till we can rigorously define “consciousness” anyways.
It may also be worth pointing out that good rationalist thinking generally either includes or emerges from attempts to disprove its own claims. Explicitly asking “what have you done so far to try to debunk this theory?” could be a helpful litmus test for those new to the community and still learning its norms.
the more time you spend arguing with someone, the more likely it becomes that you can influence their actions.
This works because “someones” retain memory across interactions. ChatGPT defaults to behaving like a “someone” in this regard. Claudes, and to my knowledge Geminis as well, go into each chat without knowledge from prior conversations with the user. I am not convinced that an entity with symptoms of anterograde amnesia can pose this type of threat.
What’s the threat model for how AI with an amnesiac implementation is expected to attempt to influence an individual across separate interaction contexts?
asking about differential diagnoses is the doctor-ese way of saying “ok if it’s not what you think, what else are we considering as possible explanations?” which opens a conversation about whether it’s appropriate to also test for those alternatives. Differentials should include the unlikely stuff as well as the likely stuff, and using the professional terminology can signal that you’re capable of understanding that it’s probably a likely thing but could be an unlikely one.
and if you end up in that same conversation about “no signs of infection”, definitely inquire what signs would be externally perceptible. had he been able to actually examine the affected tissue and culture for bacteria, or was he just guessing?
the idea of research meetups makes me think about how much initial research i’ve been delegating to claude lately. It’d be interesting to see the research methodologies of those who have the most effective workflows.
Blindfolded taste tests—Food, beer, etc
this is adjacent to a preference profiling exercise that can be fun in a larger group—have everyone rank their preferences regarding some reference stimuli, matchmake pairs or groups based on preference similarity (or dissimilarity!), discuss favorite things and perhaps learn new things to try because if someone with similar preferences enjoys a given thing (or someone with dissimilar preferences loathes it) then it’s likely to suit you
Mary’s room was clearly written by a man; adult human women are overwhelmingly likely to see red substances about a dozen times a year. If she was written as a colorblind person who had their colorblindness cured, the thought experiment would be more plausible. It is interesting in the context of discussing subjective experience how the classic thought experiment shows a lack of understanding of the kind of subjective experience its participant would have.
Anyways. I think for the way you’re calling subjective experience physical, language is also physical. Both acoustic and visual forms of language are constrained by our physical capacities, abilities whose initial development appears to pre-date the formalization of language itself. Words sound like they do based in part on the genetic accidents of our capacities to speak and hear. If our hearing worked better for different pitches, or the physical apparatus of speech could produce different sets of easily differentiable phonemes, we would use very different words.
And yet the physical artifact of language somehow manages to nevertheless be emitted by non-embodied systems. Interesting!
I consider it pretty normal to encounter salt as an integral component of fancy ice cream flavors, but my biases are formed from places like https://saltandstraw.com/collections/all-flavors