Mira Murati's startup emerges from stealth

Read time: under 4 minutes

Welcome back, Superhuman. Step aside PayPal mafia, the OpenAI alumni are stepping into the limelight. Dario Amodei cofounded Anthropic, Ilya Sutskever is raising a billion for his AI venture, and now former CTO Mira Murati just unveiled her new AI startup.

Today’s Insights

• Mira Murati’s startup gets a name, R1 unleashed, and OpenAI’s board shakeup

• Grok-3 proves that AI scaling isn’t finished yet

• Tutorial: How to write academic papers with Deep Research

• 5 new AI tools to boost your productivity

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

TODAY IN AI

1. Ex-OpenAI CTO’s new startup finally has a name: Since leaving OpenAI last September, Mira Murati has been quietly assembling a star-studded team — including OpenAI co-founder John Schulman and alums from Character AI and DeepMind — for her next project. With an estimated $100M to work with, Thinking Machine Labs just emerged from stealth with the noble (if a little vague) goal of making AI systems that are “more flexible, adaptable, and personalized.”

2. DeepSeek’s R1, minus the censorship: Perplexity unveiled a fine-tuned version of DeepSeek’s R1, cheekily called R1 1776, that offers the model’s powerful reasoning capabilities, but with fewer sensitivity filters. Critics say R1 ordinarily steers clear of topics that could violate China’s strict censorship rules. Perplexity’s version gets rid of those guardrails through post-training — plus, it’s available on HuggingFace so programmers can take a closer look at how it was built.

3. OpenAI races to prevent another takeover: Still reeling from Elon Musk’s $97B bid, OpenAI is thinking over a rule change that would give its non-profit board the final say over major decisions. Those special voting privileges would prevent investors like SoftBank and Microsoft from grabbing too much power. In an emergency, OpenAI could also use what’s called a “poison pill” — or discounting stocks for shareholders to make a third-party takeover much more expensive.

PRESENTED BY GOOGLE

Give your startup an unfair advantage with Google's Startup School

Our immersive training series empowers founders and startups with the knowledge you need to build and scale the next breakout product:

• Live, hands-on, and free classes

• Led by Google experts and industry leaders

• Build smarter with AI and learn how to implement pre-trained APIs

• Connect & collaborate with industry leaders

• Receive guidance from our cloud and AI experts to help you grow

Register and save your spot. We're excited to meet you on February 25th!

FROM THE FRONTIER

Grok-3 proves that AI scaling isn’t finished yet

Nvidia’s right back where it started. Just weeks after DeepSeek panic sent the company’s market capitalization tumbling, it’s already recovered nearly all of its $600B in losses.

What caused the quick turnaround? At first, it looked like DeepSeek’s R1 — which was built at a fraction of the cost of its US rivals — had blown a hole in the theory that AI innovation can only happen through huge investments and more chips. But a steady stream of releases, from OpenAI’s o3-mini to Perplexity’s new Deep Research feature, is starting to complicate that picture.

The biggest indicator: xAI’s Grok-3. It’s now considered the most powerful AI model in the world across multiple benchmarks. And it’s no coincidence that it was trained using one of the globe’s largest training clusters, a Memphis supercomputer called Colossus that’s made up of 200,000 GPUs. Next, Elon Musk wants to scale up to 1M Nvidia chips, adding at least 100K per quarter.

What it means: It’s another sign that AI scaling hasn’t hit a wall quite yet. While training techniques are important, there’s also the simple fact that more GPUs can still lead to better performance. Whether we can reach AGI through sheer spending is still an open question. But for now, it seems like R1 might have been an exception to the rule — and one that can’t be easily replicated.

THE AI ACADEMY

How to write academic papers using ChatGPT’s Deep Research

1. Go to ChatGPT and sign up with your account.

2. Select ‘o3-mini’ model and turn on ‘Deep Research’ feature.

3. Enter your prompt and press enter.

Sample Prompt:

I'm writing an academic paper on [Your Topic]. Please provide a brief overview, outline (Intro, Literature Review, Methodology, Analysis, Discussion, Conclusion), key studies, research methods, data presentation ideas, and main implications—all in formal academic language with proper citations.

1. Wait 5-30 minutes while it analyzes hundreds of sources and creates citations.

2. Review your report with clickable citations and verified academic sources.

PRESENTED BY NETGURU

Build an AI Agent PoC with a 20% discount for Superhuman readers 🔥

Have an idea for an AI-powered solution? We’re here to help!

Netguru excels at building rapid proofs of concept.

We’ve built an AI PoC for Merck – we’re sure we can build one for you too.

Validate your idea now and save 20% 🙌

AI & TECH NEWS

Everything else you need to know today

🧑‍💻 Get Real: OpenAI unveiled a new benchmark that grades how well models can perform $1M-worth of real-world coding tasks — “ranging from $50 bug fixes to $32,000 feature implementations.”

🎨 Model for Hire: Freelancing platform Fiverr’s new generative features will let gig workers create custom LLMs based on their own work, then charge clients for access to those models.

💥 Friendly Fire: A turf war is reportedly brewing within Google’s massive AI division, with teams vying for control over popular products like NotebookLM.

💼 Future of Finance: A new study from software company Datarails found that 27% of all CFO job listings this January mentioned AI, compared to just 8% during the same period last year.

💰 Funding: Hightouch, which helps sales and marketing teams organize their data with AI, raised $80M in a Series C funding round, vaulting it to unicorn status. Legal AI startup Luminance, based in Cambridge, UK, raised $75M in a round led by billionaire Steve Cohen’s Point72 Private Investments. And Humane, which became notorious for its ill-fated $500 AI Pin, has been acquired by HP for $116M — while the Pin itself will be discontinued.

PRODUCTIVITY

5 AI Tools to Supercharge Your Productivity

✅ ElevenLabs: Generate voices for podcasts, ads, videos, and more.

✅ LeadTalk: An AI tool that empowers sales and marketing teams to identify and prioritize high-fit accounts.

✅ Spinach AI: Still taking manual meeting notes in 2025? Put your meetings on autopilot and wow your team and customers. Try it here.

✅ Deplyr: Build and deploy tools for your organization quickly and securely, with agent-enabled code generation, connections to popular data stores, and built-in authentication.

✅ Basejump: An AI data analytics platform that helps you get answers in seconds and share them with your team.

^{* indicates a promoted tool, if any}

PROMPT OF THE DAY

Goal Setting

Prompt: I am [mention the problem you’re facing in detail with background context]. I am coaching a [client/student/mentee] who is looking to achieve [specific goal, e.g., improving productivity, losing weight, starting a business]. Create a personalized goal-setting framework for them, which includes SMART goals, milestones, and actionable steps. Include suggestions for tracking progress and adjusting goals if needed. I want you to [mention how you want the output in detail with examples].

_{Source: Shushant Lakhyani / Medium}

SOCIAL SIGNALS

What’s trending on socials today

🔍 Correcting the Record: OpenAI technical staffer Rex Asabor claims that the full version of o3 still beats Grok-3 across most benchmarks, but that those figures were conveniently left out of xAI’s recent livestream.

⚔️ Screaming Match: Were you “Team o3-Mini” or Team Phone-Sized Model”? Sam Altman’s poll, which we mentioned in yesterday’s newsletter, left fans bitterly divided, with users from both sides pleading their case in the comments.

💬 Grok Shock: Neuralink’s Shivon Zillis said her hour-long conversation with Grok’s soon-to-be-released voicebot “was one of the most unexpectedly rewarding hours of my life.”

💡 Best and Brightest: Data scientist Alex Reibman shared some of the coolest inventions from a recent Stanford hackathon — including one that turns junk materials into a fully-functioning braille printer and another that can transform thoughts into binary.

AI-GENERATED IMAGES

Surreal portraits

Midjourney Prompts: A surreal illustration of a woman with hair made of lines, in blue and white colors, in the style of Kazumasa Nagai. --ar 85:128 --profile 58adm17

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 opinion matters!

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