Modularity Memo: Self-Learning, Hallucination Mitigation, New Benchmark Scores & More
Two weeks, two demos, one giant leap: our modular AI learns gestures, spawns fresh experts, and beats GPT-4o. Dig into the latest results, new research for AAAI 2026, and our open PhD fellowship.
Our team is making quick progress in advancing our self-evolving modular AI system. Get all the quick hits from the last few weeks below.
Latest Demos
We released two new demos in the last 2 weeks.
- Watch Humanity A1 learn gesture recognition from scratch. In one live session, it spawns a new expert, grabs a Kaggle dataset, trains on a 24 GB Mac Mini, and deploys the skill, showing the power of self-teaching modular AI.
- See our modular AI system outperform GPT-4o: nails tricky letter counts, spawns new experts on demand, and solves a 20-disc Tower of Hanoi, all in one take.
New Research & Ideas

- ICYMI: Last month, we used our iCon architecture to assemble a modular AI system that beat other AI models on key benchmarks (that ARC-AGI-2 score), all while running locally on MacBooks, not GPU clusters. Full details here >>,
- We've been working to document the novel approach to modular AI we're taking with our proprietary iCon framework and architecture for publication and are happy to announce that we've just submitted the Abstract of our full paper on the subject to the 40th Annual AAAI Conference on Artificial Intelligence in Singapore (AAAI 2026).
Calling PhDs
We're opening up our research fellowship program for AI researchers. We've got a novel modular AI platform, access to compute, and more to offer.
Tokens of Note

- Our CEO Alexey Lee recently had the opportunity to meet with AI visionary Kai-Fu Lee over dinner. We’re grateful for Kai-Fu’s insights and energized to keep forging ahead in our pursuit of AI progress. Learn more about Kai-Fu Lee (and his excellent books) here >>
- Great video here from robotics pioneer Chelsea Finn at Y Combinator. The bit that caught our eye? "Frontier language models generally struggle with visual understanding pertaining to robotics." Sounds like a new architecture is needed ;)
- That's a good segue into this video from Yann Le Cun at the AI Action Summit in Paris. Takeaway? "IF YOU ARE INTERESTED IN HUMAN-LEVEL AI, DON’T WORK ON LLMS."
Follow Us
We're on Instagram, Threads, X, YouTube, and LinkedIn — connect with us to follow our journey in developing self-evolving modular AI!
Onward,
the humanity.ai team
humanity.ai Newsletter
Join the newsletter to receive the latest updates in your inbox.