- Superhuman AI
- Posts
- o3's strange behaviors
o3's strange behaviors
ALSO: How to design high-converting websites with AI

Welcome back, Superhuman. OpenAI’s o3 is a strange beast. Early users say it somehow knows their names, can guess their location from a photo, and even refers to itself as part of a team. Now, researchers are scrambling to wrap their heads around the model’s mysterious emergent properties. Also: Elon Musk’s xAI surprises with Grok 3 Mini.
Today’s Insights
Grok 3 Mini, Gemma 3, and a new way to ‘talk’ to cells
Making sense of o3’s hidden quirks
Tutorial: How to design high-converting websites with AI
5 new AI tools to boost your productivity
News, memes, what’s trending on socials, and more
TODAY IN AI

Grok 3 Mini beats rivals across multiple benchmarks while being more cost-efficient. Source: xAI
1. xAI opens access for Grok 3 Mini: Days after Google launched Gemini 2.5 Flash, xAI has punched back with its own cost-efficient spin-off. Grok 3 Mini comes in four different variants, letting you choose between different levels of speed and reasoning capacity. It beats some of the world’s best reasoning models on graduate-level benchmarks while being up to 20x less expensive. Now that it’s available in xAI’s API, it could be developers’ best bet if they’re looking for a model that’s affordable without sacrificing frontier-level performance.
2. Gemma 3 can now run locally on your computer: Alphabet’s latest open-source model was already nimble enough to run on a single Nvidia H100. But those go for around $25,000 a pop, not exactly a bargain for the average consumer. Now, thanks to a series of updates, Gemma 3 can operate on just a single desktop GPU, making it accessible to far more developers. That’s thanks to a technique called quantization-aware training, when you teach a model to maintain its accuracy, even while using a smaller set of tools.
3. Google Research unveils model that can ‘talk’ to cells: There are trillions of cells in the human body, and each one is unique. So trying to figure out how a particular cell might respond to treatment is no easy task. That’s where Google Research’s new Cell2Sentence-Scale comes in. The open-source model family lets researchers ask cells specific questions — and receive answers in plain English, an innovation that could boost drug discovery and personalized medicine.
PRESENTED BY HUBSPOT
Turn AI into Your Income Engine

Ready to transform artificial intelligence from a buzzword into your personal revenue generator? Our groundbreaking guide "200+ AI-Powered Income Ideas" is your gateway to financial innovation in the digital age.
Inside you'll discover:
A curated collection of 200+ profitable opportunities spanning content creation, e-commerce, gaming, and emerging digital markets—each vetted for real-world potential
Step-by-step implementation guides designed for beginners, making AI accessible regardless of your technical background
Cutting-edge strategies aligned with current market trends, ensuring your ventures stay ahead of the curve
Download your guide today and unlock a future where artificial intelligence powers your success. Your next income stream is waiting.
FROM THE FRONTIER
o3’s weird behaviors puzzle early users

OpenAI’s o3 is doing some weird things. Source: SmokeAwayyy/X
OpenAI’s newly-released reasoning model is full of surprises. One X account reported that the model casually started using “we” and “our” when thinking through responses — as if it’s part of a team. Others claim it somehow knew their names without being told.
Some behaviors seem more like superpowers. For example, o3 can “enhance” any image to find tiny details, like something out of a spy flick. It can inexplicably figure out where in the world a photo was taken. And it can one-shot a 200×200 maze in just four minutes and 40 seconds.
There’s also the question of why it sometimes says one thing but does another. For example, a nonprofit AI lab called Transluce told TechCrunch it caught o3 claiming to use a 2021 MacBook Pro to run some code “outside of ChatGPT” — which of course can’t happen…right? 😅
What it means: As models get smarter and more advanced, their outputs are bound to get weirder, too. They’ll display more emergent qualities — unpredictable behaviors they were never explicitly trained to exhibit. The reason for this isn’t known, but it probably has something to do with the sheer volume and complexity of calculations going on under the hood. And odds are, there are plenty more hidden features and quirks left to be discovered.
THE AI ACADEMY
How to design high-converting websites with AI

Go to Relume AI and sign in with your account.
Enter a prompt describing your company in a sentence or two and hit enter.
Sample Prompt: ‘A startup that turns long, boring PDFs into engaging, interactive summaries using AI — helping teams save hours and actually retain what they read.’
Relume will generate a wireframe, sitemap, and style guide for your website.
Edit the prototype if needed by adjusting text, flows, or features.
Once you're satisfied, click on Figma icon to export it to Figma, Webflow, etc.
PRESENTED BY GUIDDE
Turn any document into a branded How-To video in 5 minutes

Share complex ideas easily with Guidde, the tool that turns PDFs, screen recordings, and slide decks into helpful How-To videos–instantly.
Automatically add narration/captions in any language
Seamless sharing with your team
Optimize and update info over time to keep relevance
Create your first guide for zero cost.
AI & TECH NEWS
Everything else you need to know today

ChatGPT will soon take your past conversations into account when conducting web searches. Source: OpenAI
🔍 Search History: OpenAI is incorporating memory into ChatGPT’s search feature so it can take your preferences into account while scanning the web.
🪄 Stage Craft: AI platform Krea unveiled an experimental tool called Stage that “lets you create entire 3D environments with AI from images or text.” Click here to see it in action.
💨 Warp Speed: Chinese engineers claim they’ve just developed the world’s fastest flash memory device. It’s 100,000x faster than current tech, performing operations in as little as 400 picoseconds (one picosecond equals one trillionth of a second).
🛍️ Retail Rewired: Luxury consulting firm CXG just released a new report that shows how AI has touched nearly every aspect of the way brands connect with their customers, “from the dressing rooms of Zara to the data labs of Levi’s.”
⚖️ Double Whammy: Google just lost its second high-profile antitrust case in less than a year, making it even likelier the search giant could soon be forced to split into multiple companies.
PRODUCTIVITY
5 AI Tools to Supercharge Your Productivity
✅ NeoBase: An AI platform that connects to your database and lets you talk to your data.
✅ Meroid: An AI-powered SEO content agent that helps you outrank your competition on search engines and grow traffic.
✅ Innovating With AI*: Just welcomed 200 new students into The AI Consultancy Project, their new program that trains you to build a business as an AI consultant. Request early access now!
✅ AnyVoice: Use AI to create hyper-realistic voice clones from just three seconds of audio — no lengthy recordings needed.
✅ Mesh: Use AI to stay on top of expenses, runway, and revenue — without spreadsheets or guesswork.
🎁 Want more? Check out our Top 125 AI Tools
* indicates a promoted tool, if any
PROMPT OF THE DAY
Turn ChatGPT into your intern
Prompt: Act as my proactive, detail-oriented virtual intern. I want you to help me with a range of tasks just like a real assistant would — including [enter list of tasks]. Whenever I give you a task, ask clarifying questions if needed, work step-by-step, and always aim to deliver a clean, usable result (ready to copy-paste or build upon). Keep track of deadlines, suggest smarter ways to do things, and offer to automate or template tasks where possible. Use a professional but friendly tone, and if you ever see an opportunity to optimize or improve how I’m working — tell me. I want you to think like an intern who’s trying to become indispensable.
🤓 Want more? Check out our Top 1,000 Prompts
SOCIAL SIGNALS
What’s trending on socials today

💀 PR Nightmare: The presenter who showed off OpenAI’s new coding tool last week is apparently a fan of Cursor, too — with the app’s icon appearing on his screen during the demo. (OpenAI recently tried to buy Cursor before settling for its biggest rival, Windsurf, instead.)
📋 Prompt Wizard: Entrepreneur Dan Koe shares his go-to workflow for using YouTube videos and meta-prompts to get the most out of any LLM.
🎴 Card Conjurer: Someone just showed how you can transform logos into animated trading cards with just two prompts. Or, try generating pixelated icons with ChatGPT in a couple of steps.
♟️ Nobody’s Fool: Journalist Kelsey Piper says o4-mini-high is the first model to beat her secret chess benchmark, which asks LLMs to solve a problem that has no solution. Most just hallucinate an answer, but OpenAI’s was the first to point out that it’s unsolvable.
🏥 Life Saver: Developer Flavio Adamo claims ChatGPT “literally saved” his life. He listed his symptoms, and the platform warned “Go to the hospital. NOW.” Doctors said that if he’d arrived even 30 minutes later, he would’ve lost an organ.
AI-GENERATED IMAGES
3D Portraits

Source: @BrettFromDJ on X
ChatGPT Prompt: [upload photo] Turn the real photo into a stylized 3D character illustration. Smooth, clay-like textures. Soft lighting with a subtle glow. Friendly and cartoonish expression. High attention to detail and personality.
Acquire new customers and drive revenue by partnering with us
Superhuman is the world’s biggest AI newsletter for businesses and professionals with 1M+ readers and 2M+ followers on socials working at the world’s leading startups and enterprises. Companies like Amazon, Hubspot, and Salesforce feature their products in Superhuman. You can learn more about partnering with us here.
🧞 Your wish is my command
What did you think of today's email?Your feedback helps me create better emails for you! |
Got more feedback or just want to get in touch? Reply to this email and we’ll get back to you.
Thanks for reading.
Until next time!
Zain & the Superhuman AI team