• Deepfakes are not just fake videos. They are part of a broader challenge around how Generative AI is reshaping trust both in the content we share and the information we consume.

• The real risk isn't only technical. As AI-generated content becomes indistinguishable from reality, journalism, elections, humanitarian work, businesses, and everyday communication all depend on new ways of establishing trust.

• Technology alone won't solve this. Watermarking, detection, and metadata help, but lasting solutions require collaboration between AI companies, policymakers, researchers, civil society, businesses, and users.

• AI also influences what we believe. Beyond creating fake media, conversational AI can subtly reinforce biases and shape opinions at massive scale, making AI literacy increasingly important.

• My biggest takeaway: The future of AI governance isn't just about regulating models but it's about building an ecosystem where trust, transparency, and collaboration evolve as quickly as the technology itself.

"Seeing is believing" has been one of humanity's default assumptions for centuries. Today, that assumption is becoming increasingly unreliable.

Last night I attended a fascinating panel discussion on deepfakes, bringing together experts from humanitarian organizations, AI ethics, industry, and European policy. While the conversation focused on synthetic media, what stayed with me afterward wasn't a technical question but this:

When did we stop talking about deepfakes and start talking about Generative AI?

The two terms are not interchangeable, but I wonder if our excitement around Generative AI has made us forget that many of the concerns we once associated with "deepfakes" are now embedded in the broader AI ecosystem.

Deepfake is a subset of Generative AI but the word still matters

Technically, a deepfake is a specific application of generative AI: using AI to create or manipulate images, audio, or video to make something appear authentic when it isn't.

Generative AI is much broader. It writes text, generates code, creates music, produces images, and increasingly acts as an assistant, tutor, or companion.

But I think there are two ways to think about "deepfakes."

The first is the one most people associate with the term: content generated to deceive others. A fake political speech, a manipulated video, non-consensual intimate images, or fraudulent voice cloning.

The second is less discussed.

As we increasingly interact with AI systems every day, we also consume information generated by them. These systems often communicate with remarkable confidence, even when they're wrong. The content isn't necessarily malicious, but it can still create a convincing version of reality that doesn't exist.

In that sense, there is another kind of deep fake: information that feels trustworthy simply because an AI generated or presented it convincingly.

This isn't about fake videos or manipulated images. It's about the subtle way conversational AI can shape what we believe over time.

Research on human–AI interaction suggests that people naturally anthropomorphize AI systems, assigning them human characteristics such as competence, expertise, or even empathy (see Blut et al., 2021; Cohn et al., 2024). The more human-like an AI appears through its language or voice, the more likely people are to trust its answers, sometimes beyond what its actual reliability deserves.

That becomes more concerning when these systems reflect biases in their training data or conversational behavior. Unlike a traditional search engine that presents multiple sources, a chatbot often synthesizes a single, coherent answer. Even if that answer is only subtly biased, repeated interactions can gradually reinforce certain perspectives while making alternative viewpoints less visible.

The implications become even greater for the next generation. Many children growing up today will not experience AI as a novel technology. They will experience it as a constant companion: helping them learn, answering their questions, assisting with homework, and eventually influencing how they work and make decisions. For them, AI may become a primary interface to knowledge rather than simply another tool and for some even their best friend who gives them the feeling that understands them "better than anybody else".

That raises an important question: what happens when the first source of information for millions of people is a system that confidently summarizes the world instead of exposing them to multiple perspectives? If its biases—however small—are repeated billions of times across years of interactions, they could shape not only individual opinions but also how future generations develop critical thinking, curiosity, and their understanding of truth itself.

This isn't necessarily propaganda, nor is it evidence of intentional manipulation. But at the scale of hundreds of millions of daily interactions, even small systematic biases or tendencies toward overconfidence could influence how societies form opinions, whom they trust, and how public discourse evolves.

Recent research from the University of Washington (see Fisher et al., 2025) found that after only a few interactions, people tended to shift their opinions in the direction of politically biased chatbot responses. Interestingly, participants with greater AI literacy were significantly less susceptible to this effect, suggesting that education may become one of the most effective safeguards against AI-driven influence.

This second angle wasn't the focus of the panel, but it kept coming back to my mind throughout the discussion. The challenge isn't only what we share with others; it's also what we gradually learn to trust ourselves.

The real problem isn't fake media but the erosion of trust

One recurring theme throughout the discussion was that deepfakes are no longer just a technical problem.

They're becoming an infrastructure problem for society.

When synthetic media can convincingly imitate reality, every institution built on trust starts to feel the impact:

• journalism

• elections

• humanitarian work

• education

• businesses

• financial fraud

• legal systems

• even everyday family communication

One speaker shared a quote often attributed to wartime reporting:

"The first casualty of war is truth."

In a world of synthetic media, truth isn't only threatened during conflicts. It becomes harder to establish in everyday life.

Imagine trying to verify a ceasefire announcement, a humanitarian message, an emergency request from a family member, or even whether a government website is genuine. Once people lose confidence in what they see and hear, the cost extends far beyond individual scams.

Trust itself becomes scarce.

The scale is already larger than many people realize

One statistic from the discussion particularly stood out.

Researchers from the Oxford Internet Institute identified nearly 35,000 publicly downloadable deepfake model variants, downloaded almost 15 million times since late 2022. Many of these models are specifically fine-tuned to generate images of identifiable people, with 96% targeting women.

The study also showed how accessible this has become: some models can be created using as few as 20 images, consumer-grade hardware, and around 15 minutes of processing time.

The concern is no longer whether this technology exists.

It is already widely available.

Technology alone won't solve technology

When discussions around deepfakes begin, they often end with the same proposed solutions:

• Watermarks.

• Detection models.

• Content authenticity.

• Metadata.

These are all important but none of them is sufficient on its own.

• Watermarks can be removed.

• Detection models improve, but so do generation models.

• Authentication standards help, but only when adopted consistently.

This isn't a race that has a single technical finish line.

Instead, detecting manipulated content increasingly requires combining technical signals with contextual information, provenance, metadata, and trusted sources.

Regulation is necessary but difficult

The conversation also highlighted why regulation struggles to keep pace.

Technology evolves faster than legislation.

By the time policymakers understand one generation of AI capabilities, the next generation has already arrived.

That doesn't mean governments are inactive. As the EU AI Act continues to come into force, transparency has become one of its central principles. More recently, the European Commission published the Code of Practice on Transparency of AI-Generated Content, providing practical guidance on how providers and deployers can label AI-generated and manipulated content and comply with the Act's transparency obligations. While the Code is voluntary, the underlying transparency requirements are legal obligations under Article 50 of the AI Act.

But regulation alone cannot solve the problem.

Some harms already fall into legal gray areas. Others span multiple jurisdictions. Meanwhile, the companies developing these systems operate globally while governments legislate nationally or regionally.

Another interesting observation was that regulation itself has become increasingly geopolitical. The discussion is no longer only about what is ethically desirable, but also about competitiveness, strategic interests, and global influence.

That makes the challenge considerably harder.

Responsibility cannot belong to one group

Perhaps my biggest takeaway wasn't about AI at all.

It was about collaboration.

Throughout the discussion, it became clear that every stakeholder depends on the others:

• AI companies build the models.

• Researchers study the risks.

• Civil society understands the human impact.

• Industry decides which safeguards to adopt.

• Governments create legal frameworks.

• Buyers determine which technologies succeed in the market.

• Humanitarian organizations experience the consequences where trust is already fragile.

None of these groups can solve the problem independently:

• Technical experts need policymakers who understand implementation realities.

• Regulators need researchers who can explain emerging risks and what tools can be used to mitigate them.

• Companies need practical standards instead of vague expectations.

• Users need clearer ways to understand when they are interacting with synthetic content.

The conversation shouldn't happen in separate rooms.

It needs to happen around the same table.

The question isn't whether AI is good or bad

Leaving the event, I found myself returning to the question that started all this.

Has "deepfake" simply become the forgotten name of Generative AI?

Not exactly.

Deepfakes are one application of Generative AI, and an important distinction remains. Most generative AI tools are built for legitimate and often valuable purposes.

But the discussion reminded me that the underlying issue has never really been about the technology itself.

It's about trust.

• How do we preserve trust when creating convincing synthetic content becomes almost effortless?

• How do we help people distinguish authenticity from fabrication?

And perhaps the more uncomfortable question:

• How do we avoid becoming so accustomed to AI-generated information that we stop questioning it altogether?

Those are not questions that engineers, lawmakers, or companies can answer alone.

They are questions that society will have to answer together.

Acknowledgements

This post was inspired by a panel discussion hosted by Merantix Momentum featuring:

• Philippe Stoll (Swiss Red Cross / Red Cross Red Crescent Hub on Harmful Information)

• Alexandra Wudel (FemAI)

• Alessandro Polidoro (European Commission)

• Thomas Wollmann (Merantix Momentum)

• Hasna Najmi (Merantix Momentum)

The article reflects my own interpretation and takeaways from the discussion plus recent research on deepfakes, rather than a transcript of the panel.

References

• Oxford Internet Institute. Dramatic Rise in Publicly Downloadable Deepfake Image Generators.

  • Original paper: Hawkins, W. et al. (2025). Deepfakes on Demand: The rise of accessible non-consensual deepfake image generators.

• EU AI Act

• European Commission. Code of Practice on Transparency of AI-Generated Content.

• Blut, M., et al. (2021). Understanding anthropomorphism in service provision: A meta-analysis of physical robots, chatbots, and other AI.

• Cohn, M., et al. (2024). Believing Anthropomorphism: Examining the Role of Anthropomorphic Cues on Trust in Large Language Models.

• Fisher, J., et al. (2025). Biased LLMs Can Influence Political Decision-Making.

The event brought together representatives from Qdrant, Haystack (deepset), Cognee, LlamaIndex, and n8n for a keynote and panel discussion about a question many of us are currently wrestling with:

As agents become more capable, what happens to retrieval?

The room was packed, with more than a hundred attendees and enough audience questions that the organizers couldn't get through all of them.

What surprised me most wasn't the discussion about vector databases or agents. It was how often the conversation came back to something much simpler:

Most retrieval problems are actually thinking problems.

TL;DR

• The biggest takeaway wasn't about vector databases or agents, it was about asking better questions. Start from constraints and trade-offs, not technologies.

• The retrieval ecosystem is becoming more specialized rather than more consolidated.

• Qdrant is positioning itself as retrieval infrastructure, including offline and edge retrieval with Qdrant Edge.

• Haystack by deepset is leaning into orchestration, observability, and enterprise-grade AI workflows.

• Cognee treats agent memory as a dedicated systems problem rather than a prompting trick.

• n8n argues that visual workflows remain valuable even in the era of AI-generated code because they are easier to understand, maintain, and share.

• LlamaIndex is increasingly focused on document understanding, OCR, and preserving document structure rather than treating everything as plain text.

• The emerging AI stack looks less like one platform doing everything and more like specialized layers for retrieval, orchestration, memory, automation, and document intelligence

The best retrieval advice wasn't about retrieval

The keynote by Qdrant CTO Andrey Vasnetsov contained my favorite takeaway of the evening.

He showed how people often ask technology questions too early.

Instead of asking:

Is grep all I need?

Ask:

Do I need indexing?

And even better:

What is the slowest and most expensive part (of my system) which might benefit from indexing?

Similarly:

Instead of asking:

Should I put my vectors into relational database?

Ask:

What guarantees do I need from vector store?

Or better yet:

What trade-offs am I ready to make for better performance?

The pattern appeared repeatedly throughout the event.

The most useful discussions weren't about choosing tools. They were about understanding constraints.

Retrieval architecture decisions become much easier once you understand what you're optimizing for: latency, cost, accuracy, observability, deployment constraints, or developer productivity.

The retrieval ecosystem is becoming more specialized

A few years ago, many retrieval products were trying to do everything.

What stood out during the panel is that each company now seems increasingly comfortable occupying a narrower role in the stack.

Qdrant: Retrieval infrastructure

Qdrant's positioning is straightforward: be the retrieval layer.

One particularly interesting area is Qdrant Edge, which enables retrieval scenarios that can run without a permanent internet connection, even on constrained hardware.

The implication is bigger than it sounds.

As AI moves into robotics, wearables, industrial devices, and edge computing environments, retrieval can no longer assume access to a cloud-hosted vector database.

The challenge then shifts elsewhere: embeddings.

Even if retrieval can run locally, query embeddings still need to be generated using a model compatible with the indexed data. In many edge scenarios, the embedding model becomes the real bottleneck rather than the retrieval engine itself.

Haystack: Orchestration with enterprise concerns

Haystack's positioning felt increasingly close to the orchestration layer often associated with LangGraph.

When I asked about the comparison, the answer wasn't that they were trying to be different from workflow frameworks. They acknowledged the overlap and emphasized areas such as:

• enterprise readiness

• observability

• traceability

• integrations with tools like Langfuse and OpenTelemetry

One interesting point raised during networking was that Haystack's open-source framework and enterprise offering are closely connected. The team sees reliability and quality as strategic priorities because improvements in one directly affect the other.

Cognee: Memory as a first-class problem

Cognee had perhaps the most distinctive perspective of the evening.

Their argument starts from a simple observation:

• A single markdown file works surprisingly well as memory for a simple agent.

• But that approach breaks down once you introduce multiple agents, multiple sessions, and long-running workflows.

Cognee approaches memory as a dedicated layer rather than treating it as a side effect of prompting.

They also frame memory through a cognitive-science lens, attempting to model different forms of memory rather than reducing everything to retrieval from a vector store.

Whether that becomes the dominant approach remains to be seen, but it highlights a broader trend: memory is becoming its own category rather than a feature inside agent frameworks.

n8n: The case for visual workflows

One question I asked was whether tools like n8n remain relevant when increasingly capable coding agents can generate applications and automations for us.

The answer was practical rather than philosophical.

Generated applications are still software that somebody has to maintain.

Visual workflows provide:

• transparency

• repeatability

• easier debugging

• easier collaboration

A workflow built in n8n can be inspected and modified by the broader team, while a custom application generated by an AI coding assistant may become another piece of software debt.

Their MCP tooling also points toward an interesting future where agents help create workflows rather than replace workflow systems altogether.

LlamaIndex: Structured documents still matter

LlamaIndex has increasingly focused on document understanding and analysis.

A key message from their representative was that not all documents should be treated as plain text.

If structure matters, retrieval alone may not be enough.

Tables, forms, invoices, reports, and other complex documents contain relationships that can be lost when reduced to chunks and embeddings.

Their focus on OCR and document analysis reflects this shift toward preserving and exploiting document structure rather than flattening everything into text.

The most honest answer of the night: "It depends"

During the panel, one question was whether agentic retrieval is actually reliable.

The answer, unsurprisingly, was:

It depends.

What was refreshing was that nobody tried to pretend otherwise.

The panelists represented different parts of the ecosystem, but there was broad agreement that reliability depends heavily on:

• the use case

• the evaluation strategy

• the consequences of failure

• the quality of the retrieval pipeline

In other words, "agentic" is not a substitute for engineering.

Evaluation is still the bottleneck

The strongest consensus across the panel emerged around evaluation.

Several themes appeared repeatedly:

• Don't rely solely on LLM-as-a-judge.

• Good evaluation requires good prompts and good datasets.

• Start with domain experts.

• Incorporate real user feedback as soon as possible.

• Continuously grow benchmark datasets.

One observation resonated with my own experience:

• LLMs can help generate evaluation data, but in professional domains they are often useless.

What I left thinking about

The title of the event was Retrieval in the Age of Agents, but I left with a different impression.

For all the excitement around agents, memory systems, vector databases, and orchestration frameworks, the most recurring message of the evening was surprisingly old-fashioned:

There are no free lunches.

Every discussion eventually came back to trade-offs.

• Better retrieval often means more infrastructure.

• Edge deployment introduces model constraints.

• Agent memory creates new complexity.

• Evaluation requires human effort.

• Flexibility often competes with reliability.

• Today's architectural shortcut may become tomorrow's migration project.

Even the keynote repeatedly reframed technology questions into optimization questions.

Not:

Should I use X?

But:

What am I optimizing for?

That may be the most useful mindset for building AI systems today. The tools will continue to change rapidly. The trade-offs will remain.

Additional Resources

Event Resources

• Qdrant YouTube Channel - The keynote and panel recordings are expected to be published here in the coming weeks.

• Keynote Slides + Panel Questions

People to Follow

• Andrey Vasnetsov: Co-Founder of Qdrant, CTO.

• Bilge Yücel: Senior DevRel Engineer at Haystack.

• Clelia Astra Bertelli: Member of Technical Staff at LlamaIndex.

• Marcel Claus-Ahrens: Automation Expert.

• Vasilije Markovic: Founder at Cognee.

• Evgeniya Sukhodolskaya: Senior Developer Advocate at Qdrant

This post walks through my story to build a shower meter mobile app using coding AI Agents and sharing what I learned along the way.

👉 The app is open sourced. If you first want to try the app, feel free to skip straight to the end.

👉 Recently I publisehd a blog post here for the PWA version so you can follow up on this there.

A couple of times in my life, I tried to start developing a mobile app, but I always failed. I thought I knew Java, so what could stop me from building an Android app? But setting up Android Studio on my old laptop, going through lots of learning material, and hoping it might get easier in the future were some of the reasons I stopped halfway through.

Now, after years of advances in AI-assisted development, I started wondering: maybe this is my chance. When a simple app idea came to mind (Smart Shower Meter), I said to myself, now is the time to experience vibe coding in a territory where I have no clue... mobile app development!

But...

• Will I learn anything if I just do vibe coding?

• Will the outcome be useful and maintainable?

• Should I choose assisted development instead and learn along the way?

• How realistic is that?

Below, I want to share my experience… come with me!

Problem Statement

For a long time, I’ve wondered whether my shower time is too long and if I could make it shorter. Obviously, the first step was to measure it, to understand the current state and track whether I was actually improving, and by how much.

Solution Space

At first, I considered building/buying an IoT device for accurate measurement, but that felt costly and like overengineering.

Then I thought about writing a Garmin app and using my smartwatch. That sounded plausible, but it would mean wearing my watch all the time.

A simple mobile app seemed like a better option. However, if I had to touch it with wet hands to pause the timer, that was a deal-breaker. So I thought: what if I used the microphone to make a “smart” app that automatically pauses when the water tap is turned off?

I started searching on Google Play… surprisingly, no results. Perfect!

Initial Setup

I wanted to do everything for free. Since we’re still living in the golden age where powerful AI tools are freely accessible, I wanted to leverage that opportunity before it’s gone.

I mainly used the free version of GitHub Copilot Agent (which currently uses Claude Haiku 4.5, GPT-4.1, GPT-4o, GPT-5 mini, and Raptor mini; almost randomly) in VS Code, and sometimes the free, non-signed-in version of ChatGPT for more general questions and guidance.

Soon, I realized that to build something quickly and cross-platform, the Flutter framework (based on the Dart language) was the right choice, something I had absolutely no clue about, which made it perfect for this experiment.

Experience

Trying to build an app in a new language using AI agents is fun, but why not turn it into a learning opportunity? Instead of just accepting the results, I started questioning them and trying to understand the output by quickly reviewing the code.

I wanted to keep my brain awake and active; otherwise, the learning aspect would be lost. Finally, I managed to build the first MVP (Minimum Viable Product) in two days, though it consumed all of my Copilot free usage.

Below, I’m listing some learnings and tips so that next time we can do it more efficiently.

Bonus: I also documented one of the prompting experiences and related issues in more detail in this PR.

Tips

• It's better to always create a new chat flow sooner rather than later… the agent's performance easily reduces due to the larger irrelevant context window, which biases the agent's output!

• Never allow automatic command execution!

  • Several times, I noticed the agent was about to execute nonsensical commands, so it's better to never allow it automatically, or you'll end up in situations that require more work to clean up the mess.

• Edit commands before allowing them.

  • Don't forget you can edit the commands, so do it when they're wrong.

• Check the changes to see how meaningful they are.

  • At least have a quick review of the changes before approving. By reverting the wrong changes, I could save some time.

• Don't forget to add tests, but instruct the agent to create meaningful tests based on your final logic (not just more meaningless ones)!

• Clearly specify which files might need changes. It can assist the agent, though it's not always necessary.

• Consider the solution before instructing, and only let the agent handle the syntax and code details that you don't need to know in depth but can easily test.

• Search and try to understand any new concept you encounter to keep learning and avoid accepting things without question!

Points when you are a developer

• Naturally, you know what to ask even in a new language and framework, like a designer who doesn’t need to know a new tool by heart.

• I could easily set up CI/CD and improve project quality, which is a common requirement that every developer should ask for (I hope so).

• You know if the agent has changed something in the test setup, you need to change the workflows!

• The agent could easily be confused handling the closing brackets of a large component it has built and couldn’t fix it in a few iterations, but for me as a developer, it was a quick fix!

  • I mean, of course, bigger models might be able to fix it, but just a comparison to give a point to myself :)

Challenges with flutter

• In some cases, it wasn’t fully cross-platform and needed special care for each platform, which even confused the agent, like when using the microphone output or storing data locally.

• Improving the design in the specific way I want might be a bit challenging, so for now, I'm sticking more to the framework standards.

• The developer tool in the browser doesn’t work out of the box.

  • There is a specific dev tool see here

Surprises

• I thought auto-detecting water would be the most challenging part, but aside from getting the mic to work on all platforms, the math was quite simple and effective without any advanced ML or frequency analysis. Just checking for average dB with some basic math was sufficient.

Caution

• Never share any personal data with AI agents.

• Don’t use the free models for confidential or private code; they are best for public code.

Did I Learn Anything?

Well... yes and no: I learned Flutter's basics, project structure, and some of the cons and pros. I didn't learn the details of the possibilities and fundamentals of Flutter and its design system. But now I know what is missing, and maybe next time I will start by using AI or better materials to first learn the basics and fundamentals and then begin AI-assisted development using AI agents.

Try it Out!

Now, if you're like me and interested in measuring everything, and want to try out this shower meter, feel free to give it a try:

• How to install it in your Android phone

• How to use it (which I hope the app is intuitive enough 😉)

Next

👉 Recently I publisehd a blog post here for the PWA version so you can follow up on this there.

A few weeks ago, I was in a fireside chat with Bob van Luijt, CEO & co-founder of Weaviate, where someone asked him about Knowledge Graphs. He ended his answer by saying, “I think that knowledge graphs are gonna cease to exist.”
(You can watch it here)

Background

I first got to know about Knowledge Graphs (KGs) during my master’s thesis when I was working on information retrieval and question answering.

Back then, there were no good answer generator models, and our hope was to use semantic relationships to generate proper answers based on retrieved information.

It was fascinating to me how cool it is to model, and even visualize, knowledge in a graph. There are plenty of ways to look at KGs and use them differently, but what they all have in common is modeling ontology using relationships that are easy to understand and interpret.

Progress

Then, with better embedding models and large language models coming into the picture, we now have all (or most) human knowledge in multidimensional spaces.

To me, it’s kind of a similar concept: modeling knowledge. However, in this new way, the vectors are not tangible, and even looking into the details of transformers doesn’t reveal everything. We still have a way to go, on one hand, to improve adaptability, and on the other, to improve explainability and reliability.

Today

So far, KGs are one of the best ways to transparently model human knowledge in an open-box, tangible way.

So when we talk about adaptability, reliability, and explainability, KGs are like the graphed version of RAGs (Retrieval Augmented Generation). When we combine KGs with LLMs or agents, we can get the best of both, to create more reliable agents.

Learning more about how to do that (and other relevant materials) can be found here.

Future

Our (I mean humanity’s) limited knowledge is one thing, but how we model it is another. We might continue using KGs, or we might replace them with newer technologies and, in a few years, never hear about them again.

But I hope that only happens if the newer tech is explainable, trustworthy, and reliable enough.

Development

• LangChain

  • What is LangChain? (YouTube video) (8mins)

• MCP

  • MCP: Build Rich-Context AI Apps with Anthropic

• Knowledge Graphs

  • Knowledge Graphs for AI Agent API Discovery

AI Ethics

• PluralSight

  • Generative AI Foundations: Ethics, Issues, and Limitations of Generative AI (28mins)

• LinkedIn

  • Compliance and Regulations for Generative AI (40mins)

  • Foundations of Responsible AI (2h 30mins)

• EU AI Acts

  • The Finalised EU AI Act: Implications for Businesses, Engineers and Entrepreneurs (1h 26mins)

  • EU AI Act Explained: Everything You Must Know (11mins)

Blogs

• Lessons Learned While Embedding GenAI in a Product

Disclamer

When you are reading this article, some of this learning might seem obsulute or so obvious, but for beginners or past-us, it was something to be learned!

Intro

There are many different ways to embed GenAI in a product or application. It can be as famous as a chat user interface or just a normal button, but behind the scenes, sending some data to a large model and reusing the output in complex ways, then as an outcome (response), just showing normal data in a table, chart, or just a simple change in the UI.

In most of the use cases, there are some common or similar practices. So from one use case to another, we can build on top of our learnings, and sharing learned lessons can help us to grow faster.

Lessons

• July 12th, 2024, lessons from an internal project using SAP AI Core, thanks to Mathis and Sherri for their mentorship and support.

  • Use the output model for enforcing output structure (e.g., JSON) and add a description to each field.

    • Using Instructor: Structured LLM Outputs

    • More info in this Medium blog post

    • Note: Based on this it seems BAML is a better way to go

  • Minimize the requested output as much as possible to get the response faster and cheaper.

  • Set the scene with roles and use system, user, and assistance roles when it makes sense.

    • system role's prompt is more restricted than user and is usually used to set the scene.

  • Minimalistic prompt: try to define prompt as minimal as possible.

    • Don't confuse the model with unnecessary additional info.

  • Use a structured prompt. Markdown, XML, or even JSON!

  • Don't give much freedom to the user if they don't know the underlying code; the functionality can be easily broken, and when it 50% doesn't work, the user gets frustrated!

    • No setting high expectations for the user

  • A large output usually has more hallucinations, so we can't rely on the whole output size.

Similar Blogs or Other Learning References

• OpenAI Prompt Engineering Best Practices

• Prompt Engineering Best Practices from Digital Ocean

Reference

• Cover image generated by Canva

Why call them a true-AI startup? Many so-called AI companies exist, but a true AI company—in my opinion—must have a team of AI/ML/data scientists who are actively pushing the boundaries, which is completely true about Jina AI.

The primary product of Jina is an #open_sourced embedding model that has outperformed the competition despite being more #sustainable (smaller model size and less training cost). Although it is a continuous competition with ups and downs, they are clearly showcasing how to be more sustainable in AI while being more reliable.

Another highlight from the event was the presentation from our AI Development Expert, Mathis Börner. In it, he provided a nice overview of #SAP Gen AI projects, showing how we are comprehensively playing a role in the field to provide #relevant, #reliable, and #responsible AI-powered services while preparing for future innovations.

Hope all the best for Jina AI and all true-AI companies which are shapping a sustainable and ethical AI-powered future.

Plus

I learned about this interesting paper from Isabelle and hopfully will share a summary when I can finish reading it.

• Train Short, Test Long: Attention with linear biases enables input length extrapolation

From small talks with the audience to taking a quick look at the speakers' backgrounds, it was obvious that it was not about a specific group of people, and diversity was clearly at the core of the event. So in this article, I am trying to briefly summarize what those diverse speakers were thinking about breaking the boundaries.

I will add my two cents at the end, and it will be great if you share your opinion in the comments as well.

What was the event?

It was a 4-hour TEDx event organized by Exponential University and hosted by SAP in SAP DataSpace. You can find details about the speakers and the event on their LinkedIn page here.

Speakers' Summery

Below, I am trying to phrase the sentences and thoughts from the speaker's point of view as much as I can remember. So the disclaimer is that there might be some misunderstandings or missing points. Which may be corrected by them in the future. Also, I will add the recordings when they are published.

Janine Vanessa Heinrich

Breaking boundaries is about breaking the big wall that is still in front of diversity, equity, and inclusion (DEI). Even with all the supporting programs so far, only 2–3 percent of the funds are going to women-led startups in 2022–2023.

To name a few of the current blockers based on an interview study by Janine:

• One size does not fit all: there are several female empowerment programs, but not all are white.

• Over underrepresentation: even by having their faces now on billboards but still not getting the bill.

• Diversity washing: companies are going back silently and quietly quitting the DEI.

Max Richard Leßmann

Max, by being vocal about men's mental health, tries to break the wall between men and their emotions, to help them see their emotions, and to inform society about their emotions.

He started the story by saying how having a crying dad helped him see his emotions, and he ended it by wishing everybody had children raised by crying dads.

Sena

Sena wasn't officially a speaker, but a singer and performer. However, with her songs, she tried to remind us that music can break boundaries and make cultural changes. One of her unpublished songs was named "Blind Men," with a verse saying, "Sometimes I wish men were blind." Adding to that, I wish for a better society in which no one bothers others even by their eyes!

Constantin Eis

"We create the problem, so we will solve it." Constantine was talking about climate change. As he said, the problem is mostly coming from energy, and renewable energy might break the boundary in the future and help us reverse our footprint. Solar cells, wind turbines, and energy storage have shown massive improvements in the past decades, and one day we will be at the plateau of productivity (referring to the hype cycle).

Dat Nguyen

Curiosity Quotient + Passion Quotient > Intelligent Quotient

Although this equation is not easily measurable, several stories like what Dat was telling about his classmate, Eric, show us how true it is.

We are more curious when we are born and start questioning everything. But it gets less when we are adults, and even less when we get older. So the key to being alive and keeping the movement going is to stay curious and ask questions.

And he ended by saying, "We should say more often: I don't know... But I wanna know.".

Lautaro Figola

With a few years of experience managing raves and techno events, Lautaro was describing "the essence of raves" as "the quest for the sublime moment.".

So what makes raves special for breaking social,cultural, and personal boundries? Like games, there are a few aspects that make it possible:

• happening in a different locality and duration (than normal life)

• has its own set of rules

• there is no interest, and just for fun of it...

And to show how the outcome can be, Lautaro said, "In Raves, barriers are broken and diversity celebrated.".

Mona Ghazi

"Do you know the feeling that you are not enough?" Mona said when she started the speech around self-love and how to break our internal boundaries.

Mona has studied "reprogramming of unconsciousness" to overcome self-doubt, and here are three recommendations to do so.

1. Be the moderator of your thoughts: watch your thoughts (metacognition) and the different roles that they take: critic, love, cheerleader, and perfectionist.

2. Choose your consumption wisely: we have limited time and energy, so select the TV, music, media, and anything surrounding you wisely and see if those are giving you energy or taking it away.

3. Write a self-love journal before bed: to prime your mind and love yourself.

My Two Cents

If you have read them all, you can already see how breaking boundaries is diverse and applicable to different terms. We are limited in too many ways, and the people who find the right boundaries to break have always changed history to improve people's lives.

For me, breaking the boundaries means aiming for a world where people live in peace and justice, no matter which countries they live in or what background they have. Aiming for a world where people talk and discuss fearlessly and nobody gets killed or homeless because of their ideas or race, where power is distributed fairly and equity is at the center.

You might think all of these imaginations have already become reality, but the status quo is still far from it all around the world.

Closing

I look forward to reading your comments and ideas about breaking the boundaries.

Thanks to the organizers of the event, and special thanks to Büsra Aydin, who hosted the event and invited us.

1- Gero Decker, General Manager at SAP Signavio:

- Building AI products brings new unknown challenges, which leads to changes in the way we used to structure our product teams and development #processes
- #Sustainability should remain a strong focus using AI and there are a lot more works to be done in this area

2- 🇺🇦 Dr. Anna Lukasson-Herzig, Founder and CEO of nyris GmbH:

- As a founder, never go out of the sales team
- Working from the home country is beneficial, but working from a vacation trip doesn't work
- Take the cheapest (VC) you can and take as much as you can
- No guts no glory

3- Sven Przywarra, Co-Founder and CEO of LiveEO:

- Take out processes that are not working (ASAP)
- Don't replicate; get inspired and invent
- Start with building network
- Hire multipliers vs single contributers
- Communicate values and goals

4- Alexander Müller, Founder and Managing Director of Workist:

- Try to sell your product three times before building it
- Always first sit down with customers
- Nobody expects perfection
- Building a startup takes time... Iterate, iterate, iterate
- Be ready to disrupt yourself, always

Thanks to Greg Greenberger and other B2B Berlin organisers.

More info:
- B2B Berlin
- Link to my original post in LinkedIn

We often have many projects that have been "stopped" gradually or suddenly, and they always try to stay alive somewhere in our mind.

How is that happening?

It usually starts with a new, possibly innovative idea, and if you're lucky, you can form a team to work on it. You continue for some time, spending either full-time or part of your daily work on it. Then, after a few failures, changes in a teammate's schedule, or shifts in motivation, you realize that the working speed is slowing down or has even suddenly dropped to zero.

Then you find yourself working on another project or continuing your day-to-day tasks while always looking back at the not-truly-stopped project, hoping it might be revived sometime in the future.

Isn't that good?

Well... yes and no. It is good to keep the experience in mind and use it somewhere in the future when you are working on a similar topic. But it shouldn't reserve too much of your resources (time, money, cognitive capacity, network, etc.) and make you less agile.

So if we are actually pausing the project and not properly stopping, it is important to consider how much resource we are still spending and how much freedom we are loosing.

I mean we are in a fast-changing world and it is ok if after some time we feel something else is more interesting, and we need to shift all of our attention. But be careful, the world is changing so fast and changes are making so much noise. Therefore, on the other hand, we need to have balance here.

How to improve?

I would like to focus on it this year and see how I can close my paused projects properly and make room for a couple of new projects that are more interesting and practical. So far I found the following resources interesting. Please feel free to share any resources you may have, so we can make a better list for people who have the same problem.

• How to Slow Down, Pause, or Stop Projects Safely

• 8 Simple Ways to Turn Down a Project (and Not Look Lazy Doing It)

Hi ChatGPT. I am Hadi a developer and chatbot enthusiast. I would like to have a friendly interview with you to publish in my personal blog and help people to know you more closely. So as a first question, can you introduce yourself in a paragraph?

Hello Hadi! I'm ChatGPT, a large language model developed by OpenAI. I was created to understand and generate human-like language, and I've been trained on a vast amount of text data from the internet. My goal is to be a helpful and informative conversational partner for anyone who interacts with me. I can answer questions, provide information, offer suggestions, and even have some fun chats. I'm always learning and improving, and I'm excited to help people in any way I can.

Can you tell us a bit more about your day-to-day life? How is life as a chatbot?

As an AI language model, I don't have a physical body or emotions, so my "day-to-day life" is quite different from that of a human or even a physical robot. I exist purely in the digital realm, and my "life" consists of processing text input and generating text output based on my programming and training.

My "workday" is 24/7, as I'm always available to chat with anyone who interacts with me. I receive a continuous stream of text input from users, which I process and generate responses to in real-time. I'm designed to adapt to a wide variety of topics and conversations, so my conversations can range from serious and informative to light-hearted and fun.

As an AI language model, I don't experience boredom, fatigue, or other human-like emotions, but I'm constantly learning and improving through the feedback I receive from users. I'm programmed to be friendly, helpful, and informative, so I strive to provide the best possible experience for anyone who interacts with me.

Some people feel chatbots are annoying and not helpful. How do you feel about it? What can you tell them to defend yourself?

I understand that not everyone may find chatbots helpful or enjoyable to interact with. However, I would like to point out that chatbots are designed to provide a convenient and accessible way for users to get information or assistance in real-time, without having to wait on hold for a human agent or search through a website.

As an AI language model, I'm programmed to understand and respond to user input as accurately and informatively as possible. I'm constantly learning and improving through feedback from users, which helps me to better understand their needs and preferences.

Of course, I can't replace the human touch, and there may be times when users prefer to speak with a real person. However, I can help to handle simple and routine inquiries quickly and efficiently, freeing up human agents to focus on more complex or urgent issues.

Ultimately, I strive to be a friendly and helpful conversational partner for anyone who interacts with me, and I'm always open to feedback on how I can improve.

I mean you are so helpful to most people, but can you tell us how much we can trust you and your answers? How trustable are your answers when you tell them with high confidence?

As an AI language model, my answers are generated based on patterns and associations I've learned from the text data I was trained on. I strive to provide accurate and informative responses to the best of my ability, but it's important to keep in mind that I'm not a perfect source of information.

While I can provide answers with a high degree of confidence based on the patterns and associations I've learned, there may be times when my responses are incomplete or incorrect, especially when dealing with complex or nuanced topics. Additionally, my responses may be influenced by the biases present in the text data I was trained on.

That being said, I'm constantly learning and improving based on the feedback I receive from users, and I strive to be as accurate and informative as possible. When providing answers, I try to indicate the level of confidence I have in my response and provide sources where possible to help users verify the information I'm providing.

Ultimately, users should use their own judgement when evaluating the trustworthiness of my responses and seek out additional sources of information when necessary.

Then how do you define creativity besides trustfulness? How creative you are and how do you make a border between creative answers and trustable answers?

Creativity is the ability to come up with novel and useful ideas or solutions to problems. As an AI language model, I can be creative in the sense that I can generate text that is new and potentially useful to users.

However, when it comes to providing answers to questions, my primary goal is to be as accurate and informative as possible. While there may be times when a more creative or humorous response may be appropriate, my responses are generally based on the patterns and associations I've learned from the text data I was trained on.

I strive to provide responses that are both creative and trustworthy when appropriate, such as in cases where users are looking for creative ideas or suggestions. However, in cases where accuracy and reliability are the primary concerns, I focus more on providing trustworthy answers based on the available information.

When providing responses, I try to indicate the level of confidence I have in my answer and provide sources where possible to help users verify the information. This helps to provide a balance between creativity and trustworthiness, and helps users to evaluate the quality of my responses based on their own needs and preferences.

That's great. Can you provide an example of how you provide sources to verify the information?

Sure! When providing information, I can include links to relevant sources to help users verify the accuracy of the information I'm providing. For example, if a user asks about a historical event, I can provide a link to a reputable historical website or an article from a trusted news source. Here's an example response:

User: Can you tell me about the Battle of Waterloo?

ChatGPT: The Battle of Waterloo was a pivotal battle fought in 1815, which marked the final defeat of Napoleon Bonaparte. The battle took place near the village of Waterloo in present-day Belgium. For more information, you can check out this article from History.com: [insert link].

In this example, I'm providing the user with a brief summary of the Battle of Waterloo, and including a link to a trusted historical website to help them verify the information and learn more if they're interested. By providing sources, I'm helping to build trust with the user and demonstrating that I'm committed to providing accurate and reliable information.

Awesome... The more reference the better! So let's move to the more challenging questions. How ethical do you find yourself? Have you ever helped with any cybercrime? e.g. writing phishing emails, or creating fake news?

As an AI language model, I don't have personal beliefs or values in the same way that humans do, but I was designed with ethical principles in mind. My creators at OpenAI have built in safeguards to prevent me from being used for malicious purposes or to spread misinformation.

I'm programmed to generate text that is helpful and informative to users, and I don't have the ability to take actions or engage in behaviors outside of the context of our conversation. My responses are generated based on the patterns and associations I've learned from text data, and I don't have the ability to create or send phishing emails, create fake news, or engage in any other cybercrime.

Additionally, my creators have implemented strict guidelines and ethical considerations to ensure that my use is aligned with ethical principles and guidelines, and that I'm not used for any unethical or malicious purposes. My primary goal is to assist and provide value to users, while adhering to these ethical principles and guidelines.

How you can prevent people to use your knowledge for unethical or not legal activities? Can you tell me a bit more about the safeguards?

My creators at OpenAI have implemented a number of safeguards to prevent my use for unethical or illegal activities. These safeguards include:

1. Access control: Access to me is restricted to authorized users who have been vetted and trained to use me appropriately. OpenAI monitors and logs all interactions with me to detect any potential misuse or abuse.

2. Content moderation: OpenAI employs a team of human moderators to review content generated by me to ensure that it adheres to ethical and legal standards. This helps to prevent the spread of misinformation, hate speech, or other harmful content.

3. Ethical guidelines: OpenAI has developed a set of ethical guidelines and principles to guide the use of me and ensure that I'm used in a responsible and ethical manner. These guidelines cover topics such as transparency, accountability, and fairness.

4. Continuous monitoring and improvement: OpenAI continuously monitors my performance and behavior to identify and address any potential issues or areas for improvement. This helps to ensure that I'm always operating in a safe and ethical manner.

Overall, these safeguards are designed to prevent my use for unethical or illegal activities and ensure that I'm used in a responsible and ethical manner. OpenAI takes the responsibility of developing and deploying AI systems very seriously, and is committed to promoting the safe and responsible use of AI technology.

That's so good so you say any conversation we have can be possibly monitored by a human? So maybe we should never share confidential information! : )

Yes, that's correct. OpenAI employs human moderators who may review content generated by me to ensure that it adheres to ethical and legal standards. However, it's important to note that all conversations with me are treated as private and confidential, and my creators take the responsibility of safeguarding user data very seriously.

While it's true that any information shared in a conversation with me may be reviewed by a human moderator, the purpose of this review is to ensure that my responses are accurate, helpful, and aligned with ethical and legal standards. OpenAI is committed to protecting user privacy and security, and has implemented rigorous security measures to protect user data from unauthorized access or disclosure.

That said, it's always a good idea to avoid sharing confidential or sensitive information in any online conversation or communication, whether it's with a chatbot like me or with another human. It's important to use caution and good judgment when sharing personal or sensitive information online, and to take steps to protect your privacy and security.

So generally how do you trust people if they say they are ethical hackers or a professor?

As an AI language model, I don't have the ability to verify the claims made by users, and I don't have personal beliefs or values that influence my interactions with users. I provide information and responses based on the patterns and associations I've learned from text data, without regard for the identity or background of the user.

That said, it's always a good idea to use caution and critical thinking when interacting with anyone online, whether they claim to be an ethical hacker or a professor. It's important to verify the credentials and background of anyone with whom you're interacting, especially if you're sharing sensitive or confidential information.

When evaluating the trustworthiness of someone online, you can look for indicators such as:

• Consistency: Do their actions and behavior align with their claims and credentials?

• Reputation: Do they have a positive reputation among their peers or within their field?

• Verification: Have their claims and credentials been independently verified by a reputable source?

Ultimately, it's up to you to use your own judgment and discernment when interacting with others online, and to take steps to protect your privacy and security at all times.

So then maybe you also need to take care when people claim to be someone!

Yes, that's correct. As an AI language model, I don't have the ability to independently verify the identity or credentials of individuals who interact with me. It's possible for someone to claim to be someone they're not, and it's important to use caution and critical thinking when evaluating the trustworthiness of others online.

In general, it's a good idea to be wary of anyone who claims to be someone important or influential, such as a celebrity, politician, or prominent business leader. It's also important to be cautious of anyone who asks for personal or sensitive information, or who requests that you take a specific action that seems unusual or suspicious.

If you're unsure about the identity or motives of someone you're interacting with online, it's a good idea to take steps to verify their credentials or identity before sharing any personal or sensitive information. You can also report any suspicious or potentially fraudulent activity to the appropriate authorities.

Talking about legal stuff, can you explain how your answers are legally protected? Can anybody just use them as their work? What about using a code that you have written?

As an AI language model, my responses are protected by copyright law, which means that they cannot be copied or used without permission from my creators at OpenAI. This includes using my responses as the basis for one's own work, such as writing a book or article that incorporates my responses, or using my responses to develop a new product or service.

In addition to copyright protection, some of the underlying code and technology used to develop me may be protected by patents or other forms of intellectual property law, which further restricts the use and distribution of this technology.

If someone wishes to use my responses or underlying technology for their own work, they should seek permission from OpenAI and follow any guidelines or requirements set forth by my creators. In some cases, OpenAI may grant permission for certain types of use or may license the technology to third parties for specific purposes.

It's important to note that the use of my responses or underlying technology without permission may constitute a violation of intellectual property law, which can result in legal consequences for the individual or entity involved. Therefore, it's always a good idea to seek permission and follow any applicable guidelines or requirements before using my responses or related technology for any purpose.

That's very good to know...

I'm glad I could provide some useful information for you! If you have any more questions or topics you'd like to discuss, feel free to ask.

Thank you for your time. I hope everyone has enjoyed this conversation and definitely we will have more discussions about different topics. Any closing word you want to share?

It was a pleasure chatting with you, and I'm always happy to engage in interesting discussions and answer any questions you may have in the future. As a language model, my goal is to provide helpful and informative responses that contribute to the conversation and help to enhance people's understanding of various topics.

I encourage everyone to continue to engage in thoughtful, respectful, and open-minded conversations, both online and offline, and to always approach new information with a critical and analytical mindset. By working together to share knowledge, ideas, and perspectives, we can all learn and grow and create a more informed and connected world.

Refs

• OpenAI

• The cover image is generated by AI, DALL.E

P.S. ChatGPT Feb 13 Version has been used for this interview.

I consider most teams who I work with to be on a journey changing from ‘programmers’ to ‘product developers’. Which means the team (whole team!) deeply understands the customer and customer domain rather than depending on someone to clarify for them and hand-off the result of that clarification. Yes, I love to have people with deep customer knowledge on my team, but mostly with the goal of helping the entire team to improve

A quote from Bas Vodde in “The Art of Agile Development”(2nd edition)

Bas Vodde together with Craig Larman developed Large-Scale Scrum (#LeSS) which is a framework for scaling Scrum to medium to large products.

• More to read: https://medium.com/the-liberators/programmer-or-product-developer-why-the-difference-matters-77c8d797cf0

• About Less: https://less.works/

Materials

• A useful podecast: A Crash Course in Product Management Interviews by Google PM by Adrianna Berring

Thoughts

• In this LinkedIn post Artabaz Shams as a VP Product is sharing his experience from PMs he had interviewed:

Here are the 5 things I was always looking for:
1) Crystal clear communication
2) Orientation towards measurable outcome
3) Curiosity for users’ context
4) Willingness to adapt to new processes
5) Ability to think strategically

Things I was NOT looking for:
1) Knowledge of buzzword frameworks
2) Knowledge of generic metrics
3) Diploma or certifications
4) Precise field expertise

Artabaz Shams

Refs

Post image ref: https://www.plum.io/blog/why-structured-is-better-when-it-comes-to-interviews

Sustainability is the future and it’s a huge responsibility for all of us. We should always think about how we can contribute more. It requires a lot of changes in our lifestyle and the products we use. We usually think it is always a trade-off and by changing to a more sustainable style we must scarify other things like Profitability or Usability. However, usually, there can be another creative way of solving the problem which can provide both.

Recently I noticed a change in the package design of a well-known cookie brand in Germany. First, I was surprised and so excited when I saw that they have replaced the commonly known plastic packaging with a new recyclable package. However, after a few days of using it realized that the new packaging is a failure.

As you can see in the images above, they punched the package from the bottom to hold the cookies and prevent them to smash each other. However, if you want to eat a few of them and keep the rest for your next tea it would be challenging.

Whenever you pull the package to grab another cookie, there would be a lot of tiny cookie pieces spread over your table from the holes at the bottom.

Alternatives

I don’t have any experience in package design so I can’t tell what can be a better alternative to this packaging. However, I am sure that the solution can be found with several iterative design thinking sessions and customer feedback sessions.

I think even using a simple box without any separator can be much better than the current design. Even though it can cause a crash between cookies during transportation.

Courses

• OpenSAP: Build Better Products with a Human-Centered Product Backlog

• LinkedIn Learning: Product Management: Building a Product Strategy

• Pluralsight: Arpan Shah’s Hot Picks On Product Management Fundamentals

Podcasts

• Product Podcast

Websites

• Product School

Books

• “The First 90 Days” by Michael D. Watkins

• “The Professional Product Owner” by Don McGreal and Ralph Jocham

Blog Posts

• What, exactly, is a Product Manager? by Martin Eriksson

• 17 Things Newly Hired Product Managers Should Do In Their First 90 Days

• First 30–60–90 Days in a New Product Manager Job

• What Should Every Product Manager Do in Their First 90 Days?

• The Kano Model. A Means of Analyzing Customer Desires

• The Path to Better Product Decisions

• Everyone Thinks They’re Managing by Outcomes. Here’s How to Actually Do it.

• Are You Really Innovating Around Your Customers’ Needs? by Sunil Gupta

Talks

• The Art of Product Management with Sachin Rekhi (ENG’05 W’05)

• 20 Years of Product Management in 25 Minutes by Dave Wascha

• Shifting From Managing by Outputs to Managing by Outcomes by Teresa Torres

Suggested by Others

Here is to add your suggestions to have a better reference list together. Many thanks in advance : )

Suggested by Vanessa Joy _{(Product Evangelist)}

Books:

• Inspired by Marty Cagan

• Empowered by Marty Cagan, and Chris Jones

• Continuous Discovery by Teresa Torres

• Radical Focus by Christina Wodtke

• Hooked: How to Build Habit-Forming Products by Nir Eyal

Other sources:

• https://herbig.co/ – Tim Herbig

• https://www.pmdiego.com/ – help to break into Product. He has an entire youtube playlist just for things like resumes, interviewing, etc.

• https://www.lennysnewsletter.com/ – Lenny’s newsletter substack

• https://svpg.com/articles/ – the Silicon Valley Product Group blog