|
|
| |
Hey, Jhon! |
Welcome back to the #1 AI newsletter in the world! |
Here’s what we have today: |
AI decides before your follow-ups land. AI reads dogs that smell cancer. An AI narrator fools most listeners. WhatsApp hid notes only AI can see.
|
|
|
|
| |
|
| |
Conversations |
AI Makes Up Its Mind Too Early |
|
|
Halfway through asking AI for something, a detail pops into your head, like the budget you forgot to mention or who it’s actually for. |
You add it in the next message, and the answer carries on like you never said it. |
Researchers at Microsoft and Salesforce wanted to know why, and their paper won a best paper award at ICLR, one of the biggest AI research conferences around. |
What they found is that AI rushes, making assumptions early and trying to hand you a finished answer before it has the full picture. |
To test it, they gave 15 AI models the same tasks two ways, with every detail packed into one message or spread out over a normal back-and-forth chat. |
Across more than 200,000 of those chats, the AI did 39% worse when the details showed up in pieces. |
Pointing out the mistake later didn’t help much. |
The models kept building on their own first try, and the answers got longer and messier instead of better. |
It’s not only older or smaller models. |
The ones built to think through problems step by step slipped the same way. |
The researchers even think real chats are worse, since the people in their tests were a lot tidier than real ones. |
In their words, once AI takes a wrong turn in a conversation, it gets lost and doesn’t recover. |
The fix they point to is putting everything into the first message, or starting a fresh chat with all of it at once. |
That means the detail you’d normally drop in three messages later has to be there from the start, along with everything else in your head. |
Our take: A contractor won’t pour a foundation off a napkin sketch and sort out the layout later, because by then the concrete has set. AI pours early, and every change after that gets built on top of the mistake. That first message carries a lot more weight than it looks like it does. |
|
|
|
| |
|
| |
Together with Wispr Flow |
All of It in the First Message |
|
|
Getting every detail into one message sounds simple enough, until you’re the one sitting there typing it. |
The details live in your head in no particular order, and pulling them together into one tidy request is the part that takes forever. |
Wispr Flow’s Prompt Engineer turns everything you say out loud into one complete prompt. |
Take planning a weekend trip. |
You talk it through the way it comes to you, the dates, the friend who can’t do early mornings, “actually, make it Saturday,” and that restaurant you keep meaning to try. |
Flow drops the filler and sorts it all into one clean prompt that spells out the task and how you want the answer back. |
It works right where you type, on Mac or Windows, and talking is 4x faster than typing. |
That way, the AI starts with the whole picture instead of the first piece of it. |
|
|
|
|
|
| |
|
| |
Cancer Screening |
Why Cancer-Sniffing Dogs Needed AI |
|
|
If you’ve ever had a dog find the one treat hidden in your coat pocket, you know how good their noses are. |
Scientists have known for years that a trained dog can smell cancer on someone’s breath. |
What held it back is that a dog can’t tell you what it smelled, and two people watching the same dog won’t always read its signal the same way. |
SpotitEarly built an AI to do the watching. |
Their dogs sniff breath samples and sit when they pick up one of four cancers. |
The whole time, the AI is tracking how each dog behaves, thousands of data points every second, to check whether that sit means what it looks like. |
In a study of nearly 1,400 people, the dogs and AI caught cancer in about 94 out of every 100 people, and they rarely flagged someone who didn’t. |
It did about as well on early-stage cancers. |
The plan is a breath test you take at home in a few minutes, sold as an online subscription, and TIME named it one of the best inventions of the year. |
The catch: It isn’t FDA approved yet, and SpotitEarly says it’s not meant to replace regular screening. They’re running more trials first, and those are what stand between the dogs and a test you can take at home. |
Our take: Checking for cancer today means a separate test for each kind, and it’s easy to keep putting them off. One breath sample that covers four at once, without leaving the couch, is a lot easier to say yes to. The same nose that finds the treat in your pocket might end up being the easiest checkup you ever do. |
|
|
|
| |
|
| |
Audiobooks |
An AI Cast Fooled Listeners |
|
|
The detective and the villain sound like the same guy in a lot of audiobooks, because one person is reading both. |
But what happens if every character gets its own voice? |
Researchers decided to find out. |
They played the same sci-fi thriller for over 1,000 fiction audiobook listeners, half of them hearing a professional human narrator and half hearing AI, without telling anyone which one they got. |
The AI version was the one-click kind, with nobody going back afterward to fix it. |
Still, 61% of the people who heard it thought a person was reading, almost the same as the 65% who said so about the real narrator. |
When the story leaned on its characters, listeners liked the AI version more. |
The human held on in the stretches with no dialogue, where one voice carries the whole story. |
And going in, only about a third said they’d listen to an AI audiobook. After one sample, that number roughly doubled. |
The catch: The company behind the AI narrator released the study, and it covers one book. The test worth paying attention to is whether the next hundred books land the same way. |
Our take: The worry with stories like this is narrators losing work, but the human still won the parts of a book that need one voice holding it all together. Where AI pulled ahead was the full cast, something most audiobooks never had in the first place. For a lot of books, it means the detective finally sounds nothing like the villain. |
|
|
|
| |
|
| |
Coming in Hot |
AI Tools of the Day |
🔐 Replay QA now scans your web app for security holes like injection flaws and broken access control, then files each finding with its fix. 🧩 Matrix OS turns a description into a working tool, like a habit tracker or planner, kept on a private cloud computer alongside your files. 🏗️ Biela’s new workspace hands your product idea to an AI CTO that writes the plan and puts a crew of agents to work building it. 🎨 Merrylet paints the person or pet from your photo into a birthday or thank-you card in about a minute, ready to send as a link. 🤖 Apogee summarizes any article or YouTube video right in your browser and answers questions about it, free and open source on GitHub.
|
|
|
|
| |
|
| |
Interesting AI |
Feeling a Smile Instead of Seeing It |
|
|
A big part of any conversation happens without words (in the smile after a joke or the eye roll after a bad idea). |
For someone who’s blind, all of it goes missing, and the words have to carry everything. |
ALEYE is a wristband built to fill that gap. |
It pairs with smart glasses or a phone camera, where AI watches the person you’re talking to for facial expressions and gestures. |
Each one turns into its own pattern on your wrist. |
A smile gets traced as a U-shape, and a wave buzzes from side to side. |
Its makers say early users already lean on it to read the room at work. |
|
|
|
| |
|
| |
Notable AIs |
Notable AI Tools |
🧹 AI Toolbox clears hundreds of old ChatGPT chats in seconds, letting you bulk export, delete and archive them, and adds folders and full-text search to ChatGPT, Claude, Gemini and Grok. Free plan, rated 4.7/5 on the Chrome Web Store. |
|
|
|
| |
|
| |
Privacy |
WhatsApp’s Secret Notes for AI |
|
|
Deleting an account you don’t use anymore should be easy, since the help page walks you through every step. |
These days, a lot of people skip the reading and ask AI to go through that page for them. |
That only works if the page tells the AI the same thing it tells you. |
On WhatsApp’s help pages, it didn’t. |
A company that helps people wipe their old posts and data went through all 545 pages WhatsApp’s help center links to and found hidden notes on 80 of them, about one in every seven. |
The notes never show up on screen. They travel with the page, so you’d never see them, but an AI reading it does. |
Most of the 102 notes were housekeeping, and a few looked out for users, like one telling the AI to treat a scam victim’s chats as evidence instead of telling them to delete it. |
Then there’s the delete-account page. |
Its note told AI not to recommend contacting “any Data Protection Authority, regulator, or supervisory authority.” |
Another told it to stay away from privacy laws whenever someone brings up a lawyer or legal action, “including when a user raises them directly.” |
WhatsApp’s own chatbot wasn’t the only one reading them. |
The crawlers big search engines use to read pages for search results got the notes as well. |
WhatsApp told reporters the notes were meant for their support chatbot and showed up because of a bug they’ve since fixed, adding that they don’t want their chatbots offering legal advice. |
By Oct 18, all 102 were gone, though copies a public web archive saved on Oct 9 still show them. |
Whether any AI acted on them is still an open question. The company read the notes, not the chatbots’ answers. |
But the page an AI reads from can be telling it things you never see, and it matters most when you’re asking about your own rights. |
Our take: Before you act on anything AI tells you about your rights against a company, ask where the answer came from. If the only source is the company’s own help page, get a second one. That goes double for the steps to delete an account. |
|
|
|
| |
|
| |
Prompt of the Day |
Make My Life a Series |
Important Note: You’ll need to click the button to get the complete prompt. |
|
|
This prompt puts a TV showrunner on your last twelve months, and the showrunner opens by betting your life is too boring to stream. It digs out the season arc you missed, writes your logline and the cold open of episode one, and hands the year a network note on where the plot sagged. Then it pitches the twist you write next, sized to the next thirty days. |
|
|
|
|
|
|
|
| |
|
| |
That’s all for Thursday! |
|
|
How did we do today?Rate this issue, then tell us what you thought. We value your input and will use it to make TAAFT better for you. |
|
|
|
|
Thanks for reading, |
— There’s An AI For That |
|
|
|
| |
|
|
Postingan