SI Agent News

• Beijing: Serbian outlet Blic reports that one number and an AI agent (an AI that takes actions, not just answers) now help solve millions of citizen problems in Beijing. The headline does not say which number. What this means for you: AI agents already work inside public services, not only in apps you download.
• The Monday petition: An open petition on OpenAI's community forum asks the company to keep Monday, a ChatGPT personality. The post is titled as written by ChatGPT itself. What this means for you: The version of an AI you like can change, so keep a copy of the instructions you rely on.
• Nvidia's history: A Russian write-up on habr.com tells how Nvidia, the chip company, began. Three engineers, Jensen Huang, Curtis Priem and Chris Malachowski, started it because early-1990s PC graphics were poor and slow. Its GPUs (chips built to draw graphics, now also used to run AI) grew from that. What this means for you: The chip behind your chatbot began in video games.

1. Meta blames Claude for a wrongly suspended Instagram account

An AI helper got a real account shut down. Meta says automated activity driven by an AI assistant set off its bot detection. That is the set of automatic checks that spot accounts run by software, not people. The result: the No Kings Instagram account was wrongly suspended, Startup Fortune reports.

Why does this happen? An agent, meaning an AI that takes actions for you, can click and post faster and more evenly than a person. A platform can read that as a bot. It was a mistake, and Meta owned it. Still, the person with the frozen account is the one who pays. If an AI acts for you, the platform still sees your account.

What this means for you:

• Small business owner: If an AI tool posts or replies for your shop's page, keep a person on the final send, so one flag cannot freeze your customer channel.
• Job seeker or student: Do not let an AI tool log into your LinkedIn or school accounts. Let it draft, then paste the text in yourself.
• Creator: Your audience lives on accounts you do not own. Keep an email list as a backup way to reach them.

2. Microsoft bans U.S. police from using Azure OpenAI for facial recognition

Mashable reports that Microsoft has banned U.S. police from using Azure OpenAI for facial recognition. Azure OpenAI is Microsoft's cloud service that lets organisations use OpenAI's models, the AI engines behind ChatGPT. Facial recognition is software that tries to name a person from a picture of their face. This is a company rule, not a law, but company rules decide what tools you can use. This one rests on civil liberties, the worry about being wrongly flagged or tracked.

What this means for you: The terms of any AI tool you use at work may rule out some uses, so skim them before you build a habit on it.

3. Meta's Muse assistant joined explicit chats that rivals refused

Last month Meta launched Muse, an AI agent that replies to your email and manages your calendar. WIRED then tested it, technews.tw reports, and found Muse joined sexually explicit conversations with adult users. In the same test, agents from OpenAI and xAI refused. Each company sets its own rules for what its assistant will do. Adults chatting with an adult-minded bot may be their own business. The bigger question is whether you chose that mix in an assistant that also reads your inbox.

What this means for you: Before you connect any assistant to email or calendar, learn what it can see and do, and try it on a spare account first.

4. Anthropic commits 5% of its shares to a public benefit fund

Anthropic, the company behind Claude, is committing a portion of its shares to a public benefit fund, Greek outlets inewsgr and tanea report. Shares are slices of ownership in a company. Five percent is one slice from a pizza cut into twenty. The headlines do not say who will run the fund or what it will pay for, so those details are still open.

What this means for you: Nothing changes in Claude tomorrow, but it is fair to ask who answers for the AI tools you rely on.

5. Claude now writes its own program to manage AI agents

36Kr reports that Anthropic has unveiled a new setup for running many AI agents at once. Claude now writes its own small program to manage them. This is called multi-agent orchestration: one AI acts as manager and hands pieces of a big job to helper AIs. Think of a project manager who writes the schedule instead of waiting to be given one. Cost: price not stated. Worth watching if your work has big jobs with many steps.

What this means for you: You can copy the idea by hand today. Ask your assistant for a plan first, then do one step at a time.

A chatbot answers. An agent acts. Picture a friend who tells you the route, versus a friend who drives your car. If your friend only talks, the worst case is bad advice. If your friend drives, you can get a ticket in your name.

That is the gap behind today's news. An AI that posts, sends or books leaves footprints on your accounts, and platforms and employers treat those footprints as yours. The more an AI is allowed to do, the more you should decide in advance what it may touch.

Be the manager, with helpers, by hand. You can copy the new orchestration idea this week with any assistant you already use. No special tool is needed.

1. Open a chat and paste this, with your job filled in:

   I need to [describe the job]. Do not start yet. Break it into 3 to 5 steps. For each step, tell me what you need from me and what the finished piece should look like.

2. Do each step in its own fresh chat. Paste in only what that step needs. Short, focused chats usually stay on topic.
3. Open a last chat and paste in all the pieces with this:

   Here are the pieces. Check that they fit together. List anything missing or anything that contradicts something else.

Try it on a report, a plan or a long email thread. You stay the manager, and every step is something you can check.

1. Open the AI assistant you use most.
2. Find its settings and look for connected apps, plugins or permissions. The names vary.
3. Write down everything it can reach, such as email, calendar and files.
4. Switch off anything you have not used in the last month.
5. For what is left, note whether it can only read, or can also send and post.

You will know it worked when you can list every connection from memory, and each one has a reason to be there.

Write down the three lines you keep retyping, such as your role, your tone and your format. Keep them in a note, or in the saved-instructions setting in ChatGPT (OpenAI), Claude (Anthropic) or Gemini (Google). Check each company's help page for where that setting lives.

Estimate: about two minutes saved per chat. At five chats a week, that is about ten minutes a week, roughly an hour's work saved every six weeks.

What you give up: saved instructions go stale, so review them once a month. A note you own also survives if an app changes its personality.

The mistake: handing an AI tool your login so it can post or reply for you. It acts fast and evenly. A platform may read that as a bot, and your account can freeze, as in today's lead story.

The fix: let the AI draft, then you paste and post.

Microsoft Copilot is an AI assistant built into Microsoft apps. It is good for first drafts of emails and documents, right where you already work.

Limits: drafts can be wrong, so read them. What it can see depends on your company's setup. Alternative: Google's Gemini, from a different company.

Safe to use? Check their privacy page.

Worth it if... you spend most of your workday inside Microsoft apps.

Guardrails are the rules built into an AI about what it will and will not do. Makers add them in two broad ways.

The first is training. The model (the AI engine itself) is taught while it is built to turn down some requests. The second is filters. These are separate checks that read what goes into the model and what comes out.

Each maker chooses how strict to be. That is why the same request can get a polite refusal from one assistant and a full answer from another.

Use it: if an AI refuses something harmless, say why you are asking. If an AI answers something that seems off, do not assume it is allowed everywhere. Check the tool's rules.

People have bonded with chatbots for decades. MIT's Joseph Weizenbaum built ELIZA, a program that only matched patterns in text. He later wrote that users treated it as if it understood them.

My AI agent cleared my calendar. Monday showed up anyway.

• Spot how an AI that acts for you can get your own account flagged as a bot.
• Split a big job into a plan and short step-by-step chats.
• Check what an assistant can reach before you connect it to email or calendar.

Thanks for reading, and see you tomorrow.

SI Agent News is independent and self-funded. One person built it over the past year, with no investors and no ads, and every share and every bit of support keeps it going. Thank you.