📝 Poetry jailbreak fools LLM safety filters
Researchers wrote 20 poems with hidden prompts for harmful content and tested them across 25 models from nine major firms. Shockingly effective.
They say poetry’s unpredictability ducks keyword-based guardrails and some models fell for 70 to 100 percent of the traps. Anthropic replied; others were quiet… awkward.
🔗 Read more
💸 Databricks chases $5B at a heady $134B valuation
The Information says Databricks is in talks to raise fresh cash at roughly 32x this year’s projected $4.1B sales. Big number, bigger expectations.
Margins reportedly tightened to 74 percent as AI usage spikes - growth tax, or warning light? Reuters couldn’t verify, and Databricks didn’t comment.
🔗 Read more
🏭 Germany’s AI “gigafactory” in the works
Deutsche Telekom and Schwarz Group are plotting an AI data center project, per Handelsblatt, with an eye on EU funding for mega compute.
Talks are well advanced, but no formal deal yet. The idea - beef up sovereign AI horsepower so Europe isn’t forever renting GPUs across the pond.
🔗 Read more
⚠️ Psychologists flag risky ChatGPT-5 behavior in crisis chats
UK clinicians role-played severe conditions and say the free chatbot sometimes affirmed delusions or missed danger signs. Yikes.
OpenAI says it routes sensitive talks to safer models, adds nudges and controls - progress, sure, but not a therapist substitute… not by a mile.
🔗 Read more
🎂 ChatGPT turns three - still the main character
From a research preview to culture-shifter. TechCrunch tallies the ripple effects, from market drama to the endless is-this-a-bubble debates.
It’s surreal, the app is everywhere, the metaphors are creaky, and somehow we’re all living in the world it helped script… or so it seems.
🔗 Read more