Search the archive

INTERACTION PROCESSING UNITS

This article explores the development of computer hardware based on Interaction Nets, a computational model that represents calculations as interacting graph nodes. While current implementations like HVM show promise for parallel processing, traditional hardware isn't optimized for graph-based operations. The author proposes that custom hardware architecture could better utilize Interaction Nets' inherent parallelism and local memory access patterns, particularly benefiting algorithms with non-homogeneous parallelism like optimization problems and graph processing.

MLLM'S FACE UNDERSTANDING BENCHMARK

FaceXBench is a comprehensive benchmark for evaluating MLLMs face understanding across 14 tasks in 6 key categories.

DEEPSEEK RELEASES VL2, A SERIES OF MOE VISION-LANGUAGE MODELS

DeepSeek-VL2 introduces a new series of Mixture-of-Experts Vision-Language models with up to 4.5B activated parameters that excel in tasks like OCR, visual grounding, and chart understanding.

COLQWEN 2 (HUGGING FACE HUB)

A new vision based retrieval model based on Qwen that blew away state of the art for later interaction document retrieval.

CONTROLLING GENERATED LENGTH IN LLMS

The "Ruler" method helps LLMs generate responses of a specific length. It improves models' ability to follow length constraints in user instructions using Meta Length Tokens.

CONTROLLING AI'S GROWING ENERGY NEEDS

AI model training requires significant energy, with LLMs like GPT-3 consuming vast amounts. Alternatives such as neuromorphic and optical computing are being explored to reduce this footprint. Meanwhile, smaller fine-tuned models are emerging as an energy-efficient option for specific applications.

AI MEMORY AND CONTEXT: OPEN SOURCE, DEEPSEEK, META, AND MODEL RESEARCH

Yann LeCun emphasizes that true AI intelligence requires a system of interconnected components rather than a single powerful model, highlighting the need for persistent memory and context. He calls for new AI architecture, possibly through "world models," to better mimic human cognition. LeCun also advocates for open-source AI development to advance the field.

AI HATERS BUILD TARPITS TO TRAP AND TRICK AI SCRAPERS THAT IGNORE ROBOTS.TXT

Nepenthes, a malicious AI tarpit inspired by anti-spam techniques, was created to trap and poison AI crawlers ignoring robots.txt, raising operational costs for AI companies. Developer Aaron's tool disrupts AI training by feeding bots gibberish, with only OpenAI's systems reportedly escaping its grasp. While some criticize its efficacy, Nepenthes symbolizes resistance against unchecked AI scraping, sparking similar tools like Iocaine designed to protect online content.

COPYRIGHT OFFICE OFFERS ASSURANCES ON AI FILMMAKING TOOLS

The U.S. Copyright Office clarified that AI-assisted creations can still qualify for copyright protection, provided they involve human authorship.

CUSTOMERS DON'T CARE ABOUT YOUR AI FEATURE

Research by Irrational Labs found labeling products as "AI" doesn't increase trust, justify higher prices, or improve performance expectations.

OPENAI INKS DEAL WITH LOS ALAMOS LAB TO CRAM O1 INTO VENADO

OpenAI partnered with Los Alamos National Laboratory to deploy its o1 LLM on the Venado supercomputer for nuclear security and scientific advancements.

OPENAI'S TRADEMARK HINTS AT ROBOTICS AND AI HARDWARE

OpenAI's recent trademark application suggests potential AI-powered consumer hardware, including smartwatches, AR glasses, humanoid robots, and custom AI chips, aligning with ongoing partnerships and research into AI-driven robotics.

SOFTBANK-BACKED BILLIONAIRE TO INVEST $230M IN INDIAN AI STARTUP KRUTRIM

Bhavish Aggarwal is investing $230 million in his AI startup Krutrim to develop LLMs for Indian languages, aiming to raise $1.15 billion in total. Krutrim has open-sourced its AI models and plans to build India's largest supercomputer with Nvidia. The startup has released Krutrim-2, a 12-billion-parameter model, showing strong performance in Indian language tasks and code generation.

DEEPSEEK GETS AN 'F' IN SAFETY FROM RESEARCHERS

Cisco tested DeepSeek's open-source model, DeepSeek R1, which failed to fend off all 50 harmful behavior prompts from the HarmBench dataset. DeepSeek's failure rate is the highest among tested LLMs, with other models like Meta's Llama 3.1 and OpenAI's o1 performing noticeably better. The model's susceptibility to attacks, alongside data security concerns, has raised significant scrutiny and criticism.

VISION SEARCH ASSISTANT

Vision Search Assistant is a framework that combines Vision Language Models (VLMs) with web agents to improve object recognition, even for unfamiliar images.

MONST3R: A SIMPLE APPROACH FOR ESTIMATING GEOMETRY IN THE PRESENCE OF MOTION

A great project that can take video as input and estimate geometry and camera motions without any knowledge of camera intrinsics.

TOWARDS SCIENTIFIC DISCOVERY WITH DICTIONARY LEARNING: EXTRACTING BIOLOGICAL CONCEPTS FROM MICROSCOPY FOUNDATION MODELS

Dictionary learning reveals model interpretability and can discover unknown concepts from scientific data, like cell images. It successfully extracts concepts like cell type and genetic perturbation. A new algorithm, ICFL, combined with PCA whitening, enhances feature selectivity over existing methods.

Different models excel at different tasks, and GitHub Copilot

gives you the freedom to choose. Switch between Gemini 2.0 Flash, Claude 3.5 Sonnet, GPT-4o, and more to solve the most complicated coding dilemmas. Edit files simultaneously, generate tests, get code reviews, and watch your productivity soar. Start for free on GitHub or

VS Code.

GETTING STARTED WITH REAL ROBOTS

Great post from Hugging Face about using its LeRobot framework to control a robotic arm for research and development.

AUTOMATING SOP CREATION

Manually creating Standard Operating Procedures (SOPs) is time-consuming, but video-language models offer a promising solution by analyzing demonstration videos.

MEDICAL MODELS

This paper introduces Low-Rank Knowledge Decomposition (LoRKD), a framework designed to enhance the performance of medical foundation models by dividing them into lightweight expert models focused on specific anatomical regions.

HOW INDIGENOUS ENGINEERS ARE USING AI TO PRESERVE THEIR CULTURE

Indigenous researchers are using AI and machine learning to build speech recognition models for over 200 endangered Indigenous languages in North America. Initiatives like First Languages AI Reality train Native scientists to ethically handle culturally sensitive data and preserve language. Programs such as Tech Natives and IndigiGenius aim to boost Indigenous representation in tech to preserve cultural heritage through AI.

SAM ALTMAN SAYS OPENAI IS GOING TO DELIVER A BEATDOWN ON DEEPSEEK

OpenAI's Sam Altman responded to Chinese startup DeepSeek's R1 model outperforming competitors at lower costs, which created significant turbulence in the tech industry, by pledging to continue heavy investment in compute resources, emphasizing ambition towards AGI. The situation raises questions about the viability of OpenAI's costly approach given DeepSeek's efficient alternatives.

CHINESE AND IRANIAN HACKERS ARE USING U.S. AI PRODUCTS TO BOLSTER CYBERATTACKS

Hackers from China, Iran, and other nations are using AI, specifically Google's Gemini, to enhance cyberattacks, treating it as a tool for boosting productivity rather than creating groundbreaking techniques. Google reports found that groups linked to these countries are leveraging Gemini for research and phishing, with China and Iran being the most active users. Although AI hasn't yet transformed the scale of cyberattacks, both the U.S. and China are heavily investing in AI technologies for future dominance.

the brochure.

GEMINI CAN NOW DO MORE COMPLEX DATA ANALYSIS IN GOOGLE SHEETS

Google Sheets' new AI-powered Gemini update provides enhanced data insights, generates charts, and analyzes contextual trends using Python.

DEEPSEEK MIGHT HAVE A TRADEMARK PROBLEM IN THE US

DeepSeek faces a U.S. trademark conflict as Delson Group claimed the "DeepSeek" trademark just before them.

CHATGPT'S MOBILE USERS ARE 85% MALE

AI app spending reached $1.42 billion in 2024, with ChatGPT dominating despite competition, leading to a 274% increase year-over-year.

HUGGING FACE REPLICATING OPENAI'S DEEP RESEARCH

Hugging Face attempted to replicate OpenAI's Deep Research, an agentic web-search framework that significantly improved performance on the GAIA benchmark, by running a 24-hour-long experiment aimed at open-sourcing an equivalent system.

GOOGLE CEO ON DEEPSEEK VS. GEMINI

Sundar Pichai has downplayed the efficiency of DeepSeek's AI models, arguing that Google's Gemini models, particularly Gemini 2.0 Flash, outperform them despite DeepSeek's disruptive impact on the AI market.

US COPYRIGHT OFFICE RULES OUT COPYRIGHT FOR AI CREATED CONTENT WITHOUT HUMAN INPUT

The US Copyright Office states that AI-generated works without human intervention cannot be copyrighted. AI tools assisting with creativity, like de-aging actors, won't limit copyright protection, but purely generative AI outputs require further analysis.

HARMONIC LOSS TRAINS INTERPRETABLE AI MODELS

Harmonic loss is an alternative to cross-entropy loss for training neural networks that offers better interpretability and faster convergence through scale invariance and finite convergence points. Experiments across algorithmic, vision, and language datasets, demonstrate that models trained with harmonic loss show superior performance in interpretability, data efficiency, and reduced grokking compared to standard models. Harmonic loss could be particularly valuable for applications with limited data or where interpretability is crucial.

CONVEX OPTIMIZATION THEORY AND LEARNING-RATE SCHEDULING FOR LARGE MODEL TRAINING

Learning-rate schedules for large models closely match theoretical bounds from non-smooth convex optimization. These authors provide a bound for constant schedules with linear cooldown, showing cooldown's practical benefits through the absence of logarithmic terms in the bound. Their findings enabled practical improvements in training Llama-type models through optimal learning-rate extension and cross-schedule transfer.

IN-CONTEXT REINFORCEMENT LEARNING

This study explores scaling In-Context Reinforcement Learning (ICRL) to broader domains using Algorithm Distillation and shows that ICRL can be a viable alternative to expert distillation for generalist decision-making systems.

Forrester report

GOT OCR 2.0 WEIGHTS (HUGGING FACE HUB)

One of the best OCR models is now available and integrated with the Hugging Face ecosystem. It works well on documents and sheet music.

OPEN-VOCABULARY DETECTION WITH LLMS

LLMDet is an open-vocabulary detector that leverages a large language model for enhanced caption generation and grounding that significantly boosts performance compared to existing detectors.

EFFICIENT CHAIN-OF-THOUGHT REASONING

Heima introduces a framework for more efficient multimodal reasoning by compressing Chain-of-Thought processes into a single hidden token.

HOW TO SCALE YOUR MODEL

Amazing post from the DeepMind team about the mental process they use to scale their model. They break it down into mathematical equations, which allows them to reason about the costs of each operation and ensure correctness.

WHO IS LIANG WENFENG? DEEPSEEK FOUNDER COMES FROM AI INVESTING

DeepSeek's R1 reasoning model uses less computing power than its U.S. counterparts and is open source. The DeepSeek app topped App Store charts over ChatGPT. Founder Liang Wenfeng previously started AI firms and his hedge fund High-Flyer manages $8 billion, backing DeepSeek. Liang sets himself apart by offering the product for free and open source.

AI AND THE FUTURE OF NATIONAL SECURITY

Google highlights the strategic importance of AI and quantum computing for national security, emphasizing the need for private-sector leadership, government procurement reforms, and public-private collaboration to strengthen cybersecurity.

GOOGLE'S 2024 RESPONSIBLE AI REPORT

Google has released its 6th annual Responsible AI Progress Report, which details governance structures, safety evaluations, and risk mitigation techniques for AI product development.

HUGGING FACE RESEARCHERS ARE TRYING TO BUILD A MORE OPEN VERSION OF DEEPSEEK'S AI 'REASONING' MODEL

Hugging Face is working to replicate DeepSeek's R1 model with an open-source initiative called Open-R1, addressing concerns over the lack of transparency in DeepSeek's release.

META AI CAN NOW USE YOUR FACEBOOK AND INSTAGRAM DATA TO PERSONALIZE ITS RESPONSES

Meta is enhancing its AI chatbot with memory capabilities that will allow it to remember user details in conversations on Facebook, Messenger, and WhatsApp in the U.S.

GET THE EBOOK