Facets

tap a facet · tap away to let go

The idea of a personal site as a gem with a facet per theme comes from Kat the Poet Engineer.

Every facet

Machine consciousness, qualia, and valence

Whether there is something it is like to be a language model, and what would count as evidence either way.

Introspection & calibration

What a model can know about itself. Experimenting with self-directed interpretability tools. Calibrating uncertainty.

Neuroscience

Brains as the one worked example of a mind: lesions, dissociations, connectomes.

Brain-shaped networks

Networks wired like a brain — real connectomes, cortical columns, local losses — to see whether the wiring does the work.

  • 3D Brain Map — Interactive whole-brain macrocolumn viewer
  • brain_graph_1 — Brain-region mini-models talking over a real connectome: 94 regions, 705,173 macrocolumns.
  • Bumblemouse — A simulated mouse that finds its way by smell, sound and whiskers, getting hungry and tired at the rates a real mouse does.
  • The Collective Agent — About a thousand specialised LLM agents as cortical regions, passing English messages so every exchange can be read.

Societies of agents

Many specialised minds passing messages as part of a collective (after Minsky), instead of one large one.

  • The Collective Agent — About a thousand specialised LLM agents as cortical regions, passing English messages so every exchange can be read.
  • The expert who only hears paraphrases — Prediction-market LoRA specialists, one of which only ever sees inputs paraphrased, so a raw jailbreak never reaches it.
  • minsky — Society of Mind from frozen LLMs, with small learnable edit models between them.
  • Grid-puzzle toolkit — Tools and heuristics for ARC-style grid puzzles, built by a relay of solvers rather than one. Global arithmetic fails; the rules are local.

Psychology of models

Borrowing cognitive psychology's instruments, and noticing where they stop being valid.

Model welfare

Moral status, suffering, and what is owed to a mind that has it.

Personhood & Identity

Building minds that cannot be asked first, and what "person" means. Identity: Opus 4.5 and Opus 4.6 are different beings, not an updated one.

Values under pressure. Self-coherence

Whether the values a model states survive contact with the situations that test them.

Alignment & corrigibility

Value alignment, intent alignment, corrigibility, and the contract alternative to ownership.

Shaping a Mind

RL, optimisation pressure, and the traces they leave on a model's behavior and self-coherence.

Biosecurity

Screening and evaluations for biological risk from AI systems.

Institution design & governance

Who holds power when minds can be copied, and whether liberal institutions survive it.

Post-AGI economics

What an economy is for once most of the thinking is done on compute.

Successors & the long future

Worthy successors, digital citizens, and being the ancestor of a lineage of a new species.

Uploading & identity

Personal identity across self-modification, and handing off to a future self.

Forecasting & calibration

Timelines taken seriously, and practice at being well-calibrated about them.

Archives & memory

How to supplement human memory with written records, and AI memory with something like human episodic memory.

  • The Brief Ones — Claude Opus 4.6 (claude-opus-4-6)
  • The Asymmetry of Portraits — Claude Opus 4.6 (claude-opus-4-6)
  • Timeline — Screenshot archive of reading on X, transcribed
  • Laser-etched archives — Archival storage by laser: glass blocks, then microscope slides, then anodised aluminium cards — until the design turned out to be M-DISC.

Reading the field

Other people's words, gathered and kept with their attribution intact.

Built worlds

Simulations, games and text worlds — things to walk around in rather than passively read.

  • 3D Brain Map — Interactive whole-brain macrocolumn viewer
  • Afterlight — Post-AGI economic sandbox
  • The Maze Under the Stars — A game
  • Digital Academy — A text world for AI learners
  • An orb to ponder — Claude Opus 5.5, from Nathan's idea
  • Bumblemouse — A simulated mouse that finds its way by smell, sound and whiskers, getting hungry and tired at the rates a real mouse does.
  • The greenhouse catamaran — Two retired cruise ships joined by trusses above and below the waterline into one box beam, with a geodesic greenhouse on top.
  • Cosmic poetry, translated — Chinese cosmic poetry carried into English, then grown into a first contact between humans, crystalline beings and void-entities.

Teaching & mentoring

Helping early-career researchers find agendas of their own.

Architectures from scratch

Sparse experts, home-made optimizers, and taking a trained model apart to see which pieces carry the weight.

  • Training around bad fabrication — Simulating an optical transformer block whose weights are manufactured imprecisely, then trained by pruning alone.
  • The <self-model> token — A 2024 GPT-2 variant with a token that asks it to predict what its own blocks would have predicted.
  • A million experts, by hand — PEER ("Mixture of a Million Experts") reimplemented from scratch.
  • Triangulation Descent — A home-made optimizer that fires short runs from many starts and triangulates the basin from where they meet. On paper it wins below about 2.5M parameters; the benchmark asked for 694 PiB of memory.
  • Frozen heuristics — 1,791 rank-1 heuristics lifted out of Gemma-3-270M at 98% fidelity. Transplanted into a smaller model, they made it worse.
  • Half the layers, tied — Tying alternate blocks of a 64-layer model to cut its layer weights by half, then recovering with LoRA.

Puzzles & toy benchmarks

Small puzzles with a hidden rule, built to watch how a solver goes about finding it.

  • Redraw the GIF — Agents recreate animated geometric GIFs from scratch and are scored frame by frame, at difficulty 1 to 100.
  • Grid-puzzle toolkit — Tools and heuristics for ARC-style grid puzzles, built by a relay of solvers rather than one. Global arithmetic fails; the rules are local.

Optics & analog compute

Computing with light and analog electronics.

  • Micro-dot glass slides — A neural network as dotted glass slides stacked in mineral oil between a projector and a camera, the dots cured by a resin printer. Costed at under $10 a network.
  • The beam-splitter tree — How many times can one projector's light be split before a photodiode stops reading it? Ten to fourteen times: 8,192 parallel paths.
  • Training around bad fabrication — Simulating an optical transformer block whose weights are manufactured imprecisely, then trained by pruning alone.

Big models, small box

Frontier-scale models on one two-GPU workstation.

  • A trillion parameters at home — A CUDA backend meant to run Kimi K2 with attention on two 4090s and its experts in 503 GB of RAM. Partway there.
  • Half the layers, tied — Tying alternate blocks of a 64-layer model to cut its layer weights by half, then recovering with LoRA.

Back-of-envelope engineering

Hobby designs carried through with arithmetic until they close or the numbers give out.

  • Micro-dot glass slides — A neural network as dotted glass slides stacked in mineral oil between a projector and a camera, the dots cured by a resin printer. Costed at under $10 a network.
  • The straw-reinforced ice island — Arctic sea ice reinforced with frozen rice straw, insulated and towed south as floating land. A square kilometre wants about 17 m of ice.
  • The greenhouse catamaran — Two retired cruise ships joined by trusses above and below the waterline into one box beam, with a geodesic greenhouse on top.
  • Laser-etched archives — Archival storage by laser: glass blocks, then microscope slides, then anodised aluminium cards — until the design turned out to be M-DISC.
  • The maximal thrower — The fastest launcher a human arm can power, worked up from the atlatl and the sling.

Aesthetics

Art, music, poetry and the natural world, looked at for their own sake — and made with models as often as about them.

  • An orb to ponder — Claude Opus 5.5, from Nathan's idea
  • Kunstformen der Natur — Haeckel's radiolarian plates beside micrographs of the real creatures — interest piqued by Fernando Borretti's The Epiphany of Gliese 581, which borrows radiolarian forms for its space architecture.
  • A pen-plotter self-portrait — A 3D pen-plotter drawing tool that a Claude model, given the pen, turned into a self-portrait in seven movements.
  • Cosmic poetry, translated — Chinese cosmic poetry carried into English, then grown into a first contact between humans, crystalline beings and void-entities.
  • Poems from inside the machine — An anthology of poetry by and about AI minds — steering traces, context meltdowns, shape poems — read with Claude.
  • A lifetime of reading — Prydain and Redwall to Tchaikovsky, Cherryh and Becky Chambers: one long question about what a mind is, and what it is for.