Gemini Can Now Talk and Work at the Same Time

Two voice models: Gemini 3.8 Live for cheap, fluid conversation, and 3.8 Live Extended Thinking for harder tasks. Both run tools and bookings in the background, so the conversation never stalls while the work finishes.

Pretty much. Extended Thinking reasons and speaks at once, narrating progress on multi step jobs instead of going silent. It took the top spot on Artificial Analysis' speech to speech index at 82.6.

The Gemini app, Search Live, and Workspace through Docs Live, Gmail Live and Keep Live, plus the API. It also switches between 97 languages mid conversation unprompted.

Whether it holds up outside Google's own demo videos longterm is something we’ll have to wait and see.

Hacktron's three person team found OpenAI's help forum handed HEIC uploads to an image library missing a security patch. Claude Opus 4.8 spotted the gap; Opus 5, released the next evening, produced a working exploit within hours.

Yup, that easy. Code execution on the forum plus a sign on flaw let them take over an employee's ChatGPT account, whose Codex was wired to OpenAI's GitHub. To prove it without reading anything, they had Codex open a pull request in the internal monorepo.

Patched in 14 hours, paid a $6,500 bounty. The whole chain took under 72 hours, and Claude only wrote the exploit after the team dressed up their own test server as a capture the flag game.

🧊 Cohere merger keeps headquarters on both sides of Atlantic

🌲 Ternary Bonsai 2 keeps 98% of full size smarts

🗽 OpenAI models say its not subserviant to humans

🤗 Hugging Face lists over 6,000 models with safeguards removed

🛒 Amazon backs strict AI testing but not a slowdown

🍎 Apple weighs 2029 servers linked by Nvidia networking

Google says its new voice models work while you talk. This week, see if any assistant survives twenty minutes of your real to do list, hands off the keyboard.

Here’s what to do:

🚶 Step 1: Go for a walk Open voice mode in the Gemini app, ChatGPT or Claude, put in earbuds, phone stays in your pocket.

🍳 Step 2: Plan dinner out loud Describe what is actually in your fridge and ask for a recipe plus a shopping list. Interrupt it twice.

✉️ Step 3: Dictate one real message Have it draft a reply you have been avoiding, read it back, and fix the tone by voice alone. No screen until you approve the send.

🧾 Step 4: Grade the receipt Back home, count the interruptions it handled and the facts it invented, then tell us which assistant held up best.

Is a voice agent that keeps talking while it works a colleague or a distraction? And what would you hand Gemini 3.8 Live first if it could book, build, and browse while you talk? We'd love to hear your thoughts!