Access previous editions of the newsletter by clicking here

DeepSeek R1 Special Edition

On today’s menu:

• DeepSeek R1, an open source Chinese AI model, delivers performance on par with GPT-4 for just 5% of the cost.

• The startup sends a shockwave through financial markets, knocking down Nvidia (-17%) and other tech giants.

• Meta sets up four crisis teams to dissect an innovation that threatens its upcoming Llama 4 model.

• A new benchmark, “Humanity’s Last Exam”, challenges AI models with 3,000 expert-level questions and a $500,000 prize.

Ready? Let’s go...

THE BIG STORY

🤖 DeepSeek R1: The Chinese AI rattling Silicon Valley

Details:

• DeepSeek R1, the new Chinese AI model, matches the performance of OpenAI’s o1 model for just 5% of the cost. A feat that’s shaking up the tech giants.

• The app is sitting at the top of the download charts on the App Store, even ahead of ChatGPT.

• Meta has rolled out four crisis teams to figure out this unexpected success, even as the company plans to invest $65 billion in AI.

• The market reacted violently: a spectacular drop in shares of Nvidia (-17%), Broadcom (-17.4%) and TSMC (-9.88%).

• DeepSeek shows the power of open source: according to Meta’s Yann LeCun, it’s a win for open models over proprietary solutions.

Why I’m telling you about this ⤵️

DeepSeek’s breakthrough challenges an industry dogma: you no longer need colossal budgets or pricey Nvidia chips to build a high-performing AI. With standard components and a modest $5.6 million budget, DeepSeek proves that a leaner, more open approach can go toe to toe with the giants. This could reshape the AI landscape for good, making it more accessible and less dependent on the dominant players.

🤖 Deepseek R1: why everyone’s talking about this AI

Deepseek is making waves in the AI world with its R1 model, a free chatbot that rivals the premium versions of ChatGPT. What sets it apart? A hybrid approach to reasoning that combines objective and subjective evaluation, letting it excel at both coding and creative work.

The strengths that set Deepseek apart:

• Remarkable execution speed

• The ability to share its “inner dialogue”

• Superior performance on many benchmarks

• Optimized energy consumption

⚠️ Watch out with sensitive data: Deepseek stores all information in China and does not comply with the GDPR. For secure use, go with a local install.

Innovation through constraint: With just $6 million in investment, this Chinese startup proves you can build a high-performing AI without the colossal resources of the American giants. Their inspiring story is detailed here.

But what really sets DeepSeek R1 apart is that it’s free to use, with performance on par with ChatGPT o1. The model is also Open Source, unlike OpenAI’s closed models. You can download the model, modify it however you like, and even run it offline on your own computer. It’s a real kick in the hornet’s nest that will surely energize the industry and keep putting ever more advanced models within reach of users like us.

To try Deepseek: Head over to their platform and turn on “DeepThink” mode to take advantage of its advanced reasoning capabilities.

🔄 DeepSeek: Reactions and suspicions

Details:

• OpenAI accuses DeepSeek of using its data to develop a rival model, citing evidence of “distillation” – a technique for boosting a small model’s performance by learning from a bigger one. Source

• Meta on alert: “War rooms” set up to crack DeepSeek’s performance ahead of the release of Llama 4 in early 2025

  • 2 teams are analyzing cost-cutting strategies

  • 2 teams are studying the data sources and architecture

• What’s at stake: R1’s performance could surpass Llama 4, expected in early 2025

• Preventive measures: Microsoft and OpenAI have blocked API access for accounts suspected of being linked to DeepSeek.

Why I’m telling you about this ⤵️

This situation highlights two major issues: the crucial importance of training data in AI development, and the possibility of building high-performing models at a lower cost. DeepSeek is upending established standards, forcing the tech giants to rethink their innovation strategies.

Join the AI revolution with Five Minute AI! 🚀

Share our passion and earn exclusive rewards.

Click the Share button below, invite your friends to join our community, and unlock unique gifts with every new sign-up!

1️⃣ The Ultimate ChatGPT Handbook, packed with tips only insiders know. This guide will help you boost your productivity and simplify your daily tasks → 🎁 Unlock with 3 friends signed up

2️⃣ Five Minute AI collector mug: Featuring our adorable mascot Pixl enjoying its daily read → 🎁 Unlock with 15 friends signed up

Ready to spread AI knowledge? Let’s go! 🌟

LEARN TO MASTER AI

💻 Installing DeepSeek R1 on your computer

Find the long version of this tutorial on labo-ia.fr

DeepSeek recently released lightweight versions of its R1 model that can rival GPT-4o while running free and locally on your machine. Here’s how to get started easily:

Installation steps:

1. Download LM Studio for your operating system

2. Open the app and use the search magnifying glass to find a model (recommendation: the DeepSeek R1 Distill series)

3. In the chat interface, pick your model from the drop-down menu and start chatting

Note: These models run 100% offline, so you can use them anywhere, even without an internet connection.

Key points:

• Guaranteed privacy: your data stays on your machine

• Free to use: no subscription required

• Optimal performance: requires a fairly recent computer

• Bonus feature: you can analyze your local documents through the interface

Online alternative: If privacy isn’t your priority, you can test R1 directly on DeepSeek.com by turning on “DeepThink” mode.

WHAT THE RESEARCH SAYS

🧐 Humanity’s Last Exam: AI’s new challenge

Evaluating the performance of AI models gets harder and harder as they improve. The Center for AI Safety and Scale AI took on that challenge by creating Humanity’s Last Exam, a new benchmark that pushes the limits of testing LLMs’ academic knowledge.

Details:

• A collection of 3,000 questions written by experts from more than 500 institutions across 50 countries

• Multimodal questions combining text and images (10% of the total)

• A surprisingly low score for today’s best AI models (under 10% success rate)

• A $500,000 prize to encourage quality contributions, with rewards of up to $5,000 per question

Why I’m telling you about this
While today’s AI models routinely score above 90% on standard tests, Humanity’s Last Exam sets a new demanding benchmark for measuring AI progress. It arrives at just the right time to help us better understand the real capabilities and limits of advanced AI systems. Learn more

THE LATEST TRENDING TOOLS

💡 THE TOOLS:

• 🔍 Telescope 2.0 offers targeted, efficient search to quickly identify your relevant contacts.

• 🎨 Shapen turns your images and text descriptions into impressive 3D models with remarkable simplicity.

• 📊 Exa Twitter Wrapped provides an in-depth analysis of your Twitter profile, revealing your favorite topics and standout tweets of 2024.

• ✅ Habits helps you with habit tracking for 2025 using a minimalist approach, no sign-up required.

• 🎯 Put Logo Anywhere lets you place your logo in different visual contexts for free, though results may take a few tries.

IN THE NEWS

• A former OpenAI safety researcher shares his concern about the breakneck pace of AI development, stressing that no lab currently has a satisfactory alignment solution.

• DeepSeek breaks new ground with its DeepSeek-R1 model by favoring reinforcement learning over traditional fine-tuning, showing remarkable performance in coding and math.

• An internal Apple memo reveals its AI priorities for 2024: overhauling Siri with language models by spring 2026 and improving its existing AI models, with a rollout planned for iOS 19.4.

• Sam Altman announces that the o3-mini model upgrade will be available for free on ChatGPT, with expanded features for Plus users.

• To dig deeper into DeepSeek, check out the investigation into its alleged ties to China or follow this complete tutorial for beginners.

YOUR DAILY DOSE OF CONTENT

1/ 🚀 DeepSeek R1: The quiet revolution
A new Chinese AI model shaking up the industry by delivering remarkable performance with minimal hardware resources

How China’s New AI Model DeepSeek Is Threatening U.S. Dominance

2/ 🔬 Sam Altman’s take
The CEO of OpenAI acknowledges DeepSeek R1’s impressive capabilities, especially in mathematical reasoning

deepseek’s r1 is an impressive model, particularly around what they’re able to deliver for the price.

we will obviously deliver much better models and also it’s legit invigorating to have a new competitor! we will pull up some releases.

2:29 AM • Jan 28, 2025

83.4K Likes   6.66K Retweets  

3/ 💡 Market impact
A detailed analysis of how this efficient model could challenge the dominance of the tech giants and their reliance on costly infrastructure

Big Tech in panic mode… Did DeepSeek R1 just pop the AI bubble?

THAT’S ALL FOR TODAY

Thanks for reading the work of our purr-severing team 😼

💡 Want to contribute, swap ideas or build partnerships → [email protected] or→ @mattioo81

👍️ Like what you’re reading? Subscribe!

📺 If you prefer video content, I invite you to subscribe to my YouTube channel.

– Mathieu