What if your feed was an AI you could give instructions to?

Meta is rolling out "Your algorithm" for Facebook Feed and Reels. You can describe what you want to see in ordinary words, rather than relying only on likes and watch time. "Interested" and "Not interested" may also ask why, so you can give more precise feedback.

Try a 60-second feed reset:

1. In Facebook, open a post's three-dot menu and look for "Your algorithm."
2. Add a specific interest: "Hands-on AI tutorials for small businesses, with real steps and examples."
3. Mark a few vague AI-hype posts "Not interested" and say why.
4. Watch what changes in Feed and Reels. Keep tuning; one instruction will not magically fix everything.

This is for Facebook, not Instagram. The feature is rolling out, so it may not appear in every account yet.

The bigger shift: instead of guessing what the algorithm wants, you can tell it what you want.

Source: https://about.fb.com/news/2026/10/introducing-your-algorithm-facebook/

#AI #Facebook

New to AI Lab? Pick one useful starting point.

For everyday work: build a personal AI assistant in 30 minutes.
https://t.me/AISystemAgentLab/85

For business: see one small AI system that can turn scattered messages into useful follow-ups.
https://t.me/AISystemAgentLab/94

For builders: download the step-by-step Personal Task Bot guide, including memory and deployment.
https://t.me/AISystemAgentLab/119

Starting from zero? These five free learning paths come from the teams building AI.
https://t.me/AISystemAgentLab/82

This week we are choosing a new field test. Vote for the real task you want us to solve:
https://t.me/AISystemAgentLab/293

What to expect here: real inputs, reusable steps, honest results and the checks a human still needs to make. Pick a path and put one idea to work.

5 of 7 voters chose this: what can an AI app access after you connect your files?

Let's use ChatGPT + Google Drive as a concrete example.

Before you click Allow, check four things:

1. ACCOUNT + SCOPE. The right Google account? Read or write access? Check the consent screen; permissions can vary.

2. ACTIONS. "Always ask" changes when ChatGPT asks before using an app. It does not revoke Google access.

3. DATA USE. Check Memory, model improvement and chat retention separately. One toggle does not cover all three.

4. EXIT. Disconnect in ChatGPT, then review Google Account > Linked apps. Old chats and saved memories are not erased automatically.

Use a harmless test file, not a client document.

Official details:
https://help.openai.com/en/articles/10408842-google-app-data-controls-faq
https://support.google.com/accounts/answer/13533235?hl=en

Next: a reusable checklist for any AI app.

Before you connect an AI app to your files, run this 5-minute check.

1. ACCOUNT. Which account are you connecting? Choose one that has access only to the files needed for this task.

2. SCOPE. Read the permission screen. Can the app only read, or also create, edit and delete? Is access limited or broad?

3. DATA USE. Check retention, model-training and memory settings separately. Write down what applies to this specific app.

4. EXIT. Find both controls now: disconnect inside the AI app and revoke access in the source account. Check separately how to remove old chats or saved data.

Test with a harmless file first. If you cannot answer one of these four questions, pause before connecting real work or client data.

Our ChatGPT + Google Drive example: https://t.me/AISystemAgentLab/296

Official guides for that example:
https://help.openai.com/en/articles/20001494-connecting-and-managing-app-accounts-in-chatgpt
https://help.openai.com/en/articles/10408842-google-app-data-controls-faq
https://support.google.com/accounts/answer/13533235?hl=en

One document to summarize. An AI app asks to edit every file in your Drive. Would you tap Allow?

This is a hypothetical example, not a live account test. PAUSE until you know why that scope is needed and whether a narrower route exists.

Copy this prompt using public permission text, never tokens or private files:

"My task: [one outcome]. The app requests: [exact consent-screen wording]. Its policy says: [public excerpt or link]. Make a table: permission | needed for my task? | risk | narrower alternative | what I must verify myself. Mark unsupported claims UNKNOWN. Do not assume you can see my account or granted scopes. End with a human verification list, not an approval."

Then check the real authorization screen, data-use settings and revocation path yourself. An AI-generated table cannot grant or verify access for you.

Our ChatGPT + Google Drive walkthrough: https://t.me/AISystemAgentLab/296

Your AI app wants access. Is it asking for more than it needs?

For our next practical breakdown, we want one real reader question.

In the comments, share only:
- the app name
- the job you want it to do

Comments are public. Please do not post screenshots, permission text, email addresses, account details, files, passwords or API keys.

We'll choose one case and ask the author separately for redacted permission wording and explicit consent before using their example anonymously. Then we'll compare the requested access with the task, check official docs, explain what remains unknown, and show how to disconnect.

You don't need to connect the app just for this.

Which app should AI Lab examine?

100 reviews. Three product ideas hiding in plain sight.

I read 100 recent English-language Steam-purchase reviews for AI Roguelite, an AI storytelling game. 83 recommended it; 17 did not. The praise is clear: people love building worlds of their own. The friction is just as useful.

1. The story ignores my choices. Build a visible "story lock" for goals and forbidden actions, checked before each turn.

2. The character changes between images. Build a reusable visual reference, with a one-click option to skip images entirely.

3. Setup breaks before the adventure starts. Autosave the world draft, test model connections, and let people recover from failed generation.

These are hypotheses, not validated businesses. Next step: prototype one fix and watch five people try it. That's how reviews become a product test instead of a list of complaints.

Which one would you test first: 1, 2 or 3?

Source: https://store.steampowered.com/app/1889620/AI_Roguelite/

Your next product idea might be buried in customer complaints.

In our last post, 100 AI Roguelite reviews produced 3 hypotheses. Here's a prompt you can use on any review export. Remove names and personal details first.

Give the CSV to Codex or Claude Code:

"Group complaints by the job customers were trying to do. For each theme, show review row IDs, count distinct reviews, and include up to two counterexamples. Separate observed pain from your proposed fix. Rank 3 small product hypotheses by frequency, severity and ease of testing; explain the ranking. For each, write a one-week test and one metric. Mark weak evidence as uncertain. Do not invent demand or willingness to pay."

Example from our case: the story can override player choices -> a visible story-goal lock -> test 5 players for 10 turns with/without it; count unwanted AI actions.

If the evidence is thin, say so. AI can sort feedback. It cannot validate a market for you.

The original case: https://t.me/AISystemAgentLab/300

Your website has two AI doors. Are you closing the right one?

OpenAI announced support for a shared robots.txt training-preference signal on October 30. Important: it plans to implement it in the coming months. This is not a universal switch you can rely on today.

Today, OpenAI documents two separate crawlers:

- OAI-SearchBot helps surface pages in ChatGPT Search.
- GPTBot crawls content that may be used for foundation-model training.

Their rules are independent. You can allow search while disallowing training, or choose differently.

Practical check: open /robots.txt on your own domain. Look for both crawler names and any wildcard rules. Decide what you want for search and training separately; have your site admin review changes before publishing. Do not assume a training opt-out also removes you from AI search.

Announcement: https://openai.com/index/helping-websites-express-ai-training-preferences/
Crawler guide: https://developers.openai.com/api/docs/bots

AI can ace an exam. Can it find a kindergarten that fits your family?

On November 2, OpenAI introduced LifeVal: an evaluation of AI on everyday tasks, from childcare and budgets to government forms.

Here's the useful shift: judge an assistant by the decision it helps you make, not by how fluent its answer sounds.

Try a 10-minute test:
1. Pick one real task you've been postponing.
2. State your location, budget, deadline and non-negotiables.
3. Ask for three options, evidence links, trade-offs and what's still unknown.
4. Check the sources yourself. Approve any booking, payment or submission personally.

Copy this prompt:
"Help me [task]. My constraints: [details]. Compare three realistic options. Cite current sources for every key fact. Flag uncertainty and tell me what I must verify. Do not contact anyone, book, pay or submit anything."

The win isn't an impressive answer. It's a decision you can trust.

Source: https://openai.com/index/lifeval/

#PracticalAI #AILab

Your next app can solve one annoying problem, just for you.

Google AI Studio lets you create native Android apps from plain-English briefs. Eligible adults in select markets, including the US, can run private personal apps on their phones through Google Play without a developer account. Public publishing is separate.

Build "Where Is It?" for home or office gear.

Paste into AI Studio:
"Build a simple Android app. Track item name, location, borrower and return date. Show overdue items first. Save on-device. No login or cloud sync. Use large buttons."

Before you rely on it:
1. Add five items, close and reopen. Are they still there?
2. Enter an invalid date. Is the error clear?
3. Edit and delete an item. Does the list stay correct?

In AI Studio, choose "Build an Android app" and preview in the browser.

Build: https://aistudio.google.com/
Source: https://blog.google/products-and-platforms/platforms/google-play/google-play-updates-october-2026/

#VibeCoding #Android

The real AI research workflow is less "Eureka!" and more "show me the data."

Anthropic analyzed 207,049 science-related sessions on consumer Claude and Claude Code plans. Data analysis made up 33.7% of Claude sessions and 48.1% of Claude Code sessions. Writing was common; data gathering and hypothesis generation less so.

Usage is not proof of faster discovery. My takeaway: start with work you can inspect.

Try it on a non-sensitive spreadsheet:
1. Ask AI to chart one trend and explain its calculation.
2. Ask for missing values, outliers and alternative explanations.
3. Check five sample rows yourself before sharing the result.

Copy this:
"Analyze this table. Show the math behind each claim. Flag missing or unusual entries. Separate observations from hypotheses. Do not invent data."

Useful AI doesn't have to do the science. It can make the next human decision clearer.

Source: https://www.anthropic.com/research/how-researchers-use-claude

#AIResearch #PracticalAI

AI designed a gadget. Manufacturing exposed the gaps.

In OpenAI's internal test, two team members used GPT-6 Astra to design a device for talking to Codex in under a week. A manufacturer built a small prototype batch in under a month. But getting it ready still took several rounds of human review: the model missed production limits, tolerances and assembly details.

The lesson: a convincing render is not a manufacturable design.

Before you print, cut or order parts, ask AI:
"List every dimension, material, tolerance, fastener and assembly step this design depends on. Mark assumptions. Tell me what needs a real measurement or expert review before fabrication."

Try this on a low-risk object, like a desk cable organizer. Measure the actual space, then test-fit a cheap prototype. For load-bearing or electrical parts, get qualified review.

Source: https://openai.com/index/accelerating-engineering-for-the-physical-world/

#PracticalAI #Hardware

Before you pay for a repair, find the receipt.

That old email or paper slip may be the start of a useful warranty record. AI can organize it; AI cannot decide whether your damage is covered.

Try this with one laptop, phone or appliance:

1. Find the receipt and model label. Redact payment and personal details before sharing files with AI.
2. Extract the product, purchase date, seller, model and any warranty wording actually visible. Mark missing fields "unknown."
3. Check the seller's and manufacturer's official support pages yourself: coverage, exclusions, proof needed and claim route. Terms vary by product and location.
4. Save verified facts, receipt, support link and a reminder in one row. Repeat for items worth tracking.

Copy this prompt:
"Make a warranty record from these files. Cite the file for each field. Do not guess coverage or dates. List what I must verify on the official support page."

AI organizes the evidence. You make the claim.

#PracticalAI #LifeAdmin

1,000 AI agents. 30 hours. One checkable proof.

Anthropic reports that a team running an early version of Claude Fable 5.2 reproduced a formal proof of Fermat's Last Theorem in Lean.

Important distinction: AI did not discover a new solution. Andrew Wiles proved the theorem in the 1990s. The agents translated an existing proof into steps a computer could verify. Anthropic calls this a toy experiment; it supplied 60 target statements and 47 already-proved lemmas.

Why should anyone outside mathematics care? A persuasive AI answer is not the same as a checkable result.

Try this rule on your next AI task:
- Code: ask for a test you can run.
- Spreadsheet: ask for source rows and the formula.
- Research: ask for the exact passage behind each claim.

If the check fails, the work is not finished.

The breakthrough is not "more agents." It is "show me how to verify it."

Source: https://claude.dev/blog/safely-scaling-multi-agent-teams-with-fable-5-2/

#AIResearch #Verification

The chart looks right. The number isn't.

Google and Artificial Analysis just launched MKA: 200 tasks across 83 occupations testing whether AI agents can work through visually complex files, not just read plain text. Think slides, charts, spreadsheets and diagrams.

Here's a useful test for your own work:

1. Take a presentation and the spreadsheet behind one chart. Use files you are allowed to share with your AI tool.
2. Ask: "For this chart, list every displayed value, its source sheet and cell, the calculation and units. Flag mismatches. If you can't trace a value, mark it UNVERIFIED."
3. Check the flagged cells yourself before sharing the presentation.

This won't make AI your auditor. It can make a hidden mismatch much easier to find.

Source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/

#AIWorkflow #DataQuality