Kimi K3.1 Pro is here, plus Dessa: an eerily realistic voice model

Welcome back, Superhuman. Black October hasn’t slowed, and likely won’t for the foreseeable future. But while cyberattacks are accelerating, so is the pace of progress: Kimi and Block dropped new models, including an eerily realistic voice model that can run on your MacBook.

Today: Top labs address AI safety, how to automate restaurant reservation hunting with Grok Bot, and get the latest prompts and trending social posts.

1. Kimi releases an open-weight model that tops Fable and Astra: The Chinese lab unveiled Kimi K3.1 Pro, its most capable model yet that it claims excels in coding, 3D creation, and financial research. The new model tops Anthropic’s Fable 5.1 and OpenAI’s GPT-6 Astra to land 4th on the Artificial Analysis Intelligence Index. Kimi also released a Highspeed version, which delivers results at 2-3x speed. Full weights are expected by November 4th. Try Kimi K3.1 Pro here.

2. Jack Dorsey’s Block drops a voice model that runs on a MacBook: Dessa was trained from scratch by listening to human conversation, as opposed to analyzing written text. This lets it hold human-like conversations full of nuances like pauses, breathing, inflections, sarcasm, and more. It’s a full-duplex model, meaning it listens and talks at the same time, and is small enough to run locally on a MacBook Pro (M4 Max or better). Listen to Dessa here.

3. Anthropic and OpenAI share security updates amid Black October: As AI continues to exploit bugs in widely used software, Anthropic showed that formal software verification could be used to rewrite programs so they have mathematically proven security guarantees. But top labs are split: While OpenAI shared steps it’s taking to secure internal training environments, CEO Sam Altman said he no longer believes we’ll be able to patch every bug.

Agentic systems are leaving the demo stage, and production AI agents hold real permissions. Instructions buried in a page or tool output can put them to work. 

AWS experts cover least-privilege scoping of agent tools and autonomy, plus tracing agent decisions. 

Register now.

Sample Prompt: “Find a table for [party size] at [restaurant type / specific restaurants] in [location] on [date] between [time range]. Call restaurants if online availability is unavailable or unclear, ask about cancellations or waitlist openings, and record the restaurant, available time, seating type, deposit, and cancellation policy. Keep checking until you find a suitable option. Do not book anything without my approval.”

Picking an agent is a much bigger deal than it seems. It’s not just another software tool, it’s a long-term teammate that’ll either take your productivity to new highs — or be a daily source of frustration. Here’s each agent’s reputation (in no particular order).

Claude Duet is built for professional use, excelling at small tasks, large projects, and multitasking. The upside: it’s powered by the top overall model across most rankings, which can fuel better quality. The downside: you need a Pro or Max plan (starts at $20/month), and you have to join a waitlist.

Muse is known for its customization and is mainly used for personal tasks like shopping or booking reservations. The upside: it has a free tier, a playful design, and access to Facebook, Instagram, and WhatsApp. The downside: it may fall short on more complex use cases.

Dots strikes a middle ground between professional and personal use. It leverages OpenAI’s leading voice and computer abilities to do great work, and adds a cute personality to keep things fun. The upside: it excels at receiving voice commands. The downside: only one dot is included in Pro or Business Premium plans (starts at $100/month)

Grok Bot offers multi-bot coordination, making it easy to spin up entire teams. Think: one bot monitoring emails, one tackling projects, and one managing the others. The upside: Grok Bot can monitor X. The downside: it requires a premium plan (starts at $30/month).

Asking people to connect their financial accounts requires trust. That's why Perplexity built Personal CFO on Plaid.

Hear Jeff Grimes, Head of Capabilities Product at Perplexity, break down how Plaid's broad account coverage and permissioned data access helps them deliver more personalized guidance.

Read the full story.

🐸 Spilling Secrets: People are reverse-engineering corporate tech secrets. One developer learned why Candy Crush is so addictive, while another found out his smart TV was listening to private conversations. Both posts racked up 3M+ views.

📸 Hidden In Plain Sight: Can you spot the hidden message in this photo? Probably not. It’s one of Fable 5.1’s talents: concealing messages in images that only it can find. Here’s the prompt if you want to try it yourself (1.8M views).

🎧 Finding Lost Stuff: A man dropped his AirPods. Naturally, the case exploded and his earbuds went flying. He eventually found them, but only after creatively using ChatGPT’s image recognition abilities.

📉 Funding Fears: The AI stock boom may be coming to an end. The culprit? Debt markets are cutting off funding for new data centers, at least according to a viral exchange between venture capitalist Chamath Palihapitiya and Elon Musk (1.3M views).

📝 Building Furniture: One of the internet’s favorite use cases for AI is converting written manuals into DIY videos when building furniture. Videos like this could save you hours the next time you’re assembling a dresser (7.5K bookmarks).

👉 Click here to open the full prompt in ChatGPT or Claude

👉 Click here to open the full prompt in ChatGPT (Don’t forget to switch to Images!)

Grow customers & revenue: Join companies like Amazon, HubSpot, and Salesforce. Showcase your product to our 2 million+ readers and followers on socials. Get in touch.

Until next time — Zain, Theodore, & the Superhuman AI team.