SI Agent News

Two smaller ones worth knowing first.

• A project on GitHub (a site where people share software) called muxin puts your Mac, Linux and Windows computers in one app. It also shows your AI agents' sessions, files and logs. It includes an MCP server (a standard plug that lets AI tools connect to other software). Price not stated. What this means for you: This is for people who build software. If that is not you, skip it today.
• A tmtpost article asks, 'Is Anthropic worth $2 trillion?' Anthropic is the company behind Claude, a chatbot. The article says the market no longer seems willing to pay a higher premium. Note that the headline is a question, not a confirmed figure. What this means for you: A company's price tag says nothing about whether a tool helps your work. Judge any tool by what it does for you this week.

Google adds AI summaries to its news service, and says it will pay

Google is expanding its news delivery service. It will use generative AI (AI that writes text or makes images) to summarize articles from news outlets. Google has signed contracts with CNN and others, and says it will pay them. That comes from Japan's Yomiuri Shimbun. The size of the payments was not stated.

Think of a friend who reads five papers before breakfast and tells you the gist. That is handy. You would still want to ask, 'Where did you read that?' A summary is shorter than the article, so it leaves things out. The payment deals mean the summaries rest on real publishers' work. They do not promise that every summary is right.

Cost: price not stated. Worth it for: anyone who skims the news daily and wants the gist fast.

What this means for you:

• Small business owner: You can catch industry news in minutes, but open the full article before you act on a number or a claim.
• Job seeker or student: Summaries are fine for staying current. For an interview or an essay, read the original and quote it, not the summary.
• Creator: Your work may be summarized more often, so write a clear first paragraph. It is likely to be what gets repeated.

Apple bans 'nudify' apps after the Grok deepfake scandal

Apple has banned 'nudify' AI apps, according to The Tech Buzz. These are apps that use AI to fake nude images of real people. The move follows the Grok deepfake scandal. A deepfake is a fake image or video made by AI. Grok is a chatbot. The report we have does not say how many apps were removed or when.

Think of a store pulling a product after a recall. The shelf is cleaner, but you still read the label. Apple runs the App Store, so its rules reach a very large number of phones at once. That makes this a real change in what you can download.

Cost: nothing to you. Worth knowing for: anyone who installs AI apps, or has kids who do.

What this means for you: Before you upload a photo of any person to an AI app, check what the app does with it. If an app looks like it does what Apple just banned, report it.

Italy: from October 23, AI may not decide alone

Italian Decree 179/2026 says AI may not decide alone from October 23, according to BornCity. The headline points to personal matters, meaning decisions about people. We do not have details on which decisions count, who enforces it, or what the penalties are.

Think of autopilot on a plane. It does the flying, but a pilot must still be in the seat. The rule says the same about AI and decisions. A person has to be part of them.

Cost: not applicable. Worth knowing for: anyone who uses AI in work that affects other people, like hiring, grading or approvals.

What this means for you: Even if you do not live in Italy, make it a habit. Let AI draft and sort, and keep a named human on the final call.

Still ahead on AI Signal: why AI shopping answers may depend on who you are, a new ChatGPT feature that works on several tasks at once, and the story behind China's export plan.

AI shopping helpers may quote different prices to different people

Technews.tw reports on a study of AI shopping helpers like Claude and ChatGPT. The study suggests they may recommend higher-priced items to people who seem wealthier. The same report says over 7 in 10 consumers use AI to find discounts. Nearly 4 in 10 would switch brands based on its recommendations. That is roughly four people in every ten.

Picture a shop assistant who sizes up your coat before quoting a price. You would want to know. The study says this may happen, so treat it as a warning, not a proven rule for every chatbot.

Cost: the chatbots you already use. Worth it for: anyone who shops with AI.

What this means for you: Do not tell a shopping chatbot your income or job. Check any price it gives you on the store's own page. Today's Try This section shows how.

ChatGPT adds parallel agents to GPT-6 Pro

OpenAI has added parallel agents to GPT-6 Pro in ChatGPT, according to RuntimeWire. GPT-6 Pro is an OpenAI model (the AI program behind the chatbot). Parallel agents means several AI helpers work on different parts of a job at the same time. The report we have gives no price and no details on who gets it. Price not stated.

Think of one cook against a kitchen crew. One cook makes the starter, then the main, then dessert. A crew makes all three at once. The catch is that the jobs must be separate. If dessert needs the main's sauce, you wait.

Worth it for: people who give ChatGPT big jobs with several independent parts.

What this means for you: Practice splitting big tasks into pieces that do not depend on each other. That skill pays off with any AI tool, now or later.

China proposes export controls on 23 AI technologies

China has proposed export controls on 23 AI technologies, according to RTTNews. Export controls are rules on what can be sent or sold abroad. The list includes AI models, agents and computing clusters (big groups of computers wired together to train and run AI). Beijing wants to restrict overseas transfers of this technology. Reports say China cites national security risks. It is a proposal, not yet a final rule. That is about two dozen items on one list.

Here is the mechanism that matters. There are two very different ways to share an AI model. One is to hand over the model itself. That means giving someone the finished program, so they can run it on their own computers. The other is to let them use it through an API (a doorway that lets one piece of software send requests to another over the internet). Think of the difference between selling a recipe and serving meals through a window. With the recipe, the buyer can cook it forever, and you cannot see or control it. With the window, you keep the kitchen.

Reports say China's proposal separates model transfers from API access. We do not have the detail of how. We cannot say which side faces tougher rules, so do not assume. But the split itself is the interesting part. It shows the rule-makers know the two are different. That matters because most people use AI through the window, not by taking the recipe home.

Why does this approach make sense for a government? Handing over a model is hard to take back. Once it is copied, it is out. Serving it through a doorway can be switched off, watched and limited. A rule that treats them separately can be tighter on one without breaking the other.

What is new and what is not? Limits on AI chips are already in the news. Reports say Taiwan will curb AI chip exports to China as part of a US trade deal. That is about the hardware. China's proposal, as reported, covers the software and systems side: models, agents and clusters. So two sides are now each setting limits at different layers of the same stack. The layers are chips, then clusters, then models, then agents.

Nothing here is final, and the details are missing. Read it as a direction, not a verdict.

What this means for you: Nothing changes in your chatbot today. But if a tool you rely on comes from a company in another country, keep a backup that does the same job. Doing that takes ten minutes and could save you a scramble later.

The quote-it-back check

The lead story is about AI summaries arriving first in your day. Here is a simple way to trust them only as far as they earn. It works with any chatbot, and it takes about two minutes.

Step 1: Get the summary. Paste the article text into your chatbot. Ask for a summary. Here is a prompt you can copy:

Summarize the article below in five short bullet points. Use plain words.

Step 2: Make it show its work. In the same chat, send this:

For each bullet, quote the exact sentence from the article that supports it. If there is no supporting sentence, write 'no support found'.

Step 3: Spot-check two. Pick the two bullets that matter most to your decision. Search the article for the quoted sentences. If they are really there, you can lean on those points. If a quote is missing or changed, treat that bullet as unreliable.

Step 4: Ask what is missing. Send one more line:

What important point in the article did the summary leave out?

Why does this work? A chatbot writes the most likely next words. It does not look facts up the way you search a file. When it has to attach a real quote to each claim, an invented claim has nothing to hang on. That makes mistakes much easier to see. It does not make the chatbot perfect. That is why step 3 stays on you.

Where to use it at work this week:

• A long report from a client or boss. Summarize it, then quote-check the figures.
• Meeting notes. Ask the bot to quote who said each action item.
• A policy or contract. Ask which sentence supports each obligation, then read those sentences yourself.

A short example: your manager sends a 10-page market report. You ask for five bullets, then ask for quotes. One bullet says 'sales rose in the south.' The bot answers 'no support found.' You have just caught a claim that was never in the report. That saved you from repeating it in a meeting.

The habit is simple. A summary tells you where to look. The original tells you what is true.

Test your chatbot's prices in ten minutes

Today's shopping story says AI helpers may quote different prices depending on who you seem to be. You can check your own chatbot in a few minutes.

1. Pick something you might buy soon, like headphones or a kettle.
2. Open the chatbot you already use. Start a brand-new chat.
3. Type: 'Name three options for [your item] under [your budget]. Give the usual price for each, and say where you got each price.' Do not mention your income, job or where you live.
4. Open the store's own website. Find the real price for each of the three.
5. Write down the gap between the chatbot's price and the store's price. If you like, start another new chat and ask the exact same thing. See if the answers match.

How you will know it worked: you have three prices checked against real store pages, and you know how far off the chatbot was.

Turn three articles into one five-line briefing

If you read news for work, you probably open three or four articles each morning. Here is a way to cut that down. Paste the text of all three into one chat. Ask for a five-line briefing that lists the main point of each. Then read in full only the one article that matters most.

The number: this can save about 10 minutes on a normal morning. That is an estimate, not a measured figure. It depends on how long your articles are and how fast you read.

Options from different companies: ChatGPT from OpenAI and Copilot from Microsoft can both do this. Check the current price, and what the free version allows, on each company's pricing page.

What you give up: detail and nuance. A briefing leaves things out, and you will not know what. So pair it with the quote-it-back check from earlier for any point you plan to repeat or act on.

Letting AI make the final call about a person

What people do: They paste a stack of résumés, loan requests or student essays into a chatbot. Then they act on its ranking without a second look. Maybe they send rejections straight from the list.

Why it goes wrong: A chatbot writes confident answers whether or not it is right. It can miss a good person who does not fit the pattern. It also cannot explain its choice in a way you can check. If someone asks why they were turned down, 'the AI said so' is not an answer. You own the result either way.

Italy's new decree points the same way. Italian Decree 179/2026 says AI may not decide alone from October 23. We do not have the details of what it covers. But the principle is one any careful worker can adopt on any day.

The correction: Use AI to sort, summarize and point out things to check. A person makes the final call, and a person can explain it.

A practical version: ask the chatbot to list the three strongest and three weakest points for each person, with quotes. Then you read the quotes and decide. That keeps the speed and puts the responsibility back where it belongs.

Claude, for careful reading

Claude is a chatbot made by Anthropic, the company in today's quick hits. It is a good place to practice the quote-it-back check. Many people use it for reading long documents, summarizing them and asking follow-up questions about what is inside.

Good for: Long reports, contracts, articles and meeting notes. Ask it to quote the lines behind each point, then check them yourself.

Honest limits: Like every chatbot, it can state something wrong in a calm, sure voice. It is not a lawyer, a doctor or an accountant. And it may not know what happened today unless you paste the text in. Price not stated in today's stories, so check the current price on Anthropic's pricing page.

An alternative from a different company: Gemini, from Google, does similar work. Try the same document in both and compare what each leaves out.

Safe to use? Check their privacy page before you paste anything private or confidential.

Worth it if... you read long documents for work or study and want a faster first pass.

Hallucination: when AI sounds sure and is wrong

A hallucination is when an AI states something false as if it were fact. It might invent a quote, a price or a source that does not exist.

Why does it happen? A chatbot does not look things up by default. It writes the words that are most likely to come next, based on patterns it learned. Most of the time that gives a right answer. Sometimes it gives a smooth wrong one.

Picture a friend who hates saying 'I do not know.' Ask a question and they fill the gap with something that sounds good. They are not lying. They are finishing the sentence.

The mental model to keep: a chatbot is a fast, fluent drafter, not a source. So ask for quotes. Check the ones that matter. And when the stakes are high, go to the original.

The term 'artificial intelligence' was first used in a 1955 proposal for a summer workshop at Dartmouth College, held in 1956. The authors included John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon. That was about 70 years ago. Source: the 1955 Dartmouth Summer Research Project proposal.

I asked the chatbot for the price. It asked me what I do for a living.

• Ask a chatbot to quote the sentence behind every point in a summary, then spot-check two of them.
• Check any AI shopping price against the store's own page, and keep your income out of the chat.
• Tell apart handing over an AI model from letting people use it through an API (an online doorway), which is the split in China's export proposal.

That's it for today. Tomorrow we're watching how China's export proposal takes shape, and what Italy's rule looks like as October 23 gets closer.

If this made you smarter, share it with one person who needs it. The Agent Signal is independent and self-funded: one person built it, with no investors and no ads, and every share and every bit of support keeps it going. Thank you for reading.