Search the archive

JANUS FLOW

Janus is the multimodal model from DeepSeek. DeepSeek just released a new version that uses rectified flows (similar to Meta Movie Gen) for image generation and understanding. The results are extremely compelling.

LINK CONVERSATION WITH REFERENCE MATERIALS

Problem-Oriented Segmentation & Retrieval (POSR) is a method that divides conversations into meaningful segments and links each part to relevant reference materials, such as worksheets or meeting notes.

REALISTIC IMAGE COMPOSITING

Researchers have introduced an improved finetuning technique for generative image composition that blends a given foreground object into a new background to create lifelike images.

ANTHROPIC CUSTOMER SPOTLIGHT: STACKBLITZ

Stackblitz built an online developer tool that is highly integrated with Claude Sonnet 3.5. This post tells the story of how it achieved $4m ARR in just a few short months.

GENERATING DESCRIPTIONS FOR LONG VIDEOS

StoryTeller is a new system designed to generate coherent descriptions for long videos. The project addresses the challenges of plot consistency and character tracking across scenes.

WHY THE DEEP LEARNING BOOM CAUGHT ALMOST EVERYONE BY SURPRISE

Fei-Fei Li's creation of the massive ImageNet dataset was instrumental in the resurgence of neural networks. The dataset provided the training data necessary for breakthrough models like AlexNet. AlexNet, trained using GPUs and guided by Geoffrey Hinton's backpropagation technique, demonstrated the power of deep learning on large datasets, sparking the modern AI boom. This pivotal moment underscored the importance of combining neural networks, big data, and GPU computing for advancing AI.

SEGMENT ANYTHING INTO 3D PARTS

SAMPart3D is a powerful technique from the University of Hong Kong that allows for the segmentation of 3D objects into their semantically meaningful parts.

LUME RAISES $4.2M SEED

Lume automates data mapping with AI, streamlining mapping, cleaning, and validation of data.

JUST HAVE AI BUILD AN APP FOR THAT

AI agents are increasingly being used to quickly create functional apps for tasks like resizing SVGs.

Get the world's best Whisper inference with Baseten

BAIDU ANNOUNCES ITS OWN PAIR OF AI SMART GLASSES

Baidu launched a pair of AI-powered smart glasses that uses the company's ERNIE generative AI at its World Conference. Baidu also introduced an AI image generator and a tool for creating software without coding expertise. ByteDance's Doubao is currently China's leading AI chatbot in terms of monthly active users.

OPENAI CO-FOUNDER GREG BROCKMAN RETURNS AFTER THREE MONTHS OF LEAVE

OpenAI co-founder Greg Brockman has returned to OpenAI as president, following a sabbatical, amid significant executive departures and controversy over its shift to a for-profit model. OpenAI recently secured a $157 billion valuation in its latest funding round. The company is facing internal challenges as leaders like Mira Murati, Bob McGrew, and Lilian Weng have left.

EUROPEAN GOOGLE RIVALS PARTNER ON SEARCH ENGINE INFRASTRUCTURE TO COUNTER BIG TECH

Ecosia and Qwant are partnering to build a European search index to reduce reliance on U.S. Big Tech and enhance AI capabilities. The venture utilizes a "privacy-first" approach and aims to support advancements in AI through a new search infrastructure. Alternative search providers are facing rising API costs - this move positions them to better compete amid the growing influence of generative AI in search.

SPECULATION ON TEST TIME COMPUTE (47 MINUTE VIDEO)

This video covers o1 models, how they could potentially be replicated, and how useful they will be for various tasks in the future.

EFFICIENTNAT BOOSTS SPEED IN IMAGE GENERATION

EfficientNAT is a refined non-autoregressive Transformer model designed to enhance both the speed and quality of token-based image generation.

DETECTING GRAPH ANOMALIES

UniGAD is a new framework for graph anomaly detection (GAD) that identifies irregularities across nodes, edges, and entire graphs simultaneously.

today.

OBJECT AND ATTRIBUTE MATCHING IN IMAGES WITH TOKEN MERGING

Token Merging addresses a common issue in text-to-image models: difficulty in linking objects with their specific attributes, also known as semantic binding.

COMPRESSING MESH DATA FOR 3D GENERATION

Blocked and Patchified Tokenization (BPT) is a mesh compression method that reduces mesh sequence length by about 75%, enabling efficient generation of meshes with over 8k faces.

LEARNING FROM EXPERT DEMOS

Successor Feature Matching is a novel, non-adversarial approach to inverse reinforcement learning that skips reward function learning.

DATACHAIN

DataChain is a Pythonic data-frame library for AI that allows efficient processing and organization of unstructured data into datasets without abstracting AI models. It supports integration with AI tools like PyTorch, TensorFlow, and LLM APIs, enabling operations such as metadata generation, filtering, and vector search. The library also features built-in parallelization, out-of-memory compute, and vectorized operations on Python object fields.

OASIS: A UNIVERSE IN A TRANSFORMER

Oasis is a fully AI-generated, real-time open-world video game model that uses a 500M parameter foundation model without a game engine. It employs fast transformer inference for generating gameplay and is optimized for Etched's Sohu ASIC to achieve high frame rate efficiencies. While demonstrating significant potential, challenges such as domain generalization and long-context consistency remain.

BROWSER-USE

This open-source web automation tool allows LLMs to interact with websites via a simplified interface. It supports models like GPT-4o and Claude 3.5 Sonnet. Key features include multi-tab management, XPath extraction, and customizable actions. The tool enables seamless web navigation and data extraction. Its limitations include message length, which affects LLM speed and task repetition. Ongoing development will focus on robustness and cost reduction.

12 HOUR CUDA COURSE (YOUTUBE PLAYLIST)

A great course from freeCodeCamp on CUDA programming from start to finish.

1PRICE (PRODUCT)

1Price.co allows any company to dynamically change prices on autopilot with AI to maximize revenue per customer.

CONTEXT AUTOPILOT (PRODUCT LAUNCH)

Context Autopilot is a unified AI workspace.

Schedule a demo to learn more

QWEN 2.5 CODER 32B INSTRUCT IS HERE!

The Qwen 2.5 Coder family is a series of language models trained specifically for code. The newest 32B parameter model is better than GPT-4o while being small enough to run locally for many folks. It is on par with Claude Sonnet 3.5 in many benchmarks as well.

X PREPARES TO OPEN AI CHATBOT 'GROK' TO FREE USERS

X's AI chatbot Grok, initially available only to paying subscribers, may soon be accessible to everyone.

GOOGLE'S FLOOD FORECASTING AI TO REACH 700 MILLION PEOPLE

Google's AI-based flood forecasting is expanding to over 100 countries, covering 700 million people.

STDGEN

StdGen is a new method for generating 3D characters from a single image. It decomposes the problem into separable pieces (like hair and jackets) which improves the quality of output.

MIXTURE-OF-TRANSFORMERS: A SPARSE AND SCALABLE ARCHITECTURE FOR MULTI-MODAL FOUNDATION MODELS

The Mixture-of-Transformers (MoT) architecture introduces a sparse multi-modal transformer that decouples parameters by modality (text, images, and speech), enabling efficient processing while maintaining performance quality. Across multiple evaluations, including the Chameleon 7B and Transfusion settings, MoT achieves comparable or better performance than dense baselines while using significantly fewer computational resources - as low as 37.2% of the FLOPs for speech processing and 47.2% of the wall-clock time for image generation.

PROTEIN MODELING WITH MULTIMODAL ALIGNMENT

This study explores how to improve alignment between LLMs and protein-focused geometric deep models for better cross-modal understanding.

ALPHAFOLD 3

DeepMind has open sourced the code and weights of AlphaFold 3 for research use. It is a substantial improvement over previous state-of-the-art and could accelerate AI for science applications.

CONTINUAL LEARNING WITH ONLINE-LORA

Online-LoRA is a new framework designed to prevent catastrophic forgetting in online continual learning (OCL) by fine-tuning pre-trained Vision Transformers (ViTs) in real time without needing rehearsal buffers.

MARKER DETECTION IN DIFFICULT LIGHTING

DeepArUco++ introduces a deep learning-based approach to improve fiducial marker detection under challenging lighting conditions where traditional methods often fail.

CAN LLMS FOLLOW THREADS THROUGH NEAR-MILLION-SCALE HAYSTACKS?

Large Language Models (LLMs) with expanded context windows enable broader applications. New research across 17 leading LLMs reveals that while many models can effectively handle multiple concurrent information threads, their actual effective context limits are often shorter than their theoretical maximum context lengths. Many models exhibit "thread-safety" (handling multiple information threads simultaneously without performance degradation), though accuracy tends to decline as context windows expand toward their limits.

THE CASE FOR TARGETED REGULATION

AI advancements are rapidly improving capabilities in fields like mathematics, coding, and science, increasing both opportunities and risks. Controlled regulation is essential for managing potential misuse in areas like cybersecurity and CBRN. Anthropic's Responsible Scaling Policy calls for transparency and a careful legislative approach that balances safety with innovation.

HERMES 3

Hermes 3, fine-tuned from Llama 3.1, excels in reasoning and creativity. It demonstrates superior performance with models of 8B, 70B, and 405B parameters. The model unlocks new capabilities in AI alignment and artificial consciousness.

AI LEADS PYTHON TO TOP LANGUAGE AS THE NUMBER OF GLOBAL DEVELOPERS SURGES

Python has ascended to the top language on GitHub due to a surge in AI and machine learning projects.

GOOGLE ACCIDENTALLY LEAKED A PREVIEW OF ITS JARVIS AI THAT CAN TAKE OVER COMPUTERS

Google's new AI prototype, Jarvis, briefly appeared on the Chrome Web Store.

AI-POWERED PARENTING IS HERE AND A16Z IS READY TO BACK IT

Andreessen Horowitz is investing in AI-powered "parenting co-pilots" that use LLMs, such as Cradlewise and Nanit.

Read the whitepaper