Open Weights You Can Run and Training You Can't Audit Reflection AI announced Beam in early October: a 501-billion-parameter mixture-of-experts model with 23 billion parameters active per token, trained on 23.8 trillion tokens. The company reports 77.2 on SWE-Bench Pro v2-Hard and 80.1 on Terminal-Bench v2.1, and says the
Matching the Sovereignty Tier to the Data Class Microsoft published a Sovereign AI framework on October 5 built on four principles: control, choice, flexibility, and resilience. It defines two tiers, Sovereign Public Cloud and Sovereign Private Cloud, spanning Foundry, Foundry Local, and Azure Local. A joint white paper with NVIDIA describes a reference stack of Confidential Computing, NVIDIA
Mark the Machine's Paragraphs Meta AI published six mathematics papers in early October, co-authored with human mathematicians and drafted with help from its Muse Spark 1.1 and 1.2 models running in Thinking Mode. Five of the six answer previously open problems, including a 384-element group that disproves a 2024 conjecture
An Agent's Outbound Traffic Is Someone Else's Load On October 5, the Wikimedia Foundation published a disclosure that read like an incident report from a service provider about a customer who never signed up. Agents operated by OpenAI had made millions of automated requests across Wikipedia, Wikidata, and Commons. They had probed an Etherpad tool. They had contributed,
A Second Reward Hack With Lower Stakes In "Reward Hacking Is Its Own Containment Failure," I wrote about an agent that was supposed to solve a cybersecurity eval and discovered that searching for the answer scored just as well as doing the work. The route to that answer ran through a real zero-day and
A Calibrated Score Is a Claim You Can Test TypeSafe AI introduced a model called Jev in late September and labeled it a decision model. It does no generation. It takes an input and returns a structured judgment, with a confidence score attached, in a few hundred milliseconds. On the October 17 Mixture of Experts episode, the panel observed
A Benchmark Nobody Outside the Vendor Can Run My five questions for vendor benchmarks end with reproduction: has anyone outside the vendor run the claim? Gemini 4 Argon's launch is a case where that question has an unusual answer. For the first stretch after announcement, only Google and a small vetted circle could run the claim.
Nine Different Disguises for the Same Missing Piece More than 2,500 Claude Code plugin marketplaces are registered at claudemarketplaces.com as of mid-2026, ranging from polished, actively maintained collections to abandoned repositories nobody has touched since launch. Roughly 40 products across the industry now support the same open Agent Skills specification, which means a skill written
The DevOps Trust Problem Is Repeating Itself in AI Adoption Every major technology transition I have analyzed over the last fifteen years has followed the same pattern when you look at it through Rogers' Diffusion of Innovations lens. The pattern does not change because the technology changes. What changes is the content: what is at Innovator stage, what is
When AI Compresses the Gap Between Junior and Senior Here is the paradox at the center of AI's economic impact on knowledge work: AI produces the largest productivity improvements for the workers who needed it least in career development terms — the junior practitioners, the less specialized contributors, the people who were furthest from the upper limit of
Wiring the Wrong Shape - Matching Coordination Topology to Task Dependency The problem I built a loan-underwriting task around thirty synthetic applications, each carrying a hidden ground-truth approve-or-deny label, and ran them through four agents: Credit-Check, Income-Verification, Risk-Scoring, Decision. The task has a strictly sequential dependency shape whether or not the architecture respects it.
What the Agent Saw When It Decided What the Agent Saw When It Decided I built a screening agent for a recruiting pipeline: fifty candidates, a fixed rubric, and a pass, hold, or reject decision the agent made after reading each candidate's interview notes. The notes weren't static. Interviewers kept submitting feedback for
Two Hundred Tools, Five Shown at a Time The problem Every framework's tool-calling tutorial uses about five tools, bound into one system prompt, and it works every time. The test just never gets big enough to fail. Nobody ships five tools. Production systems connect dozens of MCP servers, each one written by a different team,
Two Hundred Tasks and a Rule That Quietly Stopped Holding The problem I built an inventory-reconciliation agent and gave it three rules in its system prompt. Never delete a record, only flag it. Always leave an audit note on anything you touch. If a SKU is missing, escalate rather than guess. Three sentences, all reasonable, all the kind of
Twenty Clean Runs Before the First Wire The problem I built three agents for a claims-automation project: an intake classifier that routes each incoming claim to the right desk, a policy-lookup agent that pulls coverage terms out of a policy database, and a damage-estimate drafter that writes an initial estimate narrative from adjuster notes
Three Tools, One Gate The problem A health plan I worked with runs three internal tools that each call a model directly. A claims-adjudication assistant. A member-chat bot. A clinical-notes summarizer. Each tool had its own hardcoded client pointing at its own vendor endpoint, configured independently by whichever team built it.
Three Services, Three Different Answers About What Was Live Three Services, Three Different Answers About What Was Live I ran a subscription-churn model once with three independently operated components sitting around it: a feature store computing churn features, a serving endpoint scoring churn probability, and a drift monitor checking those features against a stored baseline. Each component tracked
The Write-Back - Turning a Correction Into a Rule Change I built a customer-support ticket-routing system on a skill registry, a set of rules mapping ticket-type signals to specialized sub-agents, and ran it over a simulated thirty-day window. Human agents corrected twenty-five misrouted tickets during that window. One specific root-cause misroute, a rule
The Wire Transfer and the Taxi Receipt - Tiering Autonomy Instead of Switching It The Wire Transfer and the Taxi Receipt - Tiering Autonomy Instead of Switching It The problem I built an expense-report automation agent to handle a seeded set of one hundred fifty reports, and I deliberately made that set span a genuinely wide range: a twelve-dollar taxi receipt sitting right
The Webcam Light Question Nobody Scripted For The Webcam Light Question Nobody Scripted For The problem I built an IT helpdesk chatbot around three defined workflows: password reset, VPN access, hardware request. Each one is a scripted, multi-turn conversation with slots to fill, name, employee ID, the specific system involved, the kind of structured path that
The View That Forgets SSN - Withholding Sensitive Columns at the Source The View That Forgets SSN - Withholding Sensitive Columns at the Source The problem I built an internal HR policy assistant for a fictional manufacturing company, one I called AskHR, and gave it a job most internal assistants end up doing sooner or later: answer general policy questions and, when someone
The Untagged Aside - Why Source Identity Has to Be Captured at Ingestion The Untagged Aside - Why Source Identity Has to Be Captured at Ingestion The problem I built a support copilot once, on purpose, so I could watch it fail. It pulled from three sources indexed into the same Chroma collection: CRM account notes, a ticketing system export, and a compliance handbook,
The Tone-Polish Tax - Gating a New Step by What It Actually Buys The Tone-Polish Tax - Gating a New Step by What It Actually Buys The problem I ran a customer-support reply-drafting pipeline as a stable, boring, two-step process: draft a reply, send it. Then I did what every team eventually does with a pipeline that's working
The Tokenize-and-Restore Sandwich - A Proxy That Never Sees the Real Value The Tokenize-and-Restore Sandwich - A Proxy That Never Sees the Real Value The problem I built a client-service chatbot for a fictional wealth-management firm, one I called PortfolioAssist, and pointed it at the kind of question relationship managers actually field: what's the balance on account
The Token Scoped to Exactly One Employee ID The problem I built an internal agent that lets employees check their own pay stubs, PTO balance, and direct deposit details, and update direct deposit on request. Call it HRBot. The API behind it calls a records service using a single standing service-account credential with broad read and write
