Here's what caught my eye in the world of AI this week:

1. 🔑 Top Data Themes in 2025: The AI-Driven Shakeup

   Data is the new oil is so passe - AI is rewriting the rules of data in 2025, creating a battle between consolidation in the modern data stack and expansion fueled by AI advancements. Here are the biggest shifts shaping the future of data per Tomasz Tunguz, a leading Data VC who has worked with the Top Data companies in the world.

1️⃣ The Great Consolidation: Simplify or Sink

• Enterprises are reducing tool sprawl and consolidating around major platforms like Snowflake & Databricks.

• CFOs are pressuring data teams to prove ROI, pushing cost-optimized architectures like SQLMesh & Tobiko Data for up to 50% savings.

• Expect mergers and acquisitions as vendors race to own more compute power within these consolidated platforms.

💡 Action: Audit your data stack—are you using too many tools? Look for consolidation opportunities to cut costs and improve efficiency.

2️⃣ Scale-Up Architectures: Local First, Scale Later

• The rise of powerful single-machine computing allows developers to run massive AI models on their MacBooks.

• Workload-specific query engines (e.g., MotherDuck, DataFusion) are gaining traction over traditional scale-out approaches.

• Emerging data storage formats (like Iceberg) are decoupling query engines from storage, optimizing for flexibility and performance.

💡 Action: Experiment with local-first development tools like DuckDB to speed up data workflows before moving to the cloud.

3️⃣ Agentic Data: AI Will Run Your Data

• AI agents will execute the majority of SQL queries, fundamentally reshaping data engineering.

• Data modeling will become critical to prevent AI hallucinations and ensure reliable outputs.

• Data observability tools (e.g., Monte Carlo) will be essential as AI-driven analytics become a core part of enterprise decision-making.

💡 Action: Invest in data modeling & observability—garbage data in means garbage AI results out.

4️⃣ Smaller AI Models Will Win

• Smaller LLMs (10B-70B parameters) are proving nearly as accurate as massive models, but at 600X lower inference costs.

• Enterprises are prioritizing latency, cost, and efficiency, driving adoption of mid-sized AI models.

• AI cost pressures are intense—CFOs are demanding real ROI from AI investments from day one.

💡 Action: Test smaller AI models for enterprise workloads—they’re faster, cheaper, and often just as good.

5️⃣ BI & Data Governance: Centralized Control, Decentralized Access

• Self-service BI remains crucial, but enterprises need centralized governance to maintain data integrity.

• New BI approaches (e.g., Omni) allow both business users and data teams to share trusted, consistent metrics.

• AI-driven data models will act as an ORM for the data stack, enabling structured AI-powered insights.

💡 Action: Balance governance with accessibility—enable self-service BI, but ensure AI-driven analytics remain trustworthy.

Bottom Line

The world of data is evolving fast. Consolidation, AI-driven automation, and cost-efficient architectures are shaping the future. Strategic AI leaders will embrace these changes, simplify their data ecosystems, and align AI investments with real business impact.

Your Next Move: Identify where you can consolidate, optimize AI spending, and leverage new architectures to stay ahead in 2025. 🚀

Quote of the week

2. Alvin Toffler’s insights have profoundly shaped my perspective on the future. He was decades ahead of his time. My take on this quote? No matter how advanced AI becomes, human intelligence will always be the key to true innovation. AI excels at patterns, probabilities, and optimization—but the ability to think differently, challenge assumptions, and make intuitive leaps remains uniquely human. The greatest breakthroughs don’t come from logic alone; they come from bold, unconventional thinking. And for the foreseeable future, this is where AI will struggle—while visionary leaders will thrive.

3. Emerging Patterns in Building GenAI Products

As a strategic AI leader, you don’t need to be in the weeds implementing generative AI (GenAI) systems—but you do need to understand the key patterns shaping their success. This Thoughtworks article is a goldmine for AI architects and engineers navigating the transition from proof-of-concept to production, tackling challenges like hallucinations, non-determinism, and unbounded data access.

Your role is to empower your teams as they push the boundaries of what’s possible—often challenging the status quo in legal, privacy, and security. By understanding these emerging techniques, you’ll be better equipped to advocate for responsible AI adoption, remove roadblocks, and position your organization at the forefront of AI innovation.

Here’s a byte-sized breakdown of the key insights from Martin Fowler’s latest thought-provoking article, where he unpacks the emerging patterns shaping the future of GenAI in production.

1. Direct Prompting

The simplest way to use an LLM, but full of limitations.

• LLMs are only as good as their training data and lack real-time updates.

• They can hallucinate, mislead, or respond with overconfidence.

• Needs enhancements like retrieval augmentation or fine-tuning to be reliable.

Action: Don’t rely on raw LLM outputs—use techniques like Retrieval-Augmented Generation (RAG) or guardrails to ensure reliability.

2. Evals

Evaluating LLM output is crucial but complex.

• Unlike traditional deterministic software testing, LLM evals require scoring mechanisms.

• Methods include self-evaluation, LLM-as-a-judge, and human evaluation (best results come from a mix of these).

• Benchmarking helps track performance over time and assess upgrades.

Action: Integrate automated evals into your development pipeline to maintain model quality.

3. Embeddings

A foundational technique for making LLMs context-aware.

• Transforms text, images, and other unstructured data into numerical vectors.

• Enables semantic search, similarity detection, and retrieval of relevant content.

• More efficient than traditional keyword-based searches.

Action: Use embeddings to structure unstructured data and improve LLM-driven search capabilities.

4. Retrieval-Augmented Generation (RAG)

Avoids fine-tuning costs by supplying relevant context dynamically.

• Retrieves relevant document fragments before generating responses.

• Works best with structured indexing and retrieval mechanisms.

• Improves accuracy and reduces hallucinations.

Action: Implement RAG instead of fine-tuning when dealing with dynamic or domain-specific knowledge.

5. Hybrid Retrieval

A combination of keyword-based and embedding-based search.

• Embedding-based search is powerful but loses semantic nuance.

• Keyword search (TF-IDF, BM25) complements embeddings for better retrieval.

• Used in complex, large-scale information retrieval systems.

Action: Use hybrid retrieval for more precise and relevant document fetching in RAG.

6. Query Rewriting

Reframes user queries to improve search accuracy.

• LLMs generate multiple variations of a query to capture different nuances.

• Helps retrieve better documents when user queries are vague.

• Works well in domain-specific RAG applications.

Action: Implement query rewriting for better document retrieval, especially when users aren’t precise in their queries.

7. Reranker

Filters and prioritizes retrieved documents before feeding them to the LLM.

• Helps avoid “context bloat” by selecting the most relevant fragments.

• Uses deep neural models to rank search results more effectively.

• Improves RAG accuracy and response quality.

Action: Use rerankers to refine search results and reduce irrelevant context in LLM prompts.

8. Guardrails for LLMs

Protects against misleading, harmful, or sensitive outputs.

• LLMs are “gullible” and can be tricked into revealing or fabricating information.

• Guardrails enforce ethical and safety constraints on outputs.

• Essential for enterprise applications where accuracy and safety are critical.

Action: Implement guardrails like moderation APIs and prompt filtering for responsible AI deployment.

Final Thought:

This article showcases how GenAI isn’t just another software extension—it requires a shift in thinking. By combining retrieval techniques, evaluation frameworks, and safety mechanisms, businesses can build AI applications that are accurate, reliable, and scalable.

What You Can Do Next:
✅ Integrate RAG for real-world AI applications.
✅ Set up evals to continuously measure LLM performance.
✅ Experiment with query rewriting and rerankers to improve AI search accuracy.
✅ Implement guardrails to prevent AI misuse.

4. Building CERN for AI

CERN for AI aims to unite top talent and resources to develop trustworthy AI systems that benefit Europe. This initiative seeks to create a hub for advanced AI research while supporting economic growth and security. By partnering with private companies, CERN for AI will ensure its innovations lead to practical applications in society.

Summary of "Building CERN for AI – An Institutional Blueprint"

The "Building CERN for AI" report proposes creating a CERN-like institution for artificial intelligence (AI) in Europe. Inspired by the ARPA model, this initiative would focus on trustworthy general-purpose AI, bridging foundational research with real-world applications. It aims to boost Europe's competitiveness in AI, overcoming issues like fragmented markets, lack of computing power, and insufficient funding.

The institution would follow a two-track research model:

1. ARPA-style decentralized research programs working with academic and industry partners.

2. In-house Focused Research Organizations (FROs) for sensitive AI research requiring security and long-term collaboration.

Governance would balance transparency and accountability, with member state representatives overseeing strategy while allowing independent experts to guide operations. Funding would come from EU and national governments, industry partnerships, and technology licensing.

The legal structure would likely be a Joint Undertaking under Article 187 of the TFEU, ensuring flexibility while maintaining public oversight. The estimated cost for the first three years is €30–35 billion, positioning Europe as a global leader in trustworthy AI.

Key Insights

1. A Moonshot for AI Leadership in Europe

   • Europe risks falling behind in AI due to underfunding, lack of compute power, and fragmented initiatives.

   • A CERN for AI would centralize efforts, mirroring the success of CERN in physics.

2. Trustworthy AI as a Core Mission

   • Unlike commercial AI labs, this institution would focus on interpretability, reliability, and safety.

   • Ensuring AI is transparent and aligned with democratic values is key.

3. Dual Research Model for Innovation & Security

   • ARPA-style programs would allow rapid, open collaboration.

   • FRO-style in-house teams would handle sensitive AI developments securely.

4. Funding and Economic Impact

   • The initiative requires €30–35 billion over three years but could generate long-term economic benefits.

   • Funding sources include EU/national contributions, private partnerships, and licensing revenue.

5. Governance Balances Innovation & Accountability

   • A Member Representative Board (national governments) would oversee high-level decisions.

   • Two independent advisory boards (Mission Alignment & Scaling & Deployment) would ensure AI safety and strategic focus.

6. Strategic International Collaboration

   • The institution would prioritize European leadership but remain open to trusted partners (e.g., UK, Switzerland, Canada).

7. Legal Framework for Agility

   • A Joint Undertaking (TFEU Article 187) would allow CERN for AI to act like a dynamic tech startup while maintaining public accountability.

8. The Urgency of Action

   • AI is evolving rapidly; Europe must act now to remain competitive.

   • A failure to invest could leave Europe dependent on foreign AI systems with misaligned interests.

Final Thought

CERN for AI is not just about catching up—it’s about setting the global standard for safe, reliable, and transparent AI development. With the right funding, structure, and leadership, Europe could lead the next phase of the AI revolution.

Here’s Why Over 4 Million Professionals Read Morning Brew

• Business news explained in plain English

• Straight facts, zero fluff, & plenty of puns

• 100% free

See for yourself

5. AI-Ready Data: The Missing Ingredient for GenAI Success

   This is a powerful exploration of why AI-ready data is the foundation of enterprise success with GenAI. If the scale of transformation feels daunting, take it as a sign that you're on the right path. Building an AI-first company is no small feat—but that’s exactly why those who master it will define the future of business in the 21st century. The opportunity is massive, and the leaders who invest in AI-ready data today will be the ones shaping industries tomorrow. Those who don’t? Well, let’s just say history favors the bold.

   Here’s a quick, digestible summary and action plan based on the key insights:

   .

   Key Takeaways:

   .

   🔹 Most AI projects fail due to bad data – Gartner predicts that by 2026, over 60% of AI projects will be abandoned due to poor data readiness.
   .

   🔹 GenAI needs more than just "clean" data – It requires context-rich, diverse, and continuously governed data to work effectively.

   .
   🔹 The right foundation can save costs & drive innovation – Organizations prioritizing AI-ready data could cut manual data management costs by 20% annually and unlock 4x more AI use cases.

   .

   The Four-Step Playbook for AI-Ready Data:

   .

   1️⃣ Define AI-Ready Data: Assess your current data strategy against the needs of GenAI. Not all "clean" data is AI-ready.

   
   2️⃣ Get Executive Buy-In: Frame AI-ready data as an ROI driver, not just a technical requirement. Use metrics like faster decision-making and cost savings to justify investment.

   
   3️⃣ Implement & Scale: Move beyond legacy data management. Adopt flexible architectures like data fabrics, active metadata, and automated context embedding.

   4️⃣ Govern AI Data Smartly: Automate governance with policy-driven frameworks and active monitoring to ensure compliance, transparency, and AI reliability.

   Your Action Plan:

   .

   ✅ Assess your data landscape – Identify gaps that could derail AI success.

   .

   ✅ Align with leadership – Tie AI-ready data to business goals and risk reduction.
   .

   ✅ Upgrade data management – Invest in metadata management, observability, and automated reasoning to reduce AI model failures.
   .

   ✅ Govern at scale – Implement automation to reduce compliance effort by 90% while keeping AI outputs explainable.

   .

   AI success isn't just about models—it’s about the right data. What’s your plan to make AI-ready data a priority?

6. Sunday Night AI Experiment: Bringing “Data is the New Oil” to Life

   Sometimes, the best way to explore AI is to experiment with it. Last night, I decided to do just that—starting with an idea: What would a futuristic version of “Data is the New Oil” look like?

   .

   🎨 Step 1: DALLE – I started with by adding two images, my AI Quick Bytes logo for tone and color palette and an image of an Oil well. I played around with prompts, tweaking details to get the right futuristic vibe. A glowing data refinery? Digital oil rigs? A few iterations later, I had something interesting.

   .

   🎞️ Step 2: Sora – Next, I tried animating it. Adjusting scenes, testing different movements—it turned into a deep dive that stretched over an hour.

   .

   ⏳ Step 3: Wrap-Up – Eventually, I had to stop (the newsletter wasn’t going to write itself!).

   .

   🚀 Takeaway:

   .

   AI tools like ChatGPT, DALLE, and Sora are meant to be explored. The more you experiment, the more you learn what’s possible.

   .

   👉 Try it yourself: Pick a concept, create an image with DALLE, and see what happens when you bring it to life in Sora. You might surprise yourself!

7. Goldman Sachs CEO says that AI can draft 95% of an IPO prospectus in minutes

Winning the AI corporate battles isn’t just about technical expertise—it’s about mastering complexity. While subject matter expertise is crucial, many leaders underestimate the sheer challenge of implementing, maintaining, and transforming their organizations with AI.

This transformation isn’t a siloed IT project; it’s an enterprise-wide revolution that demands cross-functional collaboration and bold, empathetic leadership. Success in the AI era won’t come from technical prowess alone but from human-centric leadership that wields AI as a force multiplier—empowering, not replacing, the workforce. The real winners will be those daring enough to reinvent themselves as adaptive, continuously learning organizations.

Learn how to make AI work for you

AI won’t take your job, but a person using AI might. That’s why 1,000,000+ professionals read The Rundown AI – the free newsletter that keeps you updated on the latest AI news and teaches you how to use it in just 5 minutes a day.

Sign up to start learning.

8. A picture is worth a 1,000 words!

Check out Avi Chawla posts on RAG and you will become an expert in no time!

Until next time, take it one bit at a time!

Rob

Thank you for scrolling all the way to the end! As a bonus check out Nvidia State of AI in Financial Services.

P.S.

💎 Discover Handpicked Gems in Your Inbox! 💎

Join thousands of satisfied readers and get our expertly curated selection of top newsletters delivered to you. Subscribe now for free and never miss out on the best content across the web!

💎 Discover Handpicked Gems in Your Inbox! 💎

Here are some FREE newsletters our readers also enjoy. Explore, it helps us keep our newsletter FREE is much appreciated and it will help feed your head !