Skip to main content



I think the thing I really like about LLM-assisted coding is that it makes context switching easier.

I can be in "words mode" or "code mode", and switching between these takes time and effort. (There are more categories, but they don't change the fundamental point.)

In my job, I have to spend a lot of time in words mode, due to things like hiring and managing. Historically, this has meant that I only really get engineering work done when I have 2+ hour chunks to focus on it. But now I can often get work done in much shorter chunks, while still in words mode.

I would not like to spend all my time in words mode – I enjoy digging into the details – but it's really nice to have the option.



I think laptops should play (uniformly random) typing sounds while you type in your passwords; it's getting to be too easy to analyze the sounds and extract the contents.


Ok my new beliefs about blister prevention, after three weekends of backpacking for eight hours a day, and watching a bunch of YouTube videos:

  • blisters are caused by layers of skin delaminating, not "friction" / heat directly, though typically the delamination is due to static friction on the outer layer of skin, combined with wet skin. Dynamic friction is more likely to cause raw spots / wear straight through the skin.
  • popping them as soon as you find them is basically always the right call unless you plan to be able to avoid the activity that caused them for a week; otherwise they just keep growing as you continue to do the activity
  • blister donuts and moleskin work okay as long as you can keep them in place somehow, but they don't stick well on their own
  • leukotape, very very widely recommended, is worse than useless because the adhesive seeps through the tape and makes your skin stick to your socks even more tightly than it was before.
  • toe socks are pretty good
  • KT tape is very good
  • Vaseline/similar is pretty good as long as you can get it to stay in the right spots


Anyone have luck getting LLMs to write tests without mocks? The tests I want are often just 1-2 lines of code, but anything I get from Claude or Gemini ends up being 20-30 lines long, despite requests for conciseness, saying no mocks are needed, and seeing using real resources is ok.

(I use LLMs a lot for other stuff, but tests seem to be particularly bad.)

in reply to Satvik

I think Sonnet 4 is not *that* much smarter than 3.7 but it should be significantly more steerable and less likely to insert silly mocks
in reply to Daniel Ziegler

Sonnet 4 is tremendously more effective for my use cases, probably because I use a niche programming language (Julia). Two weeks ago I would have said LLMs make me ~10% more productive, now it looks closer to +100%.

And I'm not even committing LLM-generated code – I just use it to iterate and test on designs, then delete the code and implement from scratch manually.



so, I'm going to japan in a few weeks, to do this pilgrimage backpacking trip with a friend.

I'm very out of shape compared to the difficulty of the route (alltrails.com/explore/map/map-… : 4 days; average of 10 miles and 3200ft elevation gain)

So my plan is to train as much as I can between now and then. I've figured out this practice loop, starting from my house, that I'm going to try to work up to doing on both the 17th and 18th: alltrails.com/explore/map/kuma…

It takes a pretty cool path over the hills and down to the reservoir.

Anyone want to join for any of this? As you might guess I expect to be very slow and take lots of breaks (today I did only about half of this loop; 6-ish miles; and it took me like 4 hours)

in reply to Ben Weinstein-Raun

Yo that's really cool. I wish I was in the bay to practice with you.

Consider wearing a backpack on the trek if you aren't.



I tentatively think that rain jackets would work better if they were more like coats of feathers.

Usually rain jackets are either (a) totally waterproof, in which case you sweat and it condenses on the inside of the jacket, or (b) "breathable", in which case they fairly quickly "wet out" and the sweat actually still condenses on the inside.

Feathers work partly like a "breathable" rain jacket, in that they're porous and hydrophobic on the outermost layer, but they're also anisotropic: rain jacket material is the same in all directions, while feathers work kinda like roof shingles: The water rolls off, but there's space for air to pass underneath the feathers. This is fine because rain mostly comes from above, and anyway I bet you can make fairly complicated labyrinths of air passageways such that even splashing water is very unlikely to make it through the jacket.

in reply to Ben Weinstein-Raun

Which side do you think Gore-Tex jackets are on? I find them pretty good at both staying dry and being breathable (though not perfect at either)
in reply to Daniel Ziegler

I think they seem pretty breathable until they get wet, at which point I seem to see condensation on the inside


I think the thing that makes (much) Latin poetry most unrewarding for me is that I'm not at the stage where I can appreciate the rhythm and the meaning at the same time; I have to focus on one or the other, and in isolation neither is so great.
in reply to Daniel Filan

Of course presumably this is a skill issue to be rectified in time.


Went on a photo walk today, mostly around Berkeley, and tested out some new camera settings. Most of the photos didn't turn out as well as I hoped, but I got a few that I like after a little postprocessing in Lightroom.

flickr.com/photos/spiritfox/54…

flickr.com/photos/spiritfox/54…

flickr.com/photos/spiritfox/54…



Okay, what the heck is up with people doing deceptive things to prevent "panic"? What are the actual dangers of "panic?" I was just watching this new Veritasium video about an engineering firm discovering that their already-built Manhattan skyscraper has a 1 to 5% chance of collapsing per year, and deciding that they're not going to tell anyone about it while they spent months fixing it. The head engineer explicitly says in a recorded lecture that this was justified because people "don't need to be terrorized". Is that even plausible?

in reply to Ben Weinstein-Raun

Oof, yeah. It does sound like they had an evacuation plan if a hurricane came, but it still seems pretty indefensible to lie about the situation.
in reply to Ben Weinstein-Raun

I had similar thoughts watching this. It's not right to allow people to make the uninformed choice to enter this building.


C'mon Fabulae Syrae, this is not a very good explanation of what a harundō is.


I tried out making an unboxing video:

I'm pleased with how it turned out, though the subject matter is objectively not very interesting.



in reply to Daniel Filan

Interestingly actual Ovid seems easier to understand than the verses the textbook author wrote himself, but maybe it's because he selected the bits of Ovid to be easy to understand vs wrote a whole story in his own verse.
in reply to Daniel Filan

OK the next part was hard again so I think I'm not just becoming stronger.


ok ok the ghibli thing is very cute

i am annoyed that my local community is such that sharing this adorable picture of me + child would get people mad at me for using AI

in reply to Kevin Gibbons

openAI finally added this capability to the API, which is why I am only now playing with it, because I am much happier paying 30¢ per image I actually generate than paying $20/mo regardless of how much I use

also if you want me to generate images for you I am very happy to do that









Quest complete: Eat the traditional hotdog from each of Denmark, Norway and Sweden




in reply to Ben Weinstein-Raun

It's also very not good to use a much sharper knife than you're used to. That's part of how I cut off 2mm of my thumbtip several years ago.
in reply to Ben Weinstein-Raun

This makes me think of traffic safety studies where accident likelyhood tracks most closely with variance from average traffic speed while accident severity is often a factor of raw speed. My brain translates this to greater event likelyhood when knife behaves unexpectedly due to dullness but more severe injuries when sharp.

What also comes to mind, aside from the excellent points you've all made, is selecting the appropriate tool, sharpness, and technique for the task at hand. For the novice, flexible purpose, moderately sharp blades, with broadly applicable techniques while those with expertise further reduce injury by selecting a razor sharp Japanese blades for cutting tomatoes without needing a stabilizing hand or a garlic presses instead of dicing.



I've been getting really into pocket knives this week, and especially learning about knife steels. The biggest surprise has been that one of the several aspects of Atlas Shrugged that caused me to lose suspension of disbelief, has become much more believable in retrospect:

When I read it I felt like this whole part about the "guy invents a new metal that's just straightforwardly better than existing metals, and names it after himself" was just too farfetched.

But it turns out that actually this is just a thing that can happen. This guy Larrin Thomas basically straight-up did this with a knife steel alloy in 2021. The alloy is notably better than others for the purposes of pocket knives in almost every respect. Like, in any single dimension there are steels that do better, but this alloy is like the Hawaii of the knife steel Pareto frontier. He didn't name it after himself but I think he could have called it "Larrin Metal" if he had wanted to. He actually called it "Magnacut".

in reply to Ben Weinstein-Raun

I guess my intuition is that alloys would be "smooth" in their properties. You add more chromium and certain properties increase. Maybe they stop increasing or start reversing after a while, but it's not hard to find the optional points for each property over time.

With that intuition it seems surprising that it's hard to find new alloys that haven't already been found.

in reply to JP Addison



Have been wearing minimalist/"barefoot" sandals for the last couple days, and it feels somehow unhinged to say this, but I think they're making me feel noticeably happier?

Like, it reminds me of the thing where anosmia is linked to depression. It's like I regained a nontrivial part of my sense of "what's going on in the world".





AXRP Jason Gross


How do we figure out whether interpretability is doing its job? One way is to see if it helps us prove things about models that we care about knowing. In this episode, I speak with Jason Gross about his agenda to benchmark interpretability in this way, and his exploration of the intersection of proofs and modern machine learning.

Transcript
YouTube



Finally actually followed through on my long-standing intention to donate blood! Went pretty smoothly, and I'm ~500mL lighter. Next month hopefully I'll try to donate plasma.
in reply to Ben Weinstein-Raun

Proud of ya! Please make sure you understand the whole plasma donation process before you go forward with that. As I have said elsewhere, my experience with plasma donation was not good . . . so of course, I am concerned for you.


On the way home from the coffee shop I encountered a bright blue bird as I walked past a bush, about half a foot from me, which stared at me for a second, quietly squawked, and flew away.

Then soon afterward, in a different bush, I encountered an ooze (pictured, made of bubbly foam) trying to stealth.

I'm concerned that someone has changed the genre of my RPG, in a worrying direction. Will be avoiding taverns.

in reply to Ben Weinstein-Raun

Indigo Bunting? Western Bluebird? Scrub Jay? (Of course I don't understand the humor . . . but I bet if I did I would laugh!)




things it's the anniversary of this weekend:
- covid lockdowns
- the final alphago match
- the end of HPMOR
- the killing of Julius Caesar


don't like this

in reply to kip

Me too! Sounds like a really upsetting issue, and I hope it turns out to be easily/clearly resolvable in spite of initial indications.


Advertising, and especially targeted advertising, is widely hated. Something pretty interesting to me: insofar as I'm a rational agent, the amount a given advertiser should pay to show me their ad is positively correlated with how much I want to see that ad. On paper this sounds like an amazing situation, with positive sum trades all around.

But it's easy to observe that most ads are annoying and bad, and that people hate them. wtf is up with this? I don't have time to think about it today, but maybe someone here already knows. @Jeff Kaufman maybe? Or @Daniel Filan ?

in reply to Ben Weinstein-Raun

I would be pretty happy if ads were better. I regularly come across e.g. toddler products that I want to buy, and am very willing to spend money to save time/stress. But these things are almost never advertised to me.

I see *lots* of ads for things I already have, and lots of ads for things that are appropriate for parents of much younger or much older children. I can't actually think of the last time I bought something from an ad, which is shocking considering that I'm in a bunch of baby-related Facebook groups and often get products that people recommend there.

in reply to Ben Weinstein-Raun

Advertisers understand that humans are very manipulatable, and are very down to use dark-arts manipulation tactics. I don't suspect the correlation of values to actually be very high at all. (Intentionally annoying ads can be very effective.)

I hate having to be constantly on guard from all these attempts to hijack my attention and influence my beliefs/desires. I opt out of targeted marketing whenever I can because I don't want advertising systems to have *even stronger* memetic hooks to grab me with.



Just updated my git auto-wip-branch tool to use GPT-4o instead of 3.5. So much has changed in the 2 years since I made it! But the basic idea is still pretty great, at least for my workflow: github.com/benwr/gwipt


ugh, wtf is it with me and doctors canceling my appointments?

Like, granted, if I were a doctor I would want to cancel benwr's appointments, but how do they know?



Vitrification is an ancient and successful process for ensuring that one's brain makes it thousands of years into the future.


I wrote a plugin to sync flashcards from Logseq (open-source notes app, like a local Roam Research or a list-shaped Obsidian) to Mochi (closed-source spaced repetition app, like a remote / pretty Anki). Works quite well / has a lot of features for a two-weekend side project, though it took more effort than I hoped it would.

github.com/benwr/logseq-mochi-…



My mood tracking app says February was my worst month since I started keeping track in September. Makes sense: surgery recovery, plus GI issues, resulting in more isolation than usual. Hopefully things will improve as the days get longer.


Still trying to work out how to use superstimulus




Etymology joke


First you need to know about the fish with the scientific name of _Boops boops_, which is real. This means ‘cow face cow face’ and is not a joke. If you like you can pronounce it ‘bow-ops’ to emphasise the etymology, as in ‘coöperation’. The ‘ops’ part is the bit that means ‘face’ in Ancient Greek.

The joke is this: that the etymology of _oops_ is ‘egg face’ — from the phrase “there is egg on my face” — and is pronounced oöps. (The Ancient Greek word for egg is ‘oon’ which gives us ‘oomancy’, ‘divination by eggs’.)



in reply to kip

(this feels a little different than the tone I'd use if I wrote this post *for* this platform. I wrote it for Facebook and decided I endorse cross-posting without thinking very hard about the specific platform)


Man, given that LLMs are "dream machines" / "all they can really do is hallucinate", it's wild how much they correctly remember.

Like, Claude 3.7 correctly knows a lot about the API used to write Logseq plugins. Logseq isn't exactly obscure, but it is definitely pretty niche, and the API is based on a relatively obscure database query language and a schema designed specifically for the app.

in reply to Ben Weinstein-Raun

I think they get worse about hallucinating when you ask them for something which doesn't exist.
This entry was edited (1 year ago)
⇧