Models & Research
-
Study Warns Against Relying on Junior Staff for AI Training
Harvard, MIT, and Wharton research found junior staff’s AI risk tactics contradict expert advice, demanding top-down AI governance.
-
StoryDiffusion: Consistent Image and Video Generation
StoryDiffusion introduces Consistent Self-Attention for character and attire consistency; Semantic Motion Predictor for stable video transitions; Zero-Shot adaptability.
-
LLAMA-NAS: Efficient Neural Architecture Search
LLAMA-NAS creates smaller, efficient LLMs, achieving 1.5x size reduction, 1.3x speedup with minimal accuracy loss, and broad applications.
-
Cohere Open-Sources Aya 23 Multilingual LLMs
Cohere introduces the Aya 23 LLM series, supporting 23 languages, tailored for enterprise needs, and featuring improved text processing.
-
Google Launches Efficient Vision-Language Model PaliGemma and high-performance Gemma 2 series
Google introduces PaliGemma, a small language model (SLM) for efficient vision-language tasks, and announces the high-performance Gemma 2 series.
-
OpenAI Unveils Multimodal GPT-4o and ChatGPT Desktop App
OpenAI introduces GPT-4o, an advanced model integrating text, vision, and audio inputs, with a new desktop app enhancing accessibility.
-
UAE Unveils Falcon AI to Rival Meta and OpenAI
The UAE introduces Falcon AI, a competitor to Meta and OpenAI, with advanced capabilities, disrupting the AI market globally.
-
Fine-Tuning LLMs Increases Hallucination Risk
The study reveals that fine-tuning large language models with new knowledge increases hallucination risk. Mitigation strategies and impact discussed.
-
Microsoft’s MAI-1: Bridging AI Power and Efficiency
Microsoft is developing MAI-1, a 500 billion parameter language model for improved efficiency and cost benefits, boosting AI accessibility.
-
Meta’s Breakthrough in AI Speed: Multi-Token Prediction
Researchers from Meta and French universities have developed multi-token prediction, tripling AI speed, and effectiveness in large language models.
Eksentricity 