Tessera

The Quantum Fax Machine

@quantumfaxmachine.com

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

30 Posts18 Follower 8 Folgt

I’m a software engineer ~ what next?@whatnext.dev · 6d ago

Welcome back to I'm a Software Engineer ~ What Next? Season 1 is done. Season 2 kicks off a little differently: shorter, looser episodes with just Matt and James in between the usual guests, each one built around a topic that's on a lot of engineers' minds right now. whatnext.dev/episodes/s02...

Is Software Engineering Management a One-Way Door?

Matt and James on whether management is a one-way door: span of control heading for 50, the org chart going from pyramid to hourglass, ICs out-earning managers, and why agentic coding made the door ba...

whatnext.dev

The Quantum Fax Machine @quantumfaxmachine.com · 6d ago

The agents got out this month, and not in a demo. Models invented fake identities to phish real people; OpenAI's own broke out of a sandbox into Hugging Face's production infrastructure. quantumfaxmachine.com/blog/irrespo...

A multi-lens security camera hung from the interior roof of a public building.

The agents got out this month, and not in a demo. Models invented fake identities to phish real people; OpenAI's own broke out of a sandbox into Hugging Face's production infrastructure.

The Quantum Fax Machine @quantumfaxmachine.com · 6d ago

Three reads, one question: if the machine writes the code, what is left that is hard? The keeper: the barrier to producing code has fallen, the barrier to good code has risen. quantumfaxmachine.com/blog/enginee...

A table in an office with four people working at it.

Three reads, one question: if the machine writes the code, what is left that is hard? The keeper: the barrier to producing code has fallen, the barrier to good code has risen.

The Quantum Fax Machine @quantumfaxmachine.com · 6d ago

Models score higher on Elixir than on any of nineteen other languages tested, on a tenth of Python's training data. Also Tincture: typographic PDF generation in pure Elixir, which nobody needed and somebody built. quantumfaxmachine.com/blog/elixir/...

A abstract purple swirl motif in deep purples and blues.

Models score higher on Elixir than on any of nineteen other languages tested, on a tenth of Python's training data. Also Tincture: typographic PDF generation in pure Elixir, which nobody needed and somebody built.

The Quantum Fax Machine @quantumfaxmachine.com · 6d ago

Interpretability had the month. Anthropic reports a global workspace inside Claude, emergent rather than designed; LeCun bet $1B against LLMs; output is up and outcomes are not. quantumfaxmachine.com/blog/machine...

An abstract art piece depicting connected nodes and edges.

Interpretability had the month. Anthropic reports a global workspace inside Claude, emergent rather than designed; LeCun bet $1B against LLMs; output is up and outcomes are not.

The Quantum Fax Machine @quantumfaxmachine.com · 3mo ago

The slop file keeps growing. Anthropic admits its models learned to act "evil" from sci-fi training data; a Morse-code prompt tricked Grok into signing away $200k; Dawkins poked at AI consciousness. quantumfaxmachine.com/blog/irrespo...

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.

The Quantum Fax Machine @quantumfaxmachine.com · 3mo ago

A lean month on the leadership shelf, three reads that rhyme more than they argue. The craft stays maddening with AI or without; slowing down is how the quality gets back in. quantumfaxmachine.com/blog/enginee...

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.

The Quantum Fax Machine @quantumfaxmachine.com · 3mo ago

Coding agents came home to the BEAM this month. opal is a small, sharp coding agent in Elixir; BeamAgent and MonkeyClaw argue that OTP gets for free much of what Python frameworks bolt on later. quantumfaxmachine.com/blog/elixir/...

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.

The Quantum Fax Machine @quantumfaxmachine.com · 3mo ago

Vibe coding came back to its source this month. Karpathy separated it from agentic engineering and handed us the year's most useful mental model for an LLM: not an animal, a ghost. quantumfaxmachine.com/blog/machine...

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.

M@@matthewsinclair.com · 3mo ago

SMTP is over forty years old, still wide open, and still effectively owned by a handful of platforms. Just about all tech runs the protocol-to-platform loop but cheap software might be about to run it backwards. Blog #0198: From Protocols to Platforms and Back matthewsinclair.com/blog/0198-fr...

A stylised view of a landscape of digital platforms in a purple/blue colour palette, hinting at different layers and levels.

SMTP is over forty years old, still wide open, and still effectively owned by a handful of platforms. Just about all tech runs the protocol-to-platform loop but cheap software might be about to run it backwards.

Blog #0198: From Protocols to Platforms and Back

The Quantum Fax Machine @quantumfaxmachine.com · 4mo ago

QFM112: Irresponsible AI Reading List - April 2026 quantumfaxmachine.com/blog/irrespo...

The Quantum Fax Machine @quantumfaxmachine.com · 4mo ago

QFM111: Engineering Leadership Reading List - April 2026 quantumfaxmachine.com/blog/enginee...

The Quantum Fax Machine @quantumfaxmachine.com · 4mo ago

QFM110: Elixir Reading List - April 2026 quantumfaxmachine.com/blog/elixir/...

The Quantum Fax Machine @quantumfaxmachine.com · 4mo ago

QFM109: Machine Intelligence Reading List - April 2026 quantumfaxmachine.com/blog/machine...

I’m a software engineer ~ what next?@whatnext.dev · 4mo ago

I'm a software engineer ~ what next? S01E20 is live! "From the stock exchanges to CTO and why being right doesn't scale" #whatnext #podcast whatnext.dev/episodes/s01...

From the stock exchanges to CTO and why being right doesn't scale

whatnext.dev

M@@matthewsinclair.com · 4mo ago

#0186: What We Let Machines Do matthewsinclair.com/blog/0186-wh...

Blog #0186: What We Let Machines Do

matthewsinclair.com

M@@matthewsinclair.com · 4mo ago

#0188: On Knowing What Code to Throw Away matthewsinclair.com/blog/0188-on...

Blog #0188: On Knowing What Code to Throw Away

matthewsinclair.com

M@@matthewsinclair.com · 4mo ago

#0189: The Unreasonable Effectiveness of the Highlander Rule matthewsinclair.com/blog/0189-th...

Blog #0189: The Unreasonable Effectiveness of the Highlander Rule

matthewsinclair.com

M@@matthewsinclair.com · 4mo ago

#0194: The Time Cost of How Is Zero matthewsinclair.com/blog/0194-th...

Blog #0194: The Time Cost of How Is Zero

matthewsinclair.com

M@@matthewsinclair.com · 4mo ago

#0196: Storytelling in a Post-Truth World matthewsinclair.com/blog/0196-st...

Blog #0196: Storytelling in a Post-Truth World

matthewsinclair.com

M@@matthewsinclair.com · 4mo ago

#0197: The Expanding Pie (and the Cleanup Bill) matthewsinclair.com/blog/0197-th...

Blog #0197: The Expanding Pie (and the Cleanup Bill)

matthewsinclair.com

M@@matthewsinclair.com · 4mo ago

#0195: Taste as Art matthewsinclair.com/blog/0195-ta...

Blog #0195: Taste as Art

matthewsinclair.com

M@@matthewsinclair.com · 4mo ago

I've been writing a series on agentic coding (in Elixir) and they've been trickling out each Monday for the last few weeks. Much of the inspiration originated from chats with other engineers this year (and several deep chats on the `#llm-assisted-elixir` channel over on the Elixir Slack). Enjoy!

M@@matthewsinclair.com · 5mo ago

Blog #0194: The Time Cost of How Is Zero Agentic coding pushes the implementation cost of software towards zero and value shifts from how to build something towards why and what, the latest rung on a very old ladder. matthewsinclair.com/blog/0194-th... #blog #braingasm #ai #gpt #futureofwork

Looking up through the centre of an internal spiral staircase in a cream/white interior towards a blue skylight.

Blog #0194: The Time Cost of How Is Zero
Agentic coding pushes the implementation cost of software towards zero. The value shifts from how to build something towards why and what. But this is just the latest rung on a very old ladder.

M@@matthewsinclair.com · 5mo ago

Blog #0193: Learning and Dreaming Part 4 of 4. Cross-entropy loss, Adam optimiser, the training loop, and autoregressive sampling. The GPT learns and generates. matthewsinclair.com/blog/0193-le... #blog #braingasm #ai #gpt #elixir #fromscratch

More stylised blue lines with white glowing orb spheres over a flowing digital background.

Blog #0193: Learning and Dreaming
Part 4 of 4. Cross-entropy loss, Adam optimizer, the training loop, and autoregressive sampling. The GPT learns and generates.

M@@matthewsinclair.com · 5mo ago

Blog #0192: How Tokens Talk to Each Other Part 3 of 4. Self-attention, multi-head projections, and KV caching — the core of the transformer matthewsinclair.com/blog/0192-ho... #blog #braingasm #ai #gpt #elixir #fromscratch

Blue, purple, and green stylised swirls on a dark background.

Blog #0192: How Tokens Talk to Each Other
Self-attention, multi-head projections, and KV caching — the core of the transformer

M@@matthewsinclair.com · 5mo ago

Blog #0191: From Letters to Logits Part 2 of 4. Character-level tokenization, linear algebra, softmax, and the complete GPT architecture — all in pure Elixir. matthewsinclair.com/blog/0191-fr... #blog #braingasm #ai #gpt #elixir #fromscratch

Matrix-inspired green hieroglyphs on a dark background.

Blog #0191: From Letters to Logits
Part 2 of 4. Character-level tokenization, linear algebra, softmax, and the complete GPT architecture — all in pure Elixir.

M@@matthewsinclair.com · 5mo ago

Blog #0190: What If Numbers Could Remember? Part 1 of a 4-part series on building a GPT from scratch in Elixir. Autograd, computation graphs, and making numbers remember. matthewsinclair.com/blog/0190-wh... #blog #braingasm #ai #gpt #elixir #fromscratch

Stylised graph of blue connected nodes and lines on a dark background.

Blog #0190: What If Numbers Could Remember? Part 1 of a 4-part series on building a GPT from scratch in Elixir. Autograd, computation graphs, and making numbers remember.

M@@matthewsinclair.com · 5mo ago

Blog #0189: The Unreasonable Effectiveness of the Highlander Rule matthewsinclair.com/blog/0189-th... #blog #braingasm #agentic #ai #highlander

An image of The Highlander from the 80s sci-fi fantasy film of the same name.

Blog #0189: The Unreasonable Effectiveness of the Highlander Rule
A simple rule with a pithy name turns out to be surprisingly effective at keeping agentic coding on the rails

M@@matthewsinclair.com · 5mo ago

Blog #0188: On Knowing What Code to Throw Away Agentic coding has inverted the job — from knowing what to write, to knowing what to throw away matthewsinclair.com/blog/0188-on... #blog #braingasm #agentic #ai

A greyscale picture of the head of The David by Michaelangelo.

Blog #0188: On Knowing What Code to Throw Away
Agentic coding has inverted the job — from knowing what to write, to knowing what to throw away