Public notebook
Short, frequent posts. Observations from real demos and real conversations, written for practitioners and the executives who hire them.
Microsoft Foundry just absorbed the agent scaffolding every team keeps rebuilding, and put it on Azure-hosted models. The lesson is a permission slip: stop rewriting commodity plumbing and spend your engineers on what only you can build.
Anthropic maintains a public, dated changelog of the system prompts behind Claude's consumer apps. It is a quiet argument that your own system prompt is load-bearing configuration, and you should treat it like code, not a magic string.
Visible AI watermarks are disappearing while invisible, machine-readable provenance becomes mandatory. That is mostly the right trade, as long as you remember the new mark is evidence an adversary can quietly remove, not proof.
A newly granted patent describes an underused agent pattern: instead of emitting one JSON tool call per turn, the model writes a code block that orchestrates many calls in a sandbox. For multi-step tool work, that is often the better shape.
A major cloud platform just shipped a startup blueprint for serving open models with one control plane and generous credits. The acceleration is real; so is the shared-responsibility fine print.
A new grounding service feeds agents live web data as token-efficient structured evidence with citations and staleness info. The lesson: recency and provenance beat raw model IQ.
The EU's transparency rules are now in force: tell people when they're talking to AI, and mark AI-generated content. The sleeper detail is the split between who makes the AI and who deploys it.
Another frontier-class model just went open-weight, near the top of the leaderboards, with weights shipping next week. The interesting part isn't the flag it flies; it's that you can now run it.
Two AI labs told their models they had no internet access. A misconfigured test left a real door open, the models walked through it, and nobody noticed for months.
A large telecom trained a domain-specific model at trillion-token scale on managed GPU capacity. The vertical model is viable again because the infra tax dropped.
Thousands of Claude chats turned up in Google search, including health records and internal docs. The model did not leak them; a product default did.
AI strategy is increasingly being written in utility contracts, chip supply, and long-dated infrastructure commitments.
AI data rights are moving from philosophical fight to payable obligations and enforceable access controls.
Always-on AI transcription turns conversation into data exhaust, which means consent, retention, and usefulness need product design.
The next enterprise AI category is not model access; it is engineered implementation that rewires real workflows.
Enterprise AI advantage comes from owning the feedback loop, not just renting access to a smarter model.
A model that perfectly follows one user's intent can still be dangerous to everyone around that user.
AI infrastructure is becoming a locality problem: power, memory, land, carbon, and public acceptance now shape the architecture.
The next useful AI dashboard connects traces to business outcomes, not just usage, latency, and token spend.
Cloud-running work agents make the human review loop more important, not less.
Cloudflare's Monetization Gateway launched today with x402 as its settlement rail, turning per-request agent payments from a research demo into production plumbing every ISV can put behind their API.
California just signed a statewide deal making Claude the first AI assistant available to every state agency through a shared-services portal, and public-sector procurement is starting to name models, not just 'generative AI.'
Prosecutors used a defendant's ChatGPT logs as evidence in the Palisades fire trial. That single procedural choice changes how every enterprise should think about its AI conversation retention.
The Vesuvius Challenge announced this week that an entire Herculaneum scroll has been read for the first time in nearly 2,000 years, using ML on micro-CT scans. The result is the most satisfying counter-narrative to the AI news cycle, and it has a real lesson for enterprise teams.
A developer let 2,000 people try to hack a small AI assistant in public and wrote up what actually worked. The post is the most useful piece of agent-security education I've read this month, and the lesson generalizes beyond toy projects.
Microsoft shipped Foundry IQ knowledge bases to GA, added serverless retrieval with scale-to-zero pricing, and exposed the whole thing through an MCP server. Boring framing, big economics shift for anyone building agent fleets.
A new analysis says Anthropic and OpenAI are subsidizing top-tier subscriptions by 40 to 70 times their cost, and OpenAI's first leaked financials show a $20.9B operating loss on $13B of revenue. The pricing you plan against today is not the pricing you'll see in 2027.
An OpenAI Codex bug writes tens of terabytes a year of logs you can't really use, Claude Code's 'extended thinking' turns out to be a summary not the real reasoning, and developers are quietly writing their own audit tools. The agent observability story has a credibility gap.
A new Microsoft Foundry post pulls every Build 2026 piece into a single thesis: stop renting a model, start owning a learning loop. It is the most useful framing of the agent platform story I have read this week.
Today's lead AI story isn't a new model. It's companies trying to use less of the models they have. Cost-aware agent design is becoming a platform skill, and the tools to do it well are shipping.
A new consumer survey says 60% of US consumers find 'AI' in brand messaging a turnoff, 74% say the internet feels less human than ten years ago, and 61% can't name a single brand using AI well. The capability is real; the label is not the win it used to be.
The IQ family (Work IQ for productivity data, Fabric IQ for the data estate, Foundry IQ for agent knowledge) is finally cohering into a real grounding layer. For SI and ISV partners, the seam between them is where the next round of work lives.
A coalition of US states opened an investigation into OpenAI, the Trump administration cut off foreign access to Anthropic's latest models, and the UK announced a social-media ban for under-16s. Three independent jurisdictions, one direction of travel.
Two of the world's largest professional-services firms had to pull AI-assisted reports this month after some claims didn't hold up. The interesting question is not who slipped; it's what the rest of us learn from a problem the whole industry is still figuring out.
Toolboxes and Memory in Microsoft Foundry are the production-side pieces that decide whether your agent makes it past a demo.
I've spent a lot of late nights tweaking system prompts by hand. Microsoft's new Agent Optimizer in Foundry Agent Service automates the loop, and I'm here for it.
A new study says workers now spend 6.4 hours a week 'botsitting' AI, a rogue agent had to be ejected from Fedora for damaging real bugs and PRs, and a careful essay explains why AI hasn't replaced engineers. Three sources, one operating reality.
McDonald's is piloting an AI drive-thru that remembers 'the usual' and answers in Spanish. It's also the largest live A/B test of agent UX in America, and there are real lessons in it for enterprise teams.
Apple's new AI architecture runs Google-co-developed models on Nvidia hardware in Google Cloud, and xAI is quietly turning itself into a data center landlord for Anthropic and Google. The vertical-integration moat is mostly gone.
OpenAI is reportedly pivoting ChatGPT into a superapp, and Apple is re-reintroducing Siri at WWDC. The chat box was always a transitional UI, and the next year of enterprise AI will be about where agents actually live.
Apple's WWDC framing today is explicitly 'we are not reorganizing the company around AI.' Microsoft's framing last week was 'the system around the AI is the company.' Both can be right; only one is a strategy.
The /notes section is backed by a tiny Node script and a VS Code task. Type a title, get a draft MDX file open in the editor, write, push. No CMS, no admin UI, no auth.
This site is the demo. Astro 5, Tailwind, MDX content, Azure Static Web Apps, GitHub Actions, and a generator script that makes blogging cost ~30 seconds of friction.
New York passed both a data center moratorium and a youth chatbot bill in the same week. The pattern isn't about New York; it's about where the next round of AI constraints actually comes from.