New brain-inspired circuits • Training robots with audio • AI is funnier than humans • AI news this week
Welcome to the AI Search newsletter. Here are the top highlights in AI this week.
Advances in AI technology for improved object detection and classification
Researchers have developed a novel technology called stability diffusion-based deep generative replay (SDDGR) that enables AI to learn new information while preserving existing knowledge. This technology has transformed various aspects of daily life, making it an essential tool in smart home appliances, robotics, and medical fields, particularly in self-driving cars, enabling them to accurately recognize road objects and drive safely. SDDGR offers economic benefits by reducing data storage and processing costs through efficient data reuse, and is expected to yield significant economic benefits for businesses.
The Insane Progress of AI Video Generation
Runway Gen-3 has been fully released, and users have started to share some professional quality productions, including MV’s and short films. See a full breakdown here:
Study Employs AI to Determine Battery Composition and Conditions
An international collaborative research team has developed an image recognition technology to accurately determine the elemental composition and the number of charge and discharge cycles of a battery by examining only its surface morphology. The researchers trained the CNN-based AI to learn the surface images of battery materials, enabling it to predict the major elemental composition and charge-discharge cycle states of the cathode materials with 99.6% accuracy.
Salesforce Releases Smaller Models that Beat GPT-4o
Salesforce AI Research has introduced APIGen, a pipeline that automates the generation of high-quality datasets for function-calling applications. Using APIGen, Salesforce trained two large language models (1.3B and 6.7B parameters) that achieved impressive results on the Berkeley Function-Calling Leaderboard. The 6.7B model ranked 6th, outperforming models like GPT-4o and Gemini-1.5-Pro, while the 1.3B model surpassed GPT-3.5-Turbo and Claude-3 Haiku.
TurboType - Type Faster & Save Time
Turbotype is a Chrome extension that allows you to type faster by setting customizable shortcuts, designed to boost productivity and save time. Easily create, save, and insert keyboard shortcuts for frequently used phrases. Use it for free, forever!
Survey shows most people think LLMs such as ChatGPT can experience feelings and memories
A recent survey reveals that two-thirds of people believe that large language models (LLMs) like ChatGPT possess consciousness and can have subjective experiences such as feelings and memories. The study, conducted by the University of Waterloo, found that people who use ChatGPT more frequently are more likely to attribute consciousness to it, suggesting that the conversational style of LLMs may lead people to think they have a mind. This has implications for how people interact with AI tools, potentially leading to increased trust and social bonding, but also raises concerns about excessive trust and over-reliance on AI.
Neuromorphic dendritic network computation with silent synapses for visual motion perception
Researchers have introduced a new brain-inspired artificial dendritic neural circuit that mimics the organization and functions of the human brain. The circuit, dubbed "dendristor," replicates the morphology and function of biological dendrites, leveraging the physics of multi-gate transistors coated with an ion-doped sol-gel film to modulate the transistor's current and reflect changes in dendritic membrane potential. The new approach demonstrates remarkable energy efficiencies and the potential to detect motion utilizing fewer neurons than existing artificial neural networks.
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.
Moshi: Have a Free Realtime Conversation with AI
Moshi, a real-time speech-to-speech model developed by Kyutai, a non-profit lab in Paris, can listen and respond continuously without needing to model speaker turns or interruptions. It can engage in small talk, explain concepts, and roleplay in various emotions and speaking styles. The model can be tested online for free, and its code and model weights will be released soon.
ManiWAV: Learning Robot Manipulation from In-the-Wild Audio-Visual Data
Researchers have found that incorporating audio data into the training process of AI-based robots significantly improves their learning capabilities. In a series of experiments, the team demonstrated that adding audio to visual data enables robots to better understand and execute tasks, such as turning over a bagel in a frying pan, erasing an image on a whiteboard, and pouring dice from one cup to another. The study suggests that audio data can provide valuable cues for robots, enhancing their ability to interpret and respond to their environment.
Think you're funny? ChatGPT might be funnier
A study comparing jokes written by humans and those generated by ChatGPT found that the AI-powered tool was able to produce funnier jokes than humans in many cases. The researchers asked participants to rate the humor of jokes written by both humans and ChatGPT, and found that nearly 70% of participants rated the ChatGPT-generated jokes as funnier. The study also found that ChatGPT was able to generate jokes that were on par with those written by professional humor writers, raising concerns about the potential impact of AI on the entertainment industry.