Who Tells the Robot No? | Aleksandra Łoś of Relayn

What we talked about.

Sunny Ray welcomes Aleksandra Łoś of Relayn back to the show to ask a simple question: who tells the robot no? Łoś explains Relayn's approach of separating reasoning from execution, using a deterministic rule engine to decide and a language model only to communicate. She describes how the company has moved from early reasoning workflows to a full operating system now used with hospitals, drone companies and energy plants, and how it recently signed two partnership agreements with robotics companies in Poland. The conversation covers why the physical world has no undo button, why she sees repeated re-prompting as a kind of gambling, and why critical decisions should stay with humans. Łoś shares a local benchmark against Sonnet 4.5 on AWS Bedrock showing faster responses and far lower token usage, describes an emergency room project aimed at reducing human error, and explains why her go-to-market strategy rests on human-to-human conversations. She also reflects on lessons from her earlier exit, including the importance of a close co-founder relationship and clear communication contracts.

When AI acts in the physical world there is no undo, so Aleksandra Łoś argues rules should decide and language models should only talk.

The questions, and the answers.

What has changed at Relayn since you were last here?

Everything. Last time we were at the very beginning, building our first proprietary reasoning workflows. Today we have a full operating system working with hospitals, drone companies and energy plants, and we just signed two partnership agreements with robotics companies. We see a huge need for systems that let AI interact with the physical world, where an LLM that is right 96 or 97 percent of the time is not enough.

Why does AI that acts in the physical world need different rules than AI that answers questions?

Because there is no control Z in the physical world. You cannot rewind a mistake. Using LLMs is a lot like gambling: you re-prompt and hope it works, and that is addictive. You cannot do that with patients in a hospital, a boiler or other infrastructure, not to mention robots. A deterministic layer makes the behavior loop always the same, which builds trust and certainty.

When did you decide the language model should talk and something deterministic should decide?

My co-founder Piot has pondered this for 30 years. He studied deterministic AI in Poland, then spent 20 years in gaming before returning to AI. Our first principle is decoupling reasoning from execution. LLMs are great for commanding machines with your voice, but they are not ideal for managing the physical world. We use open source self-hosted models just to communicate prepared output, not to reason.

Why won't the frontier labs build this themselves?

That is just my bet, but they are all about LLMs and the AGI narrative. We believe the opposite: domain specific super intelligence, something good at one particular task. Once reasoning is done elsewhere, the communication model only formulates the answer, so you don't need huge models. That is why our token usage is drastically lower, because we don't use LLMs for reasoning.

What was slow or breaking in the emergency room before Relayn came in?

Information sits in legacy systems, and the environment is full of pressure and stress, so human error rates are significant. Our agent reads data from those systems and informs nurses when a patient needs pickup or another test, and alerts the doctor when results worsen. It uses the hospital's preferred communication channel. The human still makes the final decision; we reduce the potential for error.

Which lesson from your exit did you carry straight into Relayn?

The importance of teamwork and of a co-founder in close proximity, preferably speaking your native language, whom you can meet daily. Iron out communication contracts from the start: your values, where you see the company, and what it grows into. The idea might be brilliant, but most startups collapse because of a lack of human understanding.

What has worked best for getting in front of operators?

Human interaction and relationships. Many people talk about go-to-market tools, but for me it is about listening. Our partnerships and letters of intent came from meeting people who share our understanding of the technology's risks and benefits. The hospital proof of concept, an agent that finds conflicts between active substances in drugs, came from sitting with someone who lives the problem for a couple of hours.