First AI-fokus

27 June 2025 // Ove Lindström

The team behind our favorite Java conference most likely played around with ChatGPT and came up with the idea of having an AI-focused conference. As with all conferences, it focuses around the coffee breaks, so these are my coffee break reflections and highlights.

First off, the keynote Data-Empowered AI Assistants by Michael Hunger from Neo4j. Michael drilled down on the idea that while LLMs are great for helping, they are not that accurate with facts out of the box. They are more keen on helping out than being right. Since Neo4j is a company that lives off creating a nice graph database, they have done a lot of nice stuff that suits the AI arena. What they do is allow companies to create a GraphRAG that takes the LLM from “good but incompetent” to “better accuracy for the domain, but a bit more reluctant to give an answer directly”.

I had a chat with Michael after the talk and he revealed to me that the keynote was supposed to be co-presented with one of their customers/partners, but they withdrew from the talk. That was a pity, since now the keynote became a bit too much about selling Neo4j.

AI agents for the rest of us

We all know that the two groups that love AI the most are developers and investors. But what about the rest of us?

Korey Stegared-Pace from Microsoft used an analogy I really loved. He compared AI agents to Pokemon cards. It actually makes sense to look at them that way. They have different strengths and weaknesses, they can evolve, and combined with other cards, they are amazing. The core idea is that the agents are brilliant for open-ended problems that are too complex to code, due to their adaptability. Like all the personalized travel recommendations and medical diagnoses that can be more precise with the right set of agents.

Often you will have a deck of agents that solve, control, and come to a quorum with each other to find a solution.

Master the Model Context Protocol

But to handle your agent card deck and to allow them to talk to each other, you need a standard way of doing it. Enter MCP, Model Context Protocol. A “USB-C for AI” describes it perfectly. With an MCP server you can connect AI hosts like Claude and Cursor with your “agent client” and have it solve tailored problems. One of the biggest benefits of the protocol is that agents can discover other services available based on attributes. So there we have the Pokemon cards again.

There is a big “however” in this: security. You really need to be careful with what MCP servers you trust and not. And make sure to keep a human in the loop for approvals to prevent unintended actions. They may turn on life support, but not off. The one that was talking about this was Carl-William Ersgård from Meepo. Go and take a look at his presentation. It is one of the best looking presentations I have seen.

Managing AI

From the management side, My Bank Meets AI and AI Do’s and Don’ts for Managers gave us a real-world perspective from the viewpoint of institutions we have learned to trust.

The main message was that AI is a massive disruptor, and while it brings anxiety, it also creates huge opportunities. For organizations, the focus should always be on solving a clear business problem and creating value, not just adopting AI for AI’s sake. (And most of us have gotten the response “I don’t know” from the manager to our question “So what do you want the AI to do?”)

César Soto Valero and Ulf Larsson, both from SEB, stressed the importance of cross-functional collaboration from the very beginning – getting business, tech, legal, and compliance all at the table.

Also, data governance is absolutely critical: knowing where your data is, how it’s classified, ensuring its quality, and having proper traceability is paramount for trust and security. They even shared a sobering example of a loan-granting AI that failed due to bias, highlighting the need for rigorous testing and fairness.

Development Lifecycle

GitLab was at the conference with a booth and Mattias Söderberg did a quickie about Using AI Across the Dev Lifecycle. There were several talks that touched the same area and the consensus is that AI will change the way we work and do development. Tools like Copilot, Cursor, Windsurf (Codium) and so on now let you have a mob programming session on your own and allow you to vibe-code and make changes across many files at the same time.

A key distinction was made between code-centric development (where code is the source of truth, and prompts are disposable) and spec-centric (AI-native) development (where detailed specifications are the source of truth, and AI generates the code from them). Several methods of prompt engineering were presented, but I will save them for a blog of their own.

AI always feels like somebody’s watching me

Finally, for the talk Can AI See You? by Daniel Hirsch from TNG, he talked and demonstrated how multimodal AI models can combine text with other types of data like images and music and re-generate an image from a diffused source. The demos Daniel did were really cool (when they worked).

Key takeaways

1. Strategic alignment and planning: Start AI adoption now. Identify the areas where AI can truly be beneficial, rather than just doing AI. (Ulf Larsson)
2. Skill up on agent management: AI won’t take your job, but the people that can manage AIs will win over those that can’t. (Korey Stegared-Pace)
3. Embrace modularity: AI-native development is here to stay, so focus on creating clear, modular specifications that serve as the source of truth. Use AI to make sure the specifications are cohesive. (Simon Maple)
4. Use the open source, keep it local: Explore the open-source models and tools. Make sure they can be “locally” deployed so that you can get greater control over data and execution. (Cedric Cyburn)
5. Invest in data quality: The SISO principle is very much present when it comes to AI. Invest in high-quality data and proper governance and your AI projects will be more effective. (Ulf Larsson again)

The future is not just about AI helping us; it is about AI doing things autonomously. But the Terminator movies have taught us that trust is everything and there needs to be systems powered by rock-solid data and transparent, well-defined rules.

Youtube playlist

As always, all the talks were recorded. The AI-fokus 2025 Youtube playlist is where you find all the talks.