A close friend of mine came across the recently published “ARC-AGI Without Pretraining” (March 2025) and shared it with me yesterday. The paper offers a refreshing and elegant demonstration of an idea many of us in AI have long believed:
Intelligence is compression.
The authors demonstrate that a small neural network, trained from scratch during inference, can solve ARC reasoning puzzles — a widely recognized benchmark for Artificial General Intelligence (AGI) — by compressing input-output examples into a compact, generalizable rule. No pretraining. No prior memory. The model achieves ~34% on the evaluation set — not yet deployable, but it learns on the job. And that’s significant.
This paper helped me connect some dots.
In our experience, we’ve seen that pretrained LLMs, while trained on general-purpose language tasks, can still perform remarkably well in downstream security applications like spam detection — even without task-specific fine-tuning. This ability to generalize beyond their original scope offers a glimpse into the broader adaptability we associate with emergent general intelligence.
It also brought me back to 2016, when we launched WedgeARP — our “Absolute Real-time Protection” platform — in partnership with Cylance. It achieved extraordinary detection rates against never-before-seen malware, using real-time AI instead of static signature databases. It was ahead of its time — and not easy to promote.
We quickly faced deeper challenges:
- The scale: more than half a million new malware emerge every day.
- The need for speed: to stay ahead, we had to “compress” knowledge quickly.
- And longer-term, we knew we’d face concept drift and catastrophic forgetting.
In response, we published a 2022 paper (IEEE Xplore) focused on tackling scale and speed.
Our contribution: a training and inference architecture for ANN-based malware detection that could continuously retrain and converge within 9 hours, enabling us to respond daily to the evolving threat landscape.
While that paper didn’t address concept drift or catastrophic forgetting — both of which we tackled in follow-up work — it laid a foundational capability:
A system capable of keeping up with the world by learning continuously, reasoning quickly — particularly as a real-time network security system for malware prevention.
Now, with advanced research from vibrant communities like ARC-AGI, we’re seeing confirmation of something we’ve believed all along:
The future of intelligence lies in adapting quickly, learning continually, compressing structure effectively — and reasoning in real time — whether it’s for puzzles, language, or malware detection.
This is what had me writing on this quiet Sunday morning — reflecting on the long path from early real-time systems to the emerging frontiers of AGI.
Link to ARC-AGI Without Pretraining
