Showing posts with label AI News. Show all posts
Showing posts with label AI News. Show all posts

13.11.26

Kimi K3.1 Pro: "Open" Doesn't Mean Safe for Client Data


Kimi K3.1 Pro is an open AI model, so anyone can download it for free, but the download is 1.56 terabytes and it needs data center machines to run. For a virtual assistant, that means "open" won't put it on your laptop, and it won't protect your client's files.

Kimi is the chatbot from Moonshot AI, a Chinese AI company. You'll see "open" used as if it means free, private and safe all at once. It means something much narrower.

What Open Means for Kimi K3.1 Pro

Every AI model is built on weights. These are the billions of numbers a model learns during training, and they're basically its trained brain. ChatGPT's stay on OpenAI's servers. Open weights means the company lets you download them.

Moonshot posted these on Hugging Face, a site where AI companies share models. The first upload there is dated November 2, and Moonshot had promised the weights by November 4.

Moonshot calls this an open-weight model. That's different from fully open source, where you'd expect to see everything behind the project. The page lists model files and the code to run them, with no training data.

It also has its own license, the rulebook for using it. You can use, change and even sell it for free if you keep the license notice, and two more conditions are aimed at giants.

AI hosting companies making over $20 million in 12 months need a separate deal. Products with over 100 million monthly users or $20 million in monthly revenue must display the model's name. That's not you or me.

Why You Can't Run It at Home

Look at the size on that page: 1.56 TB. Moonshot lists 2.8 trillion parameters, which is just the count of those learned numbers.

[INSERT IMAGE: hf-kimi-k3-1-pro-files.jpg] The Kimi K3.1 Pro files on Hugging Face, showing the 1.56 TB size and the license file

Storage is the easy part. The setup guide from vLLM, one of the tools Moonshot recommends for running it, describes its simplest working setup on NVIDIA hardware as eight GPUs, the powerful chips AI runs on. And these are data center chips, not the kind in a gaming PC.

So when people talk about running it yourself, they mean a company with serious hardware. That's still useful, because other businesses can host the model too. Hugging Face lists Together AI as one. But that's a developer service, not something you'd open to draft a client email.

How to Try Kimi K3.1 Pro

The realistic way is Moonshot's own Kimi, on the website or the iOS and Android apps.

Is it free? It looks that way, but Moonshot doesn't say so outright. The pricing page lists a Free plan at $0, without naming K3.1 Pro in it.

On November 7, K3.1 Pro was in the model menu on kimi.ai even before logging in, though K3 was selected by default.

Moonshot doesn't state how many K3.1 Pro messages a free plan gets, and the pricing page warns that K3.1 Pro requests may queue or fail at launch.


The model menu on kimi.ai, with K3.1 Pro at the top

Paid plans start with Plus at $19 a month, or $180 billed yearly. The full 1-million-token context window, meaning how much text it can hold in one chat, is listed only for Max and Ultra.

The Client Data Question

Here's where "open" trips people up. Open weights only keep data in-house if you run the model on your own machines. If you use Kimi's app, your prompts and files go to Moonshot, like any other chatbot.

For users outside mainland China, Kimi's terms come from NOVASCENT PRIVATE LIMITED, a Singapore company using the Moonshot AI name, and they were updated November 5.

The privacy policy says it can use your prompts and files to improve its services, including training its AI models. The terms let you opt out by emailing membership@moonshot.ai, but with a catch: "Opt-out applies prospectively only." Moonshot doesn't have to remove anything already used in training.

The policy says your information may be stored outside your country, without naming where. A help page on kimi.com says mainland China, but that site's policy is written for mainland users, so I wouldn't assume either way.

The terms also say not to submit personal data or confidential third-party information you don't have the right to use. Client inboxes are full of exactly that, so you'd need real permission first.

Kimi Business is different. Its supplement says customer content isn't used for training by default, unless the customer authorizes it or the law requires it. That's a business contract, not the free app.

What This Means for VAs

I'm not singling out Kimi. These are the questions I'd ask about any AI tool before client work goes near it. Who holds what I type? Is it used for training, and can I turn that off? Does my client know I'm using it?

If you're curious about K3.1 Pro, try it with made-up details or your own non-client tasks. My 30 free VA prompts are a good place to start.

Then put the client side in writing. If a client asks why you can't just use the newest free model, my post on setting AI expectations with clients will help with that talk.

And the next time someone says a tool is "open," you'll know what to ask: open how, and where does my client's data go?

Do your clients have a rule about which AI tools you can use? 

10.11.26

DeepSeek Super Agent: Why It Thinks Like a Virtual Assistant


DeepSeek Super Agent is an open-source research prototype for software teams, so it's not something you'll use with clients this week. But its rulebook reads like a job description for a great virtual assistant, and that's why I think VAs should know about it.

Quick note: several products use the name "Superagent." This post is about DeepSeek's, the one that made AI news on October 31.

What Is DeepSeek Super Agent?

DeepSeek is the Chinese AI company behind the DeepSeek chatbot. On October 31, its team published a design article called "Exploring Super-agent Design" and put a working prototype, dsh-superagent, on GitHub. As of November 7, the project has more than 20,000 stars there, which is GitHub's version of bookmarks.

The problem it tackles is one we know well. With today's AI tools, you open a chat, give a task and keep checking on it. DeepSeek says a person can usually follow only three to five of these at once.

So the idea is one always-on agent (an AI that takes actions, not just answers questions). A part called the Router takes in every request. It passes each job to a Worker, which does that job in its own separate space. A third part, the Watchdog, checks that the Workers are really making progress.

Think of a lead VA who reads every message, hands tasks to the team and checks in on them. That's the shape of it.

Can You Use It With Clients?

Not really, and DeepSeek says so itself. The GitHub page calls it a research prototype, not an official DeepSeek product, and warns it isn't production-ready. Some ideas in the article aren't even built into this version.


The dsh-superagent page on GitHub, with DeepSeek's research prototype warning

It's also made for developers. It watches one GitHub project (the place where a team stores its code), plus Feishu, a work chat app, if a team adds it. It can answer with code changes. It runs on DeepSeek Harness, the company's open-source agent app, which is in public preview.

If you see posts calling this the tool that will run your whole business, take a breath. It isn't that yet.

The Part That Sounds Like Our Job

Here's what got my attention. The project's first document sets one rule for everything: "Take as little human attention as possible, and give as much value as possible."

Then it spells out what that means. The agent acts only when asked, or when a written rule allows it. It confirms a request once, stays quiet while it works, and comes back when there's a result or a decision to make. It asks one clear question at a time. If it fails, it says so.

If you've been a VA for more than a month, you know that list. Good VAs don't send five updates a day, and they don't bury the one question that matters in a long message. They protect the client's attention.

The design article even gives a personal assistant example. A meeting invite comes in, so the agent checks for schedule conflicts, gathers background and prepares a draft reply for the person to approve. That's a normal Tuesday for a lot of us.

The Number to Be Careful With

You may see the claim that Super Agent finished work in under 16 minutes, compared with over 29 for the usual way. Look closer before you repeat it.

In that test, the real coding was swapped for a tool that just used up time. DeepSeek was measuring how well the system juggled requests, not how good the work was. DeepSeek calls these results preliminary, and the 29 is from its slowest comparison setup. So it doesn't show that AI does a job twice as fast. One detail did stay with me: each test had about 300 messages, and only about 20 were real work requests. Spotting those 20 is a skill we practise daily.

DeepSeek also flags the risk. An agent that reads messages from everywhere can pick up untrusted instructions, and login details could leak if they reach the AI. If that isn't solved, the company says, this kind of system can only work in a small, trusted setting. For client work, that's the part I'd remember. It's also why I liked the idea behind 1Password for agents.

What This Means for VAs

This is a prototype, but the direction is clear. AI companies are designing assistants that watch messages, calendars and inboxes without being asked. I wrote about the same shift with Claude Duet.

Look at what makes DeepSeek's design work: written rules. Who may ask for things, what can happen without approval, when to interrupt and when to stay quiet. Someone has to write those rules for each business, and check what the agent does. That's work a good VA already understands.

DeepSeek's own advice is to start narrow, with proactive work only recorded at first, not acted on. That's smart for any new AI tool you try for a client.

So here's your small step this week. Write down your client's "when to interrupt me" rules: what's urgent, what can wait for the daily update, and what you can decide alone. If you're starting with someone new, my client onboarding questionnaire will help you get those answers.

Do your clients have rules like these yet? 

1.11.26

AI News This Week (Oct 25–31): What Changed for VAs


AI news this week was mostly about your inbox, your calendar and your client's slide decks. Here are the updates from October 25 to 31 that change how a virtual assistant works, and one outage you should know about.

I've grouped them by the part of your day they touch, and every date comes from the company's own announcement.

Inbox and Calendar: ChatGPT Reaches Apple Users

On October 27, OpenAI added iCloud Mail and iCloud Calendar to ChatGPT. It can find emails, catch you up on your inbox, check your schedule and update events. OpenAI says it's available on all ChatGPT plans, on the web, the desktop app and iPhone.

This matters if you have a client who lives in Apple Mail instead of Gmail or Outlook.

Then on October 30, the ChatGPT phone app started letting you stay signed in to two accounts and switch between them. Each one keeps its own chats, memory and settings. If you use your own account plus a client's Business workspace, this keeps the two from mixing.

The Outage Nobody Talked About

Here's the story I found most useful, and it got almost no coverage. Claude's status page shows its Google Calendar connector (the link that lets Claude read your calendar) had errors from October 26 to October 28, about two and a third days. A separate Gmail connector problem lasted about six and a half hours on the evening of October 27, Pacific time.


Claude's status page showing the Google Calendar connector incident, October 26 to 28

So if Claude told you "there's nothing on the calendar Tuesday" during that window, check it again by hand. And if your whole inbox routine runs through an AI connector, keep a manual fallback for days like this.

Docs, Decks and Spreadsheets

On October 29, Google said the Gemini app can now preview and edit files right inside the chat: PDFs, Docs, Sheets, slide decks and plain text. Google says it's for all users wherever the Gemini app works. Making files in Gemini isn't new. Fixing them without copying everything into Docs first is.

Google Slides also got a bigger one. Gemini in Slides can now edit a whole deck from one request, like "make the titles on all slides the same font size and color." It's English only, on paid Workspace plans and Google AI Pro or Ultra, and Google says the rollout could take more than 15 days to reach you.

Two more for the design and data side. Canva launched Brush, a tool that edits an existing design from a prompt, for Canva Pro subscribers. And Microsoft's October Copilot roundup confirmed a "Clean data" skill in Excel that finds duplicates and messy formatting, then asks you to approve each fix. That one needs Microsoft 365 Copilot.

Team Chat and Meetings

Slack's Slackbot can now run a task on its own when something happens, like a new message or an emoji reaction. Think "every time a client posts in #requests, add it to my list." Slack says what you get depends on your plan and your admin, so ask before you count on it.

Notion said on X that its Meeting Notes extension for Google Meet now labels each speaker by name, instead of just "you" and "everyone else." That makes action items much easier to assign. Notion also says its Agent can now build Excel, Word, PowerPoint and PDF files for you.

New Models and Plan Changes

OpenAI released GPT-6.1 Terra on October 26. A model is the AI brain behind the chat, and OpenAI pitches this one for "computer use" (the AI clicking through apps for you) and "clear, natural writing." It's in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, and your usage limits stay the same.

On the Claude side, two quieter changes. On October 29, Haiku 4.5 left the model picker and anyone using it was moved to Haiku 5.5, so check any saved prompts you ran on it. And on October 30, Anthropic renamed its top plans: Max 5x is now Max ($100 a month) and Max 20x is now Ultra ($200). Verified Ultra accounts no longer have the five-hour limit. The weekly limit didn't change, despite what some posts say.

The Bigger Picture

Anthropic, the company behind Claude, made its IPO filing public on October 26, planning to list on Nasdaq as "ANT." You may have seen "$2 trillion IPO" headlines. The filing doesn't set a price or a valuation yet.

Closer to home, the Foundation for Media Alternatives published a Philippine AI risk report on October 30. Job loss and "labor intensification" (more work, faster, for the same pay) came out as the most widespread risks. The Philippines still has no AI law, and the report says seven House bills are being merged into one.

What This Means for VAs

Look at the pattern. Almost every update this week moves AI deeper into the tools we already manage for clients: mail, calendars, decks, Slack, meeting notes. That's good news for us, because someone still has to set these up, check the output and notice when a connector quietly breaks for two days.

That last part is the skill. The FMA report warns about "more work, same pay," and the way around it is to sell the judgment, not the hours. A VA who can say "I connected your calendar to AI, and I double-check it every morning" is offering something a tool can't.

My pick for this weekend: if you have an Apple-using client, test the ChatGPT iCloud connection on your own account first, then decide if it's safe for theirs.

Which of these updates are you trying first? 

A virtual assistant smiling as a few big coins drop into a jar Upwork has 77,000 fewer active clients than it did a year ago. If your feed ...