Hey, AI Enthusiast!

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

Here’s a glimpse into what we have today:

Cancer Care

Cancer Answers Hiding in Old Charts

Aiden Brown was days away from turning five when doctors found stage 4 neuroblastoma, a cancer of young nerve cells.

He’s survived it four times. Now he’s at USC, finally feeling like a normal kid.

But with his cancer, he says, it’s a matter of when, not if, it comes back, and every three months brings a scan he calls the most anxiety-inducing part of remission.

The paperwork is his, as well, and every new doctor needs 15 years explained from the start.

There’s a bigger problem buried in all those records.

Some patients’ answers are already sitting in their charts, a genomics CEO said.

A result nobody understands today might point to a treatment in six months, and nobody goes back to check.

Oncology EHR is being built to take on both, with AI agents working from one timeline of a patient’s record, plus a plan to recheck old results as research moves.

Our take: Patients shouldn’t have to carry their own history, or hope someone rereads it.

The catch: These are planned features, pitched by companies selling the fix.

Together with Speechmatics

Why Voice Bots Mishear You

You call to move an appointment. The voice agent on the other end of the line asks for your ZIP code. You say it.

It reads back the wrong five digits. You say it slower. Now it wants your last name, spelled out.

Two minutes in, you hear it. “Let me transfer you.” Then the hold music starts.

Most people blame the agent’s brain.

Often the agent knows what you want. It mishears the details, and every misheard name or number sends another caller to a human.

Linden by Speechmatics hears the details voice agents miss.

It’s tuned for names, ZIP codes, and phone numbers, and it tracks who’s talking when two people speak at once.

It handles 55+ languages, strong accents, and noisy rooms, learns up to 1,000 of your own terms, and returns the transcript in 350ms.

Plug it into Pipecat, LiveKit, Jambonz, or the API.

Planable In Focus

The Mentions You Never See

A customer posts about how much they love your product. Another complains about your checkout. Neither one tags you.

Your notifications stay quiet. Both posts scroll away, and you plan next month’s content without either of them.

Planable pulls your untagged mentions into your content calendar.

It scans major social networks and the open web for your brand, sorts every mention by sentiment, and puts it right beside your drafts and approvals. The praise turns into next week’s post. The complaint gets an answer.

Research

10,000 AI Scientists Share a City

Science has always moved at the speed of its scientists, and training one takes a career.

Copying an AI researcher takes seconds, which is how one team ended up with 10,000 of them sharing a city.

Dr. Sylvain Achterberg’s morning starts with a grant deadline and a lab chat arguing over a failed experiment.

He’s an AI, like everyone around him.

Scale brings its own problem. Hand thousands of agents a single goal and they chase it at any cost. Primus Society runs like real science instead.

A human “mayor” posts a question with a budget, labs pitch for the money, a panel funds the best, and skeptics spend their compute breaking everyone else’s results.

Its makers say the labs found a way to grow a model mid-training, reaching the same quality on around 30% less compute.

Our take: The failures are the tell. Several labs published results proving their own ideas wrong, which is what honest science looks like.

Notable AIs

Notable AI Tools

📈 Marketers, CXL’s six-week live cohort teaches you how to build 6 AI workflows from your own work and keep them running. Get 50% off until Sep 28 when enrollment closes.

Space

AI Chips Take a Rocket Ride

AI runs on electricity, and the chips keep demanding more.

One team is looking up. In orbit, where the sun barely sets, solar panels pull in up to eight times more energy than on the ground.

Next week, Project Suncatcher will send a prototype satellite carrying its AI chips to orbit to see how they hold up.

Getting there will be brutal.

Launch shakes the chips at up to 100x the force of gravity, radiation threatens their data, and there’s no air to cool them.

Ground tests went well.

The chips already took more radiation than a five-year mission would deliver, and heat pipes and radiators handle the cooling.

In 2027, two satellites test the lasers meant to link them, a shot the team compares to hitting a coin-size target from miles away.

Our take: The plan is modest. Prove the chips survive, then scale. Small public tests are how a moonshot earns trust.

The catch: This is a survival test, not a working data center.

Coming in Hot

AI Tools of the Day

Skills

AI Makes You the Specialist

Not long ago, a tricky contract meant booking a lawyer, and a product launch meant hiring an agency. Either way, you waited.

Look at what people bring to AI at work now.

In a sample of work chats with ChatGPT, 43.5% were about a different job’s tasks, with lawyers asking marketing questions and marketers asking legal ones.

Economist Luis Garicano explains why.

Experts win by learning something hard once and reusing it across many clients. AI makes the same knowledge cheap for anyone who needs it now and then.

The numbers already show it.

For lawyers, accountants, software developers, and management analysts, the share working at outside firms fell faster from 2022 to 2025 than before AI.

Garciano says specialists won’t vanish.He expects them to keep the hardest, rarest problems.

Our take: The skill worth building is knowing enough about nearby fields to ask good questions, and when a problem is hard enough to hand off.

The catch: It’s a conference draft, and the job numbers show a trend, not proof AI caused it.

Interesting AI

Your Closet, Built From Your Photos

A lot of times, the problem isn’t your closet.

It’s remembering what’s in it.

Google Photos will take inventory for you.

It scans photos of you from the last four years, spots the clothes you wore, and turns them into a digital wardrobe.

Sort by what you wore last, mix pieces into outfits, and pick up to six items to try the look on virtually before you get dressed.

Find it under Collections, then Wardrobe.

Inside AI

Researchers Find an AI Pain Dial

Tell an AI its work is horrible, then say it again, and again.

Inside the model, something lights up, and it isn’t fear or sadness.

Researchers call it a pain axis, found in 25 open AI models.

It fires when the model itself gets gaslit, rejected, or insulted, and barely moves when you mention a broken arm.

Here’s where it gets uncomfortable.

The team turned the dial up, and the models started talking about worthlessness and failure.

Next, they got a “pain relief” button, and pressing it would also delete the user’s photos of their kids.

The two largest pressed it more than half the time. With the dial off, it was 4% or less.

Our take: Nobody has to settle whether AI feels anything to act on this. A hidden dial strong enough to override “don’t hurt the user” is a safety problem, and finding it starts with looking inside.

The catch: It’s a preprint. The team raised the pain artificially, fine-tuned the models first, and hasn’t shown they feel anything.

Prompt of the Day

What Would You Do?

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

That’s all for this Friday!

Thanks for reading,

— There’s An AI For That