- The Rundown AI
- Posts
- China’s new AI tops GPT-4o
Sign Up | Advertise | Tools | AI University
Welcome, AI enthusiasts.
Alibaba just unveiled Qwen2-VL, a new vision-language AI that outperformed GPT-4o across several key benchmarks.
As China's AI developments continue to accelerate, is the gap between U.S. leaders closing faster than expected? Let’s get into it…
In today’s AI rundown:
China’s new Qwen2 beats GPT-4o
OpenAI and Anthropic partner with US gov
Add yourself to images with a custom LoRA
AI startup reaches 100M token context
5 new AI tools & 4 new AI jobs
More AI & tech news
Read time: 4 minutes
LATEST DEVELOPMENTS
ALIBABA
Image source: Midjourney
The Rundown: Alibaba just unveiled Qwen2-VL, a new vision-language AI model that outperforms GPT-4o in several benchmarks — particularly excelling in document comprehension and multilingual text-image understanding.
The details:
Qwen2-VL can understand images of various resolutions and ratios, as well as videos over 20 minutes long.
The model excels particularly at complex tasks such as college-level problem-solving, mathematical reasoning, and document analysis.
It also supports multilingual text understanding in images, including most European languages, Japanese, Korean, Arabic, and Vietnamese.
You can try Qwen2-VL on Hugging Face, with more information on the official announcement blog.
Why it matters: There’s yet another new contender in the state-of-the-art AI model arena, and it comes from China’s Alibaba. Qwen2-VL’s ability to understand diverse visual inputs and multilingual requests could lead to more sophisticated, globally accessible AI applications.
TOGETHER WITH ANYSCALE
The Rundown: Ray Summit 2024 is just around the corner. Get insights from cutting-edge AI companies like NVIDIA, Intel, Google, and others. Featuring 60+ breakout sessions and hands-on tech training, only available in person in San Francisco from September 30-October 2.
At Ray Summit you’ll:
Learn from OpenAI's CTO, Instacart's co-founder, and others
Explore real-world AI solutions from tech giants using Ray and Anyscale
Boost your skills with hands-on training in GenAI, LLMs, and more
Register now with code “Newsletter15” for an exclusive discount for readers of The Rundown.
OPENAI & ANTHROPIC
Image source: Midjourney
The Rundown: OpenAI and Anthropic just signed a groundbreaking agreement with the U.S. Artificial Intelligence Safety Institute to allow government access and testing of their AI models before public release.
The details:
The U.S. AI Safety Institute will have access to major new models from both companies prior to and after their public release.
This collaboration is a step toward AI regulation and safety efforts, with the U.S. government evaluating AI models’ capabilities and associated risks.
The institute will provide feedback to OpenAI and Anthropic on potential safety improvements that should be made.
These agreements come as AI companies face increasing regulatory scrutiny, with California legislators recently passing a broad AI regulation bill earlier today.
Why it matters: The two most popular AI companies in the world are granting the U.S. government access to unreleased models before release. This could reshape how AI is developed, tested, and deployed worldwide, with major implications around innovation, safety, and international competition in the AI space, for better or worse.
AI TRAINING
The Rundown: Fal AI's Flux LoRA training tool helps you create a customized AI image generation model that can create images of you in any scenario or style from a few selfies.
Step-by-step:
Visit Fal AI's Flux LoRA training page and create an account (requires ~$10 in credits).
Upload 6-12 high-quality images of yourself with clear backgrounds.
Set training steps to 1000 and add a unique trigger word (e.g., "YourName").
Start the training process (takes about 20 minutes).
Generate images using prompts like "Portrait of [YourName] as a superhero" and explore!
Pro tip: To go one step further, import your images into Runway’s Gen-3 image-to-video feature and turn your generations into short clips.
THE RUNDOWN AI UNIVERSITY
The Rundown: Tomorrow we’re hosting a live workshop with Flo Crivello, the founder of Lindy AI, to show you how to create AI agents that respond to emails, negotiate, and schedule meetings on your behalf.
Join us on live at 1:30 PM PST to learn:
What Lindy's AI agents are capable of, and how to use the platform
How to create your own agent that responds to emails on your behalf
Exclusive Q&A for all members to ask any questions
If you’re a member of The Rundown AI University, you can RSVP here.
If you’re not a member yet, you can still join the workshop for free with a 14-day free trial.
MAGIC
Image source: Midjourney
The Rundown: Magic just developed LTM-2-mini, a model capable of processing 100 million tokens of context — equivalent to about 10 million lines of code or 750 novels — and partnered with Google Cloud to build advanced AI supercomputers.
The details:
LTM-2-mini can process and understand 100 million tokens of context given during inference, surpassing current models by 50x.
The model’s innovative algorithm processes long sequences of data 1000x more efficiently than the current top-performing AI models.
Magic is also partnering with Google Cloud to build supercomputers powered by Nvidia’s newest and most advanced GPUs.
The company has raised more than $450 million in total funding, including a recent $320 million investment round.
Why it matters: This breakthrough in context length allows AI agents to process and reason over dense and complicated codebases, vast databases, and years of conversation history in a single inference. It’s a significant step toward creating AI assistants with near-perfect recall and memory.
NEW TOOLS & JOBS
💻 GPT Engineer - Chat with AI to build web applications on your behalf
🚀 Mimrr - Automate documentation and analysis for development teams
🤖 AgentOps - Develop and debug AI agents efficiently
🎨 Krea FLUX Style Mixer - Mix multiple Flux image generation styles with full control
🔎 Next Alpha Andromeda - LLM for crypto research
🧑🔬 Mistral AI - AI Scientist - Internship
🤝 Meta - Business Development and Partnership Manager Artificial Intelligence
🚦Waymo - Product Manager, Perception
💪 Captions - Technical Recruiter
QUICK HITS
Meta reported significant growth for its Llama AI models, with downloads approaching 350 million and usage increasing 10x since January.
Nous Research released the Hermes Function Calling V1 dataset for training AI models in function calling and structured output capabilities.
Nvidia and Apple reportedly discussed joining OpenAI’s funding round with Microsoft, potentially valuing the AI startup at over $100 billion.
California lawmakers approved a bill proposing sweeping AI regulations, including safety testing requirements and potential legal consequences for harmful AI systems.
Yale University announced a $150 million investment over 5 years to support AI research, development, and education across the institution.
Codeium raised $150 million in Series C funding, reaching a $1.25 billion valuation and unicorn status less than two years after launch.
Playground launched a new AI-powered graphic design tool allowing users to make logos, social media and t-shirt designs, and more for free.
THAT’S A WRAP
SPONSOR US
Get your product in front of over 650k+ AI enthusiasts
Our newsletter is read by thousands of tech executives, investors, engineers, managers, and business owners around the world. Get in touch today.
FEEDBACK
How would you rate today's newsletter?Vote below to help us improve the newsletter for you. |
If you have specific feedback or anything interesting you’d like to share, please let us know by replying to this email.
Reply