Hey, AI Enthusiast!

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

Here’s what we have today:

Space

Claude Might’ve Found a Planet

Somewhere in seven years of NASA telescope data, there might be a planet nobody had noticed.

And someone tracked it down with Claude Code.

NASA’s telescope is called TESS, and it watches stars for tiny dips in their brightness.

When a planet passes in front of its star, it blocks a little of the light.

That dip comes back every time the planet circles around again.

Finding one usually takes a lot of specialized code. That’s where Claude Code came in.

It downloaded the data, wrote the search code, fit the dips, checked for false alarms, made the figures, and reran any test that failed.

What came up was a dip around a nearby star a bit smaller and cooler than our sun.

It repeated every 3.18 days and showed up 23 times.

If it’s a planet circling that star, it’d be about 1.4 times the size of Earth.

After it was posted about over on Reddit, people in the comments couldn’t agree on what the OP had found.

Several exoplanet researchers jumped in and said the signal looks real and makes a reasonable candidate.

They also said it’s borderline, and it might still turn out to be something else, like two stars in the background passing in front of each other.

The back-and-forth turned into one of the most useful parts of the thread, with researchers sharing how they check AI’s work.

It comes down to trying to prove the answer wrong instead of looking for reasons it’s right.

And redoing the key numbers yourself instead of trusting the summary the AI wrote.

The real test is coming.

The TESS team approved his request to watch the star up close from Oct 31st to Nov 26th.

Our take: The AI did the heavy lifting, but what made this worth taking seriously was everyone who tried to poke holes in it. Most answers you get from AI never go through anything like that. When you ask one chatbot something, you get one answer, and there’s nobody in the comments telling you where it might be wrong.

Together with Cuey

See Where Other AIs Disagree

The closest thing to a thread full of experts poking holes in an answer is hearing what the other AI models would’ve said.

Getting that usually means pasting the same question into three different chatbots and reading three long answers.

Cuey sends your prompt to several AI models at once and shows you where they agree, where they split, and what your model missed, side by side in a sidebar.

It works inside ChatGPT, Claude, Gemini, and Grok, as a free Chrome extension.

Cuey runs the same question past the other models and lines up their answers right next to the chat.

When they agree, you can move on with more confidence, and when they don’t, you know to look closer.

It runs alongside the AI you already use, and there’s nothing new to learn.

Manufacturing

Tiny Machines From a Prompt

Pretty much anything that gets mass-produced starts with a mold, a mask, or some kind of custom tooling.

And every new part needs new tooling made for it.

Shrink a part down to roughly a thousandth of an inch, and those methods stop working.

At that size, the only big option left is a chip factory.

But chip factories are built to make flat circuits, not machines that move, pump, heat, or sense.

That’s why tiny machines like those are mostly unmakeable, and each new kind can take years of custom process work.

After six years in stealth, Atomic Machines came out with their answer for this, the Matter Compiler.

It works by taking a description of what you want a device to do.

The system then designs it, checks the design, and builds it.

Then it measures what it built and carries what it learned into the next design.

AI runs the entire thing.

It even runs its own experiments on the hardware to test what the simulations predict.

The parts all come out in 3D, with moving pieces and multiple materials, and details far thinner than a human hair.

Their first machine is PrimeSwitch, a switch around three-eighths of an inch across, built for the power systems inside AI data centers.

They say it opens 1,000 times faster than a conventional power switch.

One of their investors called it the start of “vibe manufacturing.”

Our take: The loop is the interesting part of this. A normal factory gets set up once and makes the same thing for years. This one measures every part it builds and feeds what it learns into the next design (a lot like how AI models improve with more data). It’s one switch so far, though. The “hours, not years” promise still has to hold up on more than one device.

Health

AI Interviews Patients Before Visits

Think back to the last time you went to the doctor. You most likely already knew what to expect.

The doctor finally walked in, and a good chunk of your entire visit went to explaining your symptoms.

Google wanted to see what happens when an AI handles that first conversation instead.

They teamed up with doctors at Beth Israel Deaconess Medical Center to test AMIE, a research chatbot built to talk with patients and work out what might be wrong.

Before their urgent care visits, 98 patients chatted with it, while doctors watched every conversation live.

None of the conversations had to be stopped under the safety rules set before the study started.

After each chat, AMIE handed the doctor a summary of what the patient told it.

It also handed over a list of what it thought the problem could be.

Doctors said those summaries helped them prepare in 75% of cases, and changed how they approached care in more than half.

AMIE’s list of possible diagnoses also matched the doctor’s final diagnosis 90% of the time.

The study ran in The Lancet, one of the most respected medical journals in the world.

And it’s Google’s first paper in the main Lancet journal.

The Catch: It’s still a research tool, and Google says larger trials are needed before AI like this can start talking with patients at scale.

Our take: A lot of what makes a doctor visit feel rushed is the time spent collecting basic history before the real thinking starts. If an AI handles that part first, your doctor can walk in already knowing your story.

Planable In Focus

Listen First, Then Post

Writing posts with Al is fast, but a calendar full of them starts to sound like everyone else’s.

Planable and SE Ranking looked at 100K keywords and found a Reddit thread, YouTube video, or other social post in Google's top 10 for about 3 in 4 of them.

The questions your audience has are already out there, in public

Planable's free guide walks you through a 6-step workflow for finding those conversations first, then writing posts that answer them.

It also shows how their social listening picks up mentions of your brand, even when nobody tags you.

Interesting AI

This Ring Hands Jobs to AI

Pulling out your phone every time you want AI to do something adds up throughout the day.

Natura’s Interface is a smart ring that wants to help you save some time.

You press a small button on the ring, say what you need, and it’ll pass the job to an AI agent.

The answer will come back through your headphones or your phone.

Their examples are things like hearing your schedule when you wake up, reordering shampoo while you’re in the shower, or turning off the TV and setting an alarm as you’re dozing off.

At launch, it’ll connect to agents like ChatGPT and Grok.

And it’ll be able to send different jobs to different agents.

It’ll also track your heart rate, sleep, and steps, and the battery lasts about 6 to 12 days on a charge.

Shipping is planned for December or January.

Coming in Hot

AI Tools of the Day

Biology

Biohub Is Building a Virtual Cell With AI

A biology lab only has so much time.

Every experiment that ends up teaching nobody anything uses some of it up.

Mark Zuckerberg’s Biohub wants scientists to find the promising ones before they ever step into the lab.

They’re teaming up with Google DeepMind, Isomorphic Labs, Meta, the Department of Energy, and the National Institutes of Health to build what they call a “universal virtual cell.”

The idea is an AI model that predicts how a living cell behaves.

Scientists would test an experiment on a computer first and save the lab for the ones most likely to teach them something.

AI already understands individual pieces of biology, like proteins.

A whole cell is orders of magnitude more complex, though, and there isn’t enough of the right data to train on yet.

Chatbots learned from text that was already sitting on the internet, but most of what a cell model needs has never been collected.

It has to be painstakingly measured from the real world first.

That’s why the first phase is a broad map of how cells work and how they respond when something changes.

The companies involved get one year of exclusive access to the data they help create, and after that, it’s shared publicly with scientists everywhere.

Further out, Biohub’s head of science, Alex Rives, pictures a model that looks at one person’s disease, predicts what’s causing it, and points to the best way to intervene.

He doesn’t expect a long wait to see whether the bet pays off, either.

Within a year of the first big dataset, he says researchers should be able to train models and see which kinds of data make them better.

Our take: Lab work is slow because every experiment takes real time to run, and plenty of them don’t pan out. A model that tests the ideas first won’t replace the lab, but it changes which experiments make it there. And if Rives is right about the timeline, we won’t be waiting long to find out whether it works.

Open Source

7 Pro Creative Apps Rebuilt With AI

Apps like Photoshop, Premiere, and Lightroom took big teams years of work to get where they are today.

Now a dev has rebuilt free, open-source versions of seven of them, with an AI coding assistant writing the code.

The project is called ArtCraft.

PhotoCraft takes the place of Photoshop, FilmCraft goes after Premiere, and LightCraft covers Lightroom.

There are also versions of Illustrator, Acrobat, After Effects, and InDesign in the mix.

He calls it a “clean-room reimplementation,” meaning none of Adobe’s own code went into it.

The apps run on Mac, Windows, and Linux, and most of them work in your browser as well.

But they’re not all the way there yet.

LightCraft’s own page says it covers about 79% of Lightroom’s features but is only 60 to 70% of the way there as a daily replacement.

He wants all seven to match the originals within a month (a big goal).

Our take: Rebuilding software this big used to take a whole team years of work. With AI writing a lot of the code, one experienced dev got surprisingly close, and that’s a big deal. The last stretch of features is usually the hardest. The real test is whether ArtCraft gets all the way there or not.

Prompt of the Day

Live the Other Life

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

That’s all for Friday!

Thanks for reading,

— There’s An AI For That