Today:

• AI's Influence on Society

• Megalodon LLM Transformer Architecture

• SK Hynix, TSMC AI Chip

• Hugging Face Healthcare AI 

• Groq's AI Chip Sets New Performance Standard

INSANE AI VIDEO | AI Girlfriends a BILLION $ Industry | EndlessDreams, VASA-1 and AI Music

AI advancements raise concerns about safety at work due to increasingly persuasive capabilities. Large language models like Cloud 3 produce convincing outputs, sometimes indistinguishable from humans.

Despite concerns, AI's impact on various aspects of life continues to expand, prompting reflection on its role in shaping society.

WATCH THE VIDEO ON YOUTUBE

Meta challenges transformer architecture with Megalodon LLM

Meta and USC researchers introduce Megalodon, a machine learning model addressing Transformer architecture challenges. Megalodon enhances context windows for large language models (LLMs) without excessive memory usage. It surpasses Transformer models in processing extensive texts by leveraging Moving Average Equipped Gated Attention (MEGA) techniques, reducing computational complexity. Chunk-wise attention divides input sequences, enhancing parallelism and training speed. Megalodon-7B outperforms LLAMA-2-7B and matches LLAMA-2-13B on various tasks. 

Despite emerging alternatives like Mamba and liquid neural networks, Transformers remain dominant, backed by robust toolkits. Researchers aim to optimize Transformers and explore new architectures, envisioning potential successors to the Transformer era.

VENTUREBEAT

SK Hynix partners with TSMC to develop advanced AI chips

SK Hynix and TSMC are teaming up to develop cutting-edge AI chips, aiming to seize opportunities in the booming AI market. SK Hynix, the world's second-largest memory chipmaker, and TSMC, the largest contract chip manufacturer, will collaborate on next-gen high bandwidth memory (HBM) chips. 

This partnership reflects the ongoing race among global chipmakers to capitalize on AI advancements. The move signals a strategic effort by Asian chipmakers to solidify their positions in this rapidly expanding sector.

NIKKEI

Hugging Face releases a benchmark for testing generative AI on health tasks

Hugging Face introduces Open Medical-LLM, a benchmark for evaluating generative AI models in healthcare tasks. Developed with researchers from Open Life Science AI and the University of Edinburgh, it standardizes model performance assessment on medical tasks. 

Despite its value, medical experts warn against overreliance on benchmarks like Open Medical-LLM, emphasizing the need for real-world testing to ensure safe and effective deployment of AI in healthcare. The benchmark aims to drive advancements in the field and contribute to improved patient care.

TECHCRUNCH

Groq’s breakthrough AI chip achieves blistering 800 tokens per second on Meta’s LLaMA 3

Groq's AI chip showcases remarkable performance, achieving over 800 tokens per second on Meta's LLaMA 3 model, signaling a potential disruption in the AI inference landscape. The chip's Tensor Streaming Processor, optimized for deep learning operations, promises superior efficiency and speed compared to conventional GPUs. 

This breakthrough challenges Nvidia's dominance and highlights the growing importance of fast and energy-efficient AI inference for applications like chatbots and virtual assistants. Groq's success in delivering high-performance inference hardware could democratize access to advanced AI capabilities across various industries.

VENTUREBEAT

🧠RESEARCH

EdgeFusion: On-Device Text-to-Image Generation

"EdgeFusion: On-Device Text-to-Image Generation" addresses the computational challenges of Stable Diffusion (SD) in creating images from text. The authors introduce BK-SDM, a streamlined variant, and propose two novel strategies: leveraging high-quality data and refining the distillation process. Their approach achieves rapid, realistic image generation on resource-constrained devices.

TriForce: Lossless Acceleration of Long Sequence Generation with Hierarchical Speculative Decoding

"TriForce: Lossless Acceleration of Long Sequence Generation with Hierarchical Speculative Decoding" tackles the challenge of efficient long-sequence generation with large language models. By introducing hierarchical speculative decoding, TriForce reduces drafting latency by leveraging dynamic sparse key-value cache and smaller models, achieving impressive speedups without compromising generation quality. The system demonstrates scalability and robustness across different configurations.

MoA: Mixture-of-Attention for Subject-Context Disentanglement in Personalized Image Generation

"Mixture-of-Attention (MoA)" enhances personalized image generation in text-to-image diffusion models. Inspired by large language models, it balances workload between personalized and generic attention pathways, preserving the model's original capability while embedding subjects within context. MoA's routing mechanism optimizes content blend, enabling diverse compositions with disentangled subject-context control. 

Reuse Your Rewards: Reward Model Transfer for Zero-Shot Cross-Lingual Alignment

The paper explores aligning language models (LMs) across languages without abundant human preference data. It proposes using a reward model trained on one language's data and applying it to others. Results indicate success, with cross-lingual models preferred by humans in over 70% of cases. It also suggests using different-language reward models for better alignment.

OpenBezoar: Small, Cost-Effective and Open Models Trained on Mixes of Instruction Data

The paper introduces OpenBezoar models, smaller yet powerful language models (LLMs) fine-tuned for various tasks using instruction data. Utilizing a cost-effective approach, they employ synthetic data and filtering techniques, followed by sequential fine-tuning and optimization methods. Results show superior performance, even surpassing top models in certain benchmarks. 

🛠️TOP TOOLS

Private LLM - Local, Private AI Offline Chatbot for iOS & macOS

Driver AI - Explains millions of lines of code in minutes instead of months

Potion - Personalized AI videos

BoolVideo - Transform product URL, blog URL, images, videos or text into captivating videos with dynamic AI voices and great audio-visual treat

Infinity AI - Type a script. Get a movie out.

📲SOCIAL MEDIA

🗞️MORE NEWS

Meta AI is restricting election-related responses in India

Meta AI is testing its chatbot in India, blocking certain election-related queries amid concerns about misinformation. The move aligns with Meta's efforts to regulate content around elections globally. While not blocking party names, it redirects queries about specific politicians to the Election Commission's website. Updates are ongoing to enhance the AI's accuracy. TECHCRUNCH

Intel unveils largest-ever AI 'neuromorphic computer' that mimics the human brain

Intel has developed Hala Point, the world's largest neuromorphic computer, aiming to advance artificial intelligence (AI) research. Powered by over 1,000 new AI chips, it operates 50 times faster and consumes 100 times less energy than traditional systems. Hala Point, equipped with Intel's Loihi 2 processors, mimics the brain's structure and processes data differently, offering potential breakthroughs in AI efficiency and performance. LIVESCIENCE

Olympic organizers unveil strategy for using artificial intelligence in sports

The International Olympic Committee (IOC) has revealed its strategy to incorporate artificial intelligence (AI) in sports, aiming to enhance athlete identification, personalize training methods, and improve judging fairness. The IOC's plan includes using AI to address various aspects of sports, such as athlete development, training optimization, and broadcast enhancement. With the Paris Olympics approaching, the IOC is set to implement some AI projects during the games, demonstrating its commitment to embracing technology while ensuring its responsible use. AP NEWS

Langdock raises $3M with General Catalyst to help companies avoid vendor lock-in with LLMs

Langdock, a German startup, secures $3M led by General Catalyst to bridge companies and large language models (LLMs). Their chat interface allows firms to select LLMs without vendor lock-in, appealing to EU companies seeking regulatory compliance. Langdock competes with Dust, offering broader use cases and attracting notable investors. TECHCRUNCH

Google squashes AI teams together in push for fresh models

Google merges AI teams under DeepMind to enhance model development, addressing past mishaps. Responsible AI team joins, emphasizing early intervention. Standardizing launch criteria and red team testing aim to prevent inaccuracies and ensure user safety. Hardware teams also combine for efficiency. CEO warns against personal disruptions, referencing recent firings and ethical concerns in AI. THE REGISTER

Meta's newest AI-powered chatbots show off impressive features and bizarre behavior

Meta Platforms introduces advanced AI chatbots, touted by CEO Mark Zuckerberg as the most intelligent assistants available. However, initial interactions reveal quirks and limitations typical of generative AI. Meta AI, accessible on Facebook, Instagram, and WhatsApp, generates text and images, but its behavior can be unpredictable. Despite ongoing improvements, concerns persist over accuracy and appropriateness in social settings. CBS NEWS

How to Use Meta AI on Facebook, WhatsApp, Instagram and the Web

Meta's recent AI developments include the introduction of a web chatbot akin to ChatGPT, expansion of AI across its platforms, and the rollout of Llama 3, its latest model. Users can access Meta AI directly within Facebook, Messenger, Instagram, and WhatsApp by simply searching or typing "@meta AI" in group chats. Currently available in select countries, Meta AI aims to enhance user experience and enable real-time image generation. Despite initial blunders, such as mistaken responses in private groups, Meta continues to refine its AI offerings. TECH.CO

A coffee roastery in Finland has launched an AI-generated blend. The results were surprising

A Finnish coffee roastery introduced an AI-created blend, aiming to modernize the traditional craft. The "AI-conic" mix, dominated by Brazilian beans, resulted from collaboration with an AI consultancy. Despite initial skepticism, blind testing confirmed the blend's excellence, highlighting AI's potential to innovate in coffee production. AP NEWS