DAILOGUES: Deutsche Telekom is launching a new AI phone. Could you introduce the vision behind this phone?

Krzysztof Rutczynski: The phone must be considered within the context of our Magenta AI initiative. With Magenta AI, we are aiming at democratizing access to the best AI tools. Democratization here means that AI tools should not be confined to a small group of tech-savvy people with deep pockets, since many AI applications are expensive. Rather, they should be available to everyone. Magenta AI provides this level of availability to our customers. Consider the phone as an embodiment of this initiative. The phone is soon to be launched in 2025 and will have a deep integration of different AI tools, including a personal assistant to facilitate their use for every-day tasks.

DAILOGUES: Will this new phone have apps?

Krzysztof Rutczynski: Yes, the phone will still have apps. Having said that, the phone’s AI assistant will be easily activated and allow many tasks without needing to open any of the apps manually. For example, the assistant can answer general knowledge questions or queries related to Deutsche Telekom services. All the answers will have useful references. The assistant can also book restaurant tables or cabs, and it will even have the power to identify everything shown on its screen. This could be useful when you browse the internet and find something that you would like to buy but don’t know where to find it. The phone’s assistant will help you find the right shop.

DAILOGUES: In other words, the assistant combines a group of multi-modal agents that will handle the users’ queries.

Krzysztof Rutczynski: Exactly. Magenta AI combines these agents, many of which are powered by multi-modal models, on a platform that we are building in-house. Our goal is to handle the data, that the users entrust us with, responsibly. For example, we are implementing different guard rails, such as content filtering or anonymization, before passing the requests to our partners’ models. When it comes to the models, we’re trying to not only cooperate with companies from the U.S. but also from Europe.

DAILOGUES: Given that the new AI phone is powered by multiple machine learning models, is it necessary to use this technology all the time, given its high energy costs?

Krzysztof Rutczynski: This is a concerning issue, and we are trying to tackle it from different perspectives. When we implement a new feature, I advise the team to go for the simple and small models first, as it turns out that quite a few tasks can be taken care of without deploying the largest and most powerful models. When we say that we want to democratize access to AI, we are also thinking of educating the users on which tools are reasonable and informing them about over-reliance. As I’m thinking of it, it could be an interesting idea to include a reminder that shows users how much AI they’ve already used and to ask them to reconsider whether AI is still necessary for what they want to achieve. We are also experimenting with using generative AI in our own software development, weighing its benefits against possible costs, such as energy consumption.

DAILOGUES: Many developers already use AI to generate much of their code. What has your experience been like with this?

Krzysztof Rutczynski: I believe Andrej Karpathy has a nice way of framing software in light of recent advances in AI, by differentiating between software 1.0, 2.0, and 3.0. Traditional ways of programming would be software 1.0: Programmers would write scripts that are run on a personal computer or server. With 2.0, Karpathy argues that people have started to combine their code with machine learning models. Much of the previous code was rewritten and turned into machine-learning-oriented code. This code would be complemented with the models and their weights. Now we are entering software 3.0, where LLMs will play a significant role. They can be used to quickly generate new code via natural language prompts. This is very fascinating and often quite efficient. Some of the prompts are even integrated into the product and its underlying structure. When developing software 3.0, we are faced with the question of how much we want to keep the human in the loop. It’s almost as if we have an “autonomy slider” that we could adjust. If we shift this slider to “minimum”, the AI will ask us numerous questions about our framework, use case, architecture and design before writing any code, giving the developer granular control over technical choices and structure. But if we move the slider to “maximum” autonomy, the AI will generate the entire program based on a single instruction. Both our product and engineering teams are specifically experimenting with these emerging approaches of software development 3.0.

DAILOGUES: What is your idea of an AI-first company?

Krzysztof Rutczynski: I think that the AI first paradigm means that each corporate decision should be at least informed, if not solved by AI. As employees, we should ask ourselves whether a given problem can already be tackled by AI in a practical way. This requires good intuition because there might also be problems that could, in principle, be solved by some AI model, but that are not well suited to current systems yet, or for which simpler, proven solutions already exist. Employees need to develop the right mindset and become AI natives. Structured training, as well as communities around AI help to spread knowledge about AI and to share experiences. It is often a matter of details that you need to get right to successfully deploy AI. A further aspect would be that a company should also use the same AI it develops for its customers to better understand what the possibilities and limits of AI are.

DAILOGUES: Arguably, we could describe your partner companies like Perplexity as being AI-first. What has your experience been like in collaborating with such companies?

Krzysztof Rutczynski: We’ve been evaluating and ultimately working with multiple rising AI companies. We have worked with partners that deliver AI search, image, or voice models. Although each company solves different problems or operates with different modalities, I believe they share some interesting commonalities. Foremost, they have a razor-sharp focus by developing the best AI products in their domain, while using the best AI systems that they can obtain. Secondly, they set themselves ambitious problems to solve, and leverage AI as a tool for solving them. For example, Perplexity wants to become the best answer machine, taking on the paradigm of searching. Finally, we have been working with a few companies whose founders are deeply technical and understand their product all the way down to the model level – these companies attract top engineering talent and can boast high execution speed with short decision paths to ship new features.

DAILOGUES: What future trends in consumer-oriented AI do you predict?

Krzysztof Rutczynski: In general, I believe we’re starting to see a shift from simple prompt-based chats to a fuller integration of different AI tools into existing workflows with much richer user interfaces. In future, we’ll have fewer discussions about which prompts to use, and more about how much we want the human in the loop and how we can shape the interaction of AI with humans. The other trend that I foresee is an increase in high-quality voice-related products and services where AI powers the vocal interface.

We thank Krzysztof Rutczynski for the DAILOGUE.