Hey, AI Enthusiast!

Welcome back to the #1 AI newsletter on the planet!

Here’s what we have today:

Prompting

OpenAI’s Four-Part Prompt Recipe

When you ask AI for something, it’s easy to leave out important details without realizing it.

But the AI fills them in anyway, with guesses.

OpenAI has a fix, and you can find it in the company’s newest model guide.

What they say is to give the model enough direction so it never has to guess at the decisions that matter most.

The guide doesn’t stop there, though.

It walks you through exactly what that direction looks like.

They break it down into four parts:

Let’s take a look at a simple example:

“Write a toast for my sister’s wedding.”

That covers the result and you’ll get something decent, but the AI is going to invent the rest:

Each of those is a decision that matters. Leave them out, and the AI is going to make every one of them for you.

Now you might think the fix is writing more (hint: it isn’t).

OpenAI pushes back on that, as well.

One of its own devs says models read nuance so well now that piling on instructions drags results down.

The four parts beat the word count by a long shot.

Go look at the last thing you typed into a chatbot, and count how many of those parts you included.

Knowing the four parts is easy. Making sure you type all four of them, every time, is where it all falls apart.

Our take: A vague answer is easy to blame on the AI. But every blank you leave becomes a choice the model makes for you, and it has no idea what you’d have picked. The details that feel too obvious to mention are the ones it needs most.

Together with Wispr Flow

The Ask Nobody Has Time to Type

The result, who it’s for, the context, and what done looks like.

It all lives in your head. But when you start typing out what you want, those are the important details you leave out.

That’s where Flow’s Transforms can help.

Prompt Engineer turns your rough thoughts into the full ask.

It works in every AI chat app you use.

All you have to do is say what you want, the way it comes out.

“3 day trip. Kids are 6 and 9. Nothing before 9AM. We hate museums.”

Highlight it and hit one shortcut. Flow rebuilds it in place as a structured prompt with a title, role, task, context, and format.

It gets the same rambling sentence from you. The AI gets the whole assignment.

The prompt lands where you type, before you send it, that way you can edit it first.

AI Progress

The Internet Scored AI’s Old Goalposts

For ten years, people on Hacker News have been setting challenges for AI in the comments.

Things like passing a Turing test or driving a car in the snow.

Someone went back through all those old threads and pulled every challenge onto one page.

Then they let people vote on which ones AI has met.

In two days, 9,694 people cast 100,590 votes.

You can look through them, but here are a few highlights.

Writing working code from a few plain sentences (from 2020). 87% say AI does it now.

Full self-driving in every condition imaginable (from 2019). 78% say no.

A robot that cleans a toilet (from 2023). 69% say it hasn’t happened yet.

Notice a pattern?

Anything that happens on a screen is falling fast. Anything with wheels or a mop, not so much.

Our take: The challenges that sounded the hardest, like writing code, fell first. The ones that sounded easy, like cleaning a bathroom, are still standing. AI learned to code by reading code. A cleaning video, on the other hand, shows a toilet getting scrubbed, but not how hard to press or how to hold the brush, and no two bathrooms look the same.

The catch: The voters are Hacker News readers who chose to show up, not a scientific sample.

Books

Audiobook Characters Talk Back

For a lot of people, they’d rather listen to a book rather than sitting down to read it.

Normally, only one narrator voices every character in a book.

But if you have a full cast narrating multiple characters, it gets hard to keep track of who’s talking.

Audible is adding AI to help with that, and it’s going to take things one step further.

Its new character guide shows who’s talking right on the player itself as you listen.

Plus, it has spoiler-free cards for each character.

That’s not the most interesting part, though.

It’s Interactive Stories.

Instead of only listening, you’ll get to talk to characters from the book, and they’ll answer back in real time.

The first one puts you in a conversation with Renfield, Dracula’s servant.

In November, a new one will arrive along with a new Audible original.

Our take: Audiobooks have always talked at you. This is the first time the story stops and waits for your answer. It’s a bonus on a couple of titles for now, but if listeners take to it, books that only talk at you might start to feel a little one-sided.

Coming in Hot

AI Tools of the Day

Brain Implants

Neuralink Goes Weeks Without a Reset

Neuralink is one of those inventions that makes you go “WOW” the first time you see it in action.

If you’ve never seen someone move a cursor with their brain, head over to YouTube and check it out.

The cool part is how it works.

The implant picks up electrical signals from the part of the brain that controls movement. Then software, called a decoder, turns those signals into cursor moves.

But there’s a problem many people don’t know about.

Brain signals don’t stay the same. The implant’s tiny threads shift a little over time, and the neurons themselves change.

That means a decoder trained on yesterday’s signals slowly stops matching today’s, and the cursor gets less accurate.

To fix it, users go through a recalibration, where the software relearns their signals.

For some Neuralink users, that’s been about 10 minutes, every single day.

Neuralink has found a way around it using something its patients made without even trying.

Every hour they spent using the implant added to a pile of their own brain data.

Over two years, that pile grew to over 50,000 hours.

So, Neuralink trained AI on all of it, and now the software recognizes each person’s signals even as they shift.

Decoders built this way kept working for weeks instead of days. For some participants, calibration dropped to about 10 minutes a week.

Our take: Picture your phone needing 10 minutes of setup every morning before it worked right. You’d stop using it. Brain implants make headlines for what they let people do, but going weeks without a reset is what makes one feel less like an experiment and more like a device people rely on.

Interesting AI

Meta’s AI Wants Out of Your Phone

At some point in your life, maybe even today, you’ve gone digging through the cushions in your couch to find your TV remote.

Well, Meta wants to put a stop to that.

They built Home Link, a small USB-C plug that sits near your router and connects its AI, Muse, to your home network.

From there, Muse talks to your TV and speakers, plus other smart gear on your network that supports it.

People have already built add-ons for things like Philips Hue lights and Samsung TVs.

Plus, Meta opened up the code for anyone who wants to build their own Muse gadget.

Losing the remote might finally stop mattering.

Learning

A Free Course on Why AI Gets It Wrong

When you use AI enough, you start to notice it gets things wrong in weird ways.

If you don’t know why it happens, it’s hard to tell which answers to trust.

Anthropic has a free course that explains it.

It’s part of Claude Academy, the company’s free learning site, built partly on the way it trains its own staff.

The course, AI Capabilities and Limitations, walks you through why AI misfires in 13 lessons.

It covers where AI is strong, like popular, well-covered topics, and where it slips, like anything rare or recent.

It also explains why a long chat starts losing track of what you said earlier, and why a new chat doesn’t remember the last one.

You don’t need a technical background or any prior AI experience to take it.

There are also separate tracks for each Claude product, from everyday chat to Claude Code.

Our take: A wrong AI answer feels random until you know why it happened. This course turns those misses into a few limits you’ll start to recognize, and once you spot which one you hit, you know when to trust an answer and when to check it yourself.

Prompt of the Day

Do My Job in 1985

Important Note: You’ll need to click the button to get the complete prompt.

That’s all for Saturday!

Thanks for reading,

— There’s An AI For That