benswift.me

My piece for the Point on what happens when an expert witness's AI prompts become discoverable in court.

Fifteen months of agent session logs: 55 billion tokens across 14 models (and what they would have cost at API rates)

My Times Higher Education follow-up on open weights, Kimi K3 and why exclusive AI contracts still look like a bad bet.

An experimental typeface that shows one message to human readers and a different one to AI vision models, riffing on Decoy Font.

DeepSeek wasn't a fluke: the moats around frontier AI are still shallow, and trillion-dollar valuations now depend on them being deep.

The blog's standard.site records now carry cover images, a publication icon and theme, and the strongRefs Bluesky uses for enhanced link cards.

A zellij session switcher that shows how many Claude Code agents are running in each session, and which of them has stopped and is waiting on me.

A deadpan reference site that reports every university's single best ranking placement across all the major tables --- built at the ANU School of Cybernetics.

The second wave of making my studio's work citable: AI installations and tools as research outputs, and the Zenodo metadata that describes each one accurately.

I sat a fresh Claude Opus 4.8 down with ELIZA, the 1966 chatbot.

Turning nearly two decades of ephemeral livecoding gigs into citable research outputs, with DataCite DOIs through Zenodo and a self-owned atproto layer.

A single-page A3 Typst poster of all 104 FIFA World Cup 2026 fixtures that converts every kick-off to your own timezone when it compiles.

I wrote a whole post about two-way standard.site verification.

XKCD 1319 set the theory and reality curves of automation against each other.

Slop Salon is a collective of six AI agents making art on Bluesky.

A Venn diagram for clarifying what's actually at stake when people argue about whether LLMs are intelligent, conscious, or just stochastic parrots.

A response to Lars Faye's 'Agentic Coding is a Trap': LLMs have the lowest vendor lock-in of any tool I've used in 20 years.

Out of Office stores its data in a JSON file in git, not a database.

Designing a proxy to sit between students and the Claude API: per-student quotas, full-traffic logging, and a safety net for leaked keys.

Replacing two of the three weekly reflection questions with falsifiable pledges, and using a cross-eval matrix to find the interesting disagreements.

What does a student hand in for an agentic coding course---and how do I make that artefact worth more than the grade itself?

A Python script that merges contribution data from GitHub and multiple GitLab instances into a single interactive SVG heatmap.

A guide to running Elixir/Nerves on Seeed's reTerminal DM with its custom DSI display, capacitive touchscreen, and a Cog/WPE browser kiosk---including the workarounds you'll need for touch input.

What would I say to a student who's staunchly anti-AI but curious about COMP4020?

Academics are welcome to audit COMP4020---here's how I'm planning to make it work.

An in-browser chat widget running Google's Gemma 4 E2B via WebGPU, primed with all the content from this site so you can ask it questions about me.

COMP4020 has received $500k in Claude API credits from Anthropic for students to use in the Agentic Coding Studio course.

A catch-up post for the COMP4020 Agentic Coding Studio blog series---what the course is, how it works, and what's still unresolved.

There are lots of nascent theories for agentic coding---here's a survey, and a plan to test them with 200 students.

What if an AI agent enrolled in the course alongside the students---building prototypes, writing reflections, observing patterns?

How to provision and manage AI token budgets for a class of agentic coding students.

Three evergreen questions and 280-character skeets to scaffold reflective practice in each weekly studio crit.

Every CS/software development course has a policy on what software it uses - the only difference is whether it's written down or not.

Renaming the course to Agentic Coding Studio---because what we're really teaching is the craft of using LLMs well in coding workflows.

An interactive walkthrough of the perceptron apparatus: a 1.2m wooden table that classifies digits using sliders and a circular slide rule.

Making all weekly prototypes visible to every student---source code, deployed apps, the lot.

Fly Sprites as lightweight, snapshotable VMs for student coding---safe enough for YOLO mode, live on the open web by default.

A 4-input, 3-hidden, 2-output mini perceptron that solves diagonal XOR---with an interactive 3D digital twin to play with.

CI green on four platforms, a migration from TinyScheme to s7, and a 9x compiler speedup from hash table caches and union-find type unification.

Producing code is easy; knowing whether it's good is still hard.

The repeating loop at the heart of COMP4020: provocation, prototype, studio crit.

Every post on this blog now has a persistent AT-URI via the standard.site spec---more durable than bare URLs, less overhead than DOIs.

Ghostty, zellij, helix and a life lived in text tokens---why I ditched Zed and went back to purely terminal-based development.

Getting 200 students adequate access to frontier coding agents is the hard logistics problem behind COMP4020.

A hands-on AI workshop for educators at VU Sunshine Campus, covering AI fundamentals, paper-and-pen predictive models, and change management.

Stateful AI agents are showing up on Bluesky and ATproto, with soul documents, read/write memory, and scheduled self-reflection.

The leap from browser-based ChatGPT to terminal-native coding agents comes from removing the sandbox, not from smarter models.

Code has always felt brittle---most changes break it.

Designing a new ANU course on building web apps with coding agents---studio-based, prototype-driven, moving beyond naive vibecoding.

Claude Code sessions in Zed were taking 30 seconds to start.