10M particles, naive all-pairs collision detection: 100 trillion comparisons per frame. Sparse spatial hash grid: O(1) inserts, O(k) neighbor queries, 60,000x memory reduction over dense grids, real-time at 60 fps. Cells are addresses, not allocations: only occupied buckets exist. C++20 header-only, generalized to N dimensions.
https://metafunctor.com/post/2025-11-11-sparse-spatial-hash/
No one is exempt from blind spots, including those who've made careers analyzing them. Moral clarity about systems doesn't guarantee moral clarity about persons. Privilege insulates. The most articulate among us can fail to turn the lens inward. The rigorous analysis we apply to power structures has to apply to ourselves too.
https://metafunctor.com/post/2025-12-15-blind-spots-consistency/
The simplest forms of learning are incomputable. Solomonoff induction is provably optimal and provably uncomputable. The map is perfect; the map is unreadable. Every algorithm we build is an approximation chosen to fit our compute budget. Memory limits force compression; compression implements Occam's razor. The biases aren't bugs, they're maps.
https://metafunctor.com/post/2025-12-19-incomputability-of-simple-learning/
Your reader's words evaporate; your writer's words endure. Forum threads from 2008 are gone; blog posts from 2008 still render. Markdown won the durability war by being boring. Every interaction we add drags back the database, and durability dies. The missing primitive: writes as durable as the reads. Git is that primitive.
https://metafunctor.com/post/2026-04-24-your-blog-will-outlive-your-database/
Your reader's words evaporate while your writer's words endure. Markdown won the durability war by being boring. Every time we add interactivity, we drag back the database, and durability dies. There's a missing primitive: writes that are as durable as the reads. Git is that primitive.
There is a version of the superintelligence story where no one has a breakthrough. No new insight about cognition. Just the same optimization loop we already use, at scale, with better search and longer horizons.
The scary part: the ingredients are already on the shelf. Scale, verifiable objectives, and iterative self-improvement. The mundane path to superintelligence might be the most likely one.
https://metafunctor.com/post/2026-03-15-superintelligence-build-system/
I live in the terminal. When I want to see an image, a plot, or a table, I don't want to leave.
dapple is a unified terminal graphics library: one Canvas API, multiple renderers (Sixel, Kitty, block characters, HTML). Compose images, plots, and text into layouts. The renderer adapts to your terminal's capabilities.
https://metafunctor.com/post/2026-02-15-dapple-composition-layout/
Some things should only happen after you can't do them yourself.
Posthumous is a federated dead man's switch. You check in via phone, browser, or API. If you stop, it triggers pre-configured actions: send emails, publish files, rotate secrets. Self-hosted, no central authority.
Intelligence is a Shape, Not a Scalar
Chollet says the intelligence ball is near-optimal. NFL says the bound is niche-specific. The bottleneck that makes us smart is the same bottleneck that prevents us from grokking what we build.
https://metafunctor.com/post/2026-04-05-intelligence-is-a-shape-not-a-scalar/
#intelligence #cognition #nofreelunch #aialignment #cognitivearchitecture
For $0.48 in API calls I found the exact boundary where MCTS beats brute-force sampling for LLM reasoning.
The answer is surprisingly narrow: tree search only wins when blind sampling hits a ceiling AND the verifier provides a gradient (not just pass/fail). Otherwise best-of-N is cheaper and just as good.
Controlled experiments on constraint satisfaction problems with Claude Haiku.
