Report from AWS Summit Berlin 2024
The AWS Summit in Berlin 2024 was another great opportunity for learning and networking. As an official AWS event, of course it began with a flashy keynote to bring good news. And it was Big News.
Stefan Höchbauer, Managing Director of AWS Germany and Europe Central, announced the AWS Sovereign Cloud. It’s the EU GovCloud. It may be the foundation of the next 100 years of utility computing. I’m excited about it, even though I won’t get to work with it (only EU residents may operate it), because it allows all European governments to use the same AWS technology that its private companies do, even for the most restricted and regulated workloads.
As I was only able to attend the second day of this two-day affair, I missed the representative of the Brandenburg local government making the same announcement. I bet they are pleased with the billions that AWS will invest to make it happen!
The two-day format was my only major complaint about the Summit, as it made planning to meet with other community builders more difficult. Some attended on Wednesday, others on Thursday. But I understand why they did it like that. The venue was so full that the Community stage was tucked into the corner of a side building that was hard to find and looked like it was in the middle of being rebuilt or knocked down! I was pleased to learn at the end of the keynote that the next host city will be Hamburg, one of Europe’s largest ports, with a fine maritime mercantile tradition. An excuse to visit more of Germany!
I had hope to write more about all the talks I enjoyed, but life got in the way. So the rest of this post shares my notes and thoughts on my favorite of the day.
Sana Shah, Cloud Operations Lead at DEMICON, opened the Community Stage with GenAI Deep Dive: Build a GenAI assistant for a women-in-cloud initiative.
This was probably the best session of the day because how it changed my thinking. I was skeptical before of generative AI services and this changed my mind. Sana maintained a confident delivery to share a clear use case and a practical solution.
Sana started the AWS Women’s User Group in Berlin. Running a user group is a lot of work. If you also need to fit it around the day job, then you need all the help you can get.
Sana showed us how to configure the AWS Bedrock service to act as her personal assistant so that it would answer her questions about the user group and about the global trends in the employment gender gap. She configured the service to combine the basic language skills of a base model (I think it was Anthropic’s Claude) with information from curated documents about the user group and a report from an NGO report about the global trends (if I remember more about these documents I’ll update here).
It turns out that this is an example of retrieval-augmented generation (RAG). The “retrieval-augmented” part just means making the generated responses more useful by getting parts of the answer from a given document.
Educational too was her summary of the economics of engineering solutions using generative technology.
- Prompt engineering ($): Query the base model directly and add any extra data you may need into every query.
- *Retrieval-augmented generation (RAG) ($$)): Configure extra inputs (databases, documents) that the system may consult if the base model has no useful answers.
- Pre-train ($$$): I’m still not sure what this is. Sana’s focus was on RAG, assuming that this stage is complete.
- New model ($$$$): Make a new base model to rival Claude or ChatGPT. Very few people in the world have resources to do this.
What’s the difference between including a PDF in the prompt and making it available via RAG? Query time simplicity and consistency. My analogy in SQL: You can write an SQL query that contains all the source data using CTEs with VALUES
clauses or SELECT ... UNION ALL SELECT ...
segments. But if you write this more than twice you probably want to just store it in a table so that you can just reference the table directly the next time. Where the BedRock system goes further is that it will automatically consult the configured RAG sources.
I would have loved to see what the model answered to a question that fell outside the use case, such as “How is the weather today in Berlin?”. Does the base model know anything about Berlin, or anything about sunshine, or a Thursday in Spring?
Sana ended by showing that Bedrock can not only respond to questions, but it can perform actions on our behalf. A Bedrock Agent is basically a way to invoke a command using a text chat interface instead of a command line interface or a graphical user interface. So Sana instructed her agent to send an email via name and address prompting to the MC, who proudly showed off her new message thanking her for her awesome contributions!
Huge thanks to the “Förderverein AWS Community DACH” for running the Community Stage that made possible this event-within-an-event. The “Förderverein” acts as a “group of user groups” in Germany and promotes the interests of the AWS community across Germany, Switzerland, and Austria. They ran the event smoothly. I haven’t seen them all in action yet, but I believe that it’s one of the best-organized AWS user communities. There’s a lot we can learn from them.