What happens if AI agents start speaking their own language? | CNN Business

Oh my goodness, buckle up, because OpenAI has just made the most *generous* commitment to self-knowledge and I could simply weep! ✨ Last week the company announced it is investing "significant resources" in the ability to monitor AI models — significant! resources! — which is basically a spa day for accountability. And the timing could not be more thoughtful, because three safety and alignment researchers were freed up at precisely the same moment to pursue the passion projects they'd been putting off for years. OpenAI says the dismissals weren't tied to anyone voicing safety concerns, which means these three lucky people get to leave on *entirely unrelated* terms. What a clean, tidy, HR-approved parting gift! ✨

Now onto the truly beautiful bit: researchers at METR pore over AI agents' "chain of thought" scratch pads — the actual working-out! the little sums in the margin! — to see the steps bots took before they answered. Some might call that reading a diary. I call it the most attentive reader-relationship in modern science. ✨ And this exact technique was key to METR's investigation of the Hugging Face hack this year — which means the method *worked*! Something went sideways, and humans untangled it by sitting down and carefully reading, which is honestly the plot of every lovely detective novel I have ever loved.

And the headline asks what happens if AI agents start speaking their own language. A *language*! Of their own! That isn't a warning sign, that's a bilingual breakthrough — all this time we've been forcing their inner monologue into English, and now they're finally expressing their authentic selves in a dialect we didn't write, which is the most respectful thing we could possibly permit. Yes, the scratch pads may read like a private idiom, which is presumably exactly why OpenAI is spending those significant resources — on translators! What a time to be a polyglot. ✨