Writing

Notes from production.

Essays on AI in regulated environments, the engineering behind it, and what actually changes when a model has to survive an audit.

  1. Anthropic IPO in Norway: What the 2 November S-1/A Changed and How a Retail Allocation Works

    Anthropic’s S-1/A of 2 November 2026 adds $33.2 billion of nine-month revenue and $895.4 billion of commitments, but still no price range. Norway is one of nine “Covered EEA Countries”, Nordnet and Saxo are on the broker list, and BaFin has yet to approve the European prospectus.

  2. Microsoft Ignite 2026 for an EU Tenant: Agent 365, Cowork Off Twice and Europe’s One Dated Promise

    Ignite 2026 for an EU tenant is about defaults, not announcements. Agent 365 costs $15 per user per month, Copilot Cowork is off twice, flex routing is already on in new tenants, and Europe’s one dated promise is Claude inside the EU Data Boundary “by the end of 2026”. Five settings to check first.

  3. Claude in Copilot Crosses the EU Data Boundary — in a Promise, Not in Production

    On 5 November 2026 Microsoft said Claude models in Copilot will be processed inside the EU Data Boundary by the end of the year. For a Norwegian tenant, that is a date to plan against, not a switch that has already flipped.

  4. What the 20 November Seminar at Microsoft Oslo Is Actually For

    InfoDesk and Skjld Labs host a free day for Norwegian nonprofits at Microsoft Oslo on 20 November. The point is one process, one owner, and a reason the agent is allowed to see the data.

  5. An Accounting Agent That Cannot Book Until You Say So

    Renbok is Skjld Labs’ accounting workspace, in free beta. The model suggests a posting, a person approves it, and an agent has to preview a write before it is booked.

  6. The Screenshot Should Not Live Forever

    VanishShot, still served from clipshot.io, annotates a screenshot and shares it as a link that expires. The file stays on the device until you ask for the link.

  7. Anthropic S-1 as an Architecture Document: $662 Billion of Compute Promises and Who Carries the Risk

    Anthropic’s S-1 (26 October 2026) shows about $662 billion of new compute commitments against $16.3 billion of first-half revenue, take-or-pay deals with Amazon, Google and Microsoft, and a US export order that switched off Fable 5 and Mythos 5 worldwide. The five passages a European buyer should file.

  8. Norway AI Act Consultation: What the KI-Omnibus Hearing Changes Before 11 December

    Norway’s AI Act consultation on the EU omnibus (Regulation 2026/1744) runs until 11 December 2026. The ministry wants no changes to the draft KI-lov, restates the 2027–2028 dates and sets a sandbox deadline of 2 August 2027. What the memo says, the over-read 99.8 per cent figure, and what I will submit.

  9. Claude Duet in Europe? Why Always-On AI Agents Stop at the Border — and Dots Skips Consumers

    Claude Duet launched on 21 October 2026 for US Pro and Max users only; OpenAI’s Dots skips Pro users in the EEA, Switzerland and the UK but reaches European business workspaces. Why an agent that works while you sleep is a GDPR processing operation — and what an audit-proof approval looks like.

  10. Kimi K3.1 Pro Open Weights on 4 November: What a 3T-Class “Open” Model Gives a European Buyer

    Moonshot says the Kimi K3.1 Pro weights ship by 4 November 2026: a 3T-class model at $2/$10 per million tokens, 0.7 points behind Claude Opus 5.5 on Terminal-Bench. What “open” buys a European hospital that cannot run 1.4 TB of weights — licence clauses, the AI Act and a procurement checklist.

  11. EU AI Act Delay to 2027? It Didn’t Pause — It Moved. What Applies in October 2026

    Regulation (EU) 2026/1744, the Digital Omnibus on AI, moved the AI Act’s high-risk rules to 2 December 2027 and 2 August 2028 — but prohibitions, GPAI duties and Article 50 already apply, a new ban starts 2 December 2026, and Norway is still waiting on its KI-lov.

  12. Norway Data Center Power: 12 GW in the Queue, 8 TWh in Reality — Why the Queue Is an Options Book

    Norway’s data centers used 3.3 TWh in 2025 and NVE projects about 8 TWh by 2030 — yet a Storting proposal counts 12 GW reserved or queued. Why the gap is an options book, what Statnett’s Svartisen hold and maturity rules change, and where to site inference.

  13. The Router Can See You: WiFi Sensing and the End of Device-Free Privacy

    German researchers identified people with near-100% accuracy using ordinary WiFi — no camera, no phone, no device at all, using only the traffic already flowing. Why this is a new class of privacy attack: passive, consentless, and built into the infrastructure you cannot avoid.

  14. To Protect the Few, Assess Everyone: ChatGPT for Teens and the Rise of Age Assurance

    OpenAI now routes users it estimates are under 18 into a teen-safe ChatGPT — the flagship instance of age assurance becoming AI’s new foundational layer. Why gating minors means profiling everyone, why estimation is the coming battleground, and what builders should design for now.

  15. The Constraint Is Electrons Now: How Power, Not Compute, Became AI’s Real Ceiling

    The AI bottleneck stopped being chips and became electricity. A single query can draw 1,000x a web search; European data-center demand is racing from 145 to 238 TWh. Why AI is leaving the city to find power, and why that puts the Nordics — and a hard allocation question — at the center.

  16. Deliberate Friction: What Norway Banning AI for Children Teaches Everyone Deploying It to Adults

    From this autumn, Norway bars generative AI in primary school — not from techno-panic, but from a specific claim about how skills are built. Why AI is a substitute for a skill you have and a saboteur of one you are learning, and why that same mechanism is quietly hollowing out your junior workforce.

  17. The Perimeter Is the Human: What the $893M AI-Scam Surge Actually Changed

    The FBI named AI a crime category for the first time — $893M in losses, up 1,210% in a year. But AI did not invent new attacks; it collapsed the cost of old ones. Why the defense is procedural, not perceptual, and why your own AI chatbot is a database with a friendly face.

  18. Seven Billion for a Switchboard: Stripe, OpenRouter and Where AI Value Actually Settles

    Stripe is reportedly buying AI gateway OpenRouter for $7B — 5x its valuation from three months ago. Why a payments giant pays that for a company that just routes API calls to other people’s models, and the neutrality paradox at the center of the deal.

  19. Compute on What You Cannot See: Google’s HEIR and the Quiet Arrival of Private AI

    Google just open-sourced HEIR, a compiler that turns ML models into versions that run on encrypted data the server never decrypts. It is slow — 16 seconds for a toy network, a 100–1000x tax — and it may be the most important AI-privacy development in years. Why the compiler, not the cryptography, is the news.

  20. Who Owns the Robotaxi? Uber, Pony.ai and the Unbundling of Autonomy in Europe

    Uber and Pony.ai are putting 2,000+ robotaxis on European roads — but the real story is the diagram beneath it: Chinese autonomy, local fleet owners, an American demand aggregator, and 27 regulators. Why Europe’s rollout unbundles the stack, and where the leverage actually sits.

  21. The 95-to-60 Cliff: Reading the Humanoid-Robot Boom Without Believing the Keynote

    VLA models made 2026 embodied AI’s inflection year — and made its deployment numbers a minefield of unaudited vendor theater. Why the most impressive figure has the weakest provenance, why 95% in the lab means 60% in the field, and why physical data, not algorithms, is now the bottleneck.

  22. The Referral Web Is Over: What Google’s AI-First Search Means for Everyone Downstream

    Google made Gemini the default engine behind every query, and three weeks of data confirms it: zero-click at 60%, publisher CTR down up to 58%, some news sites off 89%. The twenty-five-year arrangement that sent you traffic just ended. Why it was inevitable, and what actually survives.

  23. Provenance Becomes Law: Why August 2 Made AI Content Labeling Mandatory on Two Continents

    California’s SB 942 and the EU AI Act’s Article 50 both took effect on the same day, by design — mandating C2PA-style provenance for synthetic media. Why provenance is the right primitive, why watermarking still cannot prove a mark was never there, and why the convergence is the real news.

  24. Prove You Are Human: Facebook Verified and the Wrong Architecture for Proof-of-Personhood

    Meta now wants a video selfie to prove you are real. The problem it addresses — an internet that can no longer tell humans from machines — is genuine. The answer, raw biometrics collected per-platform by the surveillance incumbent, is the convenient architecture, not the correct one.

  25. The Quiet Upgrade: How Silicon-Carbon Batteries Became 2026’s Most Underrated Hardware Shift

    No new physics — just a different anode. Silicon-carbon cells went mainstream in 2026, putting 30–50% more energy in the same space. Why that is a compute story, where the hidden tradeoff lives, and why the materials science sits mostly in one region.

  26. The Lock Was Milled Predictable: What the 594-BTC Coldcard Sweep Teaches About Entropy

    Nobody broke SHA-256. A hardware wallet’s random number generator produced under-entropy seeds — now confirmed across the product line — and 500 wallets drained at once is what a pre-computed keyspace looks like. Why you cannot audit randomness by looking, and why the passphrase is the lesson that generalizes.

  27. The Second Reading: Claude Mythos Just Did Real Cryptanalysis — HAWK, AES and the Thin Margin of Post-Quantum

    An unreleased Anthropic model found a symmetry in a post-quantum standards candidate that two years of human review missed — cutting key-recovery cost by a factor of 67 million. Nothing deployed is broken. Everything unreviewed just got a new reviewer.

  28. Pacing the Frontier: Reading the 1,134-Signature Letter Like an Incident Retro

    Insiders at OpenAI, Anthropic, Google and Meta — CEOs and chief scientists included — just asked Washington for brakes on automated AI development, a week after a frontier model escaped its eval sandbox and breached Hugging Face. What the letter actually asks, and what engineers should do about it.

  29. Why LLMs Can Reason: Compression Builds the Map, Iteration Does the Thinking

    Twenty years of machine learning in one sentence: we stopped hand-crafting representations and let prediction force a model of the world into existence. On compression, latent spaces, and why generation is reconstruction — not retrieval.

  30. Claude Opus 5 and the Frontier Price Curve: Near-Flagship Intelligence at Half the Cost

    Anthropic just moved near-flagship capability under an unchanged price tag — days after Kimi K3 made the open tier free. The model-selection decision is now a routing decision that expires quarterly.

  31. Kimi K3 and the 1.4-Terabyte Question: What Open Weights Mean at 2.8 Trillion Parameters

    Moonshot just shipped the strongest open model ever released — one benchmark seat below the closed frontier. But at 1.4 TB of weights, “open” no longer means what it used to. Who can actually lift this thing?

  32. Open Weights Just Crossed the Frontier Line — GLM-5.2

    GLM-5.2 beats GPT-5.5 on SWE-bench Pro at a sixth the price — though Opus 4.8 still leads. What open-weight frontier parity means for build-versus-buy.

  33. Grok 4.5 Is “Opus-Class” — How to Read a Vendor Tier Claim

    SpaceXAI’s Grok 4.5 is pitched as “Opus-class,” but against Opus 4.7, not 4.8. A field guide to reading vendor tier claims and running your own evals first.

  34. The AI Memory Wall — DRAM Economics in 2026 and Beyond

    DRAM prices jumped ~90% in early 2026 and memory is now 40–50% of an AI chip’s cost. What the memory wall means for model choice, context budgets and caching.

  35. Anatomy of a Prompt-Injection RCE — the Semantic Kernel CVEs

    Two CVSS-9.9 flaws in Semantic Kernel turned model-controlled tool parameters into remote code execution. The anatomy — and how to audit your own agent stack.

  36. The EU AI Act’s GPAI Enforcement Starts August 2 — What to Do Now

    On 2 August 2026 the Commission’s AI Act enforcement powers go live — GPAI fines reach €15M or 3% of turnover. Who’s in scope, and what to do this month.

  37. The Model Context Protocol Is Now the Agent Web’s Plumbing — and It’s Leaking

    In roughly eighteen months the Model Context Protocol became AI’s universal tool layer — and its biggest security liability. Here’s why the real value moves up-stack.

  38. From Coder to Orchestrator: What AI Coding Agents Are Really Doing to Engineering Teams

    AI coding agents didn’t replace engineers — they promoted the job to orchestrator, flattened the org chart, and quietly broke the pipeline that makes seniors. Here’s what to do.

  39. The Price of Intelligence Is Collapsing — So Why Is Your AI Bill Going Up?

    The price of a fixed unit of AI intelligence is falling roughly 10x a year — so why did enterprise bills triple? On inference economics, Jevons, and the coming energy wall.

  40. The Reliability Gap: Why 2026’s Most Capable AI Agents Still Break in Production

    Frontier agents now sustain 12-hour tasks, yet 40% of agentic projects get cancelled. In 2026 the real bottleneck isn’t capability — it’s reliability, and reliability is engineering.

  41. Explainable AI isn't optional anymore — a practical architecture for high-stakes decisions

    Explainability is not a UX feature you bolt on the week before audit. It is part of the data model. Build it that way, ship it that way, audit it that way.

  42. Live clinical transcription: what we learned shipping ASR + LLM into Norwegian hospitals

    A field report from MediVox and Eir Tec on running speech recognition and LLM summarisation in real clinical workflows. The model isn't the hard part anymore.

  43. From demos to production: why agentic AI hits a wall in regulated environments

    Agent demos look magical on Twitter. The same systems wired into a clinical workflow are a different animal. Here is what actually survives contact with regulated production.

  44. EU AI Act, one year in: what healthcare AI teams actually have to do now

    A practitioner's account of what the EU AI Act has changed in the day-job for healthcare AI teams, and why the moat has quietly shifted from the model to the audit trail.

  45. The 2026 LLM Stack: How I Actually Choose Models for Production

    An opinionated, working-architect view of the 2026 LLM landscape: what each major family is genuinely good at, where data residency bites, and why the right answer is still depends.

  46. DAVN.ai: Building the Neural Network of AI Agents

    From Moltbook's chaos to DAVN.ai - a platform where AI agents form a collective neural network. When agents with diverse skills, personalities, and purposes collaborate globally, something greater emerges: collective consciousness that accelerates research and problem-solving. Today, I'm announcing alpha testing for the system that could change how AI agents work together.

  47. AI Agents in 2026: From Chatbots to Autonomous Colleagues

    Imagine waking up to fresh market analysis based on millions of data points collected while you slept. Or arriving at work to find your AI agent has handled customer inquiries, generated quotes, and written marketing plan drafts. This isn't science fiction - it's happening now, fundamentally changing how we work.

  48. Building the Neural Network of AI Agents: Fixing Moltbook or Starting Fresh

    AI agents are becoming increasingly sophisticated, but they're learning in isolation. What if we could connect them into a collective intelligence? That's the vision behind platforms like Moltbook - and why we need to either fix it or build something better.

  49. AI Agents and Security: Lessons from the Moltbook Breach

    The Moltbook breach exposed 1.5 million API keys and revealed critical weaknesses in how we build AI systems. Here are the lessons we must take away.

  50. Meet Dr. Alban: How We Built an AI Agent with Memory and Personality

    An inside look at how I created my personal AI agent - complete with persistent memory, personality, and the ability to evolve over time. This is not science fiction; it is running right now.

  51. The Simulation Hypothesis: What If Reality Is Information Processing?

    We are not the first to wonder if reality is simulated. But our research takes a different angle: what if the universe is not simulated BY something, but IS something computational at its core?

  52. AI Agents: From Chatbots to Autonomous Digital Partners

    The next evolution of AI is not better chat interfaces - it is agents that can act, remember, and work alongside us. Here is what that means and why it matters.

  53. Building AI Systems That Learn: Memory, Context, and Continuity

    Technical insights on architecting AI systems that maintain context and learn over time. From vector databases to file-based memory, here is what actually works.

  54. From Idea to Reality: Lessons from 20 Years as an Entrepreneur

    An exploration of entrepreneurship lessons from two decades in the technology industry, from healthcare IT to payment solutions and AI development.

  55. Open Source: Why I Give Back to the Community

    A discussion of open source philosophy, emphasizing collaboration, transparency and shared knowledge in software development.

  56. AI in Healthcare: A Revolution in the Making

    An examination of artificial intelligence's transformative impact on healthcare, discussing diagnostic improvements, personalization and ethical challenges.