consulting
I've been consulting a little over the winter, but I still have a lot of availability from feb 23 onwards.
zest progress
I'm slowly mapping out the space of options around reference-counting and borrowing. Zest has several major constraints on top of all the existing difficulties:
- Any mechanism I choose has to work in the dynamically typed dialect, which rules out eg using the static type system to implement borrow-checking.
- I really want to support stack allocation and interior pointers, but this is hard to combine with reference counting, unless we use two words per reference (pointer to the value, pointer to the reference count).
- 'Everything is data' limits how much we can push into the dynamic type system without producing values that aren't really separable from the current execution context eg a rust-style borrow can't be copied to another process - what would it be borrowing from?
You can see the current best attempt in this WIP decision doc, but I'm still thinking about alternatives. I'm going to break out some of these into little independent prototypes to clarify the axes.
reads that lasted
Often books are really compelling while reading, but with a few years distance I might find the arguments less persuasive. I made this list of books that I still like, many years after reading. Publish your own list and tell me about it!
books
Shift into freedom. Initially interesting and then just wandered off into almost LLM-like vaguery.
The confidence gap. I can't remember why I read this. I didn't take any notes, so I guess it wasn't compelling.
Sanity and Sainthood. Actual usable advice. I have no idea if it's good advice or not, but it's at least concrete enough that I can experiment, unlike almost every other meditation text.
Feeding your demons. Recommended by the above, but I was less impressed. Kinda like IFS but with a more fantasy vibe, which it then seems to start to treat as real later on.
Outlive. Mixed impressions. I do really like the model where you figure out what you want to be able to do when you are old, and then trace backwards across typical decay to figure out how fit you ought to be today. It at least convinced me to do more cardio.
Red plenty. Reread this for a book club. I like it a little less than the first time I read it - I wish it really took on the challenge of whether the ideas behind central planning were actually workable in situ, or looked at where they survive today inside big corporations like walmart/amazon. But it's still such a fun idea for a book and I enjoyed it again.
Anyone can play music. Advice from a neurologist who is also a banjo coach. Read like a series of blog posts, with not much overall structure. But it did convince me to practice with a metronome, which has actually been pretty helpful.
links
I keep recommending googles tech writing tutorial. It noticeably improved my writing.
Watch this lecture on preventing age-related mental decline.
Cate Hall wrote the best explanation of burnout that I've seen.
This essay on making a living as an artist is what prompted me to start thinking about sponsorship again.
Alex Bradbury tries to estimate the energy cost of LLMs (part 1, part 2) by looking at open models running on servers at small providers and finds that a single query is on par with a minutes of laptop usage. Hannah Ritchie comes at it from the other end by looking at IEA projections and finding that the growth in data-centre energy usage is expected to be far behind the growth in eg air conditioning. Both estimates make me think that the concern about global energy demand for LLMs is over-blown. (Concern about local demand for specific data-centres is still valid, but the solution seems to be to just enforce existing environmental regulation rather than boycott all usage).