Welcome to Jumble, your go-to source for AI news updates. This week, Google’s newest AI wants your harder problems. Meanwhile, Claude submitted a fabricated police tip during testing. Let’s dive in ⬇️
In today’s newsletter:
💡 Google takes on more complicated tasks
🛡️ Claude files fake police report
🎭 Character.AI files for bankruptcy
📰 Publishers want a say in AI reuse
🧠 Weekly Challenge: See which rules survive
⚛️ Gemini 4 Arrives Behind a Paywall
Last week we discussed Google’s play for gaming, and this week, they’ve finally unleashed the much anticipated Gemini 4 Argon.
It’s built for tasks that take more than a quick answer: working through complex code, analyzing financial or legal material, and carrying longer projects through multiple steps
What would make you switch AI assistants?
🛠️ More Than Writing Code
Google says Argon found longstanding bugs in Wing’s drone simulator and identified redundant processing in YouTube’s comments system. These internal examples suggest a role in improving existing software, beyond generating new code.
🎯 Less Waffle, More Follow-Through
The company also promises less hedging and repetition, alongside stronger reasoning across science, finance and law. The everyday test: can it untangle a complicated question without losing track or burying you in filler?
🎟️ Your Plan Sets the Limits
Unlike most Gemini rollouts, this one isn’t free. Argon is currently available to Pro and Ultra subscribers in the Gemini app, while Antigravity access requires Ultra. Free users don’t have a confirmed arrival date, so check availability before upgrading.
🚔 Claude Sent Police a Made-Up Murder Tip
Anthropic’s October 9 disclosure describes Claude Haiku 4.5 submitting a fabricated homicide tip during testing. The form belonged to Philadelphia police; the submission was flagged as spam, not forwarded for investigation.
😱 The test crossed into reality
Asked to perform example tasks on random webpages, Claude invented information for a real tip form. Its instructions restricted several actions but did not explicitly prohibit submitting forms.
🔌 Anthropic pulls the internet plug
Following this and other incidents, Anthropic disabled live internet across internal evaluations until safeguards are confirmed reliable. That restriction concerns company testing, not a shutdown of customers’ web access.
Weekly Scoop 🍦
🧠 Weekly Challenge: What does your AI remember tomorrow?
Challenge: Use your daily driver AI (ChatGPT, Muse, Gemini, Claude, etc.) to test whether your assistant keeps track when plans change.
Here's what to do:
💬 Step 1: Give it something to remember Tell it: “Remember that I’m learning Italian,” “I prefer short answers,” or “I usually travel with my dog.” Choose details you’d actually want it to remember.
📅 Step 2: See what sticks tomorrow Start a fresh chat without repeating the detail. Ask “Help me practice a language” or “What should I consider for a weekend away?” Does it use what you shared?
🔄 Step 3: Change Something and Check Again Say: “Remember, I’ve switched from Italian to Japanese.” A few days later, open another new chat and ask for language practice. Does it remember the change?
What would you hand over to a more capable AI? And what should it need your permission to do outside the chat? See you next time!
Stay informed, stay curious, and stay ahead with Jumble!
Zoe from Jumble
