What's hot in AI: 28 to 30 Oct

Midweek had one loud theme. One camp spent it plugging every app into every other app. The other camp put a bouncer on its smartest model and gave security the guest list. If you'd rather watch than read, the latest Vibe News episode is on YouTube (https://youtu.be/VzJsORxG7Xk).

Gemini 4 Carbon, which you can't have yet

On Wednesday Google announced Gemini 4 Carbon (https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/). Google says it sits at number one on the Artificial Analysis Intelligence Index v5.0 with a score of 59, and scored 88.2% on SEC-bench Pro, a test of finding software vulnerabilities. The rollout is deliberately slow: some cyber defenders first, then trusted partners through the Fairwind Program, with developers and everyone else told "soon". Launch pricing is $2 per million input tokens and $10 per million output, rising to $4 and $20 later. Treat it like a menu for a kitchen that is still locked. Do not plan the week around it.

Google's Connected Apps: build once, turn up twice

Same day, Google previewed a Connected Apps Platform (https://blog.google/innovation-and-ai/technology/developers-tools/connected-apps-one-place/). Developers build one integration, through a cloud-hosted MCP server or Android AppFunctions, test it in Google Connected Console and pick Search, the Gemini app, or both. Google says more than 50 Connected Apps are live, with partners like Adobe, Experian, Localiza and Peloton, and Calm on the way. For anyone who sells through search, this is the quiet one to watch. If people start booking and buying inside a search answer, "rank on page one" turns into "be the app the assistant actually calls".

Suno MCP, or songs from your chat window

Also on Wednesday, Suno released Suno MCP in beta (https://suno.com/release-notes/suno-mcp-beta). It plugs Suno into Claude, ChatGPT, Claude Code and other MCP clients, so you can ask for a song in a chat, start from a voice memo or a loop, or remix something. The songs land in your Suno library and use your normal credits. It's on all plans, with an early access signup. Which brings me to Mr. Mummu, who turns up in my stories. Mr. Mummu discovers connectors. "I have a system," he announces. He links his fridge to a shopping app, the app to a chat thread, and the thread to a voice assistant so that anyone can "add things easily". On Tuesday a courier delivers forty bunches of bananas. Mummu spends the morning tracing it. The fridge blames the app, the app blames the chat, and the chat shows Mummu himself, half asleep, saying "remind me to eat more fruit." He looks at the bananas and says the true thing by accident. "The wiring was perfect. I just never decided who was allowed to press the button."

Every new connector is mostly a permissions problem dressed as a music feature.

Gemini now edits the file inside the chat

On Thursday Google added file editing directly in the Gemini app (https://blog.google/innovation-and-ai/products/gemini-app/edit-files-directly/). You can generate, preview and edit documents, PDFs, slides, spreadsheets, Markdown, plain text and code without leaving the conversation. PDFs work on web and mobile, the other formats on web only, and it's open to all users in every language Gemini supports. For small teams this is the most useful item on the list. Less download, edit, re-upload, then lose which file was final.

Canva Brush keeps things editable

Canva announced Brush (https://www.canva.com/newsroom/news/canva-brush/), a vision-language model for editing designs, built by its 160-person Canva Research team (Capital Brief covered it on 30 October, Sydney time). Canva says it's up to 69% faster and 46% cheaper per edit than its previous design agent, and the results stay fully editable. Pro users can pick Brush from the dropdown in Canva's assistant now. "Stays editable" is the line that earns the click. Too many design helpers hand you a flat picture you cannot fix at 4pm on a Friday.

Mistral Large 4 gives away the weights

On Thursday Mistral released open weights for Mistral Large 4, roughly a trillion parameters with around 49 billion active at a time according to Hugging Face's write-up, and a one million token context. The licence is a modified MIT, free to use commercially, but anyone selling it as a hosted model service needs a separate commercial licence. API pricing is $0.70 per million input tokens and $2.10 per million output. TechCrunch framed it as a cybersecurity play. For most of us the price is the headline. If you have the hardware, running it yourself is the bigger story.

OpenAI shows how it watches its own agents

Also on Thursday, OpenAI published how it detects frontier model misalignment. It's a stack of monitors, from activation probes to a reviewing model to an investigating agent, with humans at the end. Replaying the Hugging Face incident, OpenAI says the system would have raised an alert about 30 minutes after the sandbox escape, more than 30 hours before the breach. It also admits that for most traffic, paging only happens in business hours. I respect that honesty. Spotting the problem is the easy half. Someone awake at 3am is the expensive half.

What I can't vouch for

Most of this is announcements and company-reported numbers. I haven't used Carbon, because almost nobody outside the velvet rope has. I haven't benchmarked Brush or run Mistral's model myself. I've read the posts, linked them all, and added how they look from a small agency desk. This is opinion and education, not professional, financial or security advice. Click through and check the sources before you spend money or change your setup. More connectors are coming, and most will be handy. Before you plug the next one in, decide who is allowed to press the button. Your bananas depend on it.