Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Diffusion models break down hard problems, like image generation, into many smaller problems, like removing a tiny amount of noise from an image. A lot of effort has been put into single step diffusion generation, which seems to defeat the point. This post explores the paradox of diffusion distillation and outlines the many research directions available.
Instructions on how to add your own models to the Ollama repository for use.
Sponsored
Future-proof your career with AI. Attend INTERVIEW KICKSTART'S SERIES ON GENAI & MLOPS PREMIUM COURSES [19], March 3, 5:30 PM PST, with Sarfaraz Hussein (Ph.D.), Sr.ML Scientist.
Speeding up inference can open up use cases. This code suggests a way to speed up inference with Llama models on Hugging Face.
Drive-WM can generate high-quality multiview videos to predict future scenarios, enabling autonomous vehicles to make safer and more efficient driving decisions.
Get YouTube video summaries in a minute using AI.
Sponsored
Ready to save up to $2,000 on every vacation for life? Dollar Flight Club has your back—TRY THE CLUB FOR JUST $1 IF YOU SIGN UP IN THE NEXT 12 HOURS [5]. That means; Hawaii from $174, Iceland from $215, or Greece from $245 round trip. Join over 1 million members saving on travel this year.
Meta's Llama team is reportedly still working on alignment to enable the model to loosen up and answer more queries that aren’t actually controversial.
Most serverless GPU offerings require one POST request per model inference. However, this is a challenge in cases where you want to stream model outputs. Websockets make streaming possible.
Apple CEO Tim Cook announced at the company's annual shareholders meeting that the company plans to break new ground in generative AI in 2024, promising transformative opportunities for users.
An amazing paper that frames video as a potentially better way for AI to interact with human users. It shows how video models can be used as planners, agents, compute engines, and environment simulators.
Most language models use 16 or more bits to represent a parameter in a model. This results in powerful models that can be expensive to run. This work proposes a method that uses 1.58 bits per parameter - each parameter is in {-1, 0, 1}. This method matches performance exactly up to 3B parameters. No code or models have been released yet.
This research focuses on advancing multi-modality foundation models like GPT-4V in low-level visual perception tasks. The large-scale experiment gathered feedback on 18,973 images from 58,000 humans and created the Q-Pathway dataset to analyze clarity, color, and brightness.
Sponsored
What separates top-performing enterprises from laggards when it comes to GenAI adoption? Join this FREE virtual summit [12] to learn about practical hands on architecture and insights from GenAI builders, the emerging enterprise GenAI stack [12], and much more. Featuring speakers from LLAMAINDEX, COINBASE, LINKEDIN, COMCAST, P&G, ROBLOX, DATABRICKS, JPMORGAN CHASE, FIDELITY, CHEGG, and many others.
Graph diffusion policy optimization enhances graph generation models for drug design using a unique reinforcement learning technique. This approach promises better performance in creating complex and diverse graph structures and could set new standards in the field.
Dual Chunk Attention (DCA) extends the capabilities of large language models like Llama2 70B, allowing them to handle over 100k tokens without additional training. It breaks down attention computation into chunks, enhancing the model's understanding of both short and long-range context.
The Big Code project has released another version of its flagship coding model. The model has a 16k context window and was trained on 4T tokens. Its performance is strong, but still under that of DeepSeek coder.
HiGPT is an approach to learning across diverse heterogeneous graphs without the need for fine-tuning. Its integration with a novel graph tokenizer and a vast corpus of graph instructions allows it to excel in adapting to various data distributions.
PromptMM improves recommendation systems on platforms like Amazon and TikTok by using Multi-modal Knowledge Distillation. It addresses inaccuracies in user preferences and simplifies systems by distilling important features from various content types (visual, textual, or acoustic) to prevent overfitting.
A minimal code set up with TRL to tune a model to be more descriptive.
A coding style at Character AI dramatically improves the readability of shapes in tensors.
Adobe Acrobat's new AI Assistant can query documents for answers with attribution, create document summaries, offer a list of suggested questions for deeper insights, assist in drafting content for various communication needs, and provide outlines and key points so users can swiftly navigate essential document information.
Sponsored
The in-person GTC experience is back. Come connect with a dream team of industry luminaries [5], developers, researchers, and business strategists helping shape what’s next in AI and accelerated computing. From the highly anticipated keynote by NVIDIA CEO Jensen Huang [5] to 900+ inspiring sessions, 300+ exhibits, and tons of unique networking events, GTC delivers something for every technical level and interest area.
After a decade of effort and billions of dollars, Apple is canceling its unsuccessful attempt at building a car to rival Tesla. Many of the employees are expected to move to generative AI projects.
Function calling is when a language model can access tools via synthesized functions to complete its given instructions. This requires proper synthesis of arguments to pass to those functions. This leaderboard is designed to assess model capability on function calling tasks.
Music understanding typically relies on adding a modality-specific encoder to a language model. This is expensive and unstable. This work found that music understanding increased substantially without impact to base language ability when music was tokenized into ABC notation.
Bytedance has produced a system called MegaScale that can be used to train massively parallel large language models. It succeeded in training a 175B LLM on 12,288 GPUs with 55.2% Model FLOPs Utilization (MFU), which is extremely impressive. Bytedance plans to open source some aspects of the codebase.
ListT5 introduces a cutting-edge reranking method that not only improves the precision of information retrieval but also offers an efficient solution to the challenges faced by previous listwise rerankers.
EMIFF is an innovative camera-based 3D detection framework for vehicle-infrastructure cooperative object detection. It uses Multi-scale Cross Attention and Camera-aware Channel Masking to correct pose errors from camera asynchrony.
FuseChat introduces an innovative way to blend the strengths of various large language models into a single, more powerful model without the high costs of training from scratch.
ShieldLM is a safety detector for large language models that aligns with human safety standards, offers customizable detection, and explains its decisions.
This article reports on interviews with founders from companies at the forefront of AI, covering their views on the advent of AGI, how to think about LLMs, and simple tactics for founders on how to approach AI within their products.
MobiLlama is a fully open-source, transparent, small language model for use on mobile devices. Its performance is excellent for a small 0.5B parameter model - it even outperforms some larger models.
IR-QLoRA is a new approach that enhances the accuracy of quantized large language models, making them more suitable for use on devices with limited resources.
Sponsored
Raycast Pro lets you harness the power of genAI, add custom themes,
Researchers from UCSD trained humanoid robot policies to be more expressive, socially inclined, and robust. Their videos of non-choreographed dancing on grass are very impressive.
Simulatrex is an open-source project focused on Generative Agent-Based Modeling (GABM). It utilizes large language models for more accurate simulations.
14,509 books summarized in 5 bullet points using AI.
Mistral, a French AI start-up, has announced a new model for its API along with an Azure partnership. Mistral Large excels at multilingual tasks, is within a few percent of GPT-4 on MMLU, and generally has powerful performance. The startup also added a new mistral-small model to its API.
Scale AI has received a one-year contract to develop a testing and evaluation framework for the Pentagon's large language AI models. The company aims to bolster safe AI deployment and robustness within military applications. The project involves creating specialized datasets for model assessment and an iterative process to refine AI systems for secure environments. This collaboration will help the Department of Defense understand and responsibly deploy generative AI technology.
Meta has proposed a 350m parameter language model with strong reasoning performance, even approaching Llama 7B on correctness in API function calling tasks. The model has not been released, but the innovation in fixed parameter models is worth exploring.
ConceptMath is a new bilingual benchmark for testing LLMs’ math skills in both English and Chinese. It's unique because it breaks down math problems into specific concepts, allowing for a more detailed assessment of an AI's strengths and weaknesses in math.