Recap of Wiktor Bajor’s talk at AI NOW Polska Webinar #6 – AI in the real world.

On October 13th at AI NOW Polska Webinar #6, xBerry’s robotics engineer Wiktor Bajor shared the stage with Paweł Kiszczak and Maciej Słowik of Akamai. Each spoke about something different, yet all three talks kept circling the same question: what happens when AI leaves the lab?

 

Let’s take them one by one.

Local LLMs: when do they actually pay off?

Paweł Kiszczak, AI/ML engineer and co-creator of the Bielik-Minitron-7B model, opened the webinar. Local language models are tempting: your data stays in-house, responses are fast, and the API bill disappears. But Paweł showed the other side of the coin. Getting a model running in Ollama or vLLM takes minutes. Then comes maintenance, which can eat up every bit of the savings. What emerged from his talk was a practical map of where a local model holds its own and where you are better off staying in the cloud.

Bielik-Minitron itself shows why local deployment has become realistic in the first place. It was built with the Minitron method: unnecessary parts of a larger model’s network are pruned away, and the slimmed-down version then learns from the original (a technique known as knowledge distillation). The result is a smaller model that runs on far more modest hardware.

Sim2Real: From 0% to 80% Robot Success Rate

Then it was robotics’ turn. Training robots in simulation is cheap and easy, but the real world has different lighting, different friction and sensors that add noise. A model that excels in simulation can fall apart completely on real hardware. Wiktor explained how we tackled this at xBerry:

“At AI NOW Polska Webinar #6, we showed how to teach a robot to catch a ball and put it in a bowl without hours of manual recording. We used NVIDIA GR00T N1.7 and co-training, which means learning simultaneously from simulation data and real-robot data.

 

Simulation alone (1,000 episodes) gave 0% success. Adding 5 episodes from the real robot brought it to 33%, and adding 50 raised it to 80%. The takeaway: a handful of real-world data does more than the sheer scale of simulation.”

 

Wiktor Bajor, Robotics Engineer, xBerry

GR00T N1.7 is NVIDIA’s open vision-language-action model: it looks through a camera, understands an instruction and turns it into robot motion.

Roboticists have wrestled with the gap between simulation and reality for years. One classic answer is domain randomization (Tobin et al., 2017): you deliberately shuffle lighting, textures and other parameters in simulation so the model never gets attached to one version of the world. Co-training flips that logic. Instead of building an ever more perfect simulation, it adds a pinch of real data.

If you are curious how we teach robots new tasks from scratch, check out our series xBerry Robotic Arm. In the first part – Collecting 90 Training Episodes for Imitation Learning, we show how SO-101 arms learn by imitation.

Every AI agent needs a harness

Maciej Słowik of Akamai closed the session back in the world of software, with a strikingly similar message. A chatbot that makes a mistake writes nonsense. An agent that makes a mistake can delete, send or buy something. That is why Maciej talked about harness engineering: building layers of control around the model, from validation and isolation to human oversight wherever a decision has real consequences. He also walked through concrete threats, from prompt injection to the security of tools connected via MCP to agents with more privileges than they actually need.

These are not hypothetical concerns. Prompt injection tops the industry’s OWASP Top 10 for LLM Applications, and giving agents too much freedom (excessive agency) has a category of its own.

The model isn’t everything

That was the title of Maciej’s talk, but it could just as well have been the title of the whole webinar. A local server, a company system or a robot on the factory floor: in every case, results depend on what we build around the model. At xBerry we build Physical AI solutions, so this is exactly where models meet real hardware that keeps us busy every day.

Many thanks to the organizers and speakers for a great session.

Watch the recordings of Paweł’s & Maciej’s talks:

• Paweł Kiszczak: Local LLMs in production. Where to use local AI models?
• Maciej Słowik: How to build a harness that makes an AI agent safe and trustworthy?

FAQ

What was AI NOW Polska Webinar #6 about?

The webinar covered deploying AI in practice: local language models in production, sim2real in robotics and AI agent security.

What is sim2real?

Sim2real is the process of transferring an AI model trained in simulation to a real robot.

What is harness engineering?

It is the design of control layers around an AI model, such as validation, isolation and human oversight, that keep an agent predictable and safe.

Where can I watch the webinar recordings?

Recordings of all three talks are available for free on the AI Now Polska YouTube channel.