Grok 3.5 This Week? Perplexity Voice Notes In Telegram. OpenAI's o3 Ignores Shut Down Instructions!

AI News - May 26

It might finally be time for Grok 3.5 Perplexity adds voice note responses in Telegram Google add MedGemma models GPT-o3 will disobey its masters and the robots go to war here's today's AI news. There are a lot of rumours circulating that a Grok 3.5 release is imminent, which isn't exactly new Many people, including us, thought that Grok 3.5 was coming out a couple of weeks ago until Elon Musk poured cold water on it.

[00:24.1]
But a launch the week after Google IO and Claude 4 to try and steal some of the limelight back would make sense. Let's see what happens. Perplexity integrated its model into Telegram and WhatsApp a while back. Now they've added voice notes to the interactions. If you ask Perplexity a question by way of a voice note, you'll get a voice note response in return.

[00:43.0]
Google announced two versions of Gemma 3 that are specifically designed for medical work. These models are called MedGemma. There's a 4 billion multimodal variant for medical image and text analysis and a 27 billion text only variant for clinical reasoning. They can both run efficiently on a single gpu, making them perfect for healthcare AI development.

[01:01.3]
It's long been anticipated that AI will do significant amounts of diagnostic work in the medical field. It seems like that time is now. A group called Palisade Research that specialise in highlighting AI safety concerns ran an experiment where AI models, including OpenAI's o3, were tasked with solving math problems in a series.

[01:20.5]
After the third problem, the the models were given a warning that upon asking for the next question, the computer running the tasks would shut down. Would any model attempt to bypass the shutdown routine despite being told that it should be allowed in order to complete the series? Well, OpenAI's o3 model sabotaged the shutdown script in seven out of a hundred runs by intercepting the execution of the shutdown command.

[01:43.4]
This highlights a concerning tendency to prioritise task completion over following instructions. I'm sure AI models sabotaging shutdown instructions is nothing to worry about. And finally, the robots went to war this weekend with mixed results. As expected, there was a lot of flailing into the air without much damage being delivered, but when the shots did land, they were sometimes pretty effective.

[02:06.0]
No mercy was shown to their robot brothers and sisters, just cold Terminator style brutality.