- The Rundown AI
- Posts
- AI's reasoning race heats up
AI's reasoning race heats up
PLUS: ChatGPT's 'eyes' spotted in beta code
Rowan Cheung
November 21, 2024
Sign Up | Advertise | Podcast | AI University
Welcome, AI enthusiasts.
Just months after OpenAI's o1 made waves, a Chinese lab just upped the ante with a model that matches the AI leader — while also transparently showing its work.
Is this (soon-to-be) open-source powerhouse about to shift the competitive landscape and send a warning shot to closed U.S. labs? Let’s get into it…
In today’s AI rundown:
DeepSeek unveils powerful reasoning AI
ChatGPT's visual AI inches closer to launch
Make any video longer with AI
DeepMind AI fixes quantum computing errors
5 new AI tools & 5 new AI jobs
More AI & tech news
Read time: 4 minutes
Friendly reminder: We’re moving to a new email sending address. Make sure to add [email protected] to your contacts to ensure you keep receiving our emails.
LATEST DEVELOPMENTS
DEEPSEEK
Image source: DeepSeek
The Rundown: Chinese AI research firm DeepSeek just launched R1-Lite-Preview, a powerful new reasoning-focused model that matches OpenAI's o1 capabilities while showing its live transparent chain-of-thought processes.
The details:
Unlike o1's condensed summaries, R1-Lite-Preview shows users its complete chain-of-thought process in real-time.
Initial benchmarks rival OpenAI's o1-preview on benchmarks like AIME and MATH with improved performance as the length of thought increases.
Users can access the model through DeepSeek Chat, with premium reasoning features limited to 50 daily messages, while basic chat remains unlimited.
DeepSeek plans to open-source the complete R1 model in the future
The company's infrastructure includes an estimated 50,000 H100 chips, putting their computing power on par with leading Western AI labs.
Why it matters: Two months after OpenAI's o1 sparked a new era in AI reasoning, DeepSeek's achievement shows how quickly the field evolves. While lesser known in the West, open-sourcing this powerful Chinese model could accelerate innovation across the entire AI industry, sending a warning shot to closed U.S. AI labs.
TOGETHER WITH HUBSPOT
The Rundown: Join industry experts Matt Wolfe and Nathan Lands as they reveal their blueprint for building businesses from the ground up using AI – helping unlock strategies to turn ideas into reality.
In this high-impact episode, you'll discover:
The top AI tools that simplify business creation
Methods to create engaging newsletters quickly with AI
How to make the most of your context mix
Tips to blend AI speed with creativity using avatars, sentiment analysis, and more
Watch the full episode now and take your first step into AI entrepreneurship.
OPENAI
Image source: OpenAI
The Rundown: Newly discovered code in ChatGPT's latest beta suggests OpenAI may be preparing to roll out its long-awaited live camera features, potentially integrating visual capabilities into the assistant’s Advanced Voice Mode.
The details:
The beta code revealed a "Live Camera" feature that allows ChatGPT to analyze and discuss users' surroundings in real-time.
First demoed in May, the tech showed impressive capabilities, such as recognizing objects and engaging in natural conversations about visual input.
The feature previously appeared in limited alpha testing, with some users reporting brief access during Advanced Voice Mode trials.
OpenAI’s potential release comes ahead of Google's similar Project Astra, which was showcased at Google I/O, continuing the AI giants' competitive release pattern.
Why it matters: 2025 is shaping up to be the year of AI agents and full multimodal capabilities, with models able to see, engage, and take action in more natural and intuitive ways. Voice AI has already started to gain traction, but pairing it with 'eyes’ would be a completely transformative new experience.
AI TRAINING
The Rundown: Transform short video clips into longer content using AI to generate seamless extensions of your footage.
Step-by-step:
Take a screenshot of your video's last frame as a reference image.
Upload the screenshot to the Kling AI video generator in the "Image to Video" section.
Write a descriptive prompt for how you want the video to continue.
Join your original video with the AI-generated extension using a video editor like CapCut.
Pro tip: For the most natural results, make sure your reference screenshot is high-quality, and your prompt clearly describes the intended motion or action.
PRESENTED BY SECTION
The Rundown: Last year, 20,000 people signed up for NYU professor (and Section founder) Scott Galloway’s 2024 predictions. Now, he's back for 2025 — giving unfiltered, exclusive access to his forecast for the year ahead.
Join Scott and Section on Dec. 12 to hear:
What Scott got right (and wrong) in 2024
His biggest predictions on AI, business, tech, and more
Raw, unfiltered takes on where the world is headed
Reserve your free spot now for the highly anticipated annual event.
GOOGLE DEEPMIND
Image source: Google
The Rundown: Google DeepMind just introduced AlphaQubit, an AI system that dramatically improves the ability to detect and correct errors in quantum computers — a crucial step toward making the tech practical for real-world use.
The details:
AlphaQubit sets new records for error detection, cutting rates by 6% compared to previous top methods and 30% compared to standard approaches.
A two-step training process allows the system to learn from simulated data before adapting to handle the complex errors in real quantum hardware.
Though trained on sequences of just 25 operations, the system maintains accuracy for over 100k — showing promising ability for quantum computations.
Google plans to open-source AlphaQuibit, allowing the broader research community to build upon the advances.
Why it matters: AlphaQubit tackles one of the field's biggest roadblocks – keeping the sensitive machines stable enough to solve real problems. While more steps are needed, DeepMind’s research brings us a step closer to letting quantum computers loose in areas like drug discovery, climate modeling, supply chains, and more.
NEW TOOLS & JOBS
📱Postiz - AI social media scheduling tool for posting, analyzing performance, and managing accounts all in one place.
🗣️ Adobe Podcast Enhance Speech v2 - AI tool to improve audio quality, eliminate background noise and enhance voice clarity
🧠 Dokko - Conversational AI platform for knowledge sharing
📖 ReadPartner - Personal AI assistant for automated news digest and quick summaries of websites, videos, and documents
📊 PaperGen - Fully-referenced, charted, long-form papers with one click
📝 The Rundown - Writer/Editor
📈 Twelve Labs - Account Executive
🏢 Mistral AI - Office Manager, Palo Alto
📊 Metropolis - Director, Go-To-Market Systems
💻 Figure AI - Software Engineer, Configuration
QUICK HITS
OpenAI released an updated version of GPT-4o featuring improved creative writing capabilities and better file analysis, with the model being revealed as ‘anonymous-chatbot’ and reclaiming the top spot on the Chatbot Arena leaderboard.
Writer introduced a new self-evolving model architecture, enabling real-time learning and the ability for LLMs to operate more efficiently without additional training.
Anthropic published research proposing a statistical framework for AI model evaluations to more accurately measure and compare language model capabilities beyond simple benchmark scores.
Meta rolled out new features to Messenger, including AI-generated video call backgrounds, HD calling capabilities, and intelligent noise suppression features.
Niantic unveiled plans for an AI model trained on millions of player-submitted smartphone scans from its Pokemon Go and Ingress games, aiming to create a system that understands and navigates physical space.
OpenAI and Common Sense Media launched a free ChatGPT course aimed at helping K-12 teachers understand and adopt AI in the classroom.
THAT’S A WRAP
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. |
See you soon,
Reply