Search the archive

PIXEL-LEVEL CAPTION GENERATION

Pix2Cap-COCO introduces a dataset for panoptic segmentation-captioning, combining pixel-level annotations and detailed object-specific captions to advance fine-grained visual and language understanding.

WATERMARKING FRAMEWORK FOR VIDEO GENERATION

VideoShield is a watermarking framework specifically designed for popular diffusion-based video generation models. It embeds watermarks directly during video generation, eliminating the need for additional training.

IF EVEN 0.001 PERCENT OF AN AI'S TRAINING DATA IS MISINFORMATION, THE WHOLE THING BECOMES COMPROMISED, SCIENTISTS FIND

NYU researchers discovered that poisoning just 0.001% of an LLM's training data with misinformation can lead to widespread errors, posing serious risks in medical applications. The study, published in Nature Medicine, found that even corrupted LLMs perform similarly to non-corrupted ones on standard benchmarks, making such vulnerabilities hard to detect.

US TIGHTENS ITS GRIP ON AI CHIP FLOWS ACROSS THE GLOBE

The U.S. introduced new AI export controls, capping chip exports to most countries while exempting 18 allied nations, to maintain AI leadership and restrict China's access. Major cloud providers like Microsoft, Google, and Amazon can seek global authorizations under these rules. Industry players like Nvidia have criticized the regulations as overreaching.

AI MISTAKES ARE VERY DIFFERENT THAN HUMAN MISTAKES

AI systems like LLMs make fundamentally different mistakes from humans, often erring randomly and with unwarranted confidence. To address this, we need new security systems and methods to manage AI errors beyond existing human-error correction techniques. Focus areas include aligning AI behavior to human-like error patterns and developing unique mistake mitigation strategies tailored for AI.

video to hear from some Jane Street ML team members

GOOGLE FOLDS MORE AI TEAMS INTO DEEPMIND TO 'ACCELERATE THE RESEARCH TO DEVELOPER PIPELINE'

Google is consolidating its AI teams, including those working on AI Studio and Gemini APIs, under Google DeepMind to speed up AI development.

OPENAI APPOINTS BLACKROCK EXEC TO ITS BOARD

OpenAI has appointed Adebayo Ogunlesi, a senior managing director at BlackRock, to its board of directors.

OPENAI'S AI REASONING MODEL 'THINKS' IN CHINESE SOMETIMES AND NO ONE REALLY KNOWS WHY

OpenAI's o1 model sometimes processes reasoning tasks in languages like Chinese even when queries are in English.

AI-driven world with this resource bundle.

QWEN 2.5 1M CONTEXT

Really powerful and local 1M context models from the Qwen team. They show how the team gradually extended the context ability throughout training. Additionally, the Qwen team has released its inference framework built on top of vLLM which is as much as 7x faster.

ELEVENLABS RAISES $250M AT $3B VALUATION

ElevenLabs secured significant funding to expand its AI voice technology platform, emphasizing new applications in entertainment, accessibility, and virtual assistants.

DEEPSEEK CLAIMS ITS 'REASONING' MODEL BEATS OPENAI'S O1 ON CERTAIN BENCHMARKS

DeepSeek's DeepSeek-R1 reasoning model matches OpenAI's o1 on benchmarks like AIME and MATH-500. With 671 billion parameters, it offers competitive performance at a lower cost but is subject to Chinese regulation constraints. Available on Hugging Face, this release comes amid U.S.-China tensions over AI tech development and export controls.

GAUSSIAN SURFACE TRACKING AND RECONSTRUCTION

The GSTAR method presented in this work shows an amazing way to both reconstruct dynamic meshes and track every point in 3D. While it does require carefully calibrated multi-view cameras, it is a great step forward to the single view case.

TRAINING A SPEECH SYNTHESIZER

Alex Nichol from OpenAI has released a great blog post showing how to train a speech synthesizer. This method, which uses VQVAEs and autoregressive models, is common in multimodal understanding and generation.

PARAMETER-EFFICIENT FINE-TUNING FOR FOUNDATION MODELS

This survey reviews parameter-efficient fine-tuning techniques for foundation models, offering insights into methods that minimize computational costs while maintaining performance across diverse tasks.

REASONING ON LLAMA (GITHUB GIST)

This is the current minimal working replication of the reasoning models presented originally by OpenAI and published by DeepSeek. It uses format and correctness rewards across math problems. This particular snippet also shows the “aha” moment when training for sufficiently long.

1PROMPT1STORY

1Prompt1Story is a training-free approach for consistent text-to-image generations with a single concatenated prompt.

IMAGE GENERATION WITH COT

This project investigates the potential of CoT reasoning to enhance autoregressive image generation.

AI STARTUPS IN EDUCATION

Google highlights emerging startups using AI to create innovative tools for personalized learning, content generation, and student engagement in education.

3D TUMOR DATASET

AbdomenAtlas 3.0 is the first public dataset with high-quality abdominal CTs and paired radiology reports. The database includes more than 9,000 CT scans with radiology reports and per-voxel annotations of liver, kidney, and pancreatic tumors.

THE PARADOX OF SELF-BUILDING AGENTS: TEACHING AI TO TEACH ITSELF

AI agents are transforming from reactive tools to proactive systems, with potential to redefine enterprise software by collapsing traditional software stacks. Yohei Nakajima outlines four levels of autonomy for agents, highlighting advancements from fixed capabilities to anticipatory self-building. Despite their promise, these agents require safeguards to prevent misuse, with careful design and oversight being essential to balance innovation with security.

LIGHTPANDA BROWSER

Headless and lightweight browser designed for AI and automation.

AI INNOVATIONS FOR RETAILERS

Google Cloud unveils new AI tools for retailers, focusing on personalized shopping experiences, real-time inventory management, and predictive analytics.

GOOGLE AI STUDIO (PRODUCT)

Google AI Studio is the fastest way to start building with Gemini.

Attend live for the full experience!

OPENAI OPERATOR

Operator is a newly released AI agent capable of performing tasks on the web by interacting with browsers, handling actions such as typing, clicking, and scrolling. Powered by the Computer-Using Agent (CUA) model, it uses GPT-4's vision capabilities and reinforcement learning to navigate graphical interfaces and execute user-defined tasks, such as filling forms or ordering groceries.

ANTHROPIC INTRODUCES CITATIONS API

Anthropic's new Citations API helps AI models generate responses with detailed and verifiable citations, increasing the trustworthiness of outputs.

LUMA AI RELEASES RAY2

Ray2 is a large-scale video generative model that's setting a new standard in realistic visuals with natural, coherent motion and logical event sequences. It was trained on Luma's new multi-modal architecture, scaled to 10x compute of Ray1. Ray2 is initially available to paid subscribers in Dream Machine with 5 second text-to-video generations.

OPTIMIZING REASONING EFFICIENCY IN LLMS

A new fine-tuning approach reduces inference time in reasoning models while preserving accuracy, improving performance on complex tasks.

PREDICTING GENETIC PHENOTYPES

FREEFORM is a framework that uses LLMs to improve feature selection and engineering for genotype data.

SYNTHETIC BENCHMARK FOR MULTIMODAL DETECTION MODELS

LOKI is a synthetic benchmark that evaluates how good VLMs are at detecting novel and challenging items.

INTEGURU

Integuru is an AI agent that reverse-engineers platforms' internal APIs to generate Python integration code. It automates actions like downloading resources by analyzing browser network requests and dependencies. Users can provide prompts and .har files to create the required executable code.

SYNTHETIC DATA ENGINE

TabularARGN is a flexible framework designed to handle mixed-type, multivariate, and sequential datasets. It supports advanced features such as fairness-aware generation, imputation, and conditional generation on any subset of columns.

EVALUATING DISTILLED LLMS

This repository provides two complementary metrics for quantifying LLM distillation.

VIRTUOSO SMALL BEATS PHI 4

Virtuoso has the same number of weights as the newly released Phi model, scores strongly in benchmarks, and also seems slightly more useful in real world tasks where the synthetic data of Phi can fall over.

WHAT FEATURES ARE COMING NEXT TO THE GEMINI APP?

SNOWFLAKE CLAIMS BREAKTHROUGH CAN CUT AI INFERENCING TIMES BY MORE THAN 50%

Snowflake has integrated SwiftKV, a technique from its AI Research, to optimize LLM inference, improving throughput by 50% and reducing costs by up to 75% for some Llama models. SwiftKV enhances efficiency by recycling hidden states in LLM layers, reducing memory usage and computational overhead without significant loss of accuracy. The technique will be available in other models within Snowflake's Cortex AI, boosting performance in tasks like summarization and real-time AI applications.

current market. Get the guide from Dropzone AI

NVIDIA'S AI NPCS ARE A NIGHTMARE

NVIDIA's ACE technology, showcased at CES 2025, powers AI-driven NPCs for games like PUBG and ZooPunk, but often results in robotic voices and lackluster interactions.

THIS AI MOTORCYCLE HELMET PROMISES 100% BLIND-SPOT ELIMINATION

Intelligent Cranium Helmets has launched the iC-R, a smart motorcycle helmet integrating AI for enhanced safety and connectivity.

FROM GMAIL TO WORD, YOUR PRIVACY SETTINGS AND AI ARE ENTERING INTO A NEW RELATIONSHIP

AI integration in software raises privacy concerns, as many applications might use personal data without explicit consent for training AI models.

Read more and see the benchmarks on the Contextual AI blog →

OPENAI'S STARGATE PROJECT

The Stargate Project is a new company that intends to invest $500 billion over the next four years to build new AI infrastructure for OpenAI in the United States.

MICROSOFT ENDS EXCLUSIVE CLOUD PARTNERSHIP WITH OPENAI

Microsoft is no longer OpenAI's exclusive cloud provider, signaling broader opportunities for OpenAI to expand its partnerships with other providers.