## Sources

1. [Before the Structure](https://jessicatalisman.substack.com/p/before-the-structure)
2. [5 data contract implementations in the wild](https://andrewrjones.substack.com/p/5-data-contract-implementations-in)
3. [Why I’m Moving Some of My Work From Claude Code to Codex](https://aimaker.substack.com/p/codex-claude-code-workflow)
4. [🔬Scaling Past Informal AI - Carina Hong, Axiom Math](https://www.latent.space/p/axiom)
5. [Reducing the difficulty of being a data owner](https://andrewrjones.substack.com/p/reducing-the-difficulty-of-being)
6. [⚡️Satya Nadella: No Priors x Latent Space Crossover Special at Microsoft Build](https://www.latent.space/p/satya-2026)

---

### **5 data contract implementations in the wild** by Andrew Jones

*   **Main Arguments:**
    *   Data contracts are no longer just a theoretical concept; they are being actively implemented across diverse industries to solve practical data reliability and ownership challenges [1, 2].
    *   The primary value of data contracts lies in fostering trust, managing hidden dependencies, and empowering data producers with autonomy through automation [2-4].

*   **Key Takeaways:**
    *   **Miro** utilized data contracts to transform their data product reliability, reducing key pipeline downtime from **50% to nearly 1%** [2].
    *   **Nandos** implemented contracts to solve the "hidden dependency" problem, allowing them to quickly identify who is impacted by table changes or which tables are shared externally [3].
    *   **Adevinta Spain** used contracts to automate ingestion, creating a single code base capable of ingesting multiple sources in a standardized way without needing to recode for every new source [4].
    *   **GoCardless** is highlighted as the origin of the movement, where the concept was introduced four years ago to improve data quality [5].

*   **Important Details:**
    *   Implementation stacks often involve tools like **YAML, Pulumi, and BigQuery** [5].
    *   The **Open Data Contract Standard (ODCS)** continues to evolve, with version 3.1.0 introducing stricter validation and richer metadata [6].
    *   Successful data platforms require a heavy focus on **culture**, not just technical infrastructure, to ensure long-term success [7].

### **Before the Structure** by Jessica Talisman, MLS

*   **Main Arguments:**
    *   Organizing is a discipline consisting of **intentional design decisions** made before any technical structure or AI implementation exists [8, 9].
    *   The "AI economy" often fails because it prioritizes the automation of organizing over the principles of organizing, treating volume as an asset rather than ensuring intent and meaning [10, 11].

*   **Key Takeaways:**
    *   An **Organizing System** requires a collection of resources, intentional arrangement by an agent, and supported interactions [9].
    *   Resources must be **identifiable, describable, and persistent** over time before they can be effectively organized; otherwise, they are merely "noise" [12].
    *   The **Ontology Pipeline™** provides a structured framework for building knowledge infrastructures, moving from controlled vocabularies to taxonomies, thesauri, ontologies, and finally knowledge graphs [13, 14].

*   **Important Details:**
    *   **Six Design Decisions** must precede structure: What is being organized, Why, How much (degree), When (at creation or on-demand), How/By whom (human vs. algorithm), and Where [15-19].
    *   Organizing principles operate in a **middle tier** (logic) that is separate from the storage tier (resources) and the presentation tier (interactions) [20].
    *   Skipping foundational work leads to "broken atoms" where data integration fails because systems define resources differently [15].

### **How Codex Changed My Claude Code Workflow** by Wyndo and Dheeraj Sharma

*   **Main Arguments:**
    *   **Portability** of workflows is a greater competitive advantage than loyalty to a specific AI model [21, 22].
    *   GPT 5.5 has improved significantly for knowledge work, making the **Codex** app a viable alternative or companion to **Claude Code** [23, 24].

*   **Key Takeaways:**
    *   Codex allows users to work "close to the files," reducing friction at the handoff point by integrating chat, file inspection, and artifact generation (like slides or documents) in one view [25, 26].
    *   Workflows built in Claude Code can often be migrated to Codex with low friction by translating `CLAUDE.md` instructions into `AGENTS.md` and reusing existing folder structures [27, 28].
    *   **Skills** are reusable instruction packages that encode repeatable processes, while **Plugins** connect the AI to external apps like Google Drive or Gmail [29].

*   **Important Details:**
    *   Codex features **"Goals"** for long-running agent tasks and **"Automations"** for scheduled workflows [30, 31].
    *   **Security guardrails** are essential when using agents; users should grant the "least permission" necessary and maintain human approval for destructive actions [32].
    *   While folder instructions and CLI tools transfer easily between platforms, **MCP (Model Context Protocol)** servers currently require manual re-setup [30].

### **Reducing the difficulty of being a data owner** by Andrew Jones

*   **Main Arguments:**
    *   The difficulty of finding data owners stems from the heavy manual burden of data management, including observability, access control, and retention [33].
    *   Automation through a data platform, driven by the context provided in **data contracts**, is the key to making data ownership sustainable [34, 35].

*   **Key Takeaways:**
    *   Data contracts provide the necessary **context** (e.g., identifying fields containing personal data) that allows a platform to automate complex tasks like anonymization [34, 35].
    *   Reducing the "difficulty" of ownership increases the likelihood of finding responsible owners across an organization [36].

*   **Important Details:**
    *   The "messy, complex boundary" between **Software Engineering and Data Analysis** is often where the most impactful innovations occur [37].
    *   Friction in data teams is often caused by unfamiliar tooling and a fear of "doing it wrong" rather than a lack of skill [38].

### **⚡Satya Nadella: No Priors x Latent Space Crossover Special at Microsoft Build** by Satya Nadella, et al.

*   **Main Arguments:**
    *   Microsoft is positioning itself as a **Frontier Intelligence Platform**, where the goal is for customers to gain more value from the ecosystem than Microsoft captures itself [39, 40].
    *   The "frontier" is no longer just about the largest model; it involves operating at high performance by combining multi-model harnesses, rich context, and private evaluations [41, 42].

*   **Key Takeaways:**
    *   **Private Evals** and traces are becoming a company's most critical Intellectual Property (IP), allowing them to "hill climb" and switch between models without losing control [43, 44].
    *   AI is shifting the value of human capital toward **metacognition and "meta work,"** where humans manage agentic systems (e.g., the "Miles" system for Azure networking) [45, 46].
    *   SaaS is being unbundled; **Work IQ** now exposes Microsoft 365 data as a "database" that agents can query to perform cross-app tasks [47, 48].

*   **Important Details:**
    *   **Token Capital** may eventually appear on corporate balance sheets as a way to capture previously "tacit" knowledge [49, 50].
    *   Pricing models are evolving from per-user subscriptions to include **consumption-based** and eventually **outcome-based** pricing [51, 52].
    *   Microsoft's **MAI models** focus on "clean lineage" and data quality to achieve high reasoning capabilities in smaller parameter sizes [53, 54].

### **🔬Scaling Past Informal AI - Carina Hong, Axiom Math** by RJ Honicky

*   **Main Arguments:**
    *   **Formal verification** is essential for scaling AI brilliance; statistical signals (like RLHF) are insufficient for reaching true AGI [55, 56].
    *   Verified AI allows for "compounding brilliance" by building on solid, proven foundations (axioms) rather than shaky, informal ones [55, 57].

*   **Key Takeaways:**
    *   **Axiom Math** achieved a perfect 12/12 score on the Putnam exam, outperforming top undergraduates and other frontier AI systems [58].
    *   Verification acts as a **stronger reward signal** during Reinforcement Learning; instead of guessing if an answer is good, the system can use a Lean verifier to prove it is correct [59, 60].
    *   The **Verina benchmark** measures a model's ability to generate both code and a formal proof of correctness; Axiom achieved 99%, while OpenAI’s o3 achieved 4.9% [59].

*   **Important Details:**
    *   **Lean** is the language used to meticulously specify mathematical proofs for formal verification [61].
    *   **AXLE** is Axiom’s open-sourced toolkit for manipulating and validating mathematical proofs [61, 62].
    *   The "onus is back on humans" to verify AI outputs as volume increases, making automated verification a critical bottleneck to solve in fields like physics and science [63, 64].