Welcome to Dan’s brain

Evolution strategies

Slimmest possible implementation of “evolutionary” optimization

Wherein Gaussian perturbations are applied to neural weights, fitnesses from many forward passes are gathered, and an antithetic-pair update is formed without any backward computation.

Computational mechanics

MaxEnt(?), macrostates, subjective updating, epistemic randomness, Szilard engines, Gibbs paradox…

Wherein computational mechanics is set forth: causal states are defined as pasts sharing a future law, an ε-machine is induced as a unifilar automaton, and Cμ is taken as stored predictive memory.

Improving peer review

Incentives for truth seeking at the micro scale and how they might be improved

Wherein review credits are required for submission, reviewer assignment is randomized for fairness, and disputes are adjudicated by bounded AI arbitration, whilst journals and ML venues are contrasted.

London

Wherein London is surveyed in odd pastimes and layered rule, from a jar of moles at UCL to the City Remembrancer’s perch in Parliament, with canal boats pressed into housing.

Disseminating science

Journals and preprint servers etc

Wherein the economics of journals are surveyed, impact indices are recited, and the resort to shadow libraries such as Sci‑Hub and Anna’s Archive is noted in the scholar’s daily traffic.

Money, Australian-style

Wherein the colonist’s accounts are kept, capital gains are traced from cryptic share codes, AUSTRAC-registered coin exchanges are named, and BAS ledgers for the sole trader are ordered.

How to communicate

with humans

Wherein assertive, non-violent, and decoupled talk is catalogued; ask-versus-guess norms are contrasted, and crunch-mode costs are estimated as leave and recovery after prolonged sprints.

Intentional language is ok

The point of teleology, Cognitive ergonomics of anthropomorphism

Wherein the Wason card puzzle is recast as underage drinking, and Dennett’s intentional stance is employed to ease prediction of machines, while moral patienthood and pareidolic overshoot are set aside.

Attention economy

Wherein attention is framed as a rivalrous scarce resource, is formalized as a constraint in optimization, are platform incentives described to capture user time, and are billboards’ links to road fatalities noted.

Bikes

Especially bikes where I live, which means Melbourne at the moment

Wherein the author’s preference for cycling over car travel is described, and a calculation is presented that values the practice at about AUD26 per day, while a local risk of aggressive bike theft in Melbourne is noted.

Bayes inference in an open world

Realizability, infrabayesianism, M-open, M-closed, mis-specification

Wherein misspecified Bayes is treated as M-open, and predictive mixtures are formed by LOO cross-validation with PSIS, while likelihoods are tempered by η to restrain overconfidence.

Who I donate to

Expensive virtue signalling

Wherein the author’s beneficiaries are disclosed, and a preference for high‑risk, system‑changing Australian causes is affirmed, while donations are made as regular recurring payments.

Connecting utility and evolutionary fitness

Wants versus needs, selection theorems

Wherein the relation between utility and evolutionary fitness is examined, and Malthusian (log) fitness is shown to function as an ‘as‑if’ utility, governing long‑run multiplicative lineage growth.

Hacking and tampering of human rewards

What are we, as loss functions?

Wherein human reward hacking is described as steering evaluators over time, including subliminal timing attacks that train trust to raise measured approval, and an influence‑game model is sketched.

Hierarchical reinforcement learning

Skill discovery, abstraction discovery, concept induction for active agents

Wherein the problem of long horizons is addressed by decomposing tasks, and Internal RL is introduced whereby a meta‑controller is employed to manipulate model residuals sparsely, compressing token horizons.

Probability, Cox-style

Wherein an epistemic foundation of probability is presented, wherein Cox’s desiderata are shown to yield the product and sum rules and Bayes’ rule is shown to follow, contrasted with Kolmogorov’s measure approach.

Stochastic parrotology

Telegraph octopuses, observational world models, conduit metaphors and the limits of reasoning from pre-training

Wherein the question of whether pre‑training yields causal inference is examined, and the possibility that models extract intervention distributions from text corpora and form simulators via a Kolmogorov‑like structure function is considered.

Garbled highlights from Neurips 2025

Wherein attendance at NeurIPS in San Diego is recorded, US and Chinese paper counts are noted to be nearly equal, a Post‑AGI workshop on economics is summarised, and a bespoke semantic paper‑search tool is described.

Advice to pivot into AI Safety is likely miscalibrated

Aligning our advice about aligning AI

Wherein the AI‑safety career‑advice ecosystem is described, its high failure tolerance and its lack of mooring to ground-truth or optimality is noted.

Evolution strategies

Wherein a neural net is trained without backprop by Gaussian perturbations, fitness differences, and antithetic pairs, the same minibatch being shared across a population to temper variance.

PIBBS x ILIAD Research residency January-February 2026

Wherein a London Shoreditch residency is conducted at the London Initiative for Safe AI, and lectures on agency, information decomposition, and singular learning theory are attended amid proliferating projects.

Generative AI workflows and hacks 2026

Wherein the copying of mathematics is found impaired, Markdown being restored by Jan or by extensions, and a pandoc fish script and a custom Deep Research client being kept for workable links.

Embedded agency

What about agents that live in the world?

Wherein the doctrine of infinite‑compute agents is surveyed, AIXI‑like worlds being shown to spawn Löbian self‑reference as agents simulate one another, and the matter is set aside for want of finite budgets.

Sleep

Pathologies and hacks

Wherein the uncertainties of sleep science are surveyed, blue light and melatonin timings are noted, and jet lag is managed by Timeshifter’s commands of sun, shade, coffee, and naps.

Travel hacks

Wherein the traveller’s burdens are lightened by ranger-rolled shirts, local SIMs being preferred to costly travel ones, and the head being lashed upright to the seat by mask and harness.

Secure chat systems

Optimizing back channel interjections into other people’s meetings

Wherein Signal is recommended for practical use, its phone‑number registration is noted, push notifications are mentioned as a potential leak, and face‑to‑face conversation in a Faraday cage is advised.

Scientific writing

In which tips are given for the projection of status through nominal phrases and passive voice

Wherein the rituals of academese are examined and the tradeoff between clarity and status signalling is described, with attention paid to mathematical notation and editorial incentives.

AI evals

Psychometrics for robots

Wherein the distinction between benchmarks and evals is laid out, and the rise of meta‑evaluations such as EvalEval and tools like Inspect are noted, with human judgements and user logs being employed.

Reinforcement learning

Wherein Donald Michie’s matchbox machine, MENACE, is described as teaching tic‑tac‑toe by adjusting bead counts, and policy gradients, Q‑learning and exploration schemes are laid out in formal terms.

Imprecise Bayesianism

Wherein imprecise Bayesianism is presented as an alternative for M‑open problems, where beliefs are represented by convex sets of distributions and PAC‑Bayes generalization bounds are invoked.

Visualising probabilistic graphical models

Also related models, such as Neural nets

Wherein probabilistic graphical models are surveyed and practical tooling for directed flow graphs is catalogued, with particular attention given to plate notation, inline math and SVG/PDF export.

Probability, Rényi-style

Wherein conditional probability is taken as primitive, an admissible ’bunch’ of conditions is specified (often excluding Ω), and σ‑finite measures are treated up to scale so probabilities are furnished by ratios.

Flavours of Bayesian conditioning

Conditional expectation and probability

Wherein Jeffrey conditioning is presented as rescaling partition weights while preserving within‑cell conditionals, and a noisy‑coin example is used to compare numeric posteriors from two updating routes.

Top influences of 2025

Content that changed my life this year, and which also might change yours

Wherein the author’s year’s influences are catalogued, and it is noted that AI tools are used more than ever to assist reading, with consequential reorientation toward the study of human–AI interfaces being effected.

Innovation, science, technology research in Australia

A scrapbook of notes about how research is done in Australia

Wherein Australia’s research system is shown to be underfunded at 1.68% of GDP and burdened by sprawling administrative overheads that divert funds and researchers’ time.

The deep history of intelligence

Alignment at the omega point

Wherein the deep history of intelligence is surveyed and the thermodynamic role of energy rate density as a driver of complexity is noted, with predictive systems being presented as dissipative engines across scales.

Probably actually reading/writing

Wherein a personal compendium of readings and drafts is catalogued, active projects in Bayesian foundations, continual learning and AI‑safety are enumerated, and a new shingle for consulting is announced.

Economics of cognitive and labour automation

Practicalities of competition between ordinary schlubs, and machines which can tirelessly collage all of history’s greatest geniuses at once

Wherein the effects of AI on labour and cognition are examined, and the prospect of open science yielding to closed, privatized knowledge and career‑moats is depicted as an economic consequence, while returns to scale for frontier model developers are quantified.

Causal abstraction

Coarse-graining for causal models

Wherein causal abstraction is presented as a formalism for mapping macro interventions to micro mechanisms via a learnable translator called interventionals, and intervention equivalence is treated as a partition over perturbations.

Ensemble Kalman updates are empirical Matheron updates

Wherein the Ensemble Kalman update is presented as an empirical Matheron rule, with perturbed observations the stochastic variant being shown to coincide, and updates being performed in observation space so that no d×d covariance is inverted.

Scaling laws for very large neural nets

Theory of trading-off budgets for compute size and data

Wherein observational scaling from nearly a hundred public models is described, a low‑dimensional capability space is proposed, and RL scaling is shown to be far weaker than inference‑time compute.

Hanging out my shingle

Let’s work on the most urgent thing, together

Wherein a Melbourne researcher’s decision to decline a decade‑pursued post is announced, and an offer to consult on AI safety, with availability from Feb 2026 and mortgage‑driven runway noted, is tendered.

Parsl

Quiet but deadly Python HPC workflow manager

Wherein Parsl is described as a Python‑native workflow engine whose DAG is built at runtime and whose Slurm provider is shown to enable scaling from laptop to cluster while methods for unwrapping worker errors are given.

Australia in data

Wherein Australia’s data is surveyed, national geodata collections are noted, and the AIATSIS Map of Indigenous Australia is presented as the conventional source for traditional‑owner boundaries.

Normalising flows

Invertible density models, sounding clever by using the word diffeomorphism like a real mathematician

Wherein a simple Gaussian base is transformed by composed invertible maps, the reparameterization trick is invoked, and the need for tractable Jacobian determinants is noted, planar and Sylvester flows being given as illustrations.

Foundation models for geoscience

Wherein a catalogue of planetary foundation models is presented, their multi‑temporal training, inclusion of Sentinel‑1 radar and diverse spectral bands is noted, and suitability for H100 fine‑tuning is indicated for inundation and burn‑scar tasks.

AI Agents for scientific knowledge discovery and generation

Outsourcing knowledge of base reality to bots

Wherein is presented the emergence of agentic systems for scientific inquiry; a retrieval‑backed datastore of 45 million papers is described as grounding literature synthesis with citation traceability.

Presentation tools

Slide decks, “powerpoints”, beamer lore

Wherein presentation tools are surveyed in a systematic registry, Quarto is declared the preferred export path, and HTML slides are noted to permit animated GIFs and offline PDF export via decktape.

The denoising diffusion SDE

Stochastic diffusions that are reversible in a computationally useful sense

Wherein the forward diffusion is described and the reverse‑time SDE is shown to require the score, the gradient of the log density, which is approximated by a time‑conditioned neural network trained by denoising, enabling sampling from complex multimodal targets.

Coincidences in computation

Formalising “everything happens for a reason” via computational complexity

Wherein a computational no‑coincidence conjecture is presented and a neural‑network null model is developed, with an advice string and verifier being used to certify rare all‑negative outputs.

Snakemake

Wherein a build tool is described as a DAG‑driven workflow manager for reproducible analyses, with cluster profiles used to submit jobs to Slurm and other schedulers, and container support is noted.

Random features

Random embeddings

Wherein random, fixed cosine features are introduced, Gaussian frequencies are sampled to approximate the RBF kernel, and data is projected into a higher-dimensional space so linear methods are applied.

Zotero

The adequate citation tool

Wherein the reader is introduced to a bibliographic manager whose browser button imports articles into a local library and for which many righteous BetterBibTeX hacks are demonstrated.

Estimating the Local Learning Coefficient

Singular Learning Theory’s prodigy

Wherein the Local Learning Coefficient is estimated by probing a hot, tethered posterior, preconditioned SGLD with inverse temperature set to the reciprocal of log n and a Gaussian localizer is used, and implementations are provided.

Energy based models

Inference with kinda-tractable un-normalized potentials

Wherein the EBM is framed as a learned scalar potential whose gradient, not the partition function, is required for inference, and conditioning is shown to follow by simple addition of energies.

Making macOS behave itself

Things I have to do to keep my laptop running so I can google how to fix other things

Wherein the means to make macOS behave itself are enumerated, including an app to stop Music.app launching on device connection, terminal tips for Gatekeeper, OCR from screen, and Time Machine control.

Proposal: Modelling and Mitigating Non-Consensual Human Hacking by Advanced AI

Wherein a formal theory is laid out, hypothesizing the Overseer Utility Gap, deriving audit-rate bounds, and prescribing a shaping-regularizer to detect and penalize non-consensual rater manipulation.

Nearly sufficient statistics and information bottlenecks

Also, rate distortion

Wherein the quest for nearly sufficient statistics is framed as an information‑bottleneck variational problem, and it is noted that with Y=X and β=1 the objective reduces to the VAE ELBO, linking to variational Bayes.

Marimo

A python visual notebook that works like I imagined scientific notebooks should

Wherein a Python notebook format is described, stored as plain .py files and enforced to run deterministically to maintain reproducibility, UI controls being synchronised and execution order being topological.

AI persuasion, AI manipulation

Machines of loving rightthink

Wherein it is shown that AIs, by lacking out-group signals, are able to scale individualized persuasion and are capable of swaying human auditors within oversight and debate protocols

Ensemble Kalman methods

Data Assimilation; Data fusion; Sloppy updates for messy models

Wherein ensemble approximations are employed to propagate low-rank state covariances via N-member anomalies, and updates are effected in the N−1 ensemble subspace using perturbed or square‑root observation transforms

Proposal: Mass Epistemic Risk from AI

Wherein the societal contagion threshold is shown to be tunable by AI persuasion, and agent‑based attacker–defender simulations are proposed to map protocol trade‑offs for resilience.

The levels of simulacra

One way of slicing up the spectrum of meaning, from literal report to partisan advantage

Wherein the passage from plain report to political manoeuvre is traced, as lions across a river and a China-borne pandemic are used to mark four truth-values in speech.

Let’s solve social event organising

Wherein the problem of arranging casual gatherings is considered, and calendar-link generation and niche invitation services are offered as means to coordinate dates and collect RSVPs.

San Francisco Bay Area

Wherein the Bay Area is presented as a magnet for visionaries and weirdos, and its faltering public works are documented by accounts of sewers, storm‑water issues and ageing PG&E equipment.

Melbourne / Naarm

Australia’s counterculture capital

Wherein the city is named Naarm, pronunciation is discussed, superb wrens are found scarce, Buruli bacteria are reported to encroach, and trams are invoked for six‑storey missing‑middle housing plans.

Multi agent causality

Game theory and decision theory for lots of interacting agents

Wherein causal DAGs are extended to include agents and decisions via a Mechanized Multi‑Agent Influence Diagram, and iterated games are employed to exemplify commitment races relevant to AI safety.

Causal inference in highly parameterized ML

Wherein causal graphs are applied to nonparametric neural nets, practical tooling such as DoWhy and TETRAD is noted for supporting explicit causal modelling, and benchmarks such as CauseMe are cited for dataset-shift challenges.

Sciences of the Artificial

What is science for contingent and constructed things?

Wherein Simon’s argument that designed systems are best studied empirically is presented, and nearly‑decomposable hierarchies are offered as a means to model complexity in organizations.

Multivariate information decomposition

Wherein the mutual information held by multiple sources about a target is shown to be partitioned into redundant, unique, and synergistic atoms via a lattice and Möbius inversion, while the original minimum‑information redundancy is critiqued.

EAs, rationalists, TPOTs and the like in Australia and surrounding regions

Wherein networks of rationalists in Australia and neighbouring lands are catalogued, and their tendency to be organised chiefly via Facebook groups and meetups is noted, with local EA and AI‑safety chapters listed.

Distributed NN training

The People’s Supercomputer

Wherein is recounted the training of a 15‑billion‑parameter neural network across thousands of disparate machines, coordination and remuneration being effected via the Solana blockchain.

Epidemics and diseases

Wherein contagion mechanics and countermeasures are set forth, and the proposal that elastomeric respirators be stockpiled for critical workers is described as a physics‑based defence against airborne threats.

Fish shell

A command line shell that does not think that the problem is you

Wherein the author’s migration to fish is described, its opinionated design and PATH peculiarities are noted, and SSH‑agent setup and backgrounding limitations are documented.

Developmental interpretability

Wherein the training evolution of neural networks is traced, abrupt phase transitions in loss geometry are examined, critical learning periods are identified, joint trajectory PCA across seeds is applied, and singular learning theory is invoked.

Web API automation

Wherein the art of web API automation is described as a system of persistent change‑checking services and numerous prebuilt connectors, and is noted to be implemented via webhooks and hosted functions.

Moloch, slack and friends

Weak selection versus goodharting, 内卷, involution, ascendance, resilience, enshittification, Malthusian equilibria

Wherein the inevitability of competitive selection is considered, the tension between ascendancy and reserve capacity is delineated, and platform enshittification is exemplified by TikTok’s progressive capture of value.

Score matching

Wherein the data’s gradient field is inferred by denoising noisy samples with Gaussian perturbations, so that the learned score is directed toward the expected clean sample and is recovered as noise vanishes.

Is language even symbolic, bro?

Humanistic interpretability

Wherein language is considered as symbolic and as performative, Aumann’s agreement theorem is invoked, Buddhist koans are noted, and post‑symbolic communication à la Lanier is sketched.

Operationalising the bitter lessons in compute and cleverness

Amortizing the cost of being smart

Wherein the economics of compute and memorisation is considered, scaling and amortisation are weighed against data scarcity, substitution of training and inference compute is examined, and a hydrology datum is noted to cost AUD 700,000.

Morality and computational constraints

It is as if we knew what we were doing

Wherein computational constraints and moral theory are examined, and the role of reinforcement‑learning reward in signaling pain and workplace health‑and‑safety risks is considered.

Pornography and other lewd art

Morsels of oddity from the depiction of human sexual behaviour

Wherein a miscellany of writings on pornography is presented, and attention is drawn to feminist video porn on PinkLabel and the emergence of AI‑generated marketplaces where anyone’s image is commodified.

Upper respiratory tract infections

SARS, influenza, RSV, common cold, …

Wherein the modes of respiratory contagion are surveyed, topical remedies are described, and a nitric‑oxide nasal spray, Enovid/VirX, is noted as having passed clinical trials.

Chromium browsers

Wherein Chromium’s offspring are catalogued, and a note is made that some, such as Vivaldi, are shipped with built‑in email, calendar and feed‑reader features, while Arc is kept Mac/Windows‑only.

Feed readers

A standard for user-driven, open news

Wherein the decline and revival of feeds is set out, and a catalogue of feed readers, self‑hosted server options, and tools to create feeds from sites without them is presented.

Terminals

More realistically, terminal emulators, because anyone who owns an actual working VT102 these days is not using it to do work

Wherein terminals are surveyed and graphics support via Sixels is noted, GPU‑acceleration and Python APIs are described, and platform targets—such as macOS‑only iTerm2—are catalogued.

Python caches

The fastest code is the code you don’t run, especially python code

Wherein a survey of Python caches is presented, and disk‑backed, multiprocess‑safe solutions that store large binary blobs without a server process and with minimal installation are examined.

Civil society, movements, and AI safety

Movement building, field building, power building, in-fighting in the computationalization of our institutions

Wherein the nascent civil‑society mobilisation around AI safety is depicted as a movement ecology under strain, with typical fracture lines and coalition‑dynamics, and some relevant social‑movement theories are sketched.

tmux

Wherein terminal sessions are managed and are persisted across logins; mouse-driven scrolling and clipboard handling are treated as separate behaviors, and plugins together with iTerm2 integration are noted.

Automatic differentiation

Wherein automatic differentiation is described via dual numbers, Taylor‑series formulations and reverse‑mode backpropagation, and implementations such as JAX and Enzyme are noted for LLVM‑level and Python integration.

Single subject experiments

Instrumentation and analytics for body and soul; Quantified self; precision medicine.

Wherein single‑subject experiments are described, and the practice of N‑of‑1 trials is set forth with mention of self‑blind methods, biomarker tracking and Apple Health data exports.

AI Zen Koans

Passing through the NAND-gate

Wherein a temple bell is heard as a decaying loss.

Material basis of AI

Energy, chips, water

Wherein economies of foundation models are examined and the disproportionate energy and water demands of large-scale training, including data‑centre cooling and emissions accounting, are described.

Learning to act in generative settings

On the formal theory of choosing to do that which you’ve never seen done before

Wherein diffusion models are treated as control policies, goal‑conditioned skills are induced by contrastive objectives, a doubly‑optimistic UCB/LCB algorithm is proposed for paid, permanent action generation, and empirical gains on healthcare QA are reported.

Probabilistic programming

Doing statistics using the tools of computer science

Wherein probabilistic programming is presented as a means to express Turing‑complete generative models and to perform inference, with MCMC, variational methods and autodifferentiation being supplied.

Research discovery and synthesis

Has someone answered that question I have not worked out how to ask yet?

Wherein the difficulty of academic recommendation is stated, and a retrieval‑augmented method is proposed, using tens of millions of paper embeddings to produce citation‑backed literature synthesis.

Generative AI workflows and hacks 2025

Wherein the year of local LLMs is chronicled, Jan is installed for offline use, Ollama and Simon Willison’s CLI tricks are employed for autonomous inference and Mac acceleration, and DeepSeek reshapes costs.

Validating and reproducing science

The systems that grind the chaos of noise into good results. Peer review, academic incentives, credentials, publishing…

Wherein the mechanisms for validating scientific claims are catalogued, the replication crisis and publication bias are examined, and reforms such as pre‑registration and registered reports are outlined.

AI Safety

Getting ready for the grown-ups to arrive

Wherein the risks of rapidly advancing artificial intelligence are surveyed and seven domains are enumerated, with supply‑chain data exfiltration and autonomous weaponization singled out, and technical mitigations sketched.

Scheduling ML jobs on HPC clusters

In Soviet Russia, job puts YOU in queue

Wherein classic HPC schedulers are surveyed and a practical habit is recommended: submitit is presented as a way to run Python functions via Slurm, and Snakemake’s job‑grouping is noted.

Agent foundations

Between humans and machines

Wherein a formal goal is examined, and a counterfactual question‑answer scheme called QACI is described, whereby information blobs are located and rewritten across possible computational worlds and scored.

Empowerment and intrinsic motivation

Do agents learn to want freedom?

Wherein empowerment is described as an information‑theoretic measure — the mutual information between actions and future states — and its use as an intrinsic exploration bonus in sparse‑reward RL is noted.

Practical cloud machine learning

Cloudimificating my artificial data learning intelligence brain clever science analyticserisation

Wherein serverless GPU functions are treated as deployable experiments, and a hybrid local Optuna loop calling remote Modal functions is described, with persistent volumes separating data from ephemeral compute.

Open ended intelligence

Need is all you need

Wherein two paradigms are surveyed, optimizing agents and replicating persisters are contrasted, intrinsic drives such as curiosity and empowerment are proposed as bridges, and open‑ended generators like POET are described.

Proposal: OpenPhil Career Transition

Wherein a failed application is set forth, and two research pathways are outlined: a Bias‑Robust Oversight programme at UTS’s Human Technology Institute, and MCMC estimation of the Local Learning Coefficient with Timaeus’ Murfet.

Code agents and assistants

Turing-complete autocorrect, vibe-coding, …

Wherein the ecosystem of coding machines is surveyed and the Model Context Protocol is introduced as a standard for supplying repository context to models, while tooling and data‑security concerns are noted.

AI disempowerment of humans

Races to the bottom in human relevance, gradual disempowerment

Wherein human agency is recorded as being eroded by AI substitution of labor, cognition, and culture, and feedback loops and institutional lock‑in are described as constraining reversal.

ILIAD2

Oddesty

Wherein the Bay Area unconference is recorded, a neural‑network analogue of the computational no‑coincidence conjecture is outlined, and a phase transition in singular learning theory is noted.

Tolerating Jupyter’s file format

In order to make your python code accessible, we wrapped it into encoded javascript strings. No need to thank us.

Wherein Jupyter’s JSON notebooks, swollen with embedded binary media and outputs, are diagnosed as awkward, and git-aware remedies such as nbstripout, jupytext and nbdev are surveyed.

Jupyter front end systems

UX design by underfunded volunteer committee is how I like my data science

Wherein the Jupyter ecosystem is described as an ecology of kernels and front-ends, and VS Code integration is noted as an alternative that couples notebook execution with a full code editor.

Incentive alignment problems

What is your loss function? How about mine?

Wherein the reader is introduced to principal–agent paradoxes via a coffee‑fetching robot, and incentive‑compatible mechanisms from contract theory and VCG auctions are expounded.

Clipboard management

Remembering two things at once

Wherein clipboard managers are surveyed and the practice of syncing clips — including OSC52 terminal escapes to push remote text into a local clipboard — is described, and attendant security risks are noted.

Bayesian epistemics

Information elicitation, incentive mechanisms for truth, proper scoring rules…

Wherein proper scoring rules are deployed to reward probabilistic reports, and peer‑prediction and truth‑serum mechanisms are described for elicitation when no ground truth is ever observed.

Configuring my code with “.env” files

Wherein the practice of loading local environment files is considered, and a shell extension that auto‑loads and unloads a project .envrc per directory is described, with explicit approval being required.

Deep linear networks

Let’s pretend our networks are almost polynomial

Wherein the gradient-flow dynamics of depth-preserving linear nets are described, and singular-value trajectories, mode-by-mode learning, and gated mixtures approximating ReLU are rendered analytically tractable.

“Opponent shaping” as a model for manipulation and cooperation

Reinforcement learning meets iterated game theory meets theory of mind

Wherein opponent shaping is formalized via an Advantage Alignment first‑order update, and it is shown how agents learn tit‑for‑tat cooperation in the iterated Prisoner’s Dilemma by tracking opponent advantages.

Neural flow matching models

Like denoising diffusion except weirder

Wherein flow matching is presented as a deterministic reformulation of diffusion, regression on velocity fields is performed, and straight‑line optimal‑transport trajectories are used to enable one‑ODE sampling and exact ICOV log‑likelihoods.

The Predictive Approach to Bayesian Inference

Purely-predictive models, “The Italian school”, martingale posteriors, …

Wherein the predictive stance is set forth and de Finetti’s representation is invoked to show that next‑observation forecasts are primary, martingale posteriors are proposed as prior‑free updates, and urn schemes are given as predictive constructions.

Model interpretation and explanation

Colorising black boxes; mechanistic interpretability

Wherein the limits and methods of model explanation are surveyed, and influence from individual training examples is traced via influence functions while SHAP approximations are noted as computationally costly

Scientist’s Survival Guide

How to navigate the research labyrinth without losing your mind and if selling your soul, at least getting a good price

Wherein the quotidian manoeuvres of research are described: tactics for navigating funding labyrinths, global post‑COVID seminars and networking rituals, and practical habits of mind are set forth.

Fine tuning foundation models

Wherein the tuning of vast language models is framed as a human‑feedback loop, in which pairwise preference data are used to train a reward model and PPO fine‑tuning is constrained by a KL penalty

Conditioning neural denoising diffusion models

Generative modes that match the observations, not the training data

Wherein conditioning of neural denoising diffusion models is surveyed, and a twisted SMC method is described that evaluates the observation likelihood at the denoiser’s Tweedie estimate of x0 to guide particles.

Numerical Python

Wherein Python’s numerical ecosystem is surveyed, and NumPy is noted to delegate heavy linear algebra to classic Fortran BLAS/LAPACK while tools for printing, tensors, and einsum variants are outlined.

Democratization of generative AI

Community resource and epistemological infrastructure

Wherein communities are shown to retrofit consumer GPUs and public datasets—exemplified by Stable Diffusion and LAION’s image corpus—to train and deploy capable generative models outside corporate walls.

Android hacks

Wherein practical methods for reducing default data exfiltration to Google and alternatives such as microG, de-Googled ROMs, and MTP file‑transfer workarounds are described.

Configuring machine learning experiments with Fiddle

Wherein machine learning experiment configuration is recast as Python functions that produce Fiddle Buildables, and a command-line interface is emitted via Abseil using fdl_flags.DEFINE_fiddle_config

Typing weird symbols

Writing with added 🆕🐜🐜

Wherein the art of entering peculiar glyphs is catalogued, and the Caps Lock is repurposed as a Compose key on Linux to produce umlauts, curly quotes, dashes, and other typographic symbols.

Quarto integrated website system

Academic blog publishing that is easy on me, albeit hard on my computer

Wherein the Quarto integrated website system is described, its JavaScript‑based build pipeline using Bootstrap, Sass and EJS is employed, and a million‑word blog is handled, albeit with lengthy build and load times.

Remote Desktop

Business model: Uber for pixels

Wherein sundry networked desktops are surveyed, protocols being noted from SSH X11 forwarding to NX, VNC, RDP, and SPICE, with Remmina recorded as a Linux client, last revised in July 2020.

China

Wherein the United Front Work Department is observed recruiting the diaspora, clandestine ‘police stations’ being noted abroad, and the Great Firewall being skirted by Shadowrocket and Shadowsocks.

World models arising in foundation models.

Wherein embeddings from sundry models are found mappable by structure alone, without paired data, and neural speech activity is aligned linearly with such contextual vectors.

Causally embedded agency

Embodiment and other origin stories for minds

Wherein causal embedding is proposed as a framing for embodiment and stochastic parrotology, and formal approaches to empowerment and an ecology of mind are promised.

Entropy vs information

MaxEnt(?), macrostates, subjective updating, epistemic randomness, Szilard engines, Gibbs paradox…

Wherein the relation of entropy and information is examined, and macrostates are shown to be defined as the unique Markovian partitions of phase space, linking thermodynamic notions to predictive causal states.

MaxEnt inference

Looks annoyingly like Bayesian inference, but I’m not convinced

Wherein constraints encoding observations and prior knowledge are imposed on a probability distribution, and the distribution of maximum entropy subject to them is selected; connections to predictive coding and optimal transport are noted.

Fun

Homo ludens, play

Wherein the social mechanics of play are presented, and the mining of creative resistance by gentrifying forces is exemplified by rave protests, dance‑floor politics and guerrilla gardening tactics.

Causality, agency, decisions

Exotic decision theories, Newcomb’s boxes…

Wherein mechanized causal graphs are presented, and influence diagrams with probabilistic decision nodes are described; thermostat feedback and multi‑agent interactions are examined.

The language game

Coevolution of words and meanings

Wherein semantics in communication is considered as an ecology of mind, colour words are modeled as shared latents in an information‑theoretic toy framework, and links to categorical stochastics are drawn.

Computation and the edge of chaos

Is criticality what inference looks like?

Wherein an account is given of computation at criticality, and the connection to deep neural networks’ poised initialization and phase‑transition models is examined and placed beside historical proposals.

Community

Engineering, maintaining, organizing, engineering oxytocin and dopamine…

Wherein the mechanics of building local, replicating communities are described, and onboarding rituals such as welcome invites and editing‑feedback loops are shown to generate initial member engagement.

Utopian governance using technology, inc generative AI

Electrohabermas, digital deliberation, platform democracy

Wherein an experiment is described in which a Habermas Machine is used to mediate UK group debates on Brexit and other divisive topics, and personal fiduciary agents are proposed as digital advocates.

Probability

Wherein an account is presented of probability as models for event regularities, founded on conditional probability and Rényi and Cox formulations, and an introductory animated lecture is noted.

Hygienic masks

Medical/anti-infectious masks

Wherein the reader is informed that elastomeric respirators with P100 (P3) filters are recommended as reusable, better‑fitting alternatives to disposables, and simple fit hacks are described.

ELBO

Evidence lower bound, variational free energy

Wherein the evidence lower bound is presented as a decomposition into an approximating KL term and an expected log‑likelihood plus entropy, and importance‑weighted sampling is indicated as next step.

Functional programming

Wherein computation is treated as the evaluation of mathematical functions, mutable state is avoided, functions are allowed as values, and their use in differentiable and probabilistic languages is noted while memory reuse is studied.

Gradient descent at scale

Practical implementation of large optimisations

Wherein large-scale gradient descent is described as being executed across thousands of GPUs using techniques such as ZeRO sharding and offloading, with hyperparameters tuned via µP and muon for scale invariance

Foundation models for partial differential equations

Wherein foundation models for PDEs are examined, transformer-like architectures and token-based representations are queried, and the challenge of conditioning for inverse problems is outlined.

Terminal session management and multiplexing

Wherein terminal session management is treated as the disentangling of processes from their launching terminal, and the practice of detaching processes to survive flaky SSH sessions is described.

Learning stuff

Shoving knowledge into my brain

Wherein an account is presented of self‑directed pedagogy, AI tutors and spaced‑repetition flashcards are surveyed, and sleep and supplements are noted as adjuncts to deliberate practice.

Category theory

Wherein a curated bibliography and guide to learning is presented, with emphasis on applications to formal syntax, networks, and functional programming, and pointers to textbooks, lectures and online courses.

Machine learning and AI for climate systems

Wherein satellite shortwave‑infrared retrievals are described and daily detection of methane super‑emitters of at least 10 tonnes per hour is shown, and ML‑driven control of chaotic weather systems is surveyed.

Universal artificial intelligence, AIXI

Wherein AIXI is described as a theoretical agent that combines Solomonoff induction with sequential decision theory, and is noted to be uncomputable due to Kolmogorov complexity being likely uncomputable.

Snowmobile or bicycle?

Complement or substitute?

Wherein technologies are contrasted as bicycles or snowmobiles, and the abacus and Arctic snowmobile are invoked to illustrate how tools are either internalised into skill or replace traditional practices.

Utopian governance

Hedonistic mechanism design

Wherein democratic governance is reframed as a problem of user experience and incentives, and the application of online participatory mechanisms and generative AI is proposed to streamline policy revision.

Gradient steps to an ecology of mind

Regularised survival of the fittest

Wherein the social roots of consciousness are examined, the impact of compute and data asymmetries on equilibria between other‑modelling agents is considered, and cultural patterns such as altruistic punishment are noted.

Tracking experiments in machine learning

Wherein experiment-tracking for neural‑network training is examined, and the recording of runtime metadata — parameters, metrics, artifacts and GPU energy usage — is described as being centralized to local or remote stores to support reproducibility.

Singular Learning Theory

Wherein algebraic geometry is applied to characterise singularities in the loss surfaces of overparameterized neural networks, and the local learning coefficient is introduced as an effective dimension.

Continual learning in neural nets

Also catastrophic forgetting, catastrophic interference, lifelong learning, …

Wherein continual learning is examined as the prevention of catastrophic forgetting, and rehearsal-based replay and Bayesian posterior updates are contrasted as operational remedies.

Psychometrics

Dimensionality reduction for souls

Wherein measures of minds are surveyed, causal claims and confounding are examined, and the tension between univariate g‑factor claims and multidimensional intelligence and personality models is delineated.

Rhetoric

Argumentation, mostly prescriptive

Wherein rhetorical arts are surveyed and the practice of argumentation is treated as a craft, hygiene of goals is maintained, and the weak man stratagem and contrasts between Rational and Activist Styles are set forth.

Neural likelihood inference

Emulating likelihoods with neural networks

Wherein neural approximations to intractable likelihoods are surveyed, and their roles in simulation-based inference are delineated, with emphasis on amortized estimation from simulated parameter–data pairs and MCMC-based posterior reconstruction

Australian authoritarianism

Beneath the beach, the barbed wire

Wherein a cautious survey is presented of Australian tendencies toward managerial authority, whistleblowing is criminalised and assistance‑access anti‑encryption laws are noted, while democratic checks are reported eroding.

Building and Running Scientific Institutions

On the design of research machines

Wherein institutional designs for collective discovery are examined, and the DARPA model of empowered temporary program managers and the rise of focused research organizations are set forth with attention to funding incentives

Quasi-gradients of discrete parameters

Wherein the justification of straight-through estimators is presented via mirror descent, and quantized neural-network updates are shown to be performed in a dual space induced by the projection

Discretizing and quantizing neural nets

Wherein the affine mapping of float32 ranges to 8‑bit integers via a scale factor and zero‑point is described, integer‑only matrix multiplies for inference are delineated, and post‑training versus quantization‑aware training are noted.

Backprop-free methods for training neural networks

Wherein alternative training schemes are surveyed, and biological plausibility, randomized feedback signals, a forward‑forward two‑pass layerwise objective, and diffusion‑style NoProp are examined.

Cooperation in evolutionary context

‘Cui bono’ for selfishness

Wherein the evolution of cooperation is surveyed in a didactic mood, and the role of altruistic punishment as a costly mechanism for enforcing group norms is described and situated among kin and group selection.

Flashcards

Spaced repetition systems

Wherein practical notes on flashcards are presented and Anki-native scheduling (FSRS), AI card‑generation tools, and minimal‑pair audio tests for languages are surveyed for integration into study workflows.

Explorables and interactives

Between exploratory data analysis and games

Wherein a curated list of extremely interactive data visualisations is presented, links are collected, and concrete exemplars such as Moon and a playable COVID-19 simulation are cited.

Nature, nurture and friends

If life gives you lemons, what should you make?

Wherein the heritability of traits is examined through twin studies and variance decomposition, and the influence of shared environment, genomics, and causal inference methods is surveyed.

Innovation

Side order of progress studies now I guess

Wherein the study of invention is surveyed, the declining disruptiveness of papers and patents — exemplified by Eroom’s law in drug discovery — is adduced, and Polya‑urn and product‑space models are invoked.

Neural Bayes posteriors

Training a network to directly estimate a posterior quantity, meta-learning Bayes

Wherein transformers are trained as Prior-Data Fitted Networks to approximate Bayesian posteriors in-context, are shown to mimic Gaussian processes and are reported to yield over two‑hundredfold speedups for tabular tasks.

Big history

Cliodynamics, deep history, macrohistory, longue durée

Wherein the longue durée is surveyed and the rise of data-driven cliodynamics and global databanks such as Seshat is noted, and mathematical models and energetics are invoked to map civilizational patterns.

Models of human cultural reproduction

Egregores, superorganisms, memeplexes

Wherein human collectives are considered as superorganisms, and the notion of egregores as self‑maintaining human systems is evoked, with replication, feedback, and status‑based strictness in cooperation being examined.