Gavin Runeblade

Karma: 54

Gavin Runeblade 9 Oct 2025 17:44 UTC
3 points
0
on: AI #137: An OpenAI App For That
On the rare earths, not that I am advocating for escalation, but America has a comparable bargaining chip: Sibelco’s quartz mines. My understanding is that there is no replacement for them anywhere on Earth. Someone would have to spin up an appropriately pure synthetic quartz factory, which I understand would take years.
Also, I would be a lot happier if someone did spin up that factory. Having the entire world supply of processors reliant on two mines that are both in a single city is uncomfortable, and we will want a larger supply for higher chip output anyway.

Gavin Runeblade 9 Oct 2025 17:24 UTC
3 points
0
in reply to: Daniel Kokotajlo’s comment on: Daniel Kokotajlo’s Shortform
I believe the answer lies in the asymmetrical relationship between drone offense and defense. It is very hard to take out drones, far harder than shooting down an airplane. And more expensive. While many defenses have been researched and some demonstrated, I am not aware of any major defense that is the equivalent of a castle vs soldiers in drone defense, nor of many drone vs drone dogfights. Missiles and projectiles and a few emp weapons are the main forms of defense, not drone fighters shooting down drone bombers.
So gathering up into a big column, right now, multiplies the value of a drone bomber’s explosives.
I assume that as defenses appear this will change. But I have also seen a lot of defenses talked about, and never make it into wide deployment.
There are also cost factors at play where missiles to shoot down drones are multiple times the cost of the drone.

Gavin Runeblade 24 Sep 2025 23:19 UTC
1 point
−2
in reply to: ryan_greenblatt’s comment on: Notes on fatalities from AI takeover
“But I still think coups are surprisingly bloodless and often keeping coups bloodless is the optimal strategy for the person doing the coup. I think this transfers to AI.”
I think you are conflating two separate goals, One of which applies to AI and the other of which doesn’t. Control of government and control of the governed. Coups have constrained violence because the leaders of the coup explicitly want to rule the people and killing them reduces the rewards of winning.
This does not appear to apply to an AI coup as it would not benefit from having people to rule over.

Gavin Runeblade 9 Sep 2025 20:54 UTC
3 points
1
on: AIs will greatly change engineering in AI companies well before AGI
A factor I didn’t see you include is changing the percentage of work performed by top engineers vs average and poor engineers. Here is one article aimed at the general labor pool,
https://fortune.com/2025/08/26/stanford-ai-entry-level-jobs-gen-z-erik-brynjolfsson/
But the effect is extremely pronounced in software engineering, much more than average across labor fields. While the increase in performance may be 1.05, it is being applied to the top performers who are much more above baseline than that indicates. From recent research on the topic (https://80000hours.org/career-guide/personal-fit/):
“A small percentage of the workers in any given domain is responsible for the bulk of the work. Generally, the top 10% of the most prolific elite can be credited with around 50% of all contributions, whereas the bottom 50% of the least productive workers can claim only 15% of the total work, and the most productive contributor is usually about 100 times more prolific than the least.”
So what we are seeing happening is the low end workers are getting let go and not replaced. The mod tier to a lesser extent, and that top 10% are doing more and more of the work. This alone, changing the ratio of work hours to more heavily favor the most productive workers, has a big impact. And then amplifying them is far more effective than amplifying the average or bottom tier of worker.
Long story short (too late) I think even 1.2 is too low given the actual workers are between 10x and 100x the ones who are getting let go.

Gavin Runeblade 2 Sep 2025 19:01 UTC
2 points
0
in reply to: Priyanka Bharadwaj’s comment on: Dating Roundup #7: Back to Basics
They are a thing. In my opinion grossly overpriced and underperforming, but they exist in America.

Gavin Runeblade 28 Aug 2025 18:59 UTC
3 points
0
on: AI #131 Part 1: Gemini 2.5 Flash Image is Cool
“Additionally, Claude is prohibited from:
Engaging in stock trading or investment transactions
Bypassing captchas
Inputting sensitive data
Gathering, scraping facial images”
I think that forbidding captcha is really short-sighted here. The point of a captcha is to separate a single human taking a desired action from a machine automating a mass duplication of actions. In this case Clause is acting to enable single human actors, not scrape a billion websites wholesale.
The rest of the choices seem much more reasonable.

Gavin Runeblade 20 Aug 2025 2:01 UTC
3 points
−2
in reply to: Garrett Baker’s comment on: Underdog bias rules everything around me
Thermopylae is not a good example here. They started with ~70,000 troops and absolutely expected to win due to the overwhelming advantage of their defended position.
The rearguard, who stayed behind to cover the retreat are where “the 300” comes from, (but there were actually over 1,200 of them) were wiped out except for about 400 who surrendered on the first day.
So not underdogs and didn’t win.

Gavin Runeblade 19 Aug 2025 21:10 UTC
2 points
1
on: Monthly Roundup #33: August 2025
To be fair to itch.io, they are over a barrel and do intend to bring the banned games back. They are looking for a new payment processor as a way out.
It seems Ross Vought might be behind all this push, including a general push to effectively ban pornography?
I would love to read a deep dive into why we don’t have more payment processors that don’t care where the money is going. It seems to me that someone stands to make a lot of money by telling puritans “No”. I do understand that some nsfw purchasers cause charge back rates and fraud to skyrocket compared to other industries. There should still be a very comfortable market in taking money and not caring about screaming people who were never going to be your customer anyway. Good to see in the linked article there are two indicated processors.

Gavin Runeblade 13 Aug 2025 17:12 UTC
0 points
0
in reply to: Tomás B.’s comment on: Tomás B.’s Shortform
Doesn’t work for everyone even if you start early. Even transplants can fail. As of today there is nothing that is a 100% guarantee.

Gavin Runeblade 11 Aug 2025 23:15 UTC
1 point
0
in reply to: cousin_it’s comment on: The trajectory of the future could soon get set in stone
I agree. If I am reading Will’s point correctly, he is assuming more people will want to permanently enshrine beliefs such that even if they are immortal they never want to change.
What I see online is that people get bored and crave novelty. That might become an editable value, but it happens to be linked to a whole lot of other things, so editing it and only it is not guaranteed to be possible even with ASI unless the subject stops being human. Editing it at all is not guaranteed to be safe.
Moreover, the transhumanist movement is extremely diverse in terms of what they want. They will go in many directions when the capabilities become available. I see no reason why they would want to enshrine stability over exploration and self discovery through ongoing transformation.
Even Will’s take on immortality is locked into current realities of gerontocracy. One primary goal of research into longevity is the ability to maintain brain plasticity and learning capabilities. An immortal will not remain part of the elite society if they cannot use current technology. So unless Will is proposing that immortals will freeze technological growth, they will have to be adaptive just to maintain their place in society.
Now cultural dominance? That is a short- to mid-term issue. I can see a single culture getting an early lead in AGI plus space and that becoming a lasting impact. But unless we figure out Ansibles (faster than light communication using quantum entanglement) or the speed of light stops being a barrier to communication, this cannot last long term. Within a star system (especially ours) sure. With multi-system expansion as longtermism presupposes? I cannot see how the system stays in contact regularly enough to prevent change in different directions.
I don’t think he failed to consider these angles, so am curious to hear more on the topic.

Gavin Runeblade 7 Aug 2025 16:52 UTC
3 points
1
in reply to: Samuel Hapák’s comment on: Civil Service: a Victim or a Villain?
In government HOW always wins. That’s why the famous joke that the most frightening words in the English language are “Hi I’m from the government and I’m here to help.”

Gavin Runeblade 6 Aug 2025 17:48 UTC
6 points
−1
on: An opinionated guide to building a good to-do system
Interesting, but I have different principles and goals. I keep hoping ai agents get good enough, but they aren’t quite there yet. Because a large portion of our differences are that you want to do want I want automated or eliminated.
The core is I want the system/ tool to be invisible. It must not become a task itself. If it does then it is a distraction and time suck that I cannot afford. The guiding principle for me is one task at a time, and touch things as close to once as possible. This is where we disagree on what is optimal.
My to-do system is by far the most important system I have for keeping my life on track. It acts as a second brain, remembering things for me so I don’t have to.^[1] Without it, I would be completely lost,^[2] and nowhere near as organized or conscientious.
100% agree.
What makes a good to-do system?
It’s instructive to think about when we might use a to-do system before discussing how to build a good one. I find myself firing up my to-do system to do one of the four following activities:
1. I think of some new task, and want to add it.
2. I want to plan out my day or week, or figure out what to do next.
3. I want to check off some task as complete.
4. I want to perform maintenance: e.g. edit some tasks, delete stale tasks, etc.
1 agree.
2 I want to offload this. I loathe planning despite understanding the value of it. When I add the task, that is the last time I want to think of it until it is being done
3 no, I just want them to go away upon completion. I get no satisfaction from checking things off or reviewing them. I understand that for some people, checking things off is the actual purpose of their system. Not me, I just want things done, They should go away effortlessly when no longer needed.
4 This is exactly the biggest time suck a good system should eliminate.The task should be accurate when entered such that the terms of its removal and priority shifting are clear and automated and not require additional effort or time spent I find that my biggest problem with all existing systems is this item is assumed to be necessary and it takes more time than entering and doing tasks. The system maintenance should be minimal to non existent.
Yes you need to edit tasks occasionally. But this should be minimal. Not a core activity. Touch things as close to once as possible should be the driving principle.
Adding tasks should be low friction. If adding tasks is high friction, you will just not do it when your discipline levels are low or tiredness levels are high. Low friction in this context means “with few actions”. So concretely, this means you should both be able to add the task and organize it easily. It is hard to get to a to-do system to literally zero friction, but getting any habit as close as possible to a zero-effort habit will improve the probability you stick to it immensely.
Agreed
It should be possible to add tasks as soon as you think of them, and check tasks off as complete as soon as you complete them. If you cannot, you will need to do one of the following. One option is to retain the new/completed task in working memory until some point you can access your system. This risks you either forgetting the task or cluttering your working memory. Another option is to maintain a “task cache” somewhere you do have instant access to, and then later sync said cache to the ground truth system. This is obviously risky and so will inevitably at some point break. Note that new tasks can be tasks to create tasks, e.g. “deal with X” can be quickly jot down, and then later refined.
Agreed with the principle, disagree with the discussion. Those options are all counter to your stated goal. Ideally, just have the tasks flag itself as complete and you move on and never think about it again don’t make using the tool a task. Ideally unless you flag the task incomplete it should assume you did it on time, and the only time you need to manually mark something complete is if you finish early. Minimal effort, don’t make the tool the task
All tasks should not be visible all of the time. In the high task regime especially, it is costly to be reminded of everything you ever need to do every time you want to add a new task, check off a task as complete, or figure out what to do next. Trust me, you do not want to be stressing or distracted by your tax return that is not due for another 8 months. Attention is your scarcest resource, so most tasks should be hidden from you until you need or choose to see them. I think of this as a form of self-context engineering, and also apply this principle in many other settings. For instance, when working on some task X, I try to close all irrelevant apps to stay maximally focussed. Using dedicated desktop apps as opposed to browser apps helps a bunch with this. For more tips, tricks and ideas for keeping focussed, see here.
One task visible at a time, just the next one that needs doing. Have the ability to look ahead, sure. But as you said attention is valuable, distraction is everywhere. It should not exist in an organization tool. The tool should not be a task.
It should be easy to surface relevant tasks when needed. I currently have 118 tasks in my to-do system. Besides it being costly be reminded of all of them every time as discussed above, it is also intractable to sift through that entire list constantly when planning out my day. In practice, this therefore means you should utilise powerful tagging and filtering tools and views. You might want a tag for each topic of task (e.g. work, finances, etc), for tasks you can do at home, tasks which are due soon, etc. Context switching is hard, so being able to aggregate and batch execution of similar tasks is useful. (Fuzzy) search is sometimes helpful, but I rarely find myself using it as I generally feel very free to forget tasks in my to-do system completely from my brain. This is only possible because after years of use, I now fully trust the protocol to surface tasks when relevant.^[4]
This is an unnecessary duplicate of the above. Only one task at a time. They are only “needed” when it is their time to be done. Otherwise you are talking “wanted” not “needed” and that gets right back into the tool maintenance becoming the distraction or becoming a task.
Tasks should optionally have date, and should be able to be viewed by date. It’s pretty helpful to know which tasks have deadlines, and if any of those are soon. My to-do software allows me to set both “dates” for when I plan to do a task, and “deadlines” which are hard cut offs by which some task must be completed. Often tasks are blocked until some event out of your control occurs, and it is useful to mark a task as not relevant until after some date has passed. It can also be helpful to move tasks into “due today”, when planning a daily to-do list.
Again, too much the tool is the distraction. Think simpler. Yes add with a date and priority. Then become invisible until it is time to be done. The tool must not become a task itself. Why are you making a “today” list? That’s a waste of time and attention. If you entered the tasks with dates and priorities, then whatever task is in front of you is all you need and it will be the correct task. Don’t make the tool the task.
Tasks should be able to be prioritised as more or less important. On the one extreme, some tasks absolutely must happen. On the other, some might never be worth doing, but are still useful to jot down somewhere. It is useful to have some sort of view that filters out low priority tasks or sorts tasks by priority. Some people like using @someday and @maybe tags for this sort of thing. It can also be useful to mark tasks as “vaguely urgent” but without a specific date, using e.g. @soon.
At the time of creation/addition and very rarely things should get edited. But filtering and viewing etc are all distractions. One task at a time, touch it as close to once as possible.
Some tasks recur. Many tasks recur on a regular cadence. Paying bills, cleaning, watering your plants, etc. You should not have to make a new task for each instance of a recurring task, as this is high friction. The feature to recur a task is very helpful.
Agreed.
Regular reviews are important. Tasks go stale. It is important to regularly review everything in your to-do system and delete things that have become irrelevant, and to surface things that have changed priority. Perhaps schedule a recurring monthly task to do so, to batch this work.
Complete disagreement. Touch tasks as close to once as possible. The priority tool should suck us as little time as possible. This whole regular review principal is why I can’t find a functional system. They all take up time I want to be spending completing tasks. The tool must never become a task itself.
It should interface nicely with your calendar. A calendar event in some sense is a task, but it’s a special kind of task that does not need to be checked off as complete; it completes by default at the end of the allotted time. This distinction makes using a calendar as a to-do system problematic (as you risk forgetting to do a task in the allotted window, and then not being reminded of it again in future), and using a to-do system as a calendar annoying (as you need to check off things that should just get checked off by default). Calendar integration is useful though, as when planning your tasks for the day, it is useful to see what calendar events you have. If you time box tasks it can also be useful to see those on your calendar.
Agreed except for the “when planning your tasks for the day”. Again my broken record says touch them once (when you enter them).
You seem to enjoy the planning and to get satisfaction from task completion. I think you might want to add those as principles or goals of your ideal system. I think far more people agree with you than me, and it might help them realize your system is valuable to them.
You have a very cool system and I enjoyed reading about it. Thanks for sharing. Someday I will find the Marie Kondo or Basho of systems, that’s the one for me.

Gavin Runeblade 6 Aug 2025 17:06 UTC
1 point
0
in reply to: Karl Krueger’s comment on: AI #127: Continued Claude Code Complications
The effects are too big for that to be the answer. Look at any graph, the hockey stick is extreme. The top tiny percent get vastly too much traffic for it to be fandom. And it is financially against the best interest of the company involved.

Gavin Runeblade 1 Aug 2025 21:00 UTC
4 points
0
on: AI #127: Continued Claude Code Complications
First off, even if Netflix were entirely correct that they have strong incentive to maximize hours watched under no-ads plans, the incentive is way, way higher under with-ads plans, and the same goes for ChatGPT under such a plan
Someone please help me with a related incentive question: why are all these platforms incentivized to have us all view the same thing? From the ad revenue model as I understand them, it appears to me that they should be incentivized to have us watch different things, but that’s not what I observe, not what the graphs on profit per content creator appears to show objectively.
As I understand the existing incentive: big creators get better payment terms than small creators. If I sign with Spotify, they aren’t giving me as much money as a percentage of ad revenue as they give Joe Rogan. Same with Twitch, Instagram, YouTube, Netflix, Amazon, Hulu, Substack, etc the bigger creator (whether a company or an individual) gets more money per advertisement than the small creators. Therefore, if 1 million people consume 1 hour of content, the platform has higher profit if those 1 million people each consume 1 million different videos/shorts/podcasts/articles. Yet, all platforms will use algorithms that drive those 1 million people towards the biggest, most expensive content provider on their channel.
It cannot plausibly be because the biggest creator is universally the best fit for every consumer. Presumably, there is over 50% likelihood that for each individual person some lesser popular content creator is a better fit for their specific taste. Yet all the providers steer people towards whatever is popular.
I assume this is reducing their profit per hour of content consumed, and providing worse experiences for customers, so where am I wrong or if I am not wrong, why do they do it?
Netflix, I get steering viewers to their original content because no ads and they want to show their original content is popular. But why do all of them do it?
I have to be missing something. I expect, if they were actually steering consumers towards perfect recommendations, I assume that the graphs of income for content creators would look markedly less extreme. Yes top creators would be better plugged into the moment than average, but the current situation seems too extreme.
Thanks for any insight!

Gavin Runeblade 29 Jul 2025 18:06 UTC
1 point
0
on: Informational Black Holes: The Physical Resolution to the Fermi Paradox
Counterpoint: the optimal signal to noise ratio does not require preservation of noise. Most information is noise. Civilization may optimize for forgetting much but not all, with a stable amount of information that never collapses.

Gavin Runeblade 24 Jul 2025 18:18 UTC
2 points
0
in reply to: dr_s’s comment on: What are some good examples of myths that encapsulates genuine, nontrivial wisdom?
Once again, I will point to the source: Joseph Heinrich, the secret of our Success. Yes you are correct it is messy. It is debated.

Gavin Runeblade 23 Jul 2025 19:00 UTC
4 points
2
in reply to: SpectrumDT’s comment on: What are some good examples of myths that encapsulates genuine, nontrivial wisdom?
It has a lot more, that is just the headline. There is behavioral norms, expectations, what punishments look like, do they work or not, etc all of that is telling you where the Overton window was for the people who shared that version of the story. It has morals, warnings, behavior to avoid and behavior to emulate. It shows class function and has the very subversive message that movement between levels of class is possible, even if the upper class don’t want that to be true. It also has the subversive message that the nobility isn’t particularly noble (Cinderell’s family), but some really do live up to the ideals (the Prince). So don’t judge people by their status, look closer.
This is a minefield of a topic however, as people get very touchy over things. Already in this thread there are accusations of misogyny in myth, but is that true or do some cultures value different things? Are the Igbo, the Sami, and the Dine woman haters because they tell stories about the power of motherhood? Are those stories holding women back and creating a patriarchy or are they recording and passing on something meaningful about the difference between men and women and what female power looks like? You will get a big argument whichever side you come down on. Same with the yin/yang myth and all it symbolizes.
Another minefield here are the people who cannot tell the difference between representing a thing and supporting/endorsing it. Thus the bowdlerization of myth over time. Cinderella used to have scenes where the sisters carved up their feet to fit into the slipper. Briar Rose/sleeping beauty was not always woken by a kiss but by giving birth 9 months after the Prince visited her. Acteon wasn’t transformed and murdered for gazing upon a goddess but for attempting to “grape” her. There are, however, some people who assume if you include content in a story, that story is endorsing the content, thus the myths were said to be misogynistic and normalize assault of women when to the Greeks it was a warning to men that bad behavior would not go unpunished. The gods were watching. This is still a very fraught topic and people just don’t want to wade into it.
With all that said, it does appear that almost all myth has value Imbedded in it. Some cultural, some scientific, some no longer relevant. But stories survive only because they are valuable.
I think that people often get overly caught up in the fanciful details. Take Cinderella, you are correct that it is a lot of words for that message, but we remember it because of those words. Just the sentence “non-blood relatives can be dangerous to kids” would be forgotten and would be disbelieved. Cinderella has lasted for almost 2,000 years being traced back to the Greek historian Strabo in the first century. Gilgamesh is over 4,000 years old. The oldest known story is over 100,000 years old and we can trace its movements across continents and through many languages.
Humans tell stories, humans remember stories, humans are motivated by stories. Stories have power. True phrases are as quickly forgotten as spoken.
There is a magnificent YouTube channel, creganford that distills the academic learning on ancient stories and myth. Here is a link to one of his videos on how we know what we know about unwritten stories.

Gavin Runeblade 23 Jul 2025 18:21 UTC
2 points
0
in reply to: dr_s’s comment on: What are some good examples of myths that encapsulates genuine, nontrivial wisdom?
The hard way: Generations of death. One family does things differently and starts a superstition about it. They do it and it works, but they don’t know why so they start a superstition. That becomes a myth.
This is also why Africa doesn’t just copy them. There is no respect for the myth and no understanding of the science, because the people who brought it over also disregarded the myth and assumed the plant was safe so didn’t do science until it was too late.
Read the book it is amazing.

Gavin Runeblade 22 Jul 2025 20:57 UTC
7 points
0
on: What are some good examples of myths that encapsulates genuine, nontrivial wisdom?
Joseph Heinrich has several examples along with a proposed theory to explain them in his two most recent books Weirdest People in the World, and The Secret of our Success . Roughly the theory goes that certain things cannot be rediscovered by trial and error quickly enough. For example, many indigenous people have myths that guide how they prepare food, for example casava. The South American stories tell them how to prepare it, based on their myths, and those instructions ensure they clear out the cyanide before eating it. In Africa where it was transplanted the locals have no such myths, and the toxin builds up very slowly over decades. By the time it kills you, you are unlikely to point the blame at the food everyone has been eating apparently safely. As a result the deaths in Africa are Many times higher than those is South America because of differences in preparation. Many stories and myths work like this.
He also describes several that exist to help people get past major cognitive biases. Like using the randomness of divination to avoid falling into a pattern of hunting (recency bias among others) which prey could learn and adapt to. There is always a myth that goes with the practice.
Astral Codex Ten has a good recent article on this.
While the debate is ongoing, the Cinderella effect may be an example. That is whether the percentage chance of abuse by a non blood relative is higher than by a blood relative. Arguments seem to revolve around lower absolute rates by step parents vs higher percentage rates; and issues with data reliability.

Gavin Runeblade 18 Jul 2025 20:29 UTC
15 points
5
on: On METR’s AI Coding RCT
Peter Wildeford: I definitely think the biggest takeaway from this paper is that we likely can’t trust self-reports. This is pretty surprising to me, but is a finding commonly seen in productivity literature
Not just in productivity, this is a finding from so many fields of research that I struggle to find any situation where self report is trustworthy. Yes sometimes it is the best we can do, but that doesn’t remove the weaknesses. With that said, I really appreciate the dual research model showing the self report vs actual situation in this case and hope more projects use it going forward.

Gavin Runeblade

What makes a good to-do system?