Welcome to Tech Beat, your daily look at the stories shaping our digital world.
Researchers are shining new light on a phenomenon called grokking, where a tiny transformer model with just seven hundred seventy seven parameters didn't merely memorize arithmetic — it actually learned the underlying structure of modular math. The finding raises genuine questions about when and why neural networks develop real understanding versus pattern matching.
On the hardware side, iFixit has published its teardown of Apple's new iPhone Duo, cracking open the device to examine what's changed beneath the glass. Teardowns like this matter because they reveal the real engineering trade-offs Apple made — repairability, battery design, and component choices that consumers never see in a keynote but live with for years.
And for the security community, a new tool called CVE dot now is offering a free, real-time feed of software vulnerabilities specifically designed for AI agents to consume. As automated systems take on more of the work of monitoring and patching infrastructure, having machine-readable threat intelligence could meaningfully shrink the window between a vulnerability being disclosed and a patch being deployed.
Keep surfing. Tech Beat out.
