Learn to build intelligence that understands the physical world.
A free, hands-on programme for graduate and PhD students across Central Europe, anchored in Brno and Warsaw. Work on real problems in physical intelligence and world models, guided by practitioners from Analog AI, and find out whether this emerging field is where your career belongs.
100% free for selected participants | Opening lecture: week of 22 to 25 September, Brno, Warsaw and online | Registration closes 27 September | Final presentations: week of 26 to 30 October | Mostly online, flexible around your studies | In partnership with Masaryk University, Brno
Why now
The next paradigm is not another language model.
Language models reason impressively over text, but most of what happens in the world is not text. It is continuous, spatial, temporal and constantly changing, and forcing it through language loses much of what matters. That limitation is driving a shift toward world models and physical intelligence: systems that represent the world as it actually is, as an evolving network of people, places, things and events, and that use language models as one component rather than as the foundation. It is an early field with the foundational questions still open, which is exactly why it is worth learning now rather than in five years.
Who is behind this
Analog
Analog is an AI company building Physical Intelligence: technology that learns from the world itself rather than from text about it. It was founded by Alex Kipman, the technologist behind Kinect and HoloLens, and brings together engineers and researchers drawn from Microsoft, Meta, Amazon and leading research institutions. Analog is headquartered in Abu Dhabi and works globally.
The core of the platform is the Analog World Model: a near-real-time representation that unifies millions of spatial data points into a continuously learning fabric connecting people, places, things and creatures. Rather than replacing existing vision, audio, language and embodied-action models, it uses them as specialised cortexes and builds the integration layer between them, so new sensors and data streams can be absorbed continuously and analysed across domains. On top of it sit Ana, the Analog Neural Agent that reasons and acts over that world state, and Hive, the layer through which the whole system is deployed into real environments including cities, hospitals, estates and robotics fleets.
Learn more at analog.io. This programme is run by Analog practitioners directly, so the people reviewing your work are the people building the system. Delivered in partnership with Masaryk University in Brno.
What you take away
Learning by doing, on problems that do not have textbook answers.
Real problem statements – tasks derived from live research and engineering questions at Analog, not synthetic exercises with known solutions.
Techniques you cannot easily self-teach – spatio-temporal modelling, latent dynamics, multimodal grounding and real-time inference over evolving data.
Feedback from practitioners – written review at every step, plus informal mentoring from Analog practitioners during the project phase.
A portfolio piece and a certificate – a defensible project you can show, plus a certificate reflecting the phase you reached.
You keep the work you produce. Nothing here is unpaid labour on a product roadmap.
Choose your track
Three entry points, one problem space.
Physical intelligence is not a single discipline. Pick the track closest to your training, or tell us you are undecided and we will suggest one based on your placement task. You can switch tracks in later phases.
Track A
Perception and multimodal grounding
Turning raw sensory streams into structured world state.
- Vision, video and audio representation learning
- Cross-modal alignment and grounding
- Detecting entities and events from sensor input
- Suits: computer vision, signal processing, ML
Track B
Spatio-temporal dynamics and data
Modelling how the world changes, from raw streams to predictions.
- Graph and network models of real-world interaction
- Latent dynamics, forecasting, continuous learning
- Mobility, sensor and geospatial data at scale
- Suits: applied maths, physics, statistics, ML, geoinformatics, urban science, data engineering
Track C
Physical intelligence reasoning
Using language models to reason over physical-world data rather than over text.
- LLM-driven analysis of spatio-temporal events and structure
- Translating questions about the real world into analytic queries
- Evaluating where language-based reasoning holds up and where it breaks
- Suits: NLP, LLM tooling, cognitive science, anyone who has built with LLMs
How it works
Three steps. You choose how far you go.
Each step is self-contained. Stopping after any one of them is a legitimate outcome, and you still receive a certificate for what you completed.
-
Hybrid opening lecture
Optional, week of 22 to 25 September
An open talk by Stanislav Sobolevsky, Lead of World Models at Analog AI, on the shift from language models to physical intelligence, what Analog is building, and how the programme works. Held in Brno at Masaryk University and in Warsaw, and streamed online. Attending is optional and plays no part in selection.
-
Registration and placement task
Register by 27 September
Submit a short CV and a brief statement of motivation. You then choose one placement task from a set spanning different technical backgrounds and tool preferences. The task is deliberately approachable: it checks foundational reasoning about data, not advanced expertise, and you may solve it in Python, R, a spreadsheet or anything else you are comfortable with.
- Roughly one to two working days of effort
- Due within 10 days of assignment
-
Individual project
Approximately 3 weeks
Selected participants choose an individual project within their track and work on it over roughly three weeks. The phase runs online and largely asynchronously, with informal mentoring from Analog practitioners along the way and written feedback on what you submit. The problems are genuinely open, which is the point: this is where most participants report the steepest learning curve.
-
Presentations and closing session
Week of 26 to 30 October, Brno and online
Participants present their findings and prototypes to Analog practitioners and peers, followed by discussion and certificates. Presenting is part of the learning: explaining a modelling decision to people who will question it is a skill worth practising early.
Where this takes you
These skills belong to a field, not to one company.
Physical intelligence is being built in parallel by a lot of serious teams, and the shortage of people who can work in it is industry-wide. What you learn here transfers directly to any of them, and to academic research in the area.
AMI Labs – Yann LeCun's Paris-based venture, founded after he left Meta, pursuing JEPA-style world models that predict abstract future states rather than tokens. Raised roughly 1 billion USD in seed funding in March 2026.
World Labs – Fei-Fei Li's spatial intelligence company. Its Marble product generates interactive, editable 3D worlds, used among other things to train and evaluate robots in environments with real physics.
NVIDIA – the Cosmos family of open world foundation models for physical AI, now at Cosmos 3, combining vision reasoning, world generation and action prediction, with an ecosystem of robotics partners around it.
Google DeepMind – Genie generates persistent, interactive 3D environments in real time from a prompt. Waymo built a specialised world model on top of it for autonomous driving simulation.
Niantic Spatial – building a Large Geospatial Model trained on tens of billions of posed images, aiming at a shared, persistent, georeferenced map of the physical world for embodied AI.
Analog – the World Model as a living fabric of people, places, things and creatures, with Ana reasoning over it and Hive deploying it into cities, hospitals, estates and robotics fleets. Your host for this programme.
About the internship
Everyone who successfully completes the project will be considered for a six-month internship with Analog, remote or in person. We mention it because it is a real opportunity and you should know it exists.
It is not, however, how we measure whether this worked. The purpose of the programme is to train people who can do physical intelligence work anywhere in the field. If you finish with skills you did not have and decide to take them to a research group, a robotics company or one of the labs above, that is a good outcome by our definition.
Who can join
Curiosity matters more than your current toolkit.
A good fit if you are
Turning raw sensory streams into structured world state.
- A master's, doctoral or advanced bachelor's student at a university in Central Europe, including Brno, Warsaw and the wider region
- Studying CS, ML, maths, physics, geoinformatics, engineering, robotics, cognitive science or a related field
- Comfortable analysing data in some tool, at any level of sophistication
- Willing to spend a few focused days per phase alongside your studies
You do not need
- Prior experience with world models or physical intelligence
- Publications, competition rankings or a deep learning background
- A specific programming language
- To commit to every phase in advance
Working language is English. Capacity is limited, which is why the placement task exists.
Optional, no commitment
Opening lecture: from language models to physical intelligence
Before registration closes, we are giving an open hybrid talk on why the field is moving beyond text-based intelligence, what world models actually are, where current approaches fall short, and how Analog is approaching the problem. We will also walk through the programme format and answer questions live.
Come if you are curious and undecided. Attendance is entirely optional and plays no part in selection.
Presented by
Stan Sobolevsky
Head of the Digital City Lab
Principal Research Engineer and Lead of World Models at Analog AI. Research Professor at New York University and Extraordinary Professor at Masaryk University. Former AI Engineer at Meta.
Week of 22 to 25 September, exact date to be confirmed | Masaryk University, Brno | Warsaw, venue | Further venues to be announced
Streamed online | Approximately 45 min plus Q&A | Recording shared with registrants
Questions
Before you register
Is it really free?
Yes. There is no fee at any phase for selected participants. The opening lecture is open to everyone at no cost.
How much time will this take?
The placement task is about one to two working days. The individual project runs about three weeks and is designed to fit around coursework and research, roughly a few days of focused effort in total. Deadlines are firm, but how you distribute the work is up to you.
I have never worked on world models. Should I apply?
Yes. Almost nobody has, including most researchers who moved into the field recently. The placement task assesses general analytical reasoning, not familiarity with this specific area.
Do I have to attend in person?
No. The programme runs primarily online and asynchronously. The opening lecture is held in Brno and Warsaw and streamed online, and the final presentations are hybrid, so you can take part from anywhere in the region.
I am not in Brno or Warsaw. Can I still apply?
Yes. Those are simply where the in-person lectures are held. Students at universities elsewhere in Central Europe are welcome, and every phase after the lecture runs online.
What happens to the work I produce?
It stays yours. You are free to use it in your portfolio, thesis or publications, subject to any confidentiality terms on specific datasets, which we state clearly before you start a task.
Is this a recruitment process?
No. It is an experiential learning programme. Everyone who successfully completes the project will be considered for a six-month internship with Analog, so the opportunity is open rather than reserved for a handful of people. But joining Analog is not the point of the programme and not how we judge whether it worked. We are training practitioners for a field that many organisations are hiring into, and skills you can take anywhere are the actual deliverable.
Can I take part as a team?
The placement task and the project are individual, so we can give you personal feedback. Team collaboration may follow later for participants who continue working with us.
Who is behind this?
Analog, an AI company building Physical Intelligence and the Analog World Model, in partnership with Masaryk University in Brno. See About Analog above.
Which track should I pick if I mostly work with LLMs?
Track C. It is about pointing language models at physical-world data instead of text, which is a distinct and currently underexplored problem. Prior LLM tooling experience helps, but is not required.