I’m writing a book about epistemology. It’s about The Problem of the Criterion, why it’s important, and what it has to tell us about how we approach knowing the truth.
I’ve also written a lot about AI safety. Some of the more interesting stuff can be found at the site of my currently-dormant AI safety org, PAISRI.
I constantly find myself needing to give opposing advice because I’ll read something and feel like it leaves out the other side. So, someone says meditation is great, I’m like, woah, there are risks. Someone says meditation sucks or not worth doing, and I extoll its virtues.
Apparently I’m forever cursed to push people back towards the middle way. 😜
On one hand, at every individual step, these things made sense and (I have to admit) they worked, in that they pushed us over some difficult hurdles and actually got us to accomplish what seems to me like it was some useful stuff. But on the other hand, ingredients like these are what the Stockholm Syndrome is made of, and I saw that taking hold in myself and those around me.
In some Buddhist lineages, like Zen, the relationship between student and teacher is meant to be like that between child and parent. However, this is a relationship that normally develops over months and years, and at first you’re treated more like a lost child who’s a guest that might end up staying and getting adopted or might wander on. Many teachers won’t let new students attend sesshins (retreats) both because their practice might not be strong enough to handle it and because the relationship between teacher and student is not yet firmly established.
Personally, I think Zen’s cautious approach is better than throwing people into the deep end. Best I can tell, the risk of psychosis is much higher with Goenka style retreats, although I don’t have hard numbers, only anecdotal evidence and theory that suggests it should be more common.
As much as I want it to be true that progress is stalling, I think Sora and Atlas and the rest are mostly signs that OpenAI is trying to become an everything app via both vertical and horizontal integration. These are just a few of the building blocks they would need, and the business case for targeting an everything app given their valuation seems strong (in that the only way the valuation is defensible is if it seems like they will eat multiple existing business sectors or create new business sectors as big as many existing ones).
I’ve come to the conclusion that none of this would truly help me, and that, one way or another, I’m going to die anyway. What difference does it make whether I die in 60 years or in 10,000? In the end, I’ll still be dead.
The difference is 9,940 years of living! Who knows what you might get up to.
Perhaps it’s a difference of opinion, but the value of life is in the living of it, not in how it ends.
The problems with local positivism seem to me… kinda important philosophically, but less so in practice.
Yes, most of the time they don’t matter, but then sometimes they do! I think in particular the wrongness of logical positivism matters a lot if you’re trying to solve a problem like proving that an AI is aligned with human flourishing because there’s a specific, technical answer you want to guarantee but it requires formalizing a lot of concepts that normally squeak by because all the formal work is being done by humans who share assumptions. But when you need the AI to share those assumptions, things get dicier.
Oh, yes, I forgot about Kevin’s blog! I also forgot about the ToC! I’m unfortunately not the best person to be writing an accurate history, as I tend to forget details that were once quite important. I guess that’s why I promised an oral history, both because it was a talk and because you should treat this as a data point from which one might construct a complete history.
Regardless of what’s happening at Antropic, I totally believe that 90% of code could be written by AI, because 90%+ of my code is written by AI, and specifically by Claude Code.
Of course, what this means exactly has some nuance. Most of the time Claude is acting as something like a more heavily automated keyboard for me. It’s not actually doing that much on its own; it’s doing what I tell it to do under relatively close supervision. So it nominally writes 90%+ of the lines of code, but actually did anything interesting on its own a far smaller amount of the time, maybe 10%, that makes it into production without heavy rewriting.
I like this. I notice you don’t mention religion in this post, but I think one of the things religions do really well is try to provide access to all three of a scene, a clique, and a team at the same time (though I wouldn’t have known to put it this way before reading your post!).
Why I say this:
A “church” (a term I’m using loosely here to mean both the place in which people practice a religion, the people who practice that religion in that place, and the organizational entity which organizes these two and interfaces with the secular legal system) is at its most basic level a team: some people come together to make religion practice happen at a specific place and time.
It’s also a source of cliques. A church if often too large for all its members to be in a single clique, but it creates an environment with shared experiences upon which cliques can form (and formation is often aided by participation in the team aspect of a church).
A church is a scene in that it’s part of the larger goal of the religion. People often first join a church because they want to be part of the scene, and know nothing about the clique or team aspect. They then get to be part of the collective process of figuring out how to practice the religion in the modern world (which may or may not include things like evangelizing).
I’ve previously made a case that rationalists should be more religious, and being able to talk in more detailed terms about what the community benefits offered by religions are is helpful!
Typescript
325k LoC
Pretty familiar, but not like world leading expert familiar
With rare exception we limit PRs to under 800 LoC added, but it’s common to have a stack of 4-5 PRs each with 400-500 LoC added (harder to stay on deletes, but there’s usually plenty depending on the PR, and we don’t place a limit on them in a PR so I don’t think about it too hard)
Not that thorough. Rely a lot on manual testing + type system
Maybe 60% new features, 20% refactoring, and 20% bugs
I think it’s pretty interesting! In my mind we’re a pretty typical B2B SaaS product. So there’s a wide mix of stuff, but most of the work is shaping code to do what users want, not pushing the edges of computer science.
Yes. This is something I frequently try to emphasize when someone is meditation curious but not already committed to doing it. I say that for most people it’s great, but some people have trouble, and if you’re in the category of people who might have trouble (especially people with high risk of schizophrenia), then you should avoid doing it.
Sure. I mostly wrote this post because I wanted to share my process, not because I wanted to defend the headline claim of a 5x productivity improvement, so I skipped over that. I could have written a version of this post where I didn’t make a specific claim about how much productivity gains I got; maybe I should have rather than stating my belief, since that’s the thing people want to object to the most!
There probably is some argument I could make and it would probably produce a more accurate number than my 5x estimate. I’m just not very interested in making it given that I and approxmiately everyone has high confidence that AI is making programmers more productive.
It’s also likely that claude code work better for you because you’re more experienced and can basically tell claude exactly what to do when it’s stuck.
I strongly suspect this is a lot of what makes my workflow work well for me. My problem is rarely figuring out what broadly needs to be done or how I want it done, and mostly just actually making the changes I want, which is far more tedious if I have to do all the typing.
Where in Europe, if you don’t mind me asking? There’s a lot of variation between countries in Europe so I’m curious to have a more specific data point.
Ah, this is really interesting and helps explain why low fragrence would become important in contra dance but not so important elsewhere!
My theory is that, in the US, The Sort really took hold in the early 1980s when the highest marginal tax rates were radically reduced, thus making income a more efficient means of rewarding employees. The places where The Sort has the least hold are those where income taxes serve the place a soft ceiling on upper middle class incomes and cultural and political forces make either emigration to more lucrative states hard or otherwise limit economic mobility.
I’ve heard a theory that cheap labor is also why Japan is so nice. Not that Japan is a low income country, but rather that for complex structural reasons Japanese workers are underutilized, so everyone in low productivity jobs is overqualified, and it makes everything nice.
Or in short, Japan has isolated itself from The Sort.
Claude Code doesn’t work that well if you’re not an experienced programmer. I mean, it works okay, but it has no taste, so it just produces syntactically correct stuff that is only randomly useful by default. It takes active steering to get good code out of it.
Also, I’m rarely trying to solve algorithmic problems, I’m trying to build production software, which is mostly about managing abstractions and plumbing different systems together to produce something useful to customers. All the hard algorithm work happens somewhere else by someone else because, while it’s essential, it adds very little marginal value for our customers. If algorithms to do something don’t exist, the solution is often to just wait for someone to figure it out, then build a feature on top of them. When I do get to work on algorithms, it’s usually solving complex concurrent execution problems, which Claude is okay at helping with but not great on its own. Luckily, I mostly design smartly to rely on systems that manage these details for me so I don’t need to work them out all the time.
For what it’s worth, as I model the path, I’d guess that this was stream entry. That is, you had insight into the non-separateness of self for the first time. It’s expected that it doesn’t stick, as only rarely does that happen (very, very few people seem to experience sudden, persistent awakening), but it gave you a taste of it, and with dedicated practice it can become persistent, but it general requires having a string of insights over months and years that resolve various confusions that stand in the way of persistence.