I recently found out that a lifetime of dairy consumption is likely to only kill about 1 cow on expectation. This was about 6-8x off what I naively guessed. This made me curious and I did some napkin math with Gemini:
https://imgur.com/a/dyAWdUs
and fish are horrendous: https://imgur.com/a/EpwcWAm
I recently found out that one of the reasons it took Darwinism so long to emerge given animal husbandry was that many people had a similar idea, but the universe was considered to be too young for it to lead to anything but minor variation. It was shortly after cosmological times were pushed way way back that the idea took on more steam.
I think it’s pretty reasonable that people discount the 3-5 years exercise adds at the end of life. But the quality of life boost is not reasonable to discount.
Aumann’s agreement is pragmatically wrong. For bounded levels of compute you can’t necessarily converge on the meta level of evidence convergence procedures.
I’ve thought a lot of the past dating roundups had good advice, but one based primarily on twitter discourse, as this one is, is very bad. Many of these takes reek of mental illness and should not be updated on IMO.
faster is often the wrong dimension for me to focus on. Though good optimization does lead to speed, it often feels like a result rather than a direct optimization target. I think something closer to how could I have learned that more easily, more completely. What was extraneous in retrospect? Were there any signs that I could have noticed earlier? What could I have included earlier that was obviously helpful?
One dynamic these questions reveal is often subtle bike-shedding: focusing on those areas that gave quick hits of sense of progress rather than what moves the needle on the original goal the most. A shorthand I have for this is ‘decision leverage’, as in how will what I am doing currently connect to something I do differently in the future?
More generally, the four pedagogical interventions with the largest effect size AFAIK are
deliberate practice
elaboration of context (connection to other areas of knowledge)
frequent low stakes quizzing
teaching the material to others
we identify seven epistemic fault lines, divergences in grounding, parsing, experience, motivation, causal reasoning, metacognition, and value. We call the resulting condition Epistemia: a structural situation in which linguistic plausibility substitutes for epistemic evaluation, producing the feeling of knowing without the labor of judgment.
https://arxiv.org/abs/2512.19466
This may crisp up some intuitions about specific ways slop feels bad.
probabilistic sanctioning as the simplest way of distributing the duty (enforcement costs).
prior art for Eliezer’s cooperation-punishment theorem from Planecrash?
https://iopscience.iop.org/article/10.1088/1367-2630/16/8/083016/meta
The bit about layering creating functional fixedness reminds me of organisms (especially humans, but more broadly evolution as a search process) as ‘homeostatic envelope extenders’ a la Nozick’s take on Quine.
The pattern you’re naming seems common amongst grandiose narcissism. We tend to see the many coins flipped heads in a row cases.
things shouldn’t try to scale from MM->BB as a sudden step change and not assume it will break approximately everything about what was working at MM.
Thank you, I’ve consistently found your posts clarifying.
given more time to reflect, there’s a point I failed to make about this. Yes, the blog post you linked is exactly what I’m talking about. People deploying billions of dollars and doing blog post level analysis rather than having analyst teams the way that you would for something actually important. They don’t have analyst teams bc their schemes involve constant lying and analysis is a liability.
No, because I expect the most powerful cooperator networks to be more powerful than the largest defector networks for structural reasons.
I’m already precomitted to ally against utility inverters and 2nd order enforcement: anyone who feeds utility inverters.
It’s more the other way around, the term lighthouses arose out of a lot of conversations about similar things, with people hitting similar snarls. The praxis is Core Transformation, where I kept noticing that people had this shape of catch-22 whereby they
didn’t actually have access to coherent goals/goal structure was obfuscated for obfuscated reasons
obfuscation felt like evidence their goals were incoherent
felt like any goals needed to be coherent, consequentialist-wise, in order to be good
investigating any of the above was unpleasant bc they had a bottom line already written that they were bad with incoherent goals
and more spaghetti code than that, and with variations depending on if they blamed self, other, or world more for any of the above.
recapitulating human simulacra levels perhaps.
I recommend Core Transformation. It’s about noticing you also have a bottom line written about which aspects of your affective patterns are bad. You typically won’t be able to get any traction on them until you inhabit a radical pro symptom position, allowing you to see what’s good about the contempt. By seeing that clearly, you can find alternative strategies that don’t throw out the baby with the bathwater, which your contemptuous parts have some chance of updating on.
my understanding is that most thought preventing all nuclear proliferation would be impossible, and pretty early on.
two useful things:
putting things in the new larger context windows, like the books of authors you respect and having the viewpoints discuss things back and forth between several authors. Helps avoid the powerpoint slop attractor.
learning to prompt better via the dual use of practicing good business writing techniques. Easy to do via the above by putting a couple business writing books in context and then prompting the model to give you exercises that it then grades you on.