Forget Neural Nets—Axiom Wants to Build Digital Brains

Forget Neural Nets—Axiom Wants to Build Digital Brains

Large language models talk a good game, but they can’t think on their feet, this new AI architecture might finally teach machines to act like they mean it.

Forget brute-force learning. Axiom, a new AI system from Verse AI, mimics how real brains predict the world. Unlike deep reinforcement learning, which requires countless iterations, Axiom learns fast by fusing prior knowledge with real-time updates using a principle called "active inference."

It’s built on Karl Friston’s free energy theory, which suggests intelligence is about minimizing surprise through continuous prediction. That’s how Axiom outperforms traditional AI in mastering games like Drive, Hunt, and Bounce, using a fraction of the data and compute.

It’s not just about video games. CEO Gabe René claims Axiom could be the future of real-time, efficient, agentic AI, already being tested by a finance firm for market modeling. What’s compelling is the architectural shift: a digital brain, not a scaled-up mimic of one.

As François Chollet puts it, we need more bold detours from the LLM arms race. Axiom could be that. A few standout signals from this emerging path:

• Active inference merges cognition with action, not just pattern recognition.
• Smaller models mean less energy, more flexibility.
• Brain-inspired designs may edge closer to AGI than massive chatbots.

Axiom’s story reminds us that progress doesn’t always come from scaling. It can emerge from reimagining first principles. If we want machines that adapt like humans, should we be scaling transformers, or rethinking intelligence altogether?

Read the full article on Wired.

----