DeepSeek's origin story

Read time: under 4 minutes

As the dust settles around DeepSeek’s new reasoning model, we’re left with two big questions: Who’s behind the breakthrough — and what’s their secret sauce? Researchers are now racing to find out. And: Learn how to run R1 locally via LM Studio.

Today’s Insights

• Today in AI: R1’s origins and ChatGPT goes to Washington

• Tutorial: How to run DeepSeek’s R1 locally

• The unexpected winners of the DeepSeek shakeup

• Everything else you should know today

• 5 new AI tools to boost your productivity

TODAY IN AI

1. DeepSeek’s unlikely origin story: DeepSeek’s founder might be more interested in the mysteries of math than in money. Seemingly overnight, Liang Wenfeng, 40, has become one of AI’s biggest celebrities. He got into AI while still in school, where he built algorithms that could analyze the stock market.

• “People may think there’s some hidden business logic behind this, but it’s mainly driven by curiosity,” Liang has said

• Top AI engineers are still trying to figure out how Liang managed to come to the same conclusions as them — but with relatively little funding, dated Nvidia chips, and just 20 months of development time

• Mark Chen, a Chief Research Officer at OpenAI, commended DeepSeek for “independently finding some of the core ideas that we did on our way to o1”

2. ChatGPT introduces a new tailored version of ChatGPT for the US government. The move is designed to help government employees and agencies take advantage of AI in a secure and effective manner. OpenAI also shared that over 90,000 users across US federal, state, and local government agencies have sent over 18M messages on ChatGPT. Self-hosting will also enable agencies to manage their own security, privacy, and compliance requirements as needed.

3. It just got a lot easier to try out models on Hugging Face: To use a model from Hugging Face’s vast library, you first had to get it off of the repository and onto a third-party host — until now. The platform just revealed a new feature that lets you try out a model of your choice — yes, even R1 — in just a few clicks with the help of cloud providers like Together AI, Replicate, and Fal. And speaking of R1, Hugging Face is also assembling a team to recreate the reasoning model from scratch in the name of “open knowledge.” Even though the model is technically open-source, the datasets and code used to train it are still a mystery.

PRESENTED BY SPINACH

Still Taking Manual Meeting Notes in 2025? There’s an AI for That

Spinach AI is not only the most accurate AI notetaker, it also automates the work you do after the meeting:

• Writes recap emails for you based on your style and preference

• Updates your task management tool: Jira, Asana, Monday, Trello, Linear, or ClickUp

• Updates your CRM - HubSpot, Salesforce, Zoho, or Attio

• Integrates with thousands of additional tools with Zapier

Backed by Y Combinator, Zoom & Atlassian, trusted by thousands of companies.

Try the most accurate AI notetaker now

THE AI ACADEMY

How to run DeepSeek R1 locally using LMStudio

• Go to LM Studio, select your operating system, and download it.

• Now, open LM Studio and go to the ‘Discover’ tab.

• Find ‘DeepSeek R1 Distill QU 7B’ or ‘DeepSeek R1 D Llama’ and download it.

• Once downloaded, find it under ‘My Models’ and select the ‘DeepSeek R1 model’

• Click ‘New Chat’, type or paste your prompt, and click Send.

• Wait for a response (model will show detailed reasoning).

FROM THE FRONTIER

The bright spot in DeepSeek’s shakeup

Most American big tech companies had a rude awakening this week as they realized that a small Chinese startup might have beaten them at their own game — and for a fraction of the cost. But not everyone got burned. For instance, Apple was watching from the sidelines with a bag of popcorn as its stock managed to not sink, but pop despite the chaos.

The reason offers clues about where the industry’s headed: Everyone dunked on Apple for missing the AI boat early on. But now its decision to rely on third parties for most of its AI needs suddenly sounds like the most sensible approach. With R1 proving that AI might one day become widely available at very low costs, why invest billions in your own model when you can adapt an open-source one instead?

Apple wasn’t alone: 

• Meta made out fine, too, since its open-source vision meshes well with DeepSeek’s

• Even VCs were excited by the possibility that R1 might actually motivate US-based AI companies to innovate faster, with Y Combinator’s Garry Tan writing “A thousand flowers will bloom” from the shakeup

• Finally, the news could speed up enterprise’s AI adoption rates; with a lower barrier to entry, it’s now much more feasible for companies to start incorporating AI into their workflows

• Already, AWS clients are clambering to try out R1 because of its unique balance of price and performance

PRESENTED BY BARDEEN

The only AI Copilot built for GTM teams

Bardeen is the AI Copilot for GTM teams, helping you work smarter, close faster, and win more deals. Our AI-powered browser extension automates your busiest workflows—giving you more time to focus on what matters.

🚀 Get our free guide with automations used by companies like Deel, Miro, and Kearney.

AI & TECH NEWS

Everything else you need to know today

🏃 Model Marathon: Just a day after releasing an agent that can navigate your computer and phone, Chinese tech giant Alibaba is back with an LLM called Qwen2.5-Max that beats DeepSeek’s V3 and Meta’s Llama 3.1 across multiple benchmarks.

📚 Study Buddy: Android users will soon be able to access Deep Research, an agent that helps you “create comprehensive research reports on the go,” on the Gemini Advanced app.

🎬 Director’s Cut: Text-to-video platform MiniMax dropped a new “Director” model that gives users more control over their creations, complete with smoother transitions and cinematic camera angles.

🪙 Digital Dough: Elon Musk’s X brought on Visa as its first partner for the platform’s new digital wallet, which will work similarly to Zelle and Venmo.

💫 Space Texts: Apple and SpaceX are working on bringing support for Starlink on iPhones. For now, it’s only available in beta for select T-Mobile customers.

📈 Markets: Nvidia’s stock jumped 8.93% Tuesday as it began to recover from record-breaking losses over the fallout from DeepSeek’s new model. Alibaba soared 6.71% after it announced a new frontier model. And Apple finished the day up 3.65% due to stronger iPhone sales in China.

💰 Funding: Spanish startup Quibim, whose AI imaging tool can detect prostate cancer early, raised $50M in a Series A round. And Washington’s Helion Energy raised $425M in a Series F — pushing its valuation to $5.25B.

PRODUCTIVITY

5 AI Tools to Supercharge Your Productivity

✅ Bulletpen: An AI app that transforms your spoken thoughts and rambles into polished writing.

✅ Nowadays: An AI-powered copilot that will contact venues and handle negotiations for you, taking the hassle out of corporate event planning.

✅ Lido*: Turn PDFs into clean spreadsheet tables with AI. Extract data from invoices, contracts, POs, and more in seconds. Save time, reduce errors, and scale without headcount.

✅ Omakase AI: Turn your website into an AI-powered retail experience with just a URL — no coding needed.

✅ Sense: Automatically organize related resources and updates across all apps, giving you a ChatGPT-like search for your company.

^{* indicates a promoted tool, if any}

PROMPT OF THE DAY

Explore Data

Prompt: I want you to act as a data scientist and code for me. I have a dataset of [describe dataset]. Please write code for data visualisation and exploration.

_{Source: Travis Tang / GitHub}

SOCIAL SIGNALS

What’s trending on socials today

🕵️ Black Box: OpenAI founding member Andrej Karpathy dove into the mystery behind Move 37 — when an AI “discovers actions that are new, surprising, and secretly brilliant even to expert humans” — and explored why the phenomenon is becoming more common.

🤳 Silicon Signals: Sam Altman posted a selfie with Microsoft CEO Satya Nadella, hinting at the “next phase” of the companies’ partnership.

📊 Seeing is Believing: AI expert Jay Alammar created a handy illustrated guide that uses diagrams to show how DeepSeek’s R1 was built and trained.

🔍 Sixth Sense: A new study shows that if you use ChatGPT a lot, you can probably spot AI-generated writing, even if it’s filtered through “humanization” software first.

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