Search the archive

THAT CHIP HAS SAILED

A strong response from the DeepMind team about its AlphaChip work, which has been largely criticized by various academic groups while seeing wide adoption amongst their industry peers.

FASTER IMAGE AND VIDEO GENERATION

Polynomial Mixer is a faster, memory-efficient alternative to Multi-Head Attention (MHA) in diffusion models for generating images and videos.

CROSS-VIEW GEO-LOCALIZATION

Researchers have developed a framework to tackle the challenges of cross-view geo-localization, such as viewpoint differences and global-scale scenarios.

ENHANCING LLM REASONING

Researchers have introduced Additional Logic Training to improve reasoning in LLMs. The method trains LLMs to handle complex deductions with diverse rules and distractions.

PANORADAR: ROBOTS GAIN SUPERHUMAN VISION WITH AI AND RADIO WAVES

PanoRadar is an AI-powered system developed by researchers at the University of Pennsylvania that converts radio waves into 3D views, providing robots with LiDAR-like vision at a lower cost. It overcomes limitations faced by conventional sensors in environments with smoke, fog, and glass by using AI algorithms to process radio wave reflections. The research team aims to integrate PanoRadar with existing sensing technologies for enhanced multi-modal perception in robotics.

THE EVOLUTION OF THE CREATOR

Generative AI is revolutionizing the creator economy by significantly lowering production barriers, enabling creators to generate high-quality content with ease. AI-driven innovations, such as digital clones, are poised to redefine content distribution and engagement, offering new monetization opportunities by scaling creator interactions and fan transactions. As AI continues to reshape creation, distribution, and monetization, the creator economy is set to spawn a new wave of significant technology companies.

GOOGLE DEEPMIND HAS A NEW WAY TO LOOK INSIDE AN AI'S “MIND”

DeepMind has released Gemma Scope, a tool that reveals AI model patterns and decision-making processes and enhances understanding of AI outputs through mechanistic interpretability. It uses sparse autoencoders to dissect and analyze data layers, aiding in uncovering biases or errors, like erroneous numerical interpretations. By advancing model transparency, researchers hope to refine AI control and alignment and reduce risks in deployment.

spot now

AI MODEL IDENTIFIES OVERLOOKED BRAIN TUMORS IN JUST 10 SECONDS

FastGlioma is an AI model that rapidly detects residual brain tumor tissues during surgery with high accuracy.

IT'S SURPRISINGLY EASY TO JAILBREAK LLM-DRIVEN ROBOTS

RoboPAIR can successfully jailbreak LLM-driven robots by circumventing safety measures.

NVIDIA TO FUEL HUMANOID ROBOTS WITH ‘JETSON THOR'

Nvidia's Jetson Thor platform, launching in early 2025, focuses on enhancing humanoid robots with advanced AI capabilities.

AI applications, achieving significant ROI, and upcoming trends

like end-to-end automation of critical workflows.

Register today!

META HIRES SALESFORCE'S CEO OF AI, CLARA SHIH

Meta is creating a new product unit to develop AI tools for the 200 million businesses that use its apps.

ROX'S PUBLIC BETA AND $50M RAISE

Rox, an AI-driven sales productivity platform, enhances top enterprise sales reps' performance by over 30% by using AI analyst teams for tasks like planning and engagement. It integrates seamlessly into existing systems, bypassing traditional CRM inefficiencies, and has already been adopted by top-performing companies. Rox has announced a $50M raise, led by Sequoia and other major investors, to expand its reach.

GENIES LAUNCHES PARTIES FOR BRANDS AND CREATORS TO LAUNCH THEIR OWN ‘AI ROBLOX'

Genies has launched Parties, a platform that uses AI avatars to enable user-generated content and personalized gameplay experiences. Its technology stack emphasizes smart avatars, interoperability, and dynamic user interaction, facilitating limitless customization for brands. Genies leverages machine learning to support scalable, interactive ecosystems for IP partners.

GENERATIVE WORLD EXPLORER

The Generative World Explorer (Genex) is a system that can imagine exploring 3D spaces via generation and then use those updates to improve planning. It uses an ST-VAE and diffusion pass for imagination, which improves planning.

SEARCH, VERIFY, FEEDBACK

A review and reflection paper about how to incorporate verification into large language model post training. This is a new sort of post training that can be beneficial for improving reasoning in models.

1D CONVOLUTIONS FOR FASTER IMAGE SEGMENTATION

OneNet is a 1D convolutional encoder designed for efficient image segmentation. It is ideal for edge devices.

EVALUATION CONSTITUTIONS

Code to assist in evaluating constitutions based on human feedback.

VIDEO EDITING WITH STABLEV2V

StableV2V introduces a new way to edit videos while keeping shapes consistent, even when user prompts demand major transformations. This approach ensures smooth and accurate changes across video frames without losing structural integrity.

SMARTER DESCRIPTIONS OF REMOTE SENSING CHANGES

CCExpert is a new AI model designed to describe changes in images with natural language. It can identify what changed, where, and how.

GENERATIVE AI TAUGHT A ROBOT DOG TO SCRAMBLE AROUND A NEW ENVIRONMENT

LucidSim, a new system that uses generative AI and physics simulation, improves robot training by creating virtual environments that closely mimic real-world conditions. Robots trained with LucidSim outperform those using traditional simulations in tasks like parkour and navigating obstacles. This novel AI approach could lead to fully virtual training for diverse AI agents. It is being extended to more complex robotic tasks.

ROBOTIC SURGERY NEARS AUTONOMY WITH NEW AI TRAINING BREAKTHROUGH

Johns Hopkins researchers developed an imitation learning model that allows surgical robots to perform complex procedures by observing video footage of human surgeons. The model, which merges LLM architecture with kinematic language, reduces the need for manual programming, enabling quicker training of robots for various surgeries. Stanford's da Vinci Surgical System robot has already demonstrated human-comparable performance on essential tasks using this approach.

AI'S MATH PROBLEM: FRONTIERMATH BENCHMARK SHOWS HOW FAR TECHNOLOGY STILL HAS TO GO

Epoch AI's FrontierMath benchmark highlights AI's struggle with advanced mathematical reasoning. Top models like GPT-4o solve less than 2% of these complex problems. Unlike traditional math benchmarks, FrontierMath presents entirely new, challenging problems that require deep reasoning and creativity. This benchmark shows how the current AI has trouble with handling intricate multi-step logic.

stock analysis

DEEPL LAUNCHES DEEPL VOICE, REAL-TIME, TEXT-BASED TRANSLATIONS FROM VOICES AND VIDEOS

DeepL has introduced DeepL Voice for real-time audio translation, expanding beyond its core text translation service.

GOOGLE RELEASES STANDALONE GEMINI APP FOR IPHONE

CHATGPT CAN NOW READ SOME OF YOUR MAC'S DESKTOP APPS

OpenAI's ChatGPT desktop app for macOS now integrates with coding apps like VS Code and Xcode, allowing developers to send their code directly to ChatGPT for context alongside prompts.

Abstraction

→ Check out pgai on GitHub [5]: a framework for developing RAG, semantic search, and other AI applications directly in PostgreSQL

PIXTRAL LARGE

Pixtral Large is a 124B open-weights multimodal model built on top of Mistral Large 2. It is the second model in the multimodal family and demonstrates frontier-level image understanding. The model is able to understand documents, charts, and natural images while maintaining the leading text-only understanding of Mistral Large 2.

PERPLEXITY INTRODUCES A SHOPPING FEATURE

Perplexity has introduced a shopping feature for Pro users in the US, making it a one-stop solution where you can research and purchase products.

BEN AFFLECK TELLS ACTORS AND WRITERS NOT TO WORRY ABOUT AI

Ben Affleck reassured Hollywood actors and writers during an interview that AI poses little threat to their jobs due to its current limitations.

LLAVA-O1: REASONING VLMS

While this is not an actual replication of the process used to train o1, it is still a strong VLM trained on reasoning traces.

CASE STUDY OF SONNET 3.5 COMPUTER USE

This work aims to systematically test the ability of new autonomous computer use agents. It finds Claude quite powerful at traditional linear tasks.

TEACHING VIDEO MODELS TO UNDERSTAND TIME LIKE A STORY

This paper introduces NumPro, a creative solution that helps Video Large Language Models handle Video Temporal Grounding tasks.

LLAMA 3 405B INFERENCE

Cerebras has the largest Llama model, running 12x faster than comparable systems and 18x faster than some closed model API providers. It can run at almost 1k tokens per second. It will be available via API at the beginning of next year.

CLIP FOR SEMANTIC SEGMENTATION

Despite CLIP's success in open-vocabulary tasks, it struggles with semantic segmentation due to noisy features and limited resolution. Trident addresses the resolution issue with a training-free framework that combines CLIP and DINO features from sub-images and uses SAM's encoder for global feature aggregation.

CERTIFIED ROBUSTNESS

This work focuses on improving the certified robustness of smoothed classifiers by fine-tuning off-the-shelf models.

RECAPTURE

This paper from Google shows how you can change the camera viewpoint of an existing video.

NOUS RESEARCH FORGE

The Forge Reasoning API allows you to take any popular model and superpower it with a code interpreter and advanced reasoning capabilities. This has been shown to increase model performance.

WHERE'S MY ROBOT BUTLER?

Advancements in AI and robotics are accelerating the development of humanoid robots such as Atlas, Optimus, and Neo, robots aimed at fulfilling domestic tasks akin to Rosie from "The Jetsons." Challenges persist in creating cost-effective, safe, and efficient actuators. AI models are crucial for training robots to perform complex tasks autonomously. Despite progress, these humanoid robots are currently more suited for industrial settings. They may eventually transition into home environments if significant breakthroughs occur.