- Starter AI
- Posts
- Discover YOLO-world, DeepMind’s reasoning trick, Introducing: Large World Models
Discover YOLO-world, DeepMind’s reasoning trick, Introducing: Large World Models
Maintaining order is key even for LLMs.
Sthefania
February 26, 2024
Hello, Starters!
As we embrace another Monday, let's shake off the weekend vibes and gear up for a fresh start. We've got some interesting AI models to cover. Get ready!
Here’s what you’ll find today:
YOLO-world: An object detection model
DeepMind's trick for better language model reasoning
Large World Model: A multimodal marvel
Humane faces shipping delays
Gemini lands on Messages
And more.
YOLO-World is an advanced open-vocabulary object detection model that harnesses lightweight detectors from the YOLO series, well known for their effectiveness in object detection, to enhance its performance, reaching unprecedented speeds.
In other words, these capabilities allow YOLO-World to recognize objects in a way that's not restricted to a specific set of categories.
Contrary to traditional detectors, this model introduces a "prompt-then-detect" paradigm that allows it to understand prompts without the need for constant training, by encoding them into an offline vocabulary, providing quicker results without sacrificing efficiency.
As part of its ongoing efforts in AI research, Google DeepMind has presented a technique that addresses the challenges language models face in logical reasoning. According to the study, the order in which the premises are presented directly impacts logical reasoning performance.
Through the use of deductive reasoning, specifically the "modus ponens," which is very easy for humans – it explains that if you have two statements and one is true, the other one must be true too – researchers found that changing the order of the information confused the models, decreasing their accuracy by more than 30%.
📄 Large World Model: A multimodal marvel (3 min)
The Large World Model (LWM) is a general-purpose large-context multimodal autoregressive model. This means that it is capable of understanding and working with different types of information in a broader context, generating predictions one step at a time. It has been trained on a large dataset of videos and books using Ring Attention.
Regular language models sometimes struggle with complicated or long-form queries. LWM stands out for its scalable training, allowing it to better handle complex tasks. For example, it can easily answer questions about an hour-long YouTube video.
📦Humane has readjusted the shipping timeframe for its long-awaited AI pin. The gadget, which has made waves across the industry, was previously slated for March. Now, the company has stated that customers with "priority access" can expect mid-April delivery.
💬The Gemini takeover doesn't stop, and now it's Android's turn, with Google announcing new updates that include the integration of Gemini into Messages. This allows users to chat, draft messages, and even schedule events without leaving the texting app.
⚡️Quick links
What did you think of today's newsletter? |