This R1 rival cost $50 to train

Read time: under 4 minutes

Welcome back, Superhuman. If DeepSeek’s claim of building an AI model for $6M seemed too good to be true, wait till you hear this one. A team of researchers at Stanford and the University of Washington say they’ve built a new reasoning model by fine-tuning an existing model for *check notes* 50 US Dollars.

Today’s Insights

• Mistral goes mobile, AI-powered job searches, and a corporate shakeup

• Stanford researchers train reasoning model in less than 30 minutes for $50

• Tutorial: How to generate images with legible text

• 5 new AI tools to boost your productivity

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

TODAY IN AI

1. Mistral’s major overhaul: Don’t count Mistral out. The Paris-based AI startup — one of Europe’s largest — is launching a mobile app, bringing its open-source Le Chat chatbot to a new group of users. The speedy model comes stocked with features like AI web search, photo generation, document analysis, and coding expertise. And stay tuned for a “memories” tool that will help it remember important details about your life and work.

2. LinkedIn wants to use AI to help you find your dream job: A job title can only tell you so much. That’s why LinkedIn is experimenting with a new tool that would be able to go a step further, finding positions that match your interests. For example, you could type something like: “Find me a role where I can use marketing skills to help the environment” — or even specify the salary range you’re aiming for. Eventually, the platform wants to use AI to spot emerging career trends, like which skills are most in demand.

3. OpenAI cofounder is onto the next adventure: AI researcher John Schulman spent nine years at OpenAI, where he was a leading architect behind ChatGPT. Five months ago, he left to join Anthropic — helmed by fellow OpenAI alum Dario Amodei. Now, Schulman is rumored to be headed to none other than ex-OpenAI CTO Mira Murati’s next venture, according to Fortune. It’s still unclear what Schulman’s role will be or what exactly the much-hyped company is working on.

PRESENTED BY EIGHT SLEEP

How I get an extra 400 hours of sleep per year

As co-founder and CEO of Superhuman AI, I’m always looking to optimize my routine. So how do I work 60-80 hours per week and get a full night’s rest, every night?

I hack my sleep with The Pod by Eight Sleep.

The Pod is the most simple and effective way to fall (and stay) asleep. Since using it, my deep sleep has increased by more than 1 hour each night – that's almost 17 days of extra sleep each year:

• High-tech mattress cover + app that tracks your sleep data

• Automatically cools/heats each side of the bed to maintain optimal temp range

• The most advanced version, the Pod 4 Ultra, can even reduce snoring

Get your best night’s sleep with $350 off Pod 4 Ultra and $200 off Pod 4 with code: ‘SUPERHUMANAI’.

FROM THE FRONTIER

Researchers just built a competitive reasoning model by fine-tuning an existing one for $50

Researchers at Stanford and the University of Washington say they managed to build a top model for less than the price of a nice dinner — and in just 26 minutes. Their process might shed some light on how DeepSeek’s R1 was able to blow past the competition a few weeks ago.

How’d they do it?

• They gathered “1,000 carefully curated questions,” then asked Google’s Gemini Thinking Experimental to answer each of them.

• Like other reasoning models, this version of Gemini explains the thought process behind each of its decisions.

• The researchers collected that data, then showed it to Qwen, a small, open-source model developed by Chinese tech giant Alibaba.

• By closely observing Gemini’s workflow, the model learned to think just like it; the training process reportedly took less than a half hour.

• The result: A model called s1 that’s competitive with OpenAI’s o1 and DeepSeek’s R1 across coding and math.

What it means: The results are welcome news to open-source advocates, proving that the “moat” around cutting-edge AI is getting easier to hop over. On the other hand, if virtually anyone can build their own LLM with a couple dozen GPUs, it could throw a wrench in the industry’s entire business model. It also supports the idea that DeepSeek might have used a similar method to train R1, which was built on a much tighter budget than its US counterparts.

THE AI ACADEMY

How to generate images with legible text

Watch the tutorial here or follow the instructions below:

• Go to Ideogram’s website and sign up.

• Write in your prompt containing long text in inverted commas and select the aspect ratio, model, and color palette of your choice.

• You can choose from 5 distinct styles: General, Realistic, Design, 3D, and Anime. In case of textual images choose Design as a style.

• Once you have selected everything, click on the Generate button and wait for a few seconds.

• You’ll get your image with legible text generated within seconds.

PRESENTED BY GALILEO

Everything you need to know about AI agents

Galileo just dropped a 100-page ebook on AI agents, so you can create powerful, reliable agents like an expert:

• Matching the right agentic framework for your use case

• Evaluating/improving performance

• Identifying failure points and production issues

Start learning today.

AI & TECH NEWS

Everything else you need to know today

⏱️ Quick Fix: Lyft announced it’ll use Anthropic’s Claude model to handle customer service requests, speeding up resolution times by 87%.

📱 Image to App: GitHub’s Copilot tool can now automatically convert mockups and images into code, complete with an interactive interface and alt text. That’s part of a bigger update that gives Copilot a variety of new agentic capabilities.

👀 Shared Vision: OpenAI announced you can now share your Canvas — a custom workspace designed for coding and other complex tasks — with collaborators. Plus, Canvas now works with o1. Finally, OpenAI is updating o3-mini’s chain of thought for both free and paid users.

🔊 Sound Check: ElevenLabs is opening up access to its Studio tool, a feature that lets users edit long-form AI-generated content, like audiobooks and podcasts.

🏈 Main Event: OpenAI will reportedly broadcast its first-ever TV commercial in the Super Bowl this Sunday, joining the likes of Microsoft, Meta, xAI, and Anthropic.

PRODUCTIVITY

5 AI Tools to Supercharge Your Productivity

✅ Basedash: An AI-native Business Intelligence platform that lets you create dashboards and instantly understand your customers using natural language.

✅ Jam AI: Record a bug in your code and let AI write a bug report and repro steps for you.

✅ Bardeen*: AI Copilot for GTM teams, helping you work smarter, close faster, and win more deals. This powerful browser extension connects your tools and automates your busiest workflows—giving you more time to focus on what matters.

✅ Trendtracker: Evaluate a trend’s growth, impact, and relevance instantly, with real-time personalized insights.

✅ Concierge: Connect to your favorite apps with just a few clicks, select any model, then simply ask whatever you want with natural language.

^{* indicates a promoted tool, if any}

PROMPT OF THE DAY

Establishing Routines

Prompt: I want to establish a daily routine to help manage my ADHD. Can you suggest a morning and evening routine that will help me stay organized and focused throughout the day?

_{Source: Millennium Medical Associates}

SOCIAL SIGNALS

What’s trending on socials today

🎬 Picture This: Pika shared some examples of how its new Pikadditions tool lets you easily insert any object, person, or animal into your videos.

🎸 Bustin Jieber: A fake AI-generated Justin Bieber song has almost 2M views on YouTube.

🧠 Mind Over Matter: Neuralink revealed new details about how its three patients are “using telepathy in their daily lives” thanks to the Elon Musk-led startup’s experimental brain interface.

⚖️ Pros and Cons: Economist Kevin A. Bryan shared a helpful explanation for why ChatGPT’s deep research feature is really good at some tasks, but not so good at others.

🌎 Global Grammar: A new research paper reveals why LLMs are so good at mastering non-English languages, even when 90% of their training data is in English.

AI-GENERATED IMAGES

Tablescapes

Midjourney Prompt: A minimalist dining table setting with an empty plate, a blue-colored bottle, and two glasses on it, along with oranges, embroidered napkin, white poppies in vase, and cutlery on a white linen cloth. A delicate hand is reaching for a fork. The sunlight casts soft shadows across the scene, creating gentle highlights that accentuate textures like speckled ceramic plates and textured glassware. A sense of tranquility surrounds you, as if inviting you to slow down your fast-paced photography in the style of a minimalist artist. Frontal view photograph

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