Startup delivers personalized AI podcasts

Welcome back, Superhuman. If presidents get a daily briefing, why shouldn’t you? That’s the premise behind a new app that can turn your emails, calendar items, and interests into a custom morning podcast. And its being developed by the team that brought us NotebookLM’s Audio Overviews.

P.S. I'm joining the Lovable Shipped 6 week cohort next week, where aspiring builders learn how to build full stack apps with AI without writing any code, and get a chance to pitch investors at the end. You can join me by applying here.

Today’s Insights

• Personalized podcasts, frontier research agents, and science’s first reasoning model

• Are reasoning models really thinking — or just pretending to?

• Tutorial: How to automate any task in your browser with AI

• 5 new AI tools to boost your productivity

• News, memes, what’s trending on socials, and more

TODAY IN AI

1. Makers of NotebookLM are back with something even better: A group of ex-Google engineers and designers who left to start their own company last year have finally shown off what they’ve been working on. Huxe is a “personal audio companion” similar to NotebookLM’s Audio Overviews — except it can bring together your emails, calendar items, and news interests to deliver personalized daily briefings. If you want to steer the briefing in a certain direction, you can simply ask a question, and the audio hosts will adapt on-the-fly. While the iOS app is invite-only, Huxe has been regularly posting passcodes on its X account.

2. Startup quietly ships frontier-level research model: SF-based Exa just came out of nowhere with a frontier research agent that, in its words, “doesn’t stop until it finds what you need.” And the numbers back it up: Exa Research scores a 94.9% on OpenAI’s SimpleQA, which tests how factually reliable models are. That’s better than Gemini 2.5 Flash and Perplexity Deep Research — and it’s cheaper, too. The Y Combinator-backed startup says its edge comes from a unique neural search model that can understand the full context of your question, instead of just matching keywords like traditional LLMs. You can try it via Exa’s API dashboard.

3. The first ‘reasoning model’ built for science: Non-profit AI lab FutureHouse just unveiled ether0, which can generate new molecules based on the desired qualities you give it. For the first time, it can show its train of thought, helping scientists understand how it landed on a particular chemical formula. Meanwhile, MIT and pharma company Recursion just open-sourced Boltz-2, a model that can supposedly predict how a drug will interact with a particular protein in just 20 seconds, compared to the usual 6-12 hours. Both breakthroughs could help pave the way for new treatments — or new uses for medicines that are already on the market.

PRESENTED BY WALL STREET PREP

Are You Still Just Asking ChatGPT to Write Emails?

Despite all the hype, most professionals are just beginning to scratch the surface of AI’s business impact.

The AI for Business & Finance Certificate Program from Wall Street Prep + Columbia Business School Exec Ed helps you go deeper.

Learn from leaders who are driving AI transformation at BlackRock, OpenAI, Morgan Stanley, and more. Plus, tap into the program network to learn how others are applying AI in their role.

No coding experience needed—you’ll get step-by-step guidance from Columbia Business School faculty.

👉 Save $300 with code SUPERHUMAN + $200 off for early enrollment before June 16.

Save Your Seat

FROM THE FRONTIER

Apple researchers call AI ‘reasoning’ into question

There was a brief moment last year when it looked like traditional LLMs were finally hitting a performance wall. Then, reasoning models like OpenAI’s o1 and DeepSeek’s R1 came along. They could tackle more complicated prompts by breaking problems down into smaller steps.

The thought was that reasoning models would just keep getting better and better as we throw more compute at them. Eventually, they’d be able to answer questions about things they’d never been trained on — and make new discoveries. But a group of Apple researchers just put that theory to the test.

The study: Researchers gave both reasoning and non-reasoning models a chance to solve a series of puzzles. With the simplest puzzle, non-reasoning models performed best, since they’re less prone to overthinking. Meanwhile, just as you’d expect, reasoning models did better with more complicated, medium-tier puzzles.

But here’s the kicker: With the most difficult, top-tier puzzles, reasoning models’ capacity to “think” completely dropped off. In fact, they threw in the towel earlier than usual, and didn’t even bother trying to come up with the correct answer. Giving them more compute didn’t seem to help either.

What it means: The study suggests that instead of being fundamentally unique, reasoning models work just like their non-reasoning counterparts — by finding patterns in the data they’ve already been trained on. While we shouldn’t read too much into a single study, it does suggest AGI could be further away than we’d thought — and that we’ll need to come up with different approaches and architectures to truly take AI to the next level.

THE AI ACADEMY

How to automate tasks in your browser with AI

• Go to Nanobrowser extension, download it, and add it to your browser.

• Once added, open Nanobrowser extension, go to Settings > Models.

• Select Gemini as your model and enter your API key.

• To get an API Key, go to AI Studio and click on ‘Get API key‘.

• Copy the API key and paste it into the Nanobrowser settings.

• You can now give Nanobrowser any task you'd normally do manually on the web.

Sample Prompt: Go to arXiv and find top 10 papers on artificial intelligence in healthcare.

• It will navigate to arXiv, extract the information, and give you the results in under 3 minutes.

• You can use it to do any task like automate your social media, do research on any topic and find information.

PRESENTED BY DATADOG

What’s your ROI with OpenAI? This cheatsheet will tell you

Datadog’s OpenAI Cheatsheet gives you crystal-clear visibility into your API costs, errors, and performance.

In just a few minutes, this guide will help you…

• Estimate OpenAI costs across libraries (Python, Node.js, PHP)

• Analyze prompts and completions

• Monitor latency/throughput and catch errors fast

Download the cheatsheet here (costs nothing), and see how cost-effective your workflow is.

AI & TECH NEWS

Everything else you need to know today

📊 By the Numbers: Google’s AI Mode is getting data visualizations for financial data, letting you ask questions like “Compare the stock performance of X companies in 2024” to generate a custom interactive graph and explanation.

💰 Making History: Meta is in talks to invest up to $10B in AI data powerhouse Scale AI. Bloomberg reports that if the deal goes through, it’ll be “one of the largest private company funding events of all time.”

💬 Chat Log Standoff: OpenAI is fighting to have its users’ data kept private after a court order forced the company to hold onto chat logs “indefinitely,” instead of deleting them after 30 days.

📂 License to Train: EleutherAI just released one of the world’s largest licensed and open-domain training datasets. Packed with 800GB of content, it’ll let researchers fine-tune models without having to worry about copyright violations.

🩺 Early Warning System: UK-based startup Heartfelt Technologies unveiled an AI-powered scanner that takes thousands of photos of a patient’s feet per minute to predict how likely they are to experience heart failure, with an accuracy rate of around 83%.

PRODUCTIVITY

5 AI Tools to Supercharge Your Productivity

✅ Substrata*: Analyzes real-time human dynamics to help you sell smarter and close more deals.

✅ Flux Kontext: Modify object, transfer style, replace background, and edit texts with AI.

✅ Pylon: One tool for your ticketing, chat widget, knowledge base, AI support, customer portal, and more.

✅ PodGen: Turn web pages, YouTube videos, PDFs, and text into captivating podcasts instantly with AI.

✅ Pretty Prompt: Turn any simple text into perfect prompts and improve output in seconds.

 🎁 Want more? Check out our Top 125 AI Tools

^{* indicates a promoted tool, if any}

SOCIAL SIGNALS

What’s trending on socials today

⚔️ Divide and Conquer: Media company Every just threw Claude, Gemini, and o3 into a battle for world domination — well, at least in the strategy game Diplomacy. There were bitter backstabs, unlikely alliances, and careful negotiations. You can watch the game unfold here.

🤝 Flawless Victory: Builder Ammaar Reshi created a hyper realistic Mortal Kombat fan film using only AI tools. You can watch the video and find his tool stack here.

🤖 Bot Bonds: OpenAI’s head of model behavior, Joanne Jang, just wrote a fascinating thread about why people are starting to form emotional bonds with ChatGPT — and what the startup is doing about it. Here’s a blog post with more details.

🏊 Sink or Swim: Zapier CEO Wade Foster just shared the chart the startup uses to measure AI fluency, ranging from “unacceptable” to “transformative.” Where do you fall on the chart?

🧑‍🏫 Learn From the Best: Who better to learn Claude Code tips from than Anthropic itself? Product manager Catherine Wu shares six quick examples of how Anthropic employees use Claude Code at work. Here’s the full report.

PROMPT OF THE DAY

Optimize your LinkedIn Profile

Prompt: Act as a LinkedIn personal branding expert and copywriter. I want to optimize my LinkedIn profile to attract more [insert your goal — e.g., recruiters in tech, high-value consulting clients, etc.]. I’ve attached a PDF that includes my current profile content: headline, About section, experience, skills, and possibly recommendations.

Please analyze the profile and provide:

Headline Suggestions – Clear, scroll-stopping headlines that speak directly to my target audience and value. Give me 2–3 options.

About Section Rewrite – Rewrite the “About” section to tell a compelling story while highlighting credibility, achievements, and personality. Make sure it aligns with my goal.

Experience Section Feedback – Identify any inconsistencies, outdated phrasing, or missed opportunities to add impact. Suggest edits that use action verbs and quantifiable results.

Profile Strengths & Gaps – Point out what’s working well and what’s missing. 

Tone & Style Alignment – Ensure the voice is confident, approachable, and tailored to the audience I want to attract.

Format the output with bold headers for each section. Include examples where possible and be specific — I want this to be ready-to-copy or easy to tweak.

🤓 Want more? Check out our Top 1,000 Prompts

AI-GENERATED IMAGES

Mop-bob Tired-Pants

ChatGPT Prompt: a real sponge in pants and spongebob face is standing with a mop and a bucket of detergent on a dirty plate. all this is in the sink. you can see the faucet and part of the countertop. spongebob is washing the plate with a mop and he has a tired face, as if he is going to quit

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

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