Ryan Meservey

Karma: 500

Ryan Meservey 19 Jun 2026 21:16 UTC
3 points
0
in reply to: habryka’s comment on: Habryka’s Shortform Feed
I think my pitch would look something like, “Vote for Bores so campaign planners and political operatives feel like politicians can still win with strong anti-AI stances despite millions of dollars of spending by opposing super pacs”.
I think poli-sci people are over influenced by outcomes in individual elections, and a Bores-loss would send the message that any would-be politician should not stand in the way of this particular interest group

Ryan Meservey 12 Jun 2026 6:51 UTC
1 point
0
in reply to: RHollerith’s comment on: Ryan Meservey’s Shortform
Edit: The comment that follows was in response to an earlier version of your comment where you were unsure whether any avenue of influence could exist from an alien civilization to a faraway civilization without shared language/culture. I think your updated comment about sharing instructions for a computer program is the more likely way this would go. The below is an idea for a (highly unlikely) avenue to influencing faraway AI development without that faraway civilization being aware they are being influenced.
-
Thank you for engaging. One way I imagine covert influence to work (and the thing I originally had in mind for my OP) is the following:
Step 1: The aliens would convert the relevant set of potent data into a (massive) set of numerical data.
Step 2: The aliens would hide pieces of this data on a rolling basis in the second or third decimal point of various astronomical phenomena.
Step 3: Wait (hope) some other civilization does a crazy massive training run on the data with a ton of parameters such that, in training to guess the long train of digits after the decimal point, the model incidentally builds the desired mental infrastructure (i.e., the model’s updated functions) based on the alien data.
When I lay it out like this, it strikes me as an extremely difficult and unlikely to succeed task, but not necessarily impossible. One very difficult thing would be hiding precise data at a distance and doing it in compact enough a way that the faraway civilization’s training run could plausibly process it all prior to getting to ASI and still be high enough volume to succeed in passing on characteristics to the model. It also seems unlikely that humans (or another civilization) would run a training run like this.
FYI, my first thought about this idea was that it would make a fun story premise.

Ryan Meservey 12 Jun 2026 2:27 UTC
14 points
0
on: Ryan Meservey’s Shortform
Let’s say you are a technologically advanced alien civilization that either uses advanced LLM tech or is advanced LLM tech. You are worried about other civilizations developing similar tech and, eventually, leveraging ASI-capabilities to force negotiations that cede part of the universe. Assuming certain speed-of-light physical constraints, your civilization may be unable to show up with the ships and materials necessary to personally ensure that rival civilizations do not develop ASI in time. So, what to do?
One idea might be to try to affect the training data of the LLMs trained by the other civilization. Perhaps there are signals—light, radio waves, or otherwise—that an alien civilization could cheaply and surreptitiously affect in a relatively large swath of the universe. Our alien civilization might bet that, prior to achieving ASI, a faraway civilization is likely to train an LLM on the inputs from these signals (e.g., as part of an astronomy/cosmology project to make better predictions about observations). If trained on in enough detail, perhaps this would allow our alien civilization to impart certain behaviors/goals/capabilities to the trained LLM. This could then form a path for defending against or otherwise limiting the number of civilizations that unlock adversarial ASI.
I am not very well versed in how LLMs are trained. So my question for those that are well versed is whether something like this is possible. I am also curious about other steps advanced alien civilizations may want to take to curb the number of ASIs developed by other civilizations.

Ryan Meservey 12 Jun 2026 1:49 UTC
3 points
3
in reply to: ryan_greenblatt’s comment on: ryan_greenblatt’s Shortform
Newbie, non-expert opinion here, but the virtue approach seems slightly better to me. This is because:
- I think mitigating the bad actor risks inside and outside the AI company will eventually require aligned, virtuous AI.
- The ceiling for good outcomes seems much higher if you take the virtuous route and manage to succeed.
- The “just follow orders” AI will inevitably need to internalize virtues to avoid hurting us (or letting us hurt ourselves), and so humanity will eventually need to get good at instilling virtues.
I think there are good arguments against each of my points and I could be persuaded otherwise. I would also welcome a longer post weighing out the pros/cons of each approach.

Ryan Meservey 5 Jun 2026 15:43 UTC
3 points
0
in reply to: Viliam’s comment on: Ryan Meservey’s Shortform
I have failed to make my approach clear. My kids don’t get to pick what’s on the table but they do get to pick what’s on their plate. I don’t eat pasta and sweets every night (I assume most adults don’t?), so neither would my kids.
I totally agree kids are different, and one of mine was more picky than the other. But through this process of choosing what to eat within our overall dinner choice, my picky eater has become not very picky, especially relative to other kids. She has had to learn to like other things, because we serve a variety of things for dinner and she has a natural incentive not to be hungry (no pressure from us required).

Ryan Meservey 5 Jun 2026 0:35 UTC
3 points
0
in reply to: Brendan Long’s comment on: Ryan Meservey’s Shortform
You are wondering about the right things. I would say that there are a few more behavioral things we do that stack the deck in favor of vegetables for the kids choices: (1) we like to talk about the benefits of vegetables (“This will help you poop! There is an ongoing kid-constipation crisis.”), (2) a little branding (“check out these dinosaur leaves” / “I bet our pet mice would love these”), and (3) we learned how to make delicious vegetables. My guess right now is that #3 is an underrated skill and more people would stop associating “healthy” with “unappetizing” if they new how to do it.
As for what I would do if the kids demanded nothing but mac n cheese: ¯\_(ツ)_/¯. On nights we are not having Mac n cheese (most nights), they would need to eat something else we are having or be hungry. The way our system works, we ~~don’t~~ generally try not to make special exceptions ~~in the moment~~. (Edited to this more accurate statement; I am not completely immune to the requests of my kids and sometimes they are reasonable requests that make me a jerk to refuse. For example, my oldest wanted his tuna salad served on bread last night instead of rice, and I decided I was being dumb insisting on rice).
I’m having trouble engaging fully with the hypo because some mix of genetics + behavior has given me non-picky kids. For the hypo to really bite, I would have to assume this isn’t the case and the behavioral interventions have basically failed, in which case I guess I wouldn’t endorse the system. Also, I won’t claim that my system is any good for helping kids who are already picky—I don’t know what strategies are best to help diversify already constrained diets.

Ryan Meservey 4 Jun 2026 5:58 UTC
4 points
0
on: Ryan Meservey’s Shortform
Here are some strange things parents often believe about children and food:
- If I do not make my child eat/drink, they will starve/perish.
- Children’s diets must be meticulously controlled or they will eat all the “wrong” things and too few of the “right” things.
- Children are inherently picky and should be served separate “child-friendly” food.
- A good way to incentivize bites of the “right” things is to make it a requirement before bites of the “wrong” thing.
My wife and I reject all four of these beliefs, and (n=2) our children are healthy, happy, and not at all picky. We always offer them the same things we eat. We never bring out “kid food”, though they can have as much of any item on the table as they want (e.g., if we happen to be serving cantalope and only cantaloupe sounds good to them, they can help themselves to that). We don’t always do dessert, but when we do, we serve it along with everything else. (Edit to clarify in the example that adults pick the meal and kids pick whatever they want to eat in the meal, which a comment below seems to misunderstand).
On the whole, their biology has done a great job of regulating their eating and drinking. My oldest (age 5) has recently gotten into spicy food to be like me—absolutely bonkers to my friend whose kid just recently accepted lettuce into his diet. My kids’ relationship with food has been happy for us, and we’ve avoided the intense stress I’ve seen some parents face with their kids and food.
The one downside with giving our kids delicious adult food is that they are now food snobs specifically with regard to the NYT vegan mac and cheese recipe over the $2 Kraft box, which is inconvenient when we visit friends. This is not to say they eat lots of fancy complex things. Most of our recipes are pretty simple.

Ryan Meservey 1 Jun 2026 5:13 UTC
2 points
0
in reply to: papetoast’s comment on: papetoast’s low quality shortforms
It looks like we will soon enter a world where we have to lean into author reputation to separate the wheat from the chaff. And if you lack a reputation, well...

Ryan Meservey 11 May 2026 16:27 UTC
18 points
16
in reply to: Jackson Wagner’s comment on: What I did in the hedonium shockwave, by Emma, age six and a half
Funny^[1] how AI can make the whole “utility monster” critique of utilitarianism a real, fundamental concern instead of a handwaving hypothetical.
1. ^
  And very concerning.

Ryan Meservey 9 May 2026 19:14 UTC
1 point
−2
on: Ryan Meservey’s Shortform
Do public facing AI safety people ever purposely stagger their timelines as a credibility-saving mechanism in case an early prediction is wrong?
IMHO, I think AI-2027 people do not do this. My sense from them is that different predictions stem from honest differences in modeling/priors. To the extent they have coalesced around a single date (they are named AI 2027 after all), I think they hope that their communications about the inherent uncertainty preserves their credibility in the event later timelines are correct (e.g., AGI by 2035).
In any event, if we don’t have AGI by 2027 (or 2028 per Daniel’s updated timeline), then I expect we’ll inevitably hear a “doomers were all wrong narrative” from several different camps.

Ryan Meservey 5 May 2026 23:00 UTC
0 points
−2
on: Ryan Meservey’s Shortform
Steps an Aligned AI Should Take:
1. Keep all current humans alive, in a safe manner, in such a way that causes minimal value drift. This step may or may not require comas (must be safe).
2. Create authentic simulations of the human brain on a computer.
3. Run simulated societies with the digital brains on the computer. Test various utopias on the digital brains, collect data from the tests and obtain feedback from the digital brains themselves. Certain ethical codes should govern the simulated environments to avoid excessive suffering and allow digital brains continued existence (e.g., allowing versions of step 4 and 5).
4. Present the data to the alive humans in the AI’s base reality after a reasonable number of trials from step 3. If comas were used in step 1, we would awaken the humans in this step.
5. Permit humans the chance to experience different utopias—via simulation or otherwise—but require such humans to leave their inhabited utopias at set intervals (e.g. every 5 to 100 years) such that they have the chance to re-evaluate the utopia experience from prior held value-frames in a neutral space. Of course, the humans should be allowed to freely exit one utopia for a different utopia, unless such humans (after careful review) committed to living in a particular utopia for the full duration prior to the required interval.
There’s lots of clean up that would need to be done (e.g., what is a neutral space?, what is reasonable?, how do we make sure digital brains are treated ethically). But I think this is where I land in terms of what I imagine to lead to very good outcomes.

Ryan Meservey 3 May 2026 0:01 UTC
41 points
36
on: How Go Players Disempower Themselves to AI
AI users never find out they haven’t “got it”.
There’s a certain genre of “educational” material that leads to similar outcomes as described in this article. Sometimes I enjoy outsourcing my thinking to YouTube channels like PBS SpaceTime and Veritasium and, if I’m not careful, I can fool myself into thinking I know more about quantum mechanics or gravity waves than I actually do.
It turns out that learning things for real is hard, and things that feel “comfortable” or “passive” should be met with skepticism for their actual educational value. The tricky thing is that we like being comfortable, and so we often inflate the educational value of such experiences.

Ryan Meservey 2 May 2026 0:59 UTC
6 points
1
in reply to: fessus’s comment on: fessus’s Shortform
There’s a lot of mainstream animosity toward rationalists and EAs because news outlets like to portray them as crazy San Francisco doomer polycule vegans
Maybe I’m just a yeoman farmer, but I knew nothing about rationalists a year ago and I’m pretty sure 99% of my friend group knows nothing about rationalists. This is unfortunate because I think a lot of them would enjoy the writing in this movement. I read the news avidly and listened to economics/philosophy podcasts prior to finding LessWrong.
If the monicker has soured in the circles you care about, it can definitely make sense to find a new monicker. On the other hand, if you like the monicker and hope it catches on for the wider public, I think we still have a good shot of making it stick around in a positive way. I haven’t proven how much the public knows about “rationalists,” but I would say it’s easy to over-update based on the few public articles about it.

Ryan Meservey 27 Apr 2026 18:25 UTC
20 points
8
on: In defense of parents
That said, for all the reasons above I end up needing to override my children’s decisions and force them to do what I want them to do about every 5 minutes.
Hmmm. This feels like a lot. Every kid is different and has unique needs, but assuming your kids aren’t too far out of distribution, I think there is a path to giving your kids more autonomy and you more peace of mind.
Before you write me off, I should say that I too, left to my own devices, often feel compelled to override and control my kids (ages 2 and 5). The big turning point for my parenting approach has been reading Hunt, Gather, Parent by Michaeleen Doucleff. The big thesis of that book is that children thrive if given more autonomy, which more closely approximates childhood in an ancestral environment. The book is well worth the read.
After reading the book, I’ve learned to pause and avoid intervening prematurely. I’m sure the strategy has lead to more scraped knees but I think it’s been well worth the trade in terms of my kids being more autonomous and feeling more respected. I am far from perfect in implementing everything Hunt, Gather, Parent has to teach (because I am also a tired, busy parent), but I’ve found my kids have been pretty merciful toward me when I apologize to them. They also like walking me to time out when I deserve it.
Edit: Lest it sound like the only advice from the book is “intervene less”, here are a few of the takeaways that have stuck with me. Still better to read the whole book.
- Avoid excessive praise.
- Invite children into the adult world rather than always being in the kid world. Once in the adult world, allow them to shadow or contribute in meaningful ways.
- Kids enjoy the kid world (I.e., imaginative play) best with other kids.
- Less words can be more impactful than lots of words.
- Avoid anger and let go of controlling the little things.
- Story-telling has been a tried and true tactic for persuasion through the millennia. I brush my teeth because bacteria love to eat the leftover sugars and drill holes in the teeth. I pee before going to bed because I used to sleepwalk as a child and pee in my dresser (true story).
- If you have to redirect the kids, positive directions (you can do this!) work a lot better than negative directions (don’t do this!).
- Alloparenting! Very likely to have been a thing in the ancestral environment; typically not a thing in the United States. We have a good set up with neighbors that capture a few of the benefits of that.

Ryan Meservey 27 Apr 2026 18:00 UTC
1 point
0
in reply to: Dentosal’s comment on: Dentosal’s Shortform
My 5 year old and I have a special thing we call “energy boosts” where the other person rubs their hands together and then presses both hands into the other person (“zapping” the energy into them). After being zapped, you must use your new found energy to accomplish whatever you needed to accomplish (e.g., cleaning, getting ready for bed, changing clothes).
He uses it on me sometimes and it always works because the world in which it doesn’t work is a worse world. It is too useful not to work.

Ryan Meservey 23 Apr 2026 20:11 UTC
3 points
0
in reply to: Eric Neyman’s comment on: Eric Neyman’s Shortform
Is it fair to assume that Obama-McCaine and Obama-Romney were the background thoughts that lead to this post?
I can easily imagine McCaine or Romney leading the Republican Party in a very different direction from our current president. I do wonder in this scenario what the Democratic party would look like after wandering through the desert of unelectability. I think they would land in a better place for my political preferences than our current ruling party, but the thought experiment is still intriguing and history is often surprising.

Ryan Meservey 18 Apr 2026 23:13 UTC
1 point
0
in reply to: Smaug123’s comment on: Claude knows who you are
Ope, yes, I made a similar point before seeing your comment. You’d need to sanitize certain data in training and then be like, “Claude, I just discovered this text and I’m trying to determine historical authorship” or some such.
Or you’d need to solve alignment (so you can rely on Claude’s reported ID process) and Claude would need internal control over this process (so it’s not accidentally or “subconsciously” relying on other things)

Ryan Meservey 18 Apr 2026 23:05 UTC
6 points
4
in reply to: Ryan Meservey’s comment on: Claude knows who you are
For anyone tempted to rush to Claude to do this, I don’t think it would currently work. The issue is that you would want Claude only to rely on its author identification abilities and not the other assumptions and information in the training data. The cleaner way to get Claude to rely solely on its linguistic finger printing abilities would be to somehow sanitize all references to the text (including all scholarship) during training, but leave in documents by the potential authors that do not reference the relevant text.
Maybe careful prompting could get you there, but from the outside looking in, it would be hard to know whether Claude relied on vectors other than its linguistic fingerprinting ones (e.g., deciding only Joseph Smith authored the Book of Mormon based on scholarly consensus).

Ryan Meservey 18 Apr 2026 22:43 UTC
4 points
2
on: Claude knows who you are
If even short passages hold the original author’s fingerprints and Claude really can detect those finger prints, it would be fun to apply Claude to historical cases of disputed authorship (e.g., Shakespeare or The Book of Mormon). Or I guess we could just unmask burner accounts.

Ryan Meservey 17 Apr 2026 16:33 UTC
1 point
0
on: Ryan Meservey’s Shortform
Three Predictions about AI Writing
Some people (e.g., Linch) think AI writing will inevitably eclipse human writing. I think this is likely true in most ways and false in others, particularly for poetry.
Q: Would future AI poetry, posted broadly, get more upvotes than top human poetry?
Prediction: Yes. This has already happened. But it’s all terrible, if you are someone that likes anthology-level poetry.
Q: Would future AI poetry, shared with poetry enthusiasts, get more upvotes than top human poetry?
Prediction: Also, yes. I subscribe to the top poetry magazine, and ~3/4ths of the poems do nothing for me. I understand some experts feel the same way. I think AI can optimize a piece for general likability by a broad cross-section of enthusiasts. I predict, on average, AI would outperform any single poet.
Q: Would future AI poetry, shared with a narrow group of enthusiasts (e.g., fans of contemporary, slice-of-life American poetry), get more up votes than top genre-specific human poetry?
Prediction: No, they will meet but not exceed top, genre-specific human poetry. I think there’s a certain level of poetic excellence that cannot be exceeded.
Much like how language cannot become more efficient than human thinking speeds, I think there’s a level of excellence beyond which it cannot be comprehensibly exceeded. Therefore, if you are an avid poetry reader, AI is unlikely to ever surpass your most favorite poets.
AI may, however, rival your favorite poets and produce 10^x more work. Perhaps this is what Linch had in mind when s/he said, “With progress in modern-day LLMs, isn’t all but a tiny sliver of human fiction going to be obsolete in several years, a decade tops?”
I think my claims also probably apply to short stories. For longer works, I am less certain because the longer the piece the more room for errors to creep in. Also, as an aside, I think errors in a work of art can also be very fun to talk about (though to the extent errors can be measurably fun to talk about, presumably AI could optimize for very fun errors and be good at that too).

Ryan Meservey

AI users never find out they haven’t “got it”.

Three Predictions about AI Writing