SI Agent News

• ExtremeTech reports Google is building a memory hub for Gemini, its AI assistant. That means one place where the assistant keeps what it knows about you. What this means for you: check what your own assistant remembers about you. You can do that in the Try This Today step below.
• ARIA, a UK agency that funds research, says it exists to unlock science and technology breakthroughs that benefit everyone. What this means for you: it is just a name to recognise for now. There is nothing to act on.

Google has released Gemini 4 Carbon, a new AI model built with stronger cybersecurity skills.

Sina Finance reports a significant jump in those skills. Other outlets carried the same news. Pedaily's headline went further, saying it tops a ranking but not naming the ranking.

A model is the engine inside an AI app. Cybersecurity skill means it is better at finding and fixing weak spots in software. Think of a locksmith who is excellent at picking locks. That is great for defense and risky in the wrong hands. None of the coverage we saw describes a public launch, so check Google's own page before you plan around it. Price not stated.

What this means for you:

• Small business owner: ask your software vendors how they use AI to find security holes.
• Job seeker or student: basic security knowledge plus AI skills is a mix worth building.
• Creator: turn on two-step sign-in for your accounts today.

Google is building an assistant that acts, not just answers. The Verge reports Google is working on a new agentic experience for consumers. Agentic means the AI can take actions for you, not only reply. It is still in the works, so there is nothing to download. Price not stated. It is worth watching if you already use Gemini daily.

What this means for you: nothing to buy yet. Start practising how you hand a task to an assistant, because that skill carries over to every one of these tools.

One of those assistants has already shipped, in Japan. Nikkei reports Amazon Japan launched Alexa Plus on October 29. It links with three services: Uber Eats, Demae-can and Tabelog. Sanspo adds that it reaches owners of the newest Echo speakers first. Price not stated. It is worth it only if you are in Japan with a supported device.

What this means for you: this is a preview of where assistants are going. You state the goal, it handles the steps, and you still check the result.

New York is putting limits on companion chatbots. These are bots built to act like a friend. The Associated Press, via the Boston Herald, reports New York is enacting farther-reaching limits on how they behave with children. The outlets describe it differently. The Economic Times says the governor signed a law preventing minors from using companion chatbots. WGRZ describes limits on how chatbots interact with minors. So read the actual rules before you decide what they mean.

What this means for you: if kids use chatbots in your home, expect age limits and more settings. Check which apps they use, and talk with them about it.

A report says Chinese labs learned from OpenAI's answers. The Information, as carried by Reuters via TradingView and the Economic Times, says DeepSeek and Moonshot AI relied heavily on answers from OpenAI's models to train their newest models. It cites people with knowledge of the matter. The coverage we saw does not confirm it. It is a report.

What this means for you: a cheaper model may have learned from a pricier one. Compare answers across tools, and do not paste private data into tools you do not know.

US copyright law will not cover art made solely by AI. That is the headline from Mashable. The word solely matters. It points to work where AI did all the making.

What this means for you: if you sell or brand with AI images, your own edits and choices are what you can point to. For anything valuable, ask a lawyer.

Learning from another model's answers. Picture a student who studies a tutor's worked solutions instead of the textbook. The student gets good fast. But they only know what the tutor showed them. A model can do the same with another model's answers. The company trains its own model on the replies of a stronger one. It is cheaper than learning from scratch. Whether a company is allowed to do this is a separate question. Today's report does not settle it. The idea explains why a newer model can sound a lot like an older rival.

The ask-before-acting prompt. Agents will need this habit, and you can practise it in any chatbot today.

1. Write your goal in one sentence.
2. Add this: Before you do anything, list the steps you plan to take and what you need from me. Wait for my yes. When you finish, tell me what you did and what you were unsure about.
3. Read the plan and fix one thing.
4. Say yes.

Use it for drafting emails, planning a trip or sorting a to-do list. It costs nothing extra, and it keeps you in charge.

1. Open the AI app you use most.
2. Look in settings for a page called memory or personalization, if your app has one.
3. Read what it has saved about you.
4. Delete one item you do not want kept.
5. Type: What do you remember about me?

You will know it worked when you can say what the app remembers about you.

Let AI do the shortlist and you do the booking. Ask ChatGPT or Gemini for three restaurant options with a price range and a one-line reason each. Then book yourself. This is my estimate, not a measured figure. What you give up: details can be out of date, so check the opening hours.

The mistake: using an image made only by AI as your logo and assuming nobody else can use it. Today's copyright news says art made solely by AI is not covered. The fix: keep your own edits and drafts, and do not count on exclusivity.

Perplexity is an AI search app that shows the links it used. It is good for quickly checking reports like the ones above. The limit: it can still summarise a page wrongly, so open the links. An alternative from a different company is Microsoft Copilot.

Safe to use? Check their privacy page.

Worth it if... you often check claims or research topics for work.

Glossary word: guardrails. Guardrails are rules that limit what an AI model will do. Think of the lane lines and crash barriers on a road. Some are built into the model by its maker. Some are added by the app. Some now come from laws, like New York's. Guardrails explain why a chatbot refuses some requests. They also explain why a stronger model, like one with better security skills, may come with limits on who can use it. When a chatbot says no, it is often a guardrail at work, not a bug. You can usually rephrase the goal honestly and try again.

The phrase artificial intelligence comes from a proposal for a summer workshop at Dartmouth College, according to Dartmouth's own history.

I asked my AI assistant to book dinner. It confirmed three times. Finally, a colleague who double-checks.

• Spot the difference between a chatbot that answers and an agent that acts.
• Use an ask-before-acting prompt so an assistant shows its plan first.
• Check what your AI app remembers about you, and delete what you do not want kept.

Thanks for spending a few minutes with us. If today's issue helped, send it to one person who would find it useful.