DeepSeek's ‘Sputnik moment’

PLUS: Meta brings memory, personalization to AI assistant

Author

Rowan Cheung
January 28, 2025

Good morning, AI enthusiasts. The AI world is still digesting DeepSeek’s latest R1 breakthrough, but the company is already back with another launch, taking aim at the image generation arena.

With two major releases in a matter of weeks, is tech’s latest disruptor about to upend the AI power balance?

In today’s AI rundown:

  • DeepSeek launches new AI image model

  • Qwen launches AI models that control devices

  • Turn product photos into 3D objects

  • Meta AI gets a personalization upgrade

  • 4 new AI tools & 4 job opportunities

LATEST DEVELOPMENTS

DEEPSEEK

Image source: DeepSeek

The Rundown: Chinese AI startup DeepSeek just released Janus-Pro, a new open-source multimodal AI model that outperforms major image generation rivals like DALL-E 3 and StabIe Diffusion — coming on the heels of the company’s viral R1 launch.

The details:

  • The new Janus-Pro model family generates high-quality images from text descriptions, with 1B and 7B parameter models available.

  • Janus-Pro outperformed DALL-E 3 and Stable Diffusion in key industry benchmarks for image quality and accuracy, such as GenEval and DPG-Bench.

  • The models were released under an MIT license, allowing developers to freely use and modify the model for commercial projects.

  • The launch follows DeepSeek's R1 release, which achieved o1-level reasoning capabilities at far lower costs — shaking U.S. markets and the industry.

Why it matters: DeepSeek is the talk of the town, and the effects of R1 are being felt throughout markets as the world digests the reshaping of assumptions around development costs and capabilities. While the current panic may be an overreaction, the Chinese lab has raised questions about the U.S.'s perceived lead in the space.

TOGETHER WITH TURING

The Rundown: Turing’s AGI Advancement Hub provides tailored solutions to enhance your language models — helping you optimize performance, reduce costs, and align your models with real-world use cases for everything from multimodality to RLHF.

Explore how to:

  • Improve reasoning, coding, and multimodal capabilities

  • Deploy fine-tuned models for enterprise-grade performance

  • Benchmark effectively with cutting-edge evaluation methods

Visit Turing’s AGI Advancement Hub and start optimizing today.

ALIBABA

Image source: Alibaba

The Rundown: Alibaba’s Qwen team just released Qwen2.5-VL, a new family of vision-language models that can interact with computers and phones, alongside new advanced capabilities in document and video analysis.

The details:

  • The flagship 72B model outperforms GPT-4o and Claude 3.5 Sonnet on key benchmarks for document parsing and video understanding tasks.

  • The system can analyze hour-long videos and extract specific moments while processing complex documents like invoices and forms.

  • A new feature gives the AI agentic control for smartphone apps and computers, with demos including airfare booking, image editing, and code installation.

  • The smaller 3B and 7B versions are freely available, with the 72B model requiring permission for large-scale commercial uses.

Why it matters: A new ‘operator’ has entered the chat — with Qwen’s computer using vision model coming just a week after OpenAI’s hyped release. Between Qwen and DeepSeek’s massive past week of releases, the gap between open and closed and China vs. U.S. models continues to feel closer than ever before.

AI TRAINING

The Rundown: Krea AI’s real-time generation feature lets you create professional product visualizations by combining images, shapes, and AI-assisted enhancements.

Step-by-step:

  1. Access Krea AI's real-time generation feature from the Generate menu.

  2. Upload your product image and position it on the canvas.

  3. Add shapes and painted areas where you want additional elements.

  4. Fine-tune using the AI Strength slider and positioning tools.

Pro tip: Keep AI Strength low for realistic products. Increase it for creative backgrounds and effects.

PRESENTED BY INNOVATING WITH AI

The Rundown: Innovating with AI’s new program, AI Consultancy Project, trains AI enthusiasts with the skills and strategies to capitalize on the rapidly growing AI consulting market – which is set to 8x by 2032.

The program offers:

  • The tools and frameworks to find clients and deliver top-notch services

  • A 6-month plan to build a 6-figure AI consulting business

  • Students getting their first AI client in as little as 3 days

META

Image source: Meta

The Rundown: Meta just announced the rollout of new AI personalization features that let its assistant remember conversations and tap into user data across Facebook, Instagram, and WhatsApp.

The details:

  • Meta AI can now remember key details from one-on-one chats, like dietary preferences and interests, to provide more tailored responses.

  • The assistant will also access users' Facebook locations, Instagram viewing history, and other profile data for personalized recommendations.

  • The features are launching in the U.S. and Canada across Meta's platforms with no opt-out option, though specific conversation memories can be deleted.

  • ChatGPT and Gemini have also added ‘memory’ to their assistants, though limited to strictly in-chat and not Meta’s social data and integrations.

Why it matters: Meta has a wealth of social data at its disposal, and tapping into it (similar to its hyper-personalized ads) could give a unique edge with an assistant more in tune with users. However, the lack of an opt-out option feels like a major miss, especially given the company’s complicated history with user data and trust.

QUICK HITS

  • 🤖 Qwen2.5-1M - Alibaba’s updated models with 1M token context length

  • 🦙 Llama Stack - Bring GenAI applications to market with unified APIs

  • 📋 OpenAI Canvas - Now available in ChatGPT with o1 and on desktop, with new code rendering capabilities

  • ⚙️ co.dev - Turn ideas into full-stack apps using natural language

LinkedIn co-founder Reid Hoffman announced a $24.6M raise for Manas AI, a new AI-powered drug discovery platform targeting cancer treatments.

DeepSeek’s R1 surged to No.1 on Apple's App Store with 2.6M downloads, but due to reported cyberattacks, it has paused new user sign-ups outside of China.

xAI’s upcoming Grok-3 model briefly appeared to go live for some users, showcasing improved reasoning capabilities — with a full release expected this week.

Pika Labs launched v2.1 of its AI video generation model, which features advanced motion control, realistic physics, and customizable scene elements.

Apple released iOS 18.3, with the new system update turning on Apple Intelligence by default on devices — though AI summaries still remain disabled.

French government AI chatbot Lucie was suspended just days after launch following widespread errors with basic facts, despite funding from President Macron.

COMMUNITY

Join our next workshop today at 1 PM EST and discover how your musical ideas can become part of your creative toolkit with Sara Adkins, Machine Learning Engineer and Artist in Residence at Suno.

RSVP here. Not a member yet? Join The Rundown University on a 14-day free trial.

We’ll always keep this newsletter 100% free. To support our work, consider sharing The Rundown with your friends, and we’ll send you more free goodies.

That's it for today!

Before you go we’d love to know what you thought of today's newsletter to help us improve The Rundown experience for you.

Login or Subscribe to participate in polls.

See you soon,

Rowan, Joey, Zach, Alvaro, and Jason—The Rundown’s editorial team

Reply

or to participate.