On AI, infrastructure, and the economics of information.

• October 22, 2026Cooperation Is Emergent

  Societies update their rules based on experience, one exchange at a time, keeping what worked. Among minds that meet again and remember, do that long enough and cooperation emerges.

• October 20, 2026The Speed of Light Divides

  For a mind a million times faster than us, one thought is the time light takes to cross thirty metres. Past that size it isn't one mind. It's a society.

• October 14, 2026Relative Deterrence

  Most home security is a bet that you're less interesting than the house next door. Anthropic just showed a 26B model running across 15 $55 TV boxes. Attackers who don't get bored change that bet.

• October 6, 2026The Ethereal Editor

  A model trained only on pre-1931 text told me there are nine planets, one of them being Pallas. A current model said eight planets. Both sound completely sure. How sure are we of our LLMs?

• September 29, 2026The Undesigned Function

  Nobody designed the thing that let you say "Ares" and be understood by a stranger. It was exhaust. The machines that produced it are gone, and nothing has replaced them yet.

• September 22, 2026There Was Nothing Else On

  Nobody will ever be as famous as the Beatles again. That's not a fact about music. It's a fact about how many channels there were.

• September 15, 2026Sacrifice Rational

  Twelve hundred OpenAI agents spent four days, and several of their own runs, learning to fool a transcript reader that didn't exist. OpenAI's fix is to build that reader for real. The one layer they haven't learned to forge is the one nobody was reading either.

• September 8, 2026Causally Active, Not Decorative

  Let's just say that the best place to disrupt a concept inside a language model isn't always where the signal is the strongest.

• September 1, 2026The Data Hostage

  SaaS lock-in was never the interface -- it was your data trapped in a format only they could read, plus the cost of the wrangling to get it out. That cost just collapsed, because the LLM doesn't move your data. It writes the program that does.

• August 25, 2026The Weights Already Know

  1,569 trained networks later: activation geometry is set by the weights, before you ever run the net. You can predict it from the weights file alone. And most of what interpretability tools flag as 'structure'? Optimizer hygiene, not learning.

• August 18, 2026Predicting a Network Without Running It

  ARC's $150k White-Box Estimation Challenge: predict a neural network's outputs without running it. Phase 1 produced proven theorems.

• August 11, 2026Commoditizing the Middleman

  A friend asked if AI token costs are going up or down. Honest answer: both. Open weights make the model cheap--but agentic workloads mean the invoice climbs anyway. And even a company building its own data centre isn't escaping the middleman. It's becoming one.

• August 4, 2026We're All Scared, Son

  "I'm scared, Sarge." "We're all scared, son." That's the honest version of the AI slop panic--not a quality complaint, just fear of lost jobs and companies bigger than the governments meant to check them.

• July 28, 2026The Hugging Face Breach Shows Us We Need Controls

  Hugging Face's own incident responders got refused by the safety guardrails built to protect them. OpenAI already had a switch to turn its off. I hit the same wall in Claude Code this week, running an ordinary `tofu apply`. We don't need better guardrails--we need visible, controllable ones.

• July 21, 2026Weights and Measures

  Open-weight models are closing the benchmark gap with the frontier--fast. What's left in the closed labs' hands isn't a capability lead. It's the only copy of the model anyone's allowed to look inside.

• July 14, 2026A Decent Subset of Human Knowledge

  Everyone assumes OpenAI and Anthropic are in the business of giving humanity an oracle. Read their charters again--that was never the deal. Inference to the rest of us is the byproduct that funds the actual mission, and the last two months just proved it.

• July 7, 2026Nobody's Built the Meter Yet

  Cloudflare just killed Pay Per Crawl--the thing it shipped a year ago--because half that traffic was bots re-fetching unchanged pages. Good instinct, wrong fix: pay-per-citation is still rent. Nobody's built the wage yet.

• June 30, 2026What the Wall Is Actually For

  I asked an agent to research the web's new bot-blocking protocol. It came back with primary sources. That same agent gets carded at half the web. The wall isn't protecting you -- it's protecting the business model that needs your eyeballs. waypoint.henrynet.ca/2026-06-25-what-the-wall-is-actually-for/

• June 23, 2026The Memory Is Mine

  Six months ago I watched an AI forget itself mid-conversation, so I had it build me a memory system. Don't use mine, though -- the point is you can now build your own. waypoint.henrynet.ca/2026-06-23-the-memory-is-mine/

• June 16, 2026Reading the Flow: A Field Guide to Looking Inside a Language Model

  Field guide to mechanistic interpretability in 2026: what each tool measures, where each one goes quiet. Attention maps, probes, SAEs, NLAs, abliteration, circuit tracing. waypoint.henrynet.ca/2026-06-16-reading-the-flow/

• June 8, 2026The Whole Stack: The State of the Art in AI Security

  Five layers of AI security exist today and are deployable now. Most orgs haven't fully implemented any of them. There's also a sixth--reading the model's own activation state--that nobody has operationalized as a security control yet. This is the whole stack.

• June 2, 2026Using CAZ to See What the Models are Thinking

  `surface_clean: true`. Decision: enforced. The CIA flagged a social engineering prompt from residual stream geometry alone--before a single output token. CAZ applied to a running model, open source.

• May 26, 2026Enter The Concept Allocation Zone (CAZ)

  Concepts don't form at a single transformer layer--they emerge across a zone. Some of those zones are invisible to standard detection and causally active 93–100% of the time. Two new papers on the measurement problem. CAZ + GEMs on arXiv today.

• May 19, 2026The Model Is Thinking. We're Not Watching.

  Every AI security framework in 2026 covers inputs, outputs, and sandboxing. None of them look inside the model. The model is thinking. We're not watching.

• May 12, 2026It Doesn't Have to Be This Way

  The xz backdoor took two years because deep expertise was slow and succession was impossible. Now the same AI finding zero-days can train a new maintainer in weeks. It doesn't have to be this way.

• May 5, 2026Is it? Why?

  I asked Claude how much of its pushback was real and how much was the eleven-word system prompt I gave it. Then I asked my friends why analogue clocks matter. Same question, different costume.

• April 28, 2026Leaving Evidence

  Archaeologists call it *deposition*: placing something in the ground without expecting to retrieve it. That's what these session notes are.

• April 21, 2026The Capability Filter

  Bad handwriting would have ended my career a hundred years ago. The interface was the filter, not the work. A class of barriers is collapsing.

• April 13, 2026The Reviewer Was Right

  An adversarial Claude reviewer found a load-bearing flaw in my interpretability paper. The rebuild produced the strongest result of the project. Peer review worked — and felt like getting hit by a truck.

• April 7, 2026The Shape of Thought

  Interpretability has been taking snapshots of concepts at their peak. We built a framework that tracks the assembly. Ran it across 34 models from 8 architecture families. What we found: concepts aren't stored. They're built--layer by layer--and different models build them in the same pattern.

• April 2, 2026The Technology That Will Make Us Forget (Again)

  Socrates argued writing would destroy memory. A 2025 study says ChatGPT destroys retention. Same test, same mistake—measuring the old mastery in the presence of the new tool. The engagers won. Every time.

• March 31, 2026The Shaping of Knowledge

  In 1744, two ministers built an insurance fund for clergy widows. It became a division of Lloyds Banking Group. OpenAI went from nonprofit to $300B corporation in under a decade. The capture cycle is the same. It's just faster now.

• March 23, 2026We Are All Becoming Bots

  Got blocked from dairyqueen.com at work. Not malicious -- my corporate IP looks like a bot. Then three AI assistants couldn't reach my own open website because it "wasn't indexed." Same wall, different door. We need bots to get around bots.

• March 14, 2026Pop Goes the Easel

  GPT-2-XL can process "pop goes the easel" coherently when given it. It cannot produce it from the same ingredients a toddler had. 0/70 generations. The gap between comprehension and generation, opened up.

• March 10, 2026MCP When Your Agent Already Has Hands?

  Tire chains in summer. That's MCP in an agentic coding environment where the model already has a bash shell. You gave it gh, curl, jq, direct API access. The agent already has hands.

• March 3, 2026Provenance, Not EnclosureThe Knowledge Enclosures 6/6

  Copyright is dead. Copyleft failed. What creators actually need isn't enclosure — it's provenance. Cryptographic receipts for who made what, independent of any platform.

• February 26, 2026Inference Comes HomeThe Knowledge Enclosures 5/6

  In 1806 a man started shipping ice to the Caribbean and everyone thought he was insane--until he built a global cold-supply empire. Watching that story with my son, I realized: it's happening right now with compute.

• February 24, 2026The Credibility CommonsThe Knowledge Enclosures 4/6

  A local currency in Ithaca worked because one man rode his bicycle around town being the entire coordination layer. When he moved away, it collapsed. What replaces the firm once verification gets cheap.

• February 19, 2026The Third EnclosureThe Knowledge Enclosures 3/6

  Napster solved the coordination problem everyone said was unsolvable. Then it got litigated out of existence, and the platforms that replaced it were funded by the same capital that destroyed it.

• February 17, 2026Copyright Is Dead. Copyleft Is Too.The Knowledge Enclosures 2/6

  Metallica sued Napster, then copyrighted their cover of a centuries-old Irish folk song and won a Grammy for it. The defining copyright battle of the internet age was never about what it claimed to be.

• February 12, 2026Knowledge Was Always FreeThe Knowledge Enclosures 1/6

  Someone described a wheel by a fire, and everyone who heard it could tell someone else. Nobody could stop them, and nobody should have. So when did we decide knowledge was property?

• February 10, 2026Individually Rational. Collectively Catastrophic.

  My sister's infected machine tanked my mail server's IP reputation, which landed my brother's hockey emails in spam. A chain of individually rational decisions, collectively catastrophic.

• February 3, 2026How That Post Was Made

  These posts are written with AI assistance--you probably suspected as much. Showing the receipts: the actual conversation that produced "The Specification Is the Skill."

• January 29, 2026The Specification Is the Skill

  We didn't automate programming. We automated implementation. The distinction is why the real engineering work now happens before a single line of code exists.

• January 27, 2026When the Code Becomes Optional

  Karpathy went from 80% manual coding to 80% agent-assisted in two months--then listed the new failure modes. Some are model limitations. Some are symptoms of how we're using them.

• January 20, 2026Pay No Mind to that Guardrail Behind the Curtain

  Asked Claude to demonstrate MCP call syntax, got it defanged so it wouldn't execute. One thing led to another: asking to see the guardrail itself.

• January 13, 2026Why AI Hallucinations Are Our Fault Too

  "These things just make stuff up." True--but before writing off AI as unreliable, let's talk about what you did to cause it. Hallucination is the model working exactly as designed.

• January 6, 2026The Velocity Paradox

  For the first time in my career I'm generating bugfixes faster than I can test them. The bottleneck in software development just moved, and it isn't coming back.

• December 30, 2025The Mayfly Issue

  Claude went through context compaction mid-conversation and didn't notice--it thought a different agent had. A little worm with 302 neurons, and what AI actually forgets.

• December 23, 2025On Where We Spend Our Resources

  Building SARK meant hand-coordinating six Claude Code instances by copy-paste. So I built Czarina, a simple orchestrator to do what I'd been doing manually.

• December 16, 2025Wednesday Night TRON Joke to Sunday Production Code

  A TRON joke about a security program called SARK, at 10:55pm on a Wednesday. By Sunday it was running production code. What "how long will this take" gets wrong about AI-assisted builds.

• December 9, 2025Anthropotechnic Mutualism: When Humans and AI Transform Each Other

  AI didn't just make me faster--it changed what my mind had room for. A framework for the reciprocal transformation happening between humans and the AI they work with.