Mr. Gabriel Rene reports

VERSES DIGITAL BRAIN BEATS GOOGLE'S TOP AI AT "GAMEWORLD 10K" ATARI CHALLENGE

VERSES new AXIOM model is up to 60 per cent better, 97 per cent more efficient and learns 39 times faster than Google Deepmind's DreamerV3 in third party validated benchmark

Verses AI Inc. has released results that, it believes, mark a significant advancement in artificial intelligence. The company's new digital brain architecture, codenamed AXIOM (Active eXpanding Inference with Object-centric Models) and based on active inference, has demonstrated superior performance over Google DeepMind's DreamerV3, a leading model recognized for its generalization capabilities in game environments.

AXIOM outperformed DreamerV3 in the Gameworld 10K benchmark -- an advanced successor to the Atari 100K Challenge that pits AI agents in a head-to-head decathlon across 10 diverse, arcade-style environments to measure how well they can generalize across domains with real-world-like conditions given minimal data.

AXIOM -- better, faster, cheaper (and smaller)

Across the 10 games in Gameworld 10k which test a model's ability to perceive, catch, jump, avoid, etc., Verses's AXIOM exhibited superior gameplay capabilities while demonstrating significant efficiency gains and speed, all at a fraction of the size of Google's DreamerV3 without the use of neural networks, backpropagation and gradient descent used in almost all other AI's today.

Gameworld 10k performance highlights (AXIOM versus DreamerV3):

• 60-per-cent-better gameplay (normalized performance score -- 77 versus 48);
• 7.6 times more sample efficient (learned in 3,175 steps versus 24,207);
• 39 times faster in GPU (graphics processing unit) runtime (approximately 10 minutes versus roughly 370 minutes);
• 12 times cheaper to run (estimated GPU cost -- 66 cents versus $25.54);
• 400 times smaller in model size (950,000 versus 420 million parameters).

Digital brain AXIOM developed by leading neuroscientist Dr. Karl Friston

"AXIOM is being developed as the first digital brain, designed to mirror the modular structure and dynamic processes of our own brains -- what we refer to as 'biomimetic,'" said Dr. Karl Friston, Verses chief scientist, AXIOM project leader and one of the world's most cited neuroscientists. "While conventional AI methods, like deep learning, excel at pattern recognition, AXIOM has the ability to learn in a manner similar to humans. It doesn't merely process data; it develops an understanding of its world and how it operates within that world, which I believe will enable it to seek out experiences that massively enhance learning. AXIOM's capacity to understand, plan and grow like a brain offers a natural and efficient new paradigm for developing genuinely intelligent agents."

Third party validation

Verses submitted the AXIOM paper, mathematical proofs and source code for independent review to Soothsayer Analytics, an international data science advisory, research and development, and AI certification firm trusted by Fortune 1000 and Global 2000 clients.

"Soothsayer has validated the claims of AXIOM achieving significant gains over leading deep reinforcement learning models," said Akshay Deshpande, Soothsayer senior director of AI and lead researcher on the project. "Our evaluation found that AXIOM's efficiency stems from its integration of variational Bayesian inference with active inference, enabling more human-like learning with fewer interactions and faster convergence."

The findings are supported by Dr. Haritima Chauhan, associate professor and project co-lead, and Varun Agrawal, research assistant at Soothsayer.

This research paper has been submitted to arXiv and includes test results against multiple leading models, along with detailed mathematical proofs and test code, which are made available under an academic licence, enabling replication and broader community engagement. A white paper outlining the results is available on the Verses website.

Expert commentary

Dr. David Bray, PhD, chair of the Accelerator and distinguished fellow at the Stimson Center, senior fellow with the Institute for Human-Machine Cognition, an expert with MIT Horizon, the Oxford Internet Institute and Harvard's Leadership for a Networked World Program, was granted early access to the paper.

"In many critical fields, there's an urgent need for models that are more reliable, less constrained by past training sets and more adaptable to a changing world," said Dr. Bray. "Active inference offers a more transparent, energy-efficient and generalizable approach to intelligence -- one that mirrors how humans learn, adapt and simplify. The sample efficiency of AXIOM is particularly noteworthy, showing how it can generalize from fewer examples, much like human cognition, using far less computation. And by using Bayesian model reduction to prune complexity, this work captures a fundamental principle of intelligent behaviour: simplify where possible, but not at the expense of understanding. This is the kind of intelligence we need for systems to thrive in complex, real-world environments across both public and private sectors."

Implications for AI and Genius

"As a cognitive computing company, Verses's mission has always been to translate insights from the human brain into more effective and trustworthy AI solutions, not for $100-billion data centres, but for 100 billion devices," said Gabriel Rene, chief executive officer of Verses. "Using the brain as a blueprint provided us a breakthrough to design an entirely new class of AI, one that learns in real time instead of being trained on past data. With AXIOM, we're not making agentic solutions better by making them bigger; we're making them smarter. And that's exactly what our customers need: intelligence that fits on the edge, learns on the fly and makes decisions they can trust. We look forward to integrating AXIOM into upcoming Genius releases to empower the enterprise with agents that are smarter, more reliable, efficient and explainable."

Notes to editors:

1. Dr. Friston's recent citation count can be found on the Research website.
2. Dr. Bray is a globally recognized expert in AI, data and emerging technologies, recognized as one of the "top 24 Americans changing the world" by Business Insider for his work in national security, counter-bioterrorism and public health, who has served in multiple executive leadership roles across government and academia, and has led two successful bipartisan commissions on the future of AI, cybersecurity, biotechnology, commercial space and quantum computing.
3. The AXIOM approach leverages structured reasoning over brute-force learning. It combines biologically inspired modules for perception, action and prediction, with variational Bayesian inference and active inference principles. The result is a more efficient and adaptive intelligence system.
4. AXIOM's breakthrough arises from its use of Bayesian model reduction, core functional priors and active-inference-driven planning, allowing for rapid generalization and cross-task performance with minimal data. Unlike large foundation models, Genius operates in real time, at the edge, and can learn from fewer than 3,200 steps.
5. Further details can be found on arXiv (awaiting publication link) and at the company's website.

Atari is a registered trademark of Atari Interactive Inc. Nothing herein is intended to communicate or imply a sponsorship or endorsement by Atari, or any affiliation therewith, as no such sponsorship, endorsement or affiliation exists.

We seek Safe Harbor.