Insane AI video generator, Stable Diffusion 3 is out, AI learns language by itself, & more!
Welcome to the AI Search newsletter. Here are the top highlights in AI this week.
Breakthrough Technique Enhances Reasoning Abilities of Large Language Models
Researchers have developed a new technique called Natural Language Embedded Programs (NLEPs) that enables large language models to solve complex reasoning tasks with higher accuracy. NLEPs involve prompting a language model to generate a Python program to solve a user's query, and then output the solution as natural language. This approach achieves higher accuracy on a range of symbolic reasoning tasks, such as tracking shuffled objects or playing a game of 24, and even outperforms task-specific prompting methods. The technique also improves transparency, as users can inspect the program to understand how the model reasoned about the query.
New Sora-level Video Generator Available Now
Luma AI has launched its Dream Machine, a browser-based AI video generator that allows users to create stunning videos from text prompts, without any waiting list or restrictions. This innovative tool can produce high-quality, 5-second videos in just two minutes, with impressive image quality and coherent motion. Although still in its early stages, Dream Machine has the potential to revolutionize the video creation process.
New Technique Enables AI to Better Map 3D Space Using 2D Cameras
Researchers have developed a new technique called Multi-View Attentive Contextualization (MvACon) that improves the ability of artificial intelligence (AI) programs to map 3D spaces using 2D images captured by multiple cameras. MvACon is a plug-and-play supplement that can be used with existing vision transformer AI systems, allowing them to better identify objects, track their speed and orientation, and create a more accurate representation of 3D space.
ChatLLM by Abacus
ChatLLM by Abacus AI is an integrated platform that allows enterprises to use multiple LLMs, deploy custom agents, and collaborate with team members. Choose from state-of-the-art LLMs such as GPT-4o, Claude 3 Opus, and their new open source Smaug.
Researchers Use AI to Autonomously Identify Zero-Day Security Flaws
Scientists from the University of Illinois Urbana-Champaign has successfully utilized a hierarchical planning with task-specific agents (HPTSA) method with GPT-4 to autonomously identify zero-day security flaws. By leveraging multiple instances of a modified GPT-4 as agents, the HPTSA method significantly improved the efficiency of finding vulnerabilities, achieving a 550% increase in efficiency compared to other real-world applications.
Stable Diffusion 3 Now Available
Stable Diffusion 3 has been released! The model can be used for free for non-commercial projects, while commercial projects with less than $1 million in annual revenue or fewer than 1 million users can subscribe for $20/month. However, it’s outputs have not been promising. Here’s a full review:
ALOGIC
ALOGIC offers a wide range of professional home/office hardware, including their Clarity series of professional 4K touchscreen monitors, wireless chargers, docking stations, and more. Enhance your productivity and elevate your digital lifestyle with ALOGIC!
New AI Algorithm Learns Language by Watching Videos
Researchers have developed an AI algorithm called DenseAV that can learn human language by watching videos, without any prior knowledge of written language. The algorithm uses a "contrastive learning" approach, comparing audio and visual signals to identify matching patterns and learn the meaning of words and phrases. Trained on 2 million YouTube videos, DenseAV has demonstrated the ability to understand the connection between words and objects, and even distinguish between similar sounds, such as a dog's bark and the word "dog".
Open-Source Robot Model for Versatile Object Manipulation
Researchers have developed an open-source generalist model for robot object manipulation, called Octo. This model can control various types of robots and enable them to perform different tasks, such as picking up objects, closing drawers, and wiping tables. Octo is trained on a large dataset of robotic manipulation trajectories and can process diverse sensory inputs, including images, robot joint readings, and language instructions. The model has been tested on nine different robotic systems and has shown promising results, demonstrating its flexibility and generalizability.
Apple and OpenAI Not Paying Each Other
According to a Bloomberg report, Apple and OpenAI have not exchanged payments as part of their partnership to integrate ChatGPT into Apple devices. Apple believes the exposure OpenAI will receive from this partnership is valuable enough, and is exploring future revenue-sharing deals. The partnership is not exclusive, and Apple is in talks with other AI providers, including Anthropic and Google. Apple plans to offer users a range of third-party AI services, similar to how Safari supports different search engines.