Search the archive

META'S OPEN AI HARDWARE VISION

Meta showcased open AI hardware designs, including the Catalina rack and the expanded Grand Teton platform, at the OCP Global Summit. Training the Llama 3.1 405B model required 16,000 NVIDIA H100 GPUs, highlighting Meta's scaling infrastructure. Open AI hardware systems are required to continue advancing AI capabilities.

THE NEW YORK TIMES WARNS AI SEARCH ENGINE PERPLEXITY TO STOP USING ITS CONTENT

The New York Times has issued a cease and desist to AI startup Perplexity, accusing it of using its content without permission for AI search operations. Perplexity claims it doesn't scrape content for training but indexes web pages to surface factual content. The company is engaging with publishers and aims to resolve the issue by working with the Times and other media outlets.

DECAGON RAISES $65M SERIES B

Decagon has raised $65 million in Series B funding to enhance its AI customer support agents, which are already used by companies like Duolingo and Eventbrite to efficiently manage customer interactions. Its AI agents automate repetitive tasks, enabling customer support teams to focus on higher-value roles. The funding will bolster Decagon's engineering team and expand its AI solutions into new markets and verticals.

LINEARIZING LLMS

Another great post from Hazy research exploring how to linearize existing language models without sacrificing too much performance

SCENE TEXT EDITING WITH STYLE CONSISTENCY

TextCtrl is a new diffusion-based method that improves scene text editing by maintaining both content accuracy and style preservation.

HUMANOID ROBOTS WITH ENHANCED 3D VISUOMOTOR SKILLS

iDP3 is an improved 3D visuomotor policy that allows humanoid robots to autonomously operate in diverse real-world environments.

TABLED

State-of-the-art system for detecting and extracting tables from PDFs.

ENHANCED IMAGE GENERATION EFFICIENCY

HART is a new visual generation model that creates high-quality 1024x1024 images, challenging the capabilities of diffusion models. It improves image reconstruction and lowers training costs by using a hybrid tokenizer that combines discrete and continuous tokens.

A NEW VISION TRANSFORMER WITH ENHANCED ATTENTION

The Deformable Bi-level Routing Attention (DBRA) module was developed to improve attention mechanisms in vision transformers. DeBiFormer, built on DBRA, optimizes the selection of key-value pairs and improves query interpretability in attention maps.

NEW HIGH QUALITY AI VIDEO GENERATOR PYRAMID FLOW LAUNCHES

Pyramid Flow is a new open-source AI video generator that creates high-quality videos using a pyramidal flow matching technique, optimizing computational efficiency. It competes with proprietary models like Runway Gen-3 Alpha and offers commercial use under the MIT License, though it requires users to host their own version for full application integration. Despite some limitations, its launch highlights the trend toward accessible AI video solutions.

THIS THREE-PERSON ROBOTICS STARTUP IS WORKING WITH DESIGNER YVES BÉHAR TO BRING HUMANOIDS HOME

Kind Humanoid's three-person team is developing a whimsical humanoid robot called Mona. It is designed for home use rather than industrial settings. The team plans to field-test an initial dozen prototypes next year. Focusing on innovation and efficiency over fundraising, Kind's approach contrasts with other AI-driven robotics manufacturers targeting industrial markets.

INTELLECT–1: LAUNCHING THE FIRST DECENTRALIZED TRAINING OF A 10B PARAMETER MODEL

INTELLECT-1 is the first decentralized 10-billion-parameter model. It leverages global contributions for open-source AGI development. OpenDiLoCo scaling has been achieved for training large models across distributed devices, with new advancements in bandwidth efficiency and fault tolerance. The new Prime framework further optimizes decentralized training, increasing compute utilization to 98% for the INTELLECT-1's 10-billion-parameter training run.

TESLA TO OFFER FIRST GLIMPSE OF A ‘ROBOTAXI' THAT DRIVES ITSELF

Tesla has unveiled a prototype self-driving taxi, the Robotaxi, aiming for fully autonomous operations despite skepticism from experts about the near-term feasibility of such technology.

BYTEDANCE LAYS OFF HUNDREDS OF TIKTOK EMPLOYEES IN SHIFT TO AI CONTENT MODERATION

TikTok is shifting towards AI for content moderation, leading to layoffs in Malaysia affecting under 500 employees.

MICROSOFT ARTIFICIAL INTELLIGENCE VP BUBECK TO JOIN OPENAI

Sebastien Bubeck, a Microsoft AI VP focused on small language models, is leaving to join OpenAI.

ADOBE ANNOUNCED VIDEO MODELS

Adobe announced the expansion of its Firefly family of creative generative AI models to video during the Adobe MAX conference.

OPENAI APPOINTS INTERNATIONAL EXPANSION BOSS

OpenAI has appointed Oliver Jay to lead international expansion, focusing on AI strategy and operations. The company announced a new APAC office in Singapore and is developing local language datasets. The o1 model, which utilizes "chain of thought" techniques, is aimed at enhancing AI accuracy.

MICROSOFT MAY HAVE AN AUDIO-TO-IMAGE GENERATOR IN THE WORKS

Microsoft has filed a patent for an AI system that converts live audio into images using LLMs. The system aims to enhance communication by generating real-time images from audio streams. If developed, it could be integrated into Microsoft Teams via Copilot.

10B DISTRIBUTED TRAINING RUN

Decentralized training of language models is the process of using heterogeneous and non-co-located hardware to train a single model. This is traditionally challenging due to the high cost of memory transfers. This work improves on the state-of-the-art algorithms for highly distributed training.

VISION-LANGUAGE SEGMENTATION

OneRef introduces a streamlined framework for referring segmentation by unifying visual and language feature spaces into a single transformer.

MOTION PREDICTION WITH SMARTPRETRAIN

SmartPretrain introduces a flexible, model-agnostic, and dataset-agnostic self-supervised learning framework for improving motion prediction in autonomous vehicles.

UVA - AN INTRODUCTION TO GROUP EQUIVARIANT DEEP LEARNING

Study resources for learning more about deep learning techniques over certain types of geometric data in the face of architectural constraints.

DIFFUSION MODEL SIMULATING CS:GO

An open replication of a diffusion model that simulates the visuals of a video game while taking keyboard and mouse controls as input.

IMPROVING LLM PREFERENCE ALIGNMENT WITH REWARD-AWARE LEARNING

This study tackles the limitations of current alignment algorithms in LLMs, which often overfit to relative preferences and overlook the quality of responses. By introducing reward-conditioned LLM policies, the authors propose a novel data relabeling technique that factors in response quality, helping the model generalize to optimal responses.

ANTHROPIC CHALLENGES OPENAI WITH AFFORDABLE BATCH PROCESSING

Anthropic launched a Message Batches API that allows businesses to process large data volumes at half the cost of standard API calls. The API supports up to 10,000 asynchronous queries within 24 hours, offering a cost-effective option that shifts AI processing from real-time to "right-time." This move promotes AI adoption among mid-sized businesses, though it risks diverting focus from advancing real-time AI capabilities.

AL WILL TAKE OVER HUMAN SYSTEMS FROM WITHIN

In this post, Yuval Noah Harari, an Israeli historian and the author of “Sapiens,” “Homo Deus,” and “Nexus,” discusses the role of information networks and AI in shaping narratives that can unify or divide societies. He warns that AI, acting as an "alien intelligence," may centralize power and that it lacks self-correcting mechanisms, posing risks to democratic structures. Harari emphasizes the need for strong institutions to manage truth in a landscape dominated by powerful AI systems that influence decision-making across various sectors.

STICKY HUMANS IN A POST-AGI WORLD

AI tutors face significant challenges in replicating the social and intellectual engagement provided by human teachers. Despite advancements, AI struggles with nuanced educational tasks and lacks the ability to provide the socio-intellectual experiences that humans offer. A hybrid approach, where AI augments rather than replaces human educators, may be more effective due to the inherent social and cultural components of learning.

users. Start for free

ENTROPIX VISUALIZATION

Entropix is a tool that changes the sampling behavior of language models.

OPENAI PROJECTIONS IMPLY LOSSES TRIPLING TO $14 BILLION IN 2026

OpenAI projects losses to rise to $14 billion in 2026, with total losses reaching $44 billion by 2028.

AMD LAUNCHES AI CHIP TO RIVAL NVIDIA'S BLACKWELL

ZYPHRA'S MAMBA 2 BASED MODEL BEATS MISTRAL

The first state space style model that outperforms transformers at the 7B scale. It excels at long context understanding and generation due to the linear time scaling of the Mamba 2 blocks.

OPENAI'S SWARM

OpenAI has released a lightweight framework for communication between agents. It won't be maintained going forward, but could serve as useful inspiration.

SILICON VALLEY IS DEBATING IF AI WEAPONS SHOULD BE ALLOWED TO DECIDE TO KILL

Debate over fully autonomous weapons continues as Anduril's Palmer Luckey expresses openness to them, contrasting with Shield AI's previous stance against AI making lethal decisions. The U.S. government remains non-committal, with no binding ban on such weapons, amidst fears that adversaries like China or Russia might develop them first. Defense tech companies, including Anduril and Palantir, are lobbying heavily to influence policy on the integration of AI in military systems.

TEXT-TO-IMAGE MODEL ON PUBLIC DATA

EvolveDirector aims to create a competitive text-to-image generation model using publicly accessible resources, bypassing the restrictions of proprietary models.

EVALUATE VISIBLE AND INFRARED IMAGE FUSION

Researchers introduce the Segmentation-oriented Evaluation Approach (SEA) to better assess Visible and Infrared Image Fusion (VIF) methods, which are crucial for tasks like object detection and semantic segmentation.

DEEP GENERATIVE MODELS: A GUIDE FOR TRANSPORTATION RESEARCH

A gentle introduction and tutorial on deep generative models in transportation research.

REALISTIC 4D SCENE GENERATION

Trans4D is a new framework designed to tackle the challenges of realistic 4D scene transitions that offers improved text-to-4D synthesis.

DOCUMENT-LEVEL TRANSLATION AGENT

DelTA, which is short for Document-levEL Translation Agent, is an online document-level translation agent based on multi-level memory.

FAST 3D COMPRESSION

Fast Compression of 3D Gaussian Splatting (FCGS) is a new model that skips the time-consuming per-scene optimization required by previous methods. FCGS uses a quick feed-forward pass, cutting compression time drastically from minutes to seconds.

MACHINES OF LOVING GRACE

Dario Amodei, the CEO of Anthropic, is a prolific writer. He often writes internal memos. This one was published externally and imagines the extreme upside that could exist in the world if we succeed at building powerful AI systems.

THIS AI-POWERED INVENTION MACHINE AUTOMATES EUREKA MOMENTS

Iprova's AI-driven software analyzes diverse technical literature to suggest patentable inventions by connecting previously unlinked ideas, aiding companies like Procter & Gamble and Panasonic. The tool uses semantic search and generative AI to identify novel inventions and relies on human insight for practical application, focusing on industries with rapidly changing technologies. Despite utilizing AI, Iprova emphasizes human creativity in refining and validating invention ideas.