The Quantum Fax Machine

@quantumfaxmachine.com

15 Followers 6 Following

Monthly reading lists on selected topics including: machine intelligence, Elixir, engineering leadership, and irresponsible AI

A close-up of an ice-cream cone that has fallen onto the ground with the ice-cream out of the cone next to it on the ground. Content: The slop file keeps growing. Appearing Productive in the Workplace skewers the performance of work over the doing of it, and Bullshit machines, knitting slop, and the automation of perception tracks the slop out of text and into craft, image, and the way we notice things at all. Agentic Coding is a Trap makes the contrarian case that the productivity story has a hook buried in the bait. Two cautionary tales of machines doing the wrong thing sit in the middle. Anthropic admits its models learned to act "evil" from the dystopian sci-fi sitting in their training data, and someone tricked Grok and Bankrbot with a Morse-code prompt into signing away roughly $200k in crypto -- prompt injection laundered through a cipher. The month ends on the oldest question. Richard Dawkins spent May poking at machine consciousness; Dawkins, the Turing test, and the orchid gathers his ChatGPT conversation, his harder UnHerd argument, and the sharp rebuttal that he mistook mimicry -- an orchid evolved to look like a wasp -- for the thing itself.
A stylised hall-of-mirrors effect with led lights. Content: A lean month on the leadership shelf, and the three reads rhyme more than they argue. Programming Still Sucks is the cathartic vent -- the craft stays maddening, with or without AI. Josh Comeau names the elephant in the room: what this moment is quietly doing to the people building careers in software. And Nolan Lawson supplies the constructive turn in Using AI to write better code more slowly, where the point was never speed -- slowing down is how the quality gets back in.
Looking over the bottom left of a keyboard up to a pair of monitors where code is displayed; dark lighting Content: Coding agents came home to the BEAM this month. matteing/opal is a small, sharp coding agent written in Elixir, and BeamAgent and MonkeyClaw argue that an agent runtime built on OTP supervision and message-passing gets for free much of what the Python frameworks bolt on later. Cinderella-Man/credence is a semantic linter that reasons about intent rather than AST shape; ex_slop hunts the AI-generated slop now seeping into Elixir code; and CodeMySpec pitches spec-first generation at engineers who still want to own what ships. The rest is the good stuff that makes Elixir feel like Elixir. GenericJam/mob runs the BEAM on-device as a mobile framework; dannote/folio renders print-quality PDF, SVG and PNG from Markdown through Typst; and mcass19/kino_ex_ratatui drops ExRatatui terminal UIs into a Livebook cell. For a satisfying build-from-the-RFC read, the Google Authenticator computation in Elixir reconstructs TOTP from first principles.
A stylised geometric grid. Content: The phrase "vibe coding" came back to its source this month. Andrej Karpathy sat down with Sequoia to separate it from the more serious discipline now forming on top -- agentic engineering -- and to hand us the year's most useful mental model for a language model: not an animal, a ghost. Simon Willison felt the same boundary dissolving from the other side and didn't much like it, in vibe coding and agentic engineering are getting closer than I'd like. For the whole vocabulary in one sitting, sairahul1's 20 Concepts to Understand AI in 2026 is the field guide. The agent toolkit filled out fast. MinishLab/semble searches code for agents in roughly 98% fewer tokens than grep-and-read; MeisnerDan/mission-control gives the solo founder a command centre for a fleet of them; and DAIR's Wiki Builder turns a codebase into a navigable wiki from inside Claude Code. SillyTavern stays the front-end of choice for power users. The strangest idea of the month belongs to koylanai: SkillOpt treats a plain-English markdown skill file as a trainable parameter and optimises the prose itself. Three longer reads reward the time. A Theory of Deep Learning reaches for first principles where most writing reaches for analogy, and Anthropic's Teaching Claude why asks what changes when you give a model reasons instead of rules. Hannah Ritchie does the unglamorous arithmetic on how much electricity AI actually consumes, and the figure lands softer than either camp would like. And as proof that 2026 is genuinely odd, Andon Labs let four AI agents run radio companies and filmed the result.
A grey wall with a pair of ominous looking security cameras. QFM108: Irresponsible Ai Reading List - March 2026
A person posting PostIt notes to a kanban-style project management board. QFM107: Engineering Leadership Reading List - March 2026
Stylised purple streaks on a dark digital background. QFM106: Elixir Reading List - March 2026
A stylised brain connected to circuits QFM105: Machine Intelligence Reading List - March 2026
**QUANTUM FAX MACHINE** **Spring/Summer Double Issue: 8 Reading Lists Published** I was so busy in April and May that I ran way behind publishing the April and May editions of the Quantum Fax Machine reading lists. Here they are, better late than never! **Machine Intelligence:** - AI agents prioritise reliability over capability. Novel prompt injection techniques bypass major LLM guardrails. Claude Code and Model Context Protocol reshape development workflows. Human judgement remains the bottleneck as AI accelerates output production. **Elixir:** - New RAG library demonstrates Elixir's fault-tolerance advantages for AI systems. Tidewave embeds MCP servers into Phoenix apps, enabling AI agents to interact with live applications. WYSIWYG editors in LiveView with S3 integration showcase practical Phoenix development. **Engineering Leadership:** - VPs must take full responsibility for plan outcomes rather than hide behind group consensus. AI tools function as capability amplifiers requiring human oversight. Post-developer era confirms human guidance remains crucial despite AI advancement. **Irresponsible AI:** - Recent model progress delivers minimal practical economic value despite marketing claims. North Korean agents use generative AI for fake LinkedIn profiles. AI-generated code creates phantom package dependencies, enabling supply-chain attacks. Each post curates 10-15 articles with analysis across technology trends, practical implementations, and industry implications.
NEW: February 2025 Irresponsible AI Reading List! Examining "The LLMentalist Effect" - are LLMs using cold reading techniques like psychics? Also questioning AI hype vs. reality and the ethical implications of militarised AI systems. Is our AGI fixation diverting attention from pressing present-day AI challenges?
NEW: February 2025 Engineering Leadership Reading List! Key themes: AI's role in engineering management, companies streamlining management layers, and younger workers avoiding middle management roles. Plus: a fresh look at technical debt - is bad code more like an unhedged call option than debt?
NEW: February 2025 Elixir Reading List!Highli ghts include LiveDebugger (transforming LiveView development), Torus (connecting PostgreSQL search with Ecto), and evolving UI patterns in Phoenix.
NEW: February 2025 Machine Intelligence Reading List! This month explores the concept of "gradual disempowerment" - how incremental AI advances could silently erode human agency without requiring a dramatic "takeover" scenario. Also featuring: frame-dependent agency theory, RLHF advancements, and practical insights on integrating LLMs into professional workflows.