In this month’s newsletter:
As 2025 comes to a close, here’s an update on VERSES’ progress.
At the beginning of this year we knew that our vision of universal intelligence, delivered through AI in both the physical and virtual worlds, was right. And that our teams and technology delivered a unique and exciting approach.
But because we are so different from other AI companies, many people questioned whether we could successfully launch and sell a product or continue to develop further breakthroughs.
Today, I’m proud to share what we’ve accomplished.
Our commercialization approach
Our product, Genius™, launched on April 30.
During our initial launch, we’ve discovered that our active inference models are ideally suited to data scientists and software developers facing challenges resulting from volatility, uncertainty, complexity, and ambiguity. As a result, we focused our efforts on rolling out Genius to a small number of “lighthouse” customers.
These customers typically encounter problems that cost them millions of dollars a year and where alternative approaches have been impractical or simply failed. By collaborating closely with these customers, we are focusing our product development on the tools that drive the most value.
As these solutions are developed and commercially validated with lighthouse customers, we are incorporating these features into Genius to make them available to more customers within that vertical. This approach ensures that each product update provides direct value for key markets. We've reduced our internal costs and streamlined our product offering to stay laser-focused on increasing revenue by targeting these key verticals in 2026.
In finance, we are developing the Genius THINK module to understand, model, and drive higher returns without high volatility. As the result of the progress we made, and the impact Genius had, starting with our initial relationship in late 2024, this customer upgraded to Genius in June and further expanded our work together in September.
For the Spatial Web, we are working to help develop the infrastructure that will allow AI to operate successfully in the physical world. By developing a common language, governance, and address system, the physical world will start to become as easily navigable as today's web.
We continue to work with our partner Analog to bring this into real world applications, using Genius SHARE.
And in robotics, we are developing our Genius SENSE and ACT modules to enable robots that can perceive in real time and operate without costly pre-training.
Genius’ capabilities are, of course, able to be combined. A robot in the physical world needs SENSE and ACT to perform its tasks, but it also can use the Spatial Web built in through SHARE, so that it can communicate and collaborate in a team with other robots and agents.
Our research team is constantly working to advance our technology and ensure that we have a constant pipeline of improvements for our lighthouse customers. For example, our work on AXIOM in early 2025 allows us to build successful solutions for our financial services lighthouse customer. Similarly, our six-year championing of Spatial Web standards has given us unrivaled expertise in the key underpinning technologies of HSML, HSTP, and the Universal Domain Graph and their applicability to Smart City projects.
Other developments this year include:
Gabriel René
Founder and CEO, VERSES
This month, Karl explains his work on modeling complex systems that show stochastic chaos, such as weather patterns and financial markets. By uncovering their underlying dynamics, researchers may be able to anticipate sudden events, for example hurricanes or market crashes. While short-term fluctuations can be predicted within a limited window and medium-term trends remain difficult, long-term patterns—such as climate—offer a more stable backdrop for forecasting. For instance, in finance, deep models can capture slow-moving factors, like overall market confidence, which helps make sense of day-to-day price changes.
Karl also explains the difference between inference and learning. Inference is what the brain does in real time to figure out what’s happening right now, using pre-existing models of the world. Learning is slower—it updates the brain’s connections and parameters over time. He argues that traditional machine learning can’t truly “learn to be curious,” because real curiosity depends on recognizing what you don’t know and seeking information to reduce that uncertainty. Machine-learning systems, by contrast, usually rely on fixed patterns from past data. A robot exploring a city wouldn’t realize it was lost and wouldn’t actively look for information to figure out where it was. In short: you can’t get smarter if you don’t know what you don’t know.
Finally, Karl explains that active inference approaches will be much more efficient than current deep learning methods. These decide where to focus, which improves sample efficiency. They also greatly cut down the unsustainable energy use linked to training large models.
Watch the full interview on our YouTube channel.
You can also catch up on everything we are doing on our subreddit.
NeurIPS, one of the most influential AI conferences globally, was last week.
Here are papers that were accepted from the VERSES team, from the 21,575 papers submitted this year:
VERSIAN Hampus Linander had his paper accepted:
Learning Chern Numbers of Multiband Topological Insulators with Gauge Equivariant Neural Networks
This paper shows how enforcing symmetries from fundamental physics makes deep learning feasible for complex quantum states that overwhelm standard neural networks. It advances the simulation of topological quantum materials.
VERSIAN Professor Chris Buckley, in collaboration with the University of Sussex, had two papers accepted:
A Closer Look at NTK Alignment: Linking Phase Transitions in Deep Image Regression
This paper provides a blueprint for understanding why deep image models learn certain features quickly and struggle with others, and therefore provides tools for improving computer vision.
µPC: Scaling Predictive Coding to 100+ Layer Networks
This paper addresses the challenge of scaling predictive coding to very deep networks. Up to now that has been difficult. However the µPC Parameterization approach outlined in this paper can successfully train networks over 100 layers deep on standard classification tasks with competitive performance.
Certain information included in this newsletter contains statements that are forward-looking, such as statements related to our flagship product, Genius, the commercial application of Genius and our research developments, as well as developments in the AI sector. Such forward looking information involves important risks and uncertainties that could significantly affect anticipated results in the future and accordingly, such results may differ materially from those expressed in any forward looking statements made by or on behalf of VERSES.