{"id":11249,"date":"2026-06-01T09:14:35","date_gmt":"2026-06-01T09:14:35","guid":{"rendered":"https:\/\/i-spark.nl\/?p=11249"},"modified":"2026-07-15T09:25:49","modified_gmt":"2026-07-15T09:25:49","slug":"lets-debrief-data-ai-issue-5","status":"publish","type":"post","link":"https:\/\/i-spark.nl\/en\/blog\/lets-debrief-data-ai-issue-5\/","title":{"rendered":"Let’s debrief : data & AI | Issue 5, May 2026"},"content":{"rendered":"\n

You know that feeling when a project stops being exciting and starts being real? That’s what May felt like.<\/p>\n\n\n\n

Anthropic shipped a model that flags its own mistakes. OpenAI published a document mapping their safety practices to laws. Databricks made catalog commits production-ready. Snowflake connected its AI agents to Microsoft Copilot Studio. Power BI delivered visual calculations after years of community requests.<\/p>\n\n\n\n

All of it is the kind of work that makes the stuff built on top of it hold up. And for anyone working in data and AI day to day, that’s more useful than another benchmark announcement.<\/p>\n\n\n\n

Let’s get into it.<\/p>\n\n\n\n

Anthropic <\/strong>introduces Claude Opus 4.8<\/strong><\/h2>\n\n\n\n

Anthropic closed out May with a model upgrade, and the most interesting improvements are behavioural.<\/p>\n\n\n\n

Claude Opus 4.8<\/a> builds on Opus 4.7 with gains across coding, reasoning, and agentic tasks. <\/p>\n\n\n\n

Early testers claim that it’s more honest about what it doesn’t know. It’s more likely to flag uncertainties, push back when a plan doesn’t make sense, and catch its own mistakes before they become your problem. Anthropic’s own evaluations put it at around four times less likely than its predecessor<\/a> to let flaws in code pass unremarked. <\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Latest Opus models, comparison table (Anthropic)<\/em><\/p>\n\n\n\n

Speaking of speed, fast mode for Opus 4.8 runs at 2.5 times the normal speed and is now three times cheaper than it was for previous models. It’s available today at the same price as Opus 4.7.<\/p>\n\n\n\n

Dynamic workflows, effort control and system entries inside the messages array<\/h3>\n\n\n\n

Additionally, three new features were launched alongside the new model. <\/p>\n\n\n\n

The first is dynamic workflows<\/strong> in Claude Code, currently in research preview for Enterprise, Team, and Max plans. Claude can now plan a task, spin up hundreds of parallel subagents in a single session, and verify the outputs before reporting back. The example Anthropic gives is codebase-scale migrations across hundreds of thousands of lines of code, handled end to end with the existing test suite as the quality bar. <\/p>\n\n\n\n

The second is effort control<\/strong> on Claude.ai and Cowork. You now choose how much thinking Claude puts into a response. Higher effort equals deeper reasoning and better answers. Lower effort is for faster responses and a slower rate limit consumption. <\/p>\n\n\n\n

Finally, Developers can update Claude’s instructions mid-task without breaking the prompt cache or routing through a user turn, due to the Messages API now accepting system entries inside the messages array<\/strong>.  Useful for agentic systems where permissions, context, or token budgets need to change as the task runs.<\/p>\n\n\n\n

OpenAI: Codex on Mobile & Governance Framework<\/strong><\/h2>\n\n\n\n

A practical and strategic update from OpenAI this month.<\/p>\n\n\n\n

On the practical side, Codex is now in the ChatGPT mobile app<\/a>. When you connect to any device where Codex is running, the app loads the live state from that environment. From your phone, you can work across active threads, review outputs, approve commands, change models, or start something new. Files, credentials, permissions, and local setup stay on the device where Codex is operating. At the same time, updates flow back to your phone, including screenshots, terminal output, diffs, test results, and approvals.<\/p>\n\n\n\n

Codex uses a secure relay layer that keeps trusted machines reachable across devices without exposing them directly to the public internet. That relay also keeps active session state and context synced across any device where you’re signed in with ChatGPT. <\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Codex on a mobile device (OpenAI), May 2026<\/em><\/p>\n\n\n\n

On the strategic side, OpenAI published its Frontier Governance Framework<\/strong><\/a> this month. It outlines how their safety and security practices align with emerging legal requirements, including California’s Transparency in Frontier AI Act and the EU AI Act’s Code of Practice for General Purpose AI. <\/p>\n\n\n\n

The framework covers risk assessment and mitigation across cyber offence, CBRN risks, harmful manipulation, and loss of control, as well as model reporting, security risk management, incident response, external expert input, and framework updates. It’s a public document worth reading if you navigate AI governance requirements or building on top of OpenAI models.<\/p>\n\n\n\n

Snowflake: Security and smarter AI functions<\/strong><\/h3>\n\n\n\n

First, the AI Security Trust Centre <\/a>is now fully available for production use. It gives you centralised security intelligence for AI workloads, and what makes it practical is the Cortex Code integration. You can remediate network policies, untangle RBAC configurations, and run security audits through natural language, without writing SQL or filing a ticket. If you are trying to get a handle on AI governance without adding overhead to your security team, this is worth exploring. <\/p>\n\n\n\n

Second, Snowflake now integrates natively with Microsoft Copilot Studio via a managed MCP server. That means Cortex agents can connect directly into Copilot Studio workflows<\/a>, which is significant for organisations already running Microsoft’s ecosystem. Instead of choosing between Snowflake’s AI capabilities and Microsoft’s, you can use both together. <\/p>\n\n\n\n

Third, Cortex AI Function Studio<\/a> is now in public preview. It’s designed to take the manual work out of building production-ready AI functions on unstructured and multimodal data. It automates prompt engineering, model selection, benchmarking, and optimisation, so teams can move faster without spending weeks on experimentation overhead. Worth enabling and testing if your team is building on Cortex. <\/p>\n\n\n\n

Power BI: Visual Calculations and Reports get a new default<\/strong><\/h3>\n\n\n\n

Three things that were either in preview or long-requested are now fully available for production use.<\/p>\n\n\n\n

The biggest one is Visual Calculations and Custom Totals, hitting general availability. Visual Calculations let you write running sums, moving averages, per cent-of-parent calculations, and other patterns directly inside a chart, without creating new DAX measures in your semantic model. It keeps models cleaner and makes report authors more self-sufficient. Custom Totals, which lets you control exactly what the total row shows in tables and matrices, comes along with it. Both have been among the most requested features in the Power BI community for years, so it\u2019s great they\u2019re now available.<\/p>\n\n\n\n

Second, PBIR is now the default report format, replacing PBIX. If you’ve been following along since Issue 1, you’ll remember we flagged this transition back in January. It’s now here. Existing reports still work, but any team still treating Power BI files as unmanaged binaries is accumulating technical debt that will be unpleasant to resolve later<\/em>. Now is a good time to review your setup. <\/p>\n\n\n\n

Third, Copilot summary shortcuts are now available on the report ribbon and visual header. One click gets you an AI-powered summary of a full report or a specific visual, with Copilot surfacing key trends, changes, and insights automatically. It’s the most visible change for end users, and the one business stakeholders will ask about first.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

The new Summarize button highlighted in the report ribbon (PowerBI, May 2026)<\/em><\/p>\n\n\n\n

Databricks: May 2026 Updates<\/strong><\/h3>\n\n\n\n

Databricks had a strong May, with several things moving from preview into production-ready and a few new capabilities.<\/p>\n\n\n\n

The Lakeflow Pipelines Editor is hitting general availability. It’s an agent-first pipeline-editing experience with Genie Code built directly in your code, the Genie Code chat, the pipeline graph, and metrics. For data engineers doing pipeline work, this is a great quality of life upgrade. <\/p>\n\n\n\n

Real-time mode in Lakeflow Spark Declarative Pipelines is now also in public preview. That opens the door to use cases like fraud detection, personalisation, and instant decision-making that weren’t practical in Databricks before. It uses a new update_flow decorator and supports stateful aggregations without requiring a watermark. <\/p>\n\n\n\n

Unity Catalog becomes the coordination layer for managed Delta tables, enabling multi-statement, multi-table transactions and stronger governance across engines. Products that read or write to Unity Catalog managed tables all support catalog commits. <\/p>\n\n\n\n

A question for you<\/strong><\/h3>\n\n\n\n

Every tool in this issue shipped something that makes AI more capable, autonomous, or embedded in your existing stack. <\/p>\n\n\n\n

But here’s the thing. A more capable model is only as good as the data it reasons over. A real-time pipeline is only as reliable as the governance layer underneath it. An AI agent that acts on your behalf is only as trustworthy as the permissions and controls you’ve put in place. <\/p>\n\n\n\n

So grab your rubber duck and have that conversation. Or take a colleague for a coffee and talk it through:<\/p>\n\n\n\n

The AI is getting better at checking its own work, but who’s checking the foundations it’s building on?<\/strong><\/p>\n","protected":false},"excerpt":{"rendered":"

You know that feeling when a project stops being exciting and starts being real? That’s what May felt like. Anthropic shipped a model that flags its own mistakes. OpenAI published a document mapping their safety practices to laws. Databricks made catalog commits production-ready. Snowflake connected its AI agents to Microsoft Copilot Studio. Power BI delivered […]<\/p>\n","protected":false},"author":17,"featured_media":11254,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[8,432],"tags":[],"class_list":["post-11249","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-blog","category-data-ai-newsletter"],"acf":[],"_links":{"self":[{"href":"https:\/\/i-spark.nl\/en\/wp-json\/wp\/v2\/posts\/11249","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/i-spark.nl\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/i-spark.nl\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/i-spark.nl\/en\/wp-json\/wp\/v2\/users\/17"}],"replies":[{"embeddable":true,"href":"https:\/\/i-spark.nl\/en\/wp-json\/wp\/v2\/comments?post=11249"}],"version-history":[{"count":2,"href":"https:\/\/i-spark.nl\/en\/wp-json\/wp\/v2\/posts\/11249\/revisions"}],"predecessor-version":[{"id":11265,"href":"https:\/\/i-spark.nl\/en\/wp-json\/wp\/v2\/posts\/11249\/revisions\/11265"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/i-spark.nl\/en\/wp-json\/wp\/v2\/media\/11254"}],"wp:attachment":[{"href":"https:\/\/i-spark.nl\/en\/wp-json\/wp\/v2\/media?parent=11249"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/i-spark.nl\/en\/wp-json\/wp\/v2\/categories?post=11249"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/i-spark.nl\/en\/wp-json\/wp\/v2\/tags?post=11249"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}