🧳 On my way back from yesterday’s “Amazon Ad Tech Summit 2025” in Munich. Looking back at a packed day with loads of insights, new connections and reunion with dear colleagues.

💡 The ability to derive insight from data is driving revenue and Generative AI in Advertising and Marketing is an actual thing 🙃 . Learned also a lot about the offering of Amazon Ads and it was absolutely fascinating to hear all the value customer’s feedback in the breaks and during the panels.

Last week I had the pleasure of joining Zalando and the Zalando account team at Amazon Web Services (AWS) for a joined community event. During the opening keynote it was coined as “bringing re:invent home” and that is how it felt indeed. I loved the vivid discussions, engineers rolling up their sleeves and dive deep and the inspiring talks from customers and colleagues alike.

Like always it was a pleasure to present alongside John Mousa, which feels like an invite to his living room for having a good conversation. Together we dived into how one can evolve prompt trial-and-error pain towards a sound engineering approach as a backbone for building production-grade applications utilizing the power of generative AI. On the way we touched on some relevant launches in Amazon Bedrock from last re:invent.

Only made it through 2 out of the 5 hours of this insightful podcast between Lex Fridman, Dylan Patel and Nathan Lambert where the dive deep into Deep Seek’s AI model architectures, highlight and explain innovations. Later they dive into politics and good insights into chip factories on the back of US export regulations.

Highly recommend podcast - it will take a while though to bump up my training volume until I can consume those in just a single run. But you can consume it in bites :)

👀 Just watched the interview of Anthropic CEO Dario Amodei at WSJ Journal House Davos. Highly recommended! While obviously not mentioning any exact targeted launch dates for features, he gave some good insight in his thinking of policies over politics, safety and navigate-able challenges for the future of work. Mindset of collaboration over of replacement. 30 mins of time well invested.

A few things I noted: 🗺️ Product Roadmap • Web browsing capabilities coming in next 3-6 months • Memory features and virtual collaborator functionality in development • Voice interaction planned for future releases • Focus remains on enterprise solutions while expanding consumer offerings 🏗️ AI Development • Anthropic’s revenue grew 10x in 2024, reaching ~$1B • Partnership with Amazon to deploy hundreds of thousands of Trainium 2 chips • Novel approach to reasoning capabilities, focusing on continuous improvement rather than separate models 📈 Future of Work • AI expected to match or exceed human capabilities in most tasks by 2027 • Focus on complementary AI deployment rather than replacement • Emphasis on comparative advantage - humans leveraging the 10% AI can’t do 👓 Industry Perspective • Call for serious dialogue about AI’s societal impact • Strong focus on responsible scaling and security testing • Bipartisan approach to AI policy and regulation

📖 “Building your own RAG system is like deciding to build your own email server in 2024. Sure, you could do it. But why would you want to?” - Alden Do Rosario in his article “Dear IT Departments, Please Stop Trying To Build Your Own RAG” (https://lnkd.in/ep9ZNJzq) on medium. Love it. Highly recommended read.

🎡 Don’t reinvent the wheel! The trap of building something, which on the first glance looks so simple, but then we you get into it you discover layers of hidden complexity.

⚖️ In my conversation with customers, I don’t get tired in highlighting how important it is to decompose application’s needs into the different use cases and choose the optimal fitting foundation models per use case to optimise functional fit, cost and frugal resource utilisation.

🔐 This is all good, but obviously increases the complexity of the overall solution slightly and leaves us with the need to define guardrails across different foundation models to secure our applications . Here Amazon Bedrock guardrails come to the rescue. The allow to define Guardrails spanning across different foundation models. Directly integrated into Bedrock are just integrated via API in to your application. In my re:invent recap(https://lnkd.in/dkqPBQi3), I highlighted how new functionality in Amazon Bedrock Guardrails makes this approach even more powerful. But a Markus points out - it also got now so much more affordable. So no excuses 😉

🎡 So, re:invent 2024 is in the books. Just did my final (famous last words 😉 ) edits in my personal overview article trying to touch on all the significant announcements around (Gen) AI. For me it was an experiment on how to keep up-to-date and turned out to be a good approach. I hope you as a reader could also take something away from this.

👀 Now it is time to lift the head again and dive deeper into the stuff outside of the Gen AI bubble…

3️⃣ Good Morning! Third Day 1 of re:invent is in the books. It took a second coffee this morning as Swami was on fire in his yesterday’s keynote. But, I updated I updated my All things (Gen) AI article with the most important (just my view) announcements.

🎡 Flexibility of Model choice is key for the success of your Gen AI applications. With Amazon Bedrock Marketplace the choice got significantly broader. Another crucial aspect of the success is performance: both in terms of latency and cost. Again new features in Bedrock easy your optimisation efforts. Easing the integration into structured, both relational and graph databases, and building automated intelligent document processing empower the successful architecture pattern of RAG even more. But have a look into the article 🙂