Tess report - 2026.1

January 16 - January 31

extrapolating from current trends, I will write many millions of blog posts by 2027

what I did

I've been home this whole stretch, without visitors or major happenings. I've been splitting my time between coding and miscellaneous other things. I did a lot of work on my (no renamed) C utility library, Cleat. in particular I wrote a serialization/deserialization framework inspired by Rust's serde, a JSON parser and deserialization pattern matching for it. I haven't polished or documented all that yet, so it doesn't go in the shipped category.

what I shipped

two blog posts, which you've probably already seen if you're here:

• the plurality spectrum
• the F in foom stands for frog boiling

what's on my mind

as foreshadowed in my recent writing, I may be entering an LLM mania arc.

I've done a lot of the emotional work to let go of coding, at least the sort of coding that resembles the pastime I've spent half my life doing. I likely wont stop entirely, and if it is somehow useful in the future I will certainly have the skills to go back, but I no longer believe my highest leverage over the world is typing out C or JavaScript on my keyboard. the investments I've made into skills and tech related to this is therefore a near total loss, but the road to Dyson spheres is paved with failed lifes works. that's just part of the game.

with this out of the way I can focus on what's exciting, and that's finding effective and efficient ways to do software engineering that leverage LLMs. I'm quite behind, in the sense I haven't done much of this since the early days of GPT-4, but I have the impression this isn't the sort of field where that matters significantly.

until recently I wanted to cut deep into the problem, attempting to use transformers but partially or fully leapfrog LLMs. I would not be surprised if someone does manage this, but I'm not well positioned to and recent frontier LLMs are too smart to ignore anymore. I do want to challenge the paradigms of modern harnesses and agents. in particular I think I might be able to get better, much cheaper and possibly faster results than Claude Code style systems. this hypotheses may take some time and significant effort/expense to test, but that's my next project.

links

• suspiciously precise floats, or, how I got Claude's real limits - good research and writeup on Anthropic's mysterious pricing
• A Social Filesystem - accessible intro to the philosphy of the AT protocol. I think a lot of these ideas are very correct, and how more things should be built (both that do and do not use AT itself)
• Rotating the Space: On LLMs as a Medium for Thought - about a way to think about the thing that LLMs do. LLMs as little guys with thoughts and feelings and motivations is a very common way to think about them, that is clearly flawed in a lot of ways. we badly need popular alternative frames.
• Eons of Utopia - an ice bath of perspective