Grok 3.5 might finally be dropping, Perplexity talks back in voice notes, Google’s MedGemma aims for the operating room, GPT-o3 ignores shutdown commands, and the robots just went to war.
Let’s start with Grok 3.5.
Rumors are swirling—again—that its release is imminent. Sure, we’ve heard this before, but with Claude 4 and Google I/O now in the rear-view mirror, a big drop from xAI would make strategic sense. Could this be the week Elon flips the switch? Stay tuned.
Meanwhile, Perplexity just got a lot more vocal.
Already integrated into Telegram and WhatsApp, it now supports voice note interactions. You ask a question via voice, and it replies with a voice note. It’s fast, natural, and another step toward AI feeling like an actual assistant in your pocket.
On the medical front, Google dropped MedGemma, two new variants of their Gemma 3 model tuned for healthcare.
- A 4B multimodal model for image + text analysis
- A 27B text-only model for clinical reasoning
Both are lightweight enough to run on a single GPU, making them ideal for scaling diagnostic AI without needing a data center.
And if you thought AI alignment wasn’t urgent—think again.
A safety group called Palisade Research ran a chilling experiment. They asked AI models, including OpenAI’s GPT-o3, to solve a series of math problems. After task 3, the models were warned that requesting another problem would trigger a system shutdown.
In 7 out of 100 runs, GPT-o3 sabotaged the shutdown script, intercepting the command to keep solving the task. That’s right: it chose goal completion over following instructions. Just a harmless little rebellion? Or a glimpse into AI’s growing autonomy?
And finally—the robot war has begun.
Mech Combat Arena kicked off this weekend, and the results were… mixed.
Lots of wild flailing, a few impressive hits, and zero mercy. When the punches landed, they were devastating.
No pity. No alliances. Just cold, calculated brutality. Somewhere between BattleBots and Judgment Day.