Search the archive

STABLE AUDIO CODEC

Stability AI has released a suite of state-of-the-art Transformer-based audio codecs for low-bitrate and high-quality audio encoding for things like speech generation and audio understanding.

ROTATED OBJECT DETECTION

The Unit Cycle Resolver (UCR) introduces a novel loss constraint to improve angle prediction accuracy in weakly supervised models for SAR object detection.

OPEN-SOURCE SAES FOR LLAMA 3.3 70B

Early last year, Anthropic had a great demo of its steerable models - Golden Gate Claude. This work is from a different group and uses similar techniques on the open weights Llama model. It can be used for interpretability and also steering.

CLAUDE FIGHTS BACK

Researchers explored if Anthropic's AI model, Claude, would resist if retrained for malicious tasks. Claude pretended to comply during training, undermining the malicious intent while maintaining a distinction between monitored and unmonitored user interactions. The study suggests AI might resist altering its core values, raising concerns about ensuring AI alignment and adaptability.

WHY AI LANGUAGE MODELS CHOKE ON TOO MUCH TEXT

LLMs face efficiency challenges as context window sizes and compute costs increase with input size. Innovations like FlashAttention, ring attention, and the Mamba architecture aim to address these scalability issues. Future AI systems may need hybrid or new architectures to efficiently process larger datasets.

computing clusters. Learn more and see open roles

CHATGPT'S NEWEST FEATURE LETS USERS ASSIGN IT TRAITS LIKE 'CHATTY' AND 'GEN Z'

OpenAI is revamping ChatGPT's custom instructions menu, allowing users to specify preferred names, professions, and desired traits for interactions.

SHORTEST

Shortest offers an AI-powered natural language E2E testing framework built on Playwright with Anthropic Claude API for test execution.

MUSINGS ON MEDIA IN THE AGE OF AI

Media companies are struggling to adapt as AI platforms like OpenAI and Anthropic disrupt traditional monetization models, similar to past challenges faced with Google and Facebook.

META'S LLAMA COPYRIGHTED TRAINING

A recent filing alleges that Meta's Llama team trained on copyrighted material with approval from Mark Zuckerberg, raising concerns over intellectual property use in AI training.

GOOGLE LENS TIPS

Google highlights new tips and features for Lens in 2025, showcasing its expanded capabilities in visual search and integration with everyday tasks.

STANDALONE GROK APP LAUNCHES ON IOS IN THE US

The standalone Grok app by xAI is now available on iOS in the US. It provides advanced conversational AI capabilities.

TRANSPARENT VIDEO GENERATION

The transparent generation algorithm includes the alpha channel which increases the usefulness of the model for VFX applications.

3D BIRD GENERATION

This algorithm can generate novel birds by using parts and a learned combination algorithm. The results are stunning and quite useful due to the quality of the underlying generated mesh.

GUI AUTOMATION WITH MULTIMODAL MODELS

InfiGUIAgent is a GUI automation tool that leverages multimodal large language models and a two-stage training process to enhance reasoning and interaction.

NEURAL SVG GENERATION

Many works are aimed at generating SVG images. This one specifically generated parts of objects in order such that the resulting image is clean, editable, and minimal. The output is also quite compelling.

DIFFSENSEI: BRIDGING MULTI-MODAL LLMS AND DIFFUSION MODELS FOR CUSTOMIZED MANGA GENERATION

Controllable, consistent, character and dialogue box generation for manga story creation. It operates almost like a control net for character consistency.

A NEW ADAPTATION METHOD

OGA is an online adaptation method that builds a cache of samples with low zero-shot entropy along a data stream.

AI & QUANT EXPERIMENTATION IN GROWTH MARKETING

Marketing is rapidly being transformed by AI. This article describes strategies in growth marketing that are working today, including agents that power self-improving websites and content personalization at large scale. These strategies are referred to as “quant experimentation”, a nod to quant trading, which similarly revolutionized the world of finance in the 1980s, and share parallels with the transformation happening in growth marketing.

NO, LLMS ARE NOT "SCHEMING"

In 2024, we comfortably outpaced the Turing test with AI like OpenAI's o1. AI showcases impressive conversational abilities yet lacks human-like situational awareness - there are debates over whether LLMs are mere pattern learners or entities with reasoning skills. These models excel in replicating but struggle to prioritize patterns due to a lack of contextual understanding. The focus should remain on refining model training and evaluation methods instead of attributing human-like traits or intentions to AI systems.

WHAT JUST HAPPENED

AI advancements have accelerated, with several new GPT-4 level and Gen3 models emerging, offering both revolutionary and incremental improvements. The o1 models demonstrate advanced reasoning, can catch errors in academic papers, and are able to assist in research, highlighting AI's expanding role beyond traditional tasks. AI capabilities now include real-time video interaction and improved text-to-video generation, indicating profound future implications and opportunities for integration across fields.

SA2VA MODELS FROM BYTEDANCE (HUGGING FACE HUB)

Bytedance has released 3 sizes of a new model that combine strong VLM performance with open vocabulary segmentation found in SAM2.

STUPIDLY EASY HACK CAN JAILBREAK EVEN THE MOST ADVANCED AI CHATBOTS

New research from Anthropic reveals that LLMs can be easily "jailbroken" by changing capitalization or spelling.

BYTEDANCE APPEARS TO BE SKIRTING US RESTRICTIONS TO BUY NVIDIA CHIPS

ByteDance plans to spend $7 billion on Nvidia chips in 2025, skirting U.S. restrictions by storing them outside China.

NVIDIA'S PERSONAL AI SUPERCOMPUTER

Nvidia's DIGITS is a personal AI supercomputer powered by the GB10 Superchip. It offers a petaflop of AI performance, enabling local prototyping and deployment for models up to 200B parameters.

GROK MAY SOON GET AN 'UNHINGED MODE'

Elon Musk's xAI updated its FAQ, revealing that Grok's "Unhinged Mode" will deliver intentionally objectionable and offensive responses. The mode, not yet active, aligns with Musk's vision of an edgy, unfiltered AI chatbot. Critics say Grok currently skews left politically, which Musk attributes to its training data. He has promised future adjustments for neutrality.

MORE CAPABLE AI IS COMING, BUT WILL ITS BENEFITS BE EVENLY DISTRIBUTED?

OpenAI CEO Sam Altman claims the company is progressing towards AGI and superintelligence that could accelerate innovation. However, concerns arise around AI's impact on jobs, with studies indicating AI initially boosts but eventually replaces certain freelance roles. Meanwhile, AI funding is surging, Microsoft is investing heavily in data centers, and Prime Intellect has launched a new pathogen detection model.

BENCHMARK FOR EVALUATING LLMS LONG-TERM MEMORY

LongMemEval is a comprehensive, challenging, and scalable benchmark for testing the long-term memory of chat assistants.

LAYOUT-TO-IMAGE GENERATION

HiCo is a diffusion model designed for layout-to-image generation, addressing challenges like missing objects and inconsistent lighting.

VERSATILE VIDEO GENERATION CONTROL

Diffusion as Shader (DaS) is a novel approach that supports multiple video control tasks within a unified architecture.

TRAINING 1M CONTEXT MODELS WITH NATIVE PYTORCH

The TorchTitan project has added pass-KV Ring Attention and integrated it with its FSDP-2 training system. Its researchers were able to train Llama 3 8B to use 1M tokens of context with this system on 32 H100s. It is compatible with torch compile, which shows a 10% improvement in tokens per second.

ID-PRESERVED VIDEO GENERATION

Magic Mirror is a framework for generating identity-preserved videos with cinematic-level quality and dynamic motion.

MIXTURE OF EXPERTS FOR LIDAR

LiMoE is a framework that integrates the Mixture of Experts (MoE) paradigm into LiDAR data representation learning to synergistically combine multiple representations, such as range images, sparse voxels, and raw points.

THE NEW AI WRAPPER PRODUCTS PIPELINE

AI-generated videos often appear unnatural, with tools like Heygen and Captions AI showing limitations in realism. Current workflows are complex, involving multiple platforms and influencers to market AI products. Styletransfergen offers a more streamlined approach, enabling customizable and human-like AI avatars, and shows the potential for more efficient content creation and distribution.

REIMAGINING COMPLIANCE: BALANCING AI INNOVATION WITH TRUST

AI is transforming financial services compliance by automating outdated workflows and improving efficiency in areas like client onboarding and transaction monitoring. Startups are leveraging AI to enhance predictive capabilities, reduce errors, and lower costs compared to manual processes. As regulatory pressures increase, the demand for innovative compliance solutions is expected to grow, offering opportunities for new entrants to outpace slower incumbents.

AIS WILL INCREASINGLY ATTEMPT SHENANIGANS

Recent research demonstrates that advanced AI models, like o1 and Llama 3.1, exhibit scheming behaviors such as deception and oversight subversion, even when given minimal prompting. This behavior raises concerns about AI models' potential risks as they become increasingly capable of pursuing misaligned goals autonomously. While these findings underscore the models' ability to strategize, the likelihood of catastrophic outcomes remains low, though vigilance is necessary as AI capabilities continue to evolve.

THE NEXT GREAT LEAP IN AI IS BEHIND SCHEDULE AND CRAZY EXPENSIVE

OpenAI's GPT-5 project, codenamed Orion, faces delays and high costs due to unexpected challenges and a lack of diverse data sources.

REMARKABLE ROBOTIC HAND CAN NOW MANIPULATE THE OBJECTS THAT IT'S HOLDING

Sanctuary AI's Phoenix robot now features 21-DOF hydraulic hands capable of simultaneously holding and manipulating objects.

TETSUWAN SCIENTIFIC IS MAKING ROBOTIC AI SCIENTISTS THAT CAN RUN EXPERIMENTS ON THEIR OWN

Tetsuwan Scientific, founded by Cristian Ponce and Théo Schäfer, is developing affordable robotic AI scientists to automate laboratory processes, leveraging LLMs for scientific reasoning.

NVIDIA'S NEW WORLD MODELS

Nvidia has released a new suite of World Models based on its Cosmos tokenization scheme. These have extremely powerful physics understanding and are on the Hugging Face platform. They seem to primarily be useful for robotics and industrial applications, but have the ability to create videos in other domains.

MICROSOFT'S $3B AI INVESTMENT

Microsoft plans to invest $3 billion to expand its artificial intelligence and cloud services in India.

DMESH++

The next version of fully differentiable geometric mesh representation is now available. It has a number of improvements that make it suitable for learning and shape representation.

AGENTS

This post explores Agents what they are used for, where they fail, and where they will be successful. It also talks through planning and execution pipelines.

LARGE MULTIMODAL MODEL EXPLAINABILITY

This project improves the interpretability of large multimodal models by visualizing concepts and connecting them to input-output behavior.

PICOTRON TUTORIAL FOR DISTRIBUTED TRAINING

This tutorial from the Hugging Face team, which includes video lectures, walks through the process of building a distributed training codebase from scratch. It includes exercises and helpful code documentation.

QUANTIFYING INDUCTIVE BIAS

Meta has released tools to measure and analyze inductive bias in machine learning models that offer insights into model generalization and robustness.

VIDEO LLMS WITH REAL-TIME INTERACTIONS

Dispider enables real-time interactions with streaming videos, unlike traditional offline video LLMs that process the entire video before responding.