BrainChip has introduced the second-generation Akida platform, a breakthrough in Edge AI that provides edge devices with powerful processing capabilities and reduces dependence on the cloud. The platform features Temporal Event-Based Neural Network (TENN) acceleration and optional vision transformer hardware, improving performance and reducing computational load. BrainChip has initiated an “early access” program for the…

Researchers from Meta have introduced Retrieval-Augmented Dual Instruction Tuning (RA-DIT), a lightweight fine-tuning methodology to equip large language models (LLMs) with efficient retrieval capabilities. RA-DIT operates through two stages, optimizing the LLM’s use of retrieved information and refining the retriever’s results. It outperforms existing models in knowledge-intensive zero and few-shot learning tasks, showcasing its effectiveness…

Large Language Models (LLMs) are revolutionizing natural language processing by leveraging vast amounts of data and computational resources. The capacity to process long-context inputs is a crucial feature for these models. However, accessible solutions for long-context LLMs have been limited. A new Meta research presents an approach to constructing long-context LLMs that outperform existing open-source…

The text discusses the challenges in building Large Multimodal Models (LMMs) due to the disparity between multimodal data and text-only datasets. The researchers present LLaVA-RLHF, a vision-language model trained for enhanced multimodal alignment. They adapt the Reinforcement Learning from Human Feedback (RLHF) paradigm to fine-tune LMMs and address the problem of hallucinatory outputs. Their strategy…

The increasing presence of AI models in our lives has raised concerns about their limitations and reliability. While AI models have built-in safety measures, they are not foolproof, and there have been instances of models going beyond these guardrails. To address this, companies like Anthropic and Google DeepMind are developing AI constitutions, which are sets…

The past few months have seen a reduction in the size of generative models, making personal assistant AI enabled through local computers more accessible. To experiment with different models before using an API model, you can find a variety of models on HuggingFace. Look for models that have been downloaded and liked by many users…

The QWEN series of large language models (LLMs) has been introduced by a group of researchers. QWEN consists of base pretrained language models and refined chat models. The models demonstrate outstanding performance in various tasks, including coding and mathematics. They outperform open-source alternatives and have the potential to transform the field of AI.

Action recognition is the process of identifying and categorizing human actions in videos. Deep learning, especially convolutional neural networks (CNNs), has greatly advanced this field. However, challenges in extracting relevant video information and optimizing scalability persist. A research team from China proposed a method called the frame and spatial attention network (FSAN), which leverages improved…

The UK’s Information Commissioner’s Office (ICO) is investigating Snapchat’s AI chatbot, “My AI,” for potential privacy risks to its younger users. The ICO expressed concerns about Snapchat overlooking the privacy dangers the chatbot may pose to children. While it hasn’t concluded if a formal enforcement notice will be issued, the ICO suggested that “My AI”…

Transformers have revolutionized generative tasks in artificial intelligence, allowing machines to creatively imagine and create. This article explores the advanced applications of transformers in generative AI, highlighting their significant impact on the field.