Large language model
OpenAI’s ChatGPT-4 model has been deemed ‘lazy’ by users, sparking concerns about the future of AI. Despite OpenAI’s acknowledgment of the issue and speculation about internal safety mechanisms causing the behavior, the setback presents an opportunity for researchers to advance AI development. This highlights the challenges and importance of continuous learning in AI.
The release of smaller, more efficient AI models like Mistral’s Mixtral 8x7B has sparked interest in “Mixture of Experts” (MoE) and “Sparsity.” MoE breaks models into specialized “experts,” reducing training time and enhancing speed. Sparsity involves reducing active elements in a model, leading to less computational intensity and lower storage needs. These concepts are shaping…
LLMs like NexusRaven-V2 can interpret natural language instructions to generate code snippets, including function calls, benefiting developers by providing real-time assistance and guiding correct function invocation. The open-source model outperforms GPT-4 in function calling success rates and offers utility artifacts for seamless integration into software workflows. Nexusflow aims to advance open-source models for copilots and…
Recent advancements in auto-regressive language modeling have propelled conversational AI agents to new heights. Despite the benefits of large language models, caution is advised due to potential dangers. New input-output safeguarding tools, such as Llama Guard, aim to mitigate risks and promote responsible use of generative AI models. Purple Llama project will compile resources and…
Numerical weather prediction (NWP) models have drawbacks, prompting interest in data-driven, deep learning-based weather forecasting methods. Recent advancements include Stormer, a scalable transformer model, developed by researchers from UCLA and CMU. Stormer surpasses current techniques in accuracy for medium-range weather forecasting, outperforming Pangu-Weather and Graphcast, particularly for longer lead times. (Words: 50)
Researchers from Google DeepMind, Stanford University, and University of California, Berkeley have developed Chain of Code (CoC) to enhance code-driven reasoning of language models (LMs). CoC leverages pseudocode to improve reasoning and simulation capabilities, achieving state-of-the-art performance and broader scope of problem-solving. The approach combines advantages of code and LM’s knowledge. [50 words]
French AI startup Mistral AI secured a significant €385m or $414m in funding, led by Andreessen Horowitz and Lightspeed Venture Partners. The company focuses on open-source models, aiming to counter the emerging AI oligopoly. Its new model, Mixtral 8x7B, outperformed larger open-source models like Meta’s Llama 2 34B and even rivaled proprietary models like OpenAI’s…
Google Research introduced Generative Infinite-Vocabulary Transformers (GIVT), pioneering real-valued vector sequences for AI. This approach aims to address limitations in existing transformer models for image generation by using real-valued vectors instead of discrete tokens and exploring various sampling methods. The paper’s authors highlight GIVT’s performance and emphasize their reliance on standard deep learning techniques.
A new approach to creating mesmerizing optical illusions has emerged, eschewing assumptions about human perception by using a text-to-image diffusion model. This method generates multi-view illusions, including visual anagrams, polymorphic jigsaws, and even three to four view illusions. By sidestepping traditional assumptions, it offers a fresh and accessible tool for crafting captivating visual transformations.
The evolving landscape of AI demands a shift towards human-centric design. Don Norman emphasizes aligning AI with human instincts, while ‘Design Fiction’ helps project future usages. Scientific advancements by organizations like DeepMind and Nvidia set the groundwork, and disruptive AI usages inspired by science fiction can enhance everyday lives. Collaboration between designers and AI experts…
MIT Generative AI Week featured a flagship full-day symposium and four subject-specific symposia, aiming to foster dialogue about generative artificial intelligence technologies. The events included panels, roundtable discussions, and keynote speeches, covering topics such as AI and education, health, creativity, and commerce. The week concluded with a screening of the documentary “Another Body,” followed by…
Cognitive science studies suggest typicality is vital for category knowledge, affecting human judgment. Machine learning methods offer assurance in predictions, but considering atypicality alongside confidence improves accuracy and uncertainty quantification. Recalibration techniques with atypicality-aware measures elevate performance across subgroups. Atypicality should be integrated into models for enhanced reliability in AI.
Meta AI has introduced “Relightable Gaussian Codec Avatars,” a revolutionary method for achieving high-fidelity relighting of dynamic 3D head avatars. The approach relies on a 3D Gaussian geometry model and a learnable radiance transfer appearance model to capture sub-millimeter details and enable real-time relighting. This innovation elevates the realism and interactivity of avatar animation, marking…
Brain organoids, lab-grown mini-brains created from human stem cells, have been integrated with computers to achieve speech recognition. This innovative “Brainoware” system, described in a study in Nature Electronics, represents a shift from traditional AI using silicon chips. Despite challenges, its potential for creating energy-efficient AI hardware with human brain-like functionality is evident.
A University of Warwick study unveils an AI system, X-Raydar, trained on 2.8 million chest X-rays, demonstrating comparable accuracy to doctors in diagnosing 94% of conditions. It highlights potential for efficient diagnosis, particularly in addressing radiologist shortages. X-Raydar has been open-sourced to foster further advancements in AI medical technology.
This paper introduces LiDAR, a metric designed to measure the quality of representations in Joint Embedding (JE) architectures, addressing the challenge of evaluating learned representations. JE architectures have potential for transferable data representations, but evaluating them without access to a task and dataset is difficult. LiDAR aims to facilitate efficient and reliable evaluation.
After months of negotiations, EU lawmakers have reached a deal on the groundbreaking AI Act, introducing strict rules on transparency and ethics for tech companies, creating enforcement mechanisms, and setting up fines for noncompliance. The Act covers regulations on powerful AI models, governance mechanisms, fines for noncompliance, and bans on certain AI uses.
This blog post outlines the capabilities of diffusion models for generating custom data by using additional conditioning. It introduces methods such as Stable Diffusion Inpainting, ControlNet, and GLIGEN, and highlights how fine-tuning with the Low-Rank Optimization technique, or LoRA, can efficiently adapt these methods to specific use cases. The article emphasizes the benefits of enhancing…
The Anticipatory Music Transformer, developed by Stanford scholars, empowers composers with unique control over generative AI music composition. Differentiating itself from other tools, it focuses on symbolic music and incorporates users’ preferences. Integrated with the GPT architecture, it offers more interactive and controllable outputs. Anticipated to revolutionize music composition, it aims to make music creation…
The introduction of Large Language Models (LLMs) has brought attention to Natural Language Processing, Natural Language Generation, and Computer Vision. Researchers from Tsinghua University and Microsoft Research Asia introduced Bridge-TTS, an alternative to noisy prior models, achieving better TTS synthesis than Grad-TTS and FastGrad-TTS while demonstrating improved speed and generation quality. Find out more at…