My posts so far

1. Fool Heart Joins Halfhaven — Just an announcement post. I counted that post, but not this post, as one of my Halfhaven posts, for reasons.
2. Claude Imagine Bypassing its Security Sandbox Using Screenshot APIs — Trying out Claude Imagine and nudging it to break free from its own security sandbox, finding a creative way to make POST requests to the internet when it’s not supposed to even fetch images.
3. Trial of Claude 4.5 on Chess and Fantastical Chess — Seeing how well the new Claude model does at complex board games (poorly), and speculating on the predictability of AI progress.
4. One Does Not Simply Walk Away from Omelas — A critique of The Ones Who Walk Away from Omelas by Ursula K. Le Guin, in which I suggest the story is too optimistic and the city of Omelas as presented is a logical impossibility, at least given human nature as I know it.
5. 10 Ways to Waste a Decade — Trying my hand at a listicle. Started as notes I had about shift work and how shitty it is for the people doing it, but expanded to cover more of life’s tar pits.

I’m noticing that despite my best efforts, some of my posts are still overly verbose, and I want to keep working on that. I also sometimes write in a way that when I read it, it sounds like I was more confident about the subject that I actually was, so that’s something else I should work on.

Trying to write a blog post every 2-ish days is not easy. Normally, I would work on something until I thought it was finished. LLMs Suck at Deep Thinking Part 3 - Trying to Prove It took me weeks, what with having to coordinate with test subjects and stuff. I won’t be able to write anything like that in this time. I’ll have to focus more on short-and-sweet single-idea posts.

Some highlights from other Halfhaven writers

I certainly haven’t read all the Halfhaven posts. I’ve been struggling enough to find time to write my own. But I’ve been checking out some of them, and I thought it would be fun to call out some of the ones I liked.

• Shake Brains First (niplav) — This post had some interesting ideas, in particular I found the idea of using a Neuralink-style brain-computer interface to replay past mental states to be interesting, and wondered which past brain states I would choose to replay, if the option was available.
• Why Jhānas are Non-Addictive (Lsusr) — I didn’t watch all of Lsusr’s videos (and think the first one might be an infohazard for me), but I watched a couple. I enjoyed this discussion of the endlessly-fascinating concept of jhanas. I feel like his delivery of his script is surprisingly good. I feel better lighting and a visually interesting background could make a big difference. I felt my eyes getting bored even as my mind was entertained throughout.
• No, That’s Not What the Flight Costs (Max Niederman) — I love any post that can show me that my model of the world is wrong in some way, so I found it fascinating that airlines lose money on the actual flights, and their real revenue stream is credit card rewards. One commenter pointed out that this was overstated, and it seems like the actual flights probably aren’t actually loss leaders, but it’s still interesting how unprofitable flights are, and how much revenue comes in from credit card companies. I wouldn’t have guessed.
• Maybe social media algorithms don’t suck (Algon) — This made me realize how passive I am with my social media feeds, in particular Substack. My Substack feed has turned to garbage, and it’s because I treat the like button like it’s radioactive. Only after reading this post once, including it here, and skimming it again, did I realize I still hadn’t liked it. LessWrong doesn’t work the same as other sites, but still, I gotta start liking stuff more.
• Telling the Difference Between Memories & Logical Guesses (Logan Riggs) — Basically, sometimes people fill in the gaps in their memories with guesses, but if you pay close attention to your mind, it’s possible to (usually?) tell when you’re doing this. This feels especially important for me given that I’m writing this from a juror lounge while on jury duty. If you could instruct witnesses in this skill, I imagine you could make eyewitness testimony more reliable.

There’s a lot of activity in the Halfhaven Discord, and new people are still trickling in. The original rules as stated imply it’s too late for you to join, because you’re supposed to post at least once in every 5 day period, but in the Discord it’s been clarified that late joiners are welcome at any time. Though if they join after November 1st, they won’t be able to 100% complete the challenge. So feel free to join!