Search the archive

EXPERIMENTAL GPU META PROGRAMMING

Image processing and cropping that can be run on the GPU.

3D SCENE UNDERSTANDING

Researchers conducted a study comparing seven different visual encoding models for 3D scene understanding and found that the best model depends on the task. DINOv2 performed the best overall. Video models excelled in object-level tasks and diffusion models were best for geometric tasks. Language-pretrained models had surprising limitations.

HANDWRITING GENERATION WITH A SINGLE SAMPLE

A One-DM model that generates handwritten text mimicking any style using just a single sample.

DIFFUSION IS SPECTRAL AUTOREGRESSION

Diffusion models and autoregressive models are fundamentally similar, both using iterative refinement processes. Using Fourier transform techniques, this author shows how diffusion models operate like approximate autoregression in the frequency domain, particularly for visual data. This insight highlights potential directions for unifying generative modeling approaches across different data types.

WHY WE FEAR DIVERSE INTELLIGENCE LIKE AI

The rise of AI and diverse intelligences challenges outdated distinctions between "real beings" and machines. Instead of focusing solely on AI, people should understand and ethically engage with diverse cognitive systems, including bioengineered, robotic, and hybrid entities. Expanding our definitions of intelligence and compassion will help navigate the ethical landscape of emerging technologies.

by Specialist Staffing Group

THE NEW SHORTWAVE AI ASSISTANT

Shortwave has significantly upgraded its AI Assistant, enabling it to perform complex, multi-step tasks such as advanced searches, calendar lookups, and email analysis.

OPENAI MIGHT USE APPLE'S TSMC FOR CHIPS

OpenAI can significantly reduce operational costs with the adoption of more efficient chips, especially as its user base grows.

APPLE TAKES DIRECT AIM AT MICROSOFT'S COPILOT+ PCS IN NEW AI-FOCUSED MAC PROMOS

Apple is heavily promoting the Mac as the "best AI PC" and directly targeting Microsoft's Copilot+ PCs.

INTEL REVEALS FIRST LUNAR LAKE LAPTOP CPUS: EVERYTHING YOU NEED TO KNOW

Intel has unveiled its Core Ultra 200V lineup, previously known as Lunar Lake, boasting superior AI performance, fast CPUs, and competitive integrated GPUs for thin laptops. The processors feature eight CPU cores, integrated memory, and enhanced efficiency but are limited to 32GB RAM. Major manufacturers like Acer, Asus, Dell, and HP will launch laptops with these new chips. Reviews are pending to confirm Intel's claims.

OPENAI, STILL HAUNTED BY ITS CHAOTIC PAST, IS TRYING TO GROW UP

OpenAI is restructuring its management and organization to attract major investors like Microsoft, Apple, and Nvidia while aiming for a $100 billion valuation. The company faces internal conflicts about its mission and safety practices, leading to significant staff turnover, including key researchers joining rivals like Anthropic. Despite growing revenues and user base, OpenAI grapples with balancing profit motives and ethical concerns in advancing AI technologies.

CANVA SAYS ITS AI FEATURES ARE WORTH THE 300 PERCENT PRICE INCREASE

Canva is increasing prices for its Teams subscriptions by over 300% due to new generative AI features. Users report prices rising from $120 to $500 annually, though a discount brings it down to $300 for the first year. These changes reflect the value added by AI tools like Magic Media and Magic Expand.

DEPTHCRAFTER: CONSISTENT LONG DEPTH SEQUENCES FOR OPEN WORLD VIDEOS

By using Stable Diffusion video as a prior model, researchers were able to generate extremely precise depth information without using any camera pose or optical flow information.

MODIFYING LLMS FOR DIRECTED CHEMICAL SPACE EXPLORATION

By using supervised fine-tuning and DPO style data on open source language models, LLMs can be prompted to generate molecules that have interesting properties for drug development.

PLANNING IN NATURAL LANGUAGE IMPROVES LLM SEARCH FOR CODE GENERATION

One of the challenges of code generation is getting diverse candidate solutions. Oftentimes, even repeated sampling fails to generate enough novelty to solve a problem. However, if you plan in natural language first and brainstorm candidate solution paths, then generation is significantly more diverse and varied leading to improved code generation solutions.

guide

RUN AN LLM ON ESP32

This code shows how to run a tiny language model on an Arduino board.

DOCAI: STRUCTURED INFORMATION EXTRACTION FROM DOCUMENTS

Another example of cleverly using existing models to extract structured information from documents.

FLUXMUSIC

Text-to-music generation with a rectified flow transformer.

A PYTHON PACKAGE TO BUILD KNOWLEDGE GRAPHS

iText2KG is a Python package that uses large language models to extract entities and relationships from text, incrementally building consistent knowledge graphs.

WILL AI MAKE US OVERCONFIDENT?

Students are increasingly using AI tools like ChatGPT to solve challenging research problems, surprising educators with their rapid progress. AI-enhanced development tools, especially in coding, significantly boost ambition and productivity but come with risks of overconfidence and errors. Despite inaccuracies, AI provides valuable interactive entry points to complex tasks, potentially encouraging more proactive learning and cross-disciplinary exploration.

LLMS STRUGGLE TO EXPLAIN THEMSELVES

An interactive demo was used to test LLMs' ability to recognize and explain number sequences generated by random programs. The results showed that while LLMs often identify the correct sequence, their explanations for the patterns were frequently incorrect. This highlights the current limitations in LLMs' reasoning abilities despite their performance on standardized tests.

COLPALI AND QWEN2-VL FOR RETRIEVAL (GITHUB NOTEBOOK)

Merve has created a great resource for using language and vision models to improve retrieval.

TEXT TO X

Well curated list of “text to x” papers and projects.

COPILOT+ FEATURES ARE COMING IN NOVEMBER TO AI PCS POWERED BY INTEL AND AMD'S LATEST CHIPS

Microsoft announced that Intel's 200V processors and AMD's Ryzen AI 300 series chipset will support Copilot+ AI features starting in November.

algolia.com

OPENAI CONSIDERS $2,000 MONTHLY SUBSCRIPTION PRICES FOR NEW LLMS

OpenAI is reportedly considering subscription prices as high as $2,000 per month for the company's upcoming large language models, like Strawberry and Orion.

GOOGLE'S AI-POWERED ASK PHOTOS FEATURE BEGINS US ROLLOUT

Google Photos' new AI-powered search feature, "Ask Photos," is rolling out to select users in the U.S., allowing them to search their photos using more complex natural language queries.

ALIBABA RELEASES NEW AI MODEL QWEN2-VL THAT CAN ANALYZE VIDEOS MORE THAN 20 MINUTES LONG

Alibaba Cloud has released Qwen2-VL, a new vision-language model with enhanced visual understanding, video comprehension, and multilingual text-image processing. Qwen2-VL shows superior performance against models like Meta's Llama 3.1 and OpenAI's GPT-4o and supports various applications, including real-time video analysis and tech support. The models, available in three sizes (7B, 2B, and soon 72B), are open-source under Apache 2.0 for the smaller variants.

SGLANG 0.3

SGLang inference improvements are here in the latest release, with 7x Faster DeepSeek MLA, 1.5x Faster torch.compile, Multi-Image/Video LLaVA-OneVision, and more.

OLMO MOE

A great open MoE with best in class performance for 1B activated parameters.

ALIGNING STYLE AND TEXT IN IMAGE GENERATION

This paper introduces StyleTokenizer, a method for improving style control in text-to-image generation by aligning style representation with text prompts.

CORNELL'S APPLIED ML CLASS

Open resources for the Fall 2024 Applied ML class at Cornell.

LAMINAR

Open-source observability, analytics, evals, and prompt chains for complex LLM apps.

LONG-CONTEXT UNDERSTANDING WITH LONGLLAVA

LongLLaVA is a multimodal model designed for handling long-context tasks like video and high-resolution image understanding.

MATURING ENTERPRISE AI INFRASTRUCTURE

Interesting interview with the CEO of BentoML that talks about how to improve your enterprise tooling, ensuring that you are able to scale, but not over-engineering it at the beginning.

LLM-BASED EMBEDDING MODELS

This study investigates various designs for LLM-based embedding models, comparing different pooling and attention strategies.

OPTICAL CONNECTIVITY DIRECTLY IN THE GPU

GPU interconnect bandwidth is one of the primary bottlenecks in training large models today. Broadcom is working to integrate optical transfer directly into GPUs which would alleviate the issue substantially.

YOUTUBE IS MAKING TOOLS TO DETECT FACE AND VOICE DEEP FAKES

YouTube is developing new tools to protect artists and creators from unauthorized use of their likenesses, including AI-generated face and singing voice detection technology, with pilot programs launching early next year.

ICON (PRODUCT LAUNCH)

Icon helps brands partner with creators, turn 1 video into 20 videos with AI, and A/B test messaging to find winning ads.

GOOGLE IS WORKING ON AI THAT CAN HEAR SIGNS OF SICKNESS

Google is using AI models trained on 300 million audio samples to detect early signs of diseases like tuberculosis.

Get started with AI-ready workstations

YOU.COM RAISES $50M SERIES B

You.com, a search engine, agent platform, and knowledge base company, has raised additional funds as it continues to grow.

SAKANA RAISES $100M SERIES A

The raise will allow Sakana to continue growing its compute capabilities, hire more researchers, and generally become a leading AI lab in Japan.