Hey, AI Enthusiast!

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

Here’s what we have today:

ChatGPT

ChatGPT Answers with Intelligent UI

Last time you asked ChatGPT something a little complicated, you probably ended up scrolling through a ton of text to find what you needed.

OpenAI’s GPT-6 with Intelligent UI is trying to put an end to that.

Instead of handing you a wall of text, ChatGPT is going to start answering you with visuals, buttons you tap, forms you fill out, charts, and little interactive tools.

And it won’t leave you waiting around as much, either.

Normally, when a model “thinks” through a hard question, you either sit there watching a loading message or go do something else until it’s done.

Since this is instant, it’ll start answering you while it’s still working things out.

On questions that need a web search, OpenAI says answers start coming back 44% sooner, on average.

That’s a big change for the 1.2B+ people using ChatGPT every week.

If you’re on a paid plan, expect to have it shortly. Otherwise, you’re next in line.

There’s a catch for free users, though.

They’re getting a different version of the model, and OpenAI hasn’t said whether the new capability comes with it or not.

OpenAI also admits it still has work to do on the model’s “design judgment” (how it decides what an answer should look like).

Head over to ChatGPT and ask it for a plan, comparison, or breakdown.

If you’re on a paid plan, there’s a good chance what comes back won’t be paragraphs.

Our take: ChatGPT has mostly been something you read. Now that it’s going to be turning answers into things you tap and click, that’s a bigger shift than it sounds like. A chart or a form gets you to the part you care about way faster than scrolling ever does. It also means ChatGPT is now making design calls for you.

Together with Nebius

Run an Open Model Without the GPU

Open models like Qwen are free to download, but getting one to run well is a whole other story.

You need a high-end graphics card at minimum, and most people don’t have one sitting around.

Nebius Token Factory skips the hardware part entirely.

It runs Qwen’s 27B model for you, and all you do is call it through an API.

The model is built for coding, research, and long, multi-step agent work, and it handles tool calling.

If you’re building an agent that reads through long reports and figures out what to do next, you grab an API key, point your code at the endpoint, and it’s running.

There’s nothing to set up or deploy.

And if you want to see it in action before writing any code, the playground lets you try it right in your browser.

AI Scams

Fake Kidnapping Used AI Voice Clone

A woman was having breakfast with her husband when her phone rang with a local number.

The man on the other end told her a drug cartel had taken her daughter.

Then he asked a strange question. Was her daughter prone to panic attacks?

A second later, the woman heard her daughter crying and struggling to breathe, telling her she was sorry and that she was scared.

For the next 5.5 hours, the caller ran her day.

She had to keep her phone charged and the line open so he could hear everything, and when she muted it to talk to her husband, he started screaming at her.

He sent her to Walmart to wire money abroad (not everyone is aware that this is a red flag), with a story ready for the clerk about a relative’s medical bills.

At one point, her daughter came back on the line to say she loved her. Then another man got on the phone and demanded even more.

By the end, the woman had sent four transfers and all the money she and her husband had.

The caller finally said her daughter had been dropped off outside a Walmart, but nobody was there.

That’s when she dialed her daughter’s number herself, and she picked up her phone from work, completely fine.

The FBI has a name for this, a virtual kidnapping.

And the FTC says all a scammer needs to fake someone’s voice is a short clip of it, even one posted online.

Think about how many videos of your family are sitting on social media.

Our take: Every part of this scam works by keeping you on that one call, because the second you hang up and call your kid yourself, it falls apart. Set up a family code word before you ever need it, and treat any call that asks for secrecy and a wire transfer as fake until you’ve reached that person on a number you already have.

AI Research

AI’s Research Taste Doubles Every 3 Months

Good researchers have something that’s hard to teach.

It’s a feel for which experiment is worth running next and which one is a waste of time.

People call it research taste, and AI is picking up on it fast.

P-Zero Research, a group that tests how well AI handles AI research on its own, found that this kind of taste in top AI models has doubled roughly every three months since December 2025.

The way they measure it is pretty hands-on.

Each model gets a research problem that runs on a single graphics chip, tries its own experiments, and gets scored on how efficiently it gets to a good result.

The best model is now beating the experienced researchers P-Zero put it up against.

Neither side came up with anything new, though.

Out of 540 model attempts and 24 runs by human experts, not one used a method that wasn’t already out there in published research.

Both mostly tweaked or combined ideas that already existed.

Thinking time is what made a big difference.

Giving a model longer to think made its choices noticeably better.

They published the results and how they ran the test, but they’re keeping the tasks themselves private.

Releasing them would hand labs the exact test to train their models toward, and that’s not what P-Zero wants.

It’s the same reason a teacher doesn’t pass out the answers to a test the week before.

Our take: Picking the next experiment is the slow, expensive part of research, because every bad guess burns time. A model that guesses well, and gets better at it every few months, speeds up everything that comes after it, including building the next AI. That’s why P-Zero is so careful about what it gives away.

The catch: The tasks are quick and cheap to check, unlike real lab research, where one experiment takes weeks. The human experts also weren’t top researchers from the big AI labs, and the test doesn’t measure choosing which problems are worth working on in the first place.

Coming in Hot

AI Tools of the Day

Publishing

AI Faked Quotes in a Book About Truth

Steven Rosenbaum wrote a whole book about telling what’s real from what’s fake in the age of AI.

It went out with 26 quotes that didn’t hold up.

The book is called “The Future of Truth”, and he says he wrote every word of it himself.

AI’s only job was organizing his research, summarizing docs, and polishing the language.

But somewhere along the way, it started “fixing” quotes it thought it could make better.

It all came out when a NYT reporter asked him where a quote from an unnamed DARPA researcher came from, and he had no source to give.

Two days later, the story was on the front page of the Times.

His publisher brought in two fact-checkers to trace every quote in the book back to where it came from, and 26 of them had to be fixed, paraphrased, or cut completely.

Here’s what should worry anyone who lets AI clean up their writing.

The book went through five rounds of editing and more than 4,000 tracked changes before it was published.

Rosenbaum says the mistakes that looked off got caught, and the ones that read perfectly fine slipped right through.

He’s since put out a revised edition, and included a new chapter about what went wrong inside.

Our take: Asking AI to clean up your writing might feel harmless, but you have to remember that it has no idea which words are yours and which ones belong to someone you quoted. To the AI, it’s all text it gets to improve. So, before you send out anything AI cleaned up, make sure to check it against the original yourself (especially if it’s quotes).

Interesting AI

Make a Game by Describing It

Plenty of people have a game idea they’ve carried around for years, with no clue where to start building it.

Google’s new Playground starts with a sentence.

Describe the game you want, and AI will build a version you can play right away in your browser, on your phone or your laptop.

From there, you shape it by asking for changes to the physics, the rules, the characters, or the world, without touching any code.

Once it’s done, keep it to yourself, send the link to friends, or publish it to a gallery where other people rate it and play it.

Math

OpenAI Shows Its Work on AI Math

Every few weeks, there’s another headline about AI cracking a math problem.

What doesn’t always come with it is an easy way for anyone outside the lab to check whether it’s true.

That’s why OpenAI just released it’s latest batch of results.

They come from an internal model working on open math problems, and a lot of the proofs are written in Lean (a programming language made for math).

You write a proof in it, and a computer walks through every step and refuses to sign off on anything that doesn’t hold up, so nobody has to take OpenAI’s word for it.

And they say more of those are on the way.

It’s also showing how the model got there, with 10 summaries of its reasoning, estimates of how much computing each result took, and stats on how many problems it attempted, all on GitHub.

On average, each result took about as much computing as three hours of ChatGPT Pro thinking.

Even the way it’s being released got outside input. OpenAI worked with an independent advisory group on math and AI at the Institute for Advanced Study to set the ground rules, and it says it’ll keep updating how it announces big science results based on feedback.

It’s paying for mathematicians to dig in, as well, with workshops and conferences focused on understanding results that came from AI.

Our take: Claiming an AI solved something is easy, and handing over the proof, the reasoning, and a computer’s stamp of approval is what turns a headline into something mathematicians take seriously. If other labs start releasing results the same way, the arguments over whether AI did the work at all get a lot shorter.

The catch: OpenAI’s post doesn’t name the problems or say which proofs have already been checked in Lean, so the full picture depends on mathematicians working through the GitHub release.

Prompt of the Day

Micro-Commitment Engine

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

That’s all for Wednesday!

Thanks for reading,

— There’s An AI For That