buaya maya

October 28, 2026 | Read online

Hey, Jhon!

Welcome back to the #1 AI newsletter in the world!

Here’s what we have today:

Why a company with 70+ AIs won’t stick to one.
AI hands people with ALS sentences instead of letters.
Google’s strongest AI can stay on one job for days.
People are getting paid to walk robot dogs.

Strategy

Software Giant Benches Their Fanciest AI

Source: Reuters

Once you find an AI you like, it usually becomes the one you ask about everything.

It plans your dinners and explains the contract you don’t fully get. On most days it does fine.

But some of what you hand it is a job it was never the best pick for.

It shows up as a small annoyance.

You reword the ask a third time and settle for close enough.

Nothing in the chat tells you whether a different AI would’ve gotten it right the first time.

One of the biggest software companies in the world decided to stop settling.

SAP makes the software a lot of the world’s largest companies run on.

Behind the scenes, they use more than 70 different AI models from companies like OpenAI and Google.

After they took a closer look at the models they use, the priciest ones turned out to be overkill for some jobs.

So they pulled those models out of a few products and put in cheaper ones that do the work to standard.

You’d think a company that size would stick with the best one and call it a day.

They aren’t even trying to spend less on AI. They expect their costs to keep climbing.

The whole point is picking the right model for each job.

There’s a bigger reason as well.

Some of their customers now rule out certain models based on where they come from.

Government agencies do it, and so do companies in tightly regulated fields.

For those customers, a company with only one model to offer has nothing to sell.

You don’t need 70 AIs to get the same benefit.

When your go-to keeps missing on the same kind of task, that’s your sign the job belongs to a different one.

Our take: The best AI for a job isn’t always the most famous one or the priciest one. SAP has an entire team testing that for them. For everyone else, the next step is making it easy to switch between AIs without a team behind you.

Together with Nebius

Meta and NVIDIA's New Models in One Place

There’s a special kind of pride in a little app you put together with AI on a Sunday afternoon.

Then one day a button stops working, and you’re stuck trying to describe a broken screen to an AI in words.

A different model might be better at that part.

But trying one has always meant signing up somewhere new and redoing your setup before you even know if it helps.

Nebius Token Factory gives you three new open models (including ones from Meta and NVIDIA), through one connection.

Switching between them is a one-line change.

Meta’s model can look at a screenshot of that broken screen right next to your code and help track down what went wrong.

And NVIDIA’s Nemotron can read documents and web pages, which comes in handy when the answer is buried in a help page somewhere.

Accessibility

Full Sentences for People With ALS

Source: Drexel University

Your phone has a little habit you stopped noticing a long time ago.

You type “on my,” and before you get any further, it’s already offering “way.”

It saves you a second, maybe two.

For someone with ALS, that same idea can save minutes on a single sentence.

ALS slowly takes away control of the muscles people use to move and talk.

And many people with it end up talking through a headset that reads their brain signals instead.

The headset flashes letters on a screen one at a time.

When the one they want lights up, their brain gives off a tiny spike, and the headset catches it.

That’s one letter.

Now try to do that for a sentence while your family sits there waiting.

Researchers figured the phone trick would work here as well, only bigger.

They added AI so that it can suggest phrases instead of single letters (based on what’s being said in the room).

What used to take around 30 picks takes a handful.

It even learns as it’s being used. The more someone picks its suggestions, the more it starts to sound like them and not like some generic assistant.

It also never sends a word off the device.

The catch: Each person only tried it once, and five of the 23 testers had ALS. A follow-up this fall brings in people earlier on.

Our take: Getting your words out fast is one thing, but getting them out in your own voice is what makes a conversation feel like yours. If the bigger test holds up, it’s hard to think of a better reason to put AI on someone’s head.

Coming in Hot

AI Tools of the Day

🤳 GoFaceless drops you into the trending clips everyone is posting from one selfie and lets your friends join in with lip sync to the original song.
🧭 Athena takes a plan you worked out in ChatGPT and saves it as a goal with milestones once you approve it, so your next morning’s briefing already knows.
🎨 AdsCreator.com reads the logo and colors off your website and makes image and video ads sized for the big social and search platforms.
🌐 Reactor streams AI-generated worlds that react to you in real time, made for developers building interactive apps and robots.
🤖 Project Quarry lets you write a rule once and catch an AI agent going off script across a whole conversation, free and open source on GitHub.

Interesting AI

Race Cars Learn With AI

Source: Anki

Racing toy cars around a track used to mean you were the only one who knew where the turns were.

Anki Overdrive gives the other cars a brain of their own.

Each one is a small robot car run by AI.

You snap together a track any way you like and the cars learn it as they go.

You can race your friends or take on the AI cars. They get better the more you play.

The app lets you add weapons and swap in new cars when you want to shake things up.

The makers promise that wherever you drive, the cars will hunt you down.

Security

Google Saves Their Best AI for Defenders

Source: Google

It used to be that when a big AI company finished a new model, pretty much anyone could try it the same day.

Google’s newest one is taking a different route.

Their latest Gemini model goes to a small group of cyber defenders first.

These are the people whose job is finding holes in software and fixing them before anyone else gets there.

Google built this one to be especially good at that work.

It can spot a weak point in a program and then write the fix for it.

It also sticks with a problem for a long time.

Google’s own teams have watched it stay on one hard engineering job for days, working its way around dead ends as it went.

They trained it to cut fewer corners as well.

In their own tests, it made up about half as many unsupported claims as the model they released two weeks earlier.

The apps and sites that have your logins all run on software, and some of that software has holes nobody has found yet.

Whoever finds a hole first gets to decide what happens to it.

If a defender finds it, it gets patched. If someone with bad intentions finds it, they use it to get in.

Handing the model to defenders first gives them a head start on finding those holes.

Google says they’ll bring it to everyone else as soon as possible, once they’ve gathered feedback and tightened up the guardrails.

Our take: Holding back a new model can look like a company being overly careful. With a model built for finding holes, the order matters, because the first people to use it decide whether those holes get fixed. If Google’s plan works, most people will benefit from it without ever opening a chat window.

Robots

Robot Dogs Also Need Walkers

Source: Digital Trends

A job ad went up in San Francisco that sounded like a joke at first.

No experience needed, four to six hours of walking a day, and the dog is a robot.

A robotics startup wanted people to steer robot dogs down real city sidewalks, using a handheld remote or a phone app.

The job came with one odd duty.

If a dog overheated, the walker had to carry it back to the office.

There’s a good reason they wanted a person holding the remote.

While someone steers, the dog’s cameras and sensors record the trip.

Each curb it steps around and each crowd it weaves through becomes an example of how a person gets from one place to another.

The company feeds those walks back into their software so the dogs can learn to handle the street on their own.

It’s a lot like teaching a kid to cross the street, where you do it together a bunch of times before you ever let go of their hand.

The software running all of it is open source, so anyone building their own robot can use it.

For now, even a robot dog still needs someone to walk it.

Our take: It’s easy to read AI news and feel like people are getting written out of everything. Here, people are the whole training plan, and the robot doesn’t get far without them.

Prompt of the Day

Subtitle My Messages

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

View and copy today’s full prompt

This prompt puts subtitles on three messages you sent this week, with what you meant on the top line and what the other person most likely heard underneath. You rewrite the worst one, then subtitle your next message before it goes out.

That’s all for Wednesday!

Before you go, try the new Meta and NVIDIA models on Nebius. Give them the one job your usual AI keeps tripping over and see which one handles it best.