Search the archive

FLASH LINEAR ATTENTION

A collection of high-performance Triton kernels for state-of-the-art linear attention models and variants.

VISION-LANGUAGE MODEL EVALUATION WITH UNIBENCH

UniBench is a unified framework that simplifies the evaluation of vision-language models (VLMs) by combining over 50 benchmarks into a single implementation. It helps assess the progress of VLMs across various capabilities, from object recognition to spatial awareness.

AGENT Q

Multion has trained an agent to use self play to perform web queries. Over the course of training, it improved from 18% to 81% on a variety of web-based tasks like ordering food. It uses MCTS and DPO to improve. Researchers from Stanford participated in this work as well and a paper is available on the website. It appears to be built on the xLAM function calling model from Salesforce Research.

INTERACTIVE SEGMENTATION WITH CLICK ATTENTION

A new click attention algorithm improves interactive segmentation. This approach broadens the influence of positive clicks and reduces interference between clicks.

DOES AI DESERVE A SEAT AT THE BOARDROOM TABLE?

AI's integration into corporate governance is prompting leaders to devise robust AI strategies for data-driven decision-making. While AI offers valuable insights, especially with LLMs, challenges remain, including skill gaps and ethical concerns. A proper balance between AI and human judgment is essential for future C-suite decision-making processes.

GEMINI 1.5 FLASH PRICE DROP WITH TUNING ROLLOUT COMPLETE

Gemini 1.5 Flash has undergone a price drop, with a 78% decrease on input and 71% cut on output token costs, and its API now supports over 100 languages.

ROBOCARS PROMISE TO IMPROVE TRAFFIC EVEN WHEN MOST OF THE CARS AROUND THEM ARE DRIVEN BY PEOPLE

AI and reinforcement learning algorithms can improve mixed traffic flow - simulations show that a 5% presence of robot vehicles eliminates jams.

PREDICTION MARKETPLACE POLYMARKET PARTNERS WITH PERPLEXITY TO SHOW NEWS SUMMARIES

Polymarket has partnered with AI search engine Perplexity to integrate event-related news summaries and data visualizations into its prediction marketplace.

YOUTUBE IS TESTING A FEATURE THAT LETS CREATORS USE GOOGLE GEMINI TO BRAINSTORM VIDEO IDEAS

YouTube is trialing 'Brainstorm with Gemini,' a feature that helps creators generate video ideas and thumbnails using Google's AI. Available to selected creators for testing, this tool could differentiate YouTube from competitors by leveraging AI for content creation. The platform is evaluating creator feedback before deciding on a wider release.

OPENAI GENERATES MORE TURMOIL

OpenAI's founding team is experiencing significant turnover, with only 2 of the 11 original members currently active, as concerns grow over the organization's shift away from its initial non-profit ideals toward a more profit-driven structure. This exodus includes co-founders Greg Brockman (on sabbatical) and Ilya Sutskever (who has left), amid speculation of burnout and lucrative secondary financial rewards. The organization faces challenges as it may require a new major cash partner and anticipates delays in the release of GPT-5, while the industry considers the merits of "open" versus "closed" AI models.

FORGET MIDJOURNEY — FLUX IS THE NEW KING OF AI IMAGE GENERATION

Flux AI, by Black Forest Labs, has emerged as the latest promising open-source AI image generation tool. It is capable of running on consumer-grade laptops. It excels in rendering people and prompt adherence, outperforming competitors like Midjourney in some aspects. The model is available in Pro, Dev, and Schnell versions, with a forthcoming text-to-video model announced as open-source as well.

GEMMA SCOPE

A few weeks ago, DeepMind released a number of sparse autoencoders on the Gemma 2 suite of models. This is now the companion paper where researchers discuss the training paradigm and some interesting results.

EVENT STEREO MATCHING

Researchers propose a method to improve event stereo matching by integrating a stereo event camera with a fixed-frequency LiDAR sensor.

A NEURAL SOLVER FOR PDES

The UGrid solver is a newly developed neural solver for linear Partial Differential Equations (PDEs) that combines the strengths of U-Net and MultiGrid techniques.

for smarter process orchestration. Read it free, here.

LLM-AIDED OCR PROJECT

The LLM-Aided OCR Project improves OCR accuracy using advanced NLP and LLMs, supporting features like error correction, markdown formatting, and header/page number handling. It works with both local and API-based LLMs, offering asynchronous processing and quality assessment.

A FOUNDATION MODEL FOR ECG ANALYSIS

ECG-FM is a transformer-based foundation model designed to improve ECG analysis by reducing the need for extensive labeled data.

OPEN-VOCABULARY SEGMENTATION

ProxyCLIP is a new framework designed to improve open-vocabulary semantic segmentation by combining the strengths of CLIP models and Vision Foundation Models.

PAID APPLE INTELLIGENCE FEATURES ARE LIKELY AT LEAST 3 YEARS AWAY

Apple may eventually charge for advanced Apple Intelligence features, but this is expected to be at least three years out. Its initial AI offerings will remain free as the company develops more sophisticated functionalities. Current features, such as an updated Siri, run on-device, suggesting Apple is still catching up in AI.

KLARNA'S AI CHATBOT: HOW REVOLUTIONARY IS IT, REALLY?

Klarna integrated an AI chatbot, developed with OpenAI, that demonstrates considerable efficiency in customer service tasks, potentially reducing its support staff needs. The bot swiftly handles typical Level 1 support queries in 23 markets and 35+ languages but escalates more complex issues to human agents. While the technology saves costs and streamlines first-level support, its revolutionary impact within the business context is debatable compared to prior L1 support automation.

WHY I BET ON DSPY

DSPy is an open-source tool that can orchestrate multiple LLM calls to tackle real problems. The framework focuses on verifiable feedback for outcome measurement and is evolving to address current reliability and accessibility challenges. Despite limited reasoning capabilities, LLMs can excel as creative engines within the DSPy system.

AIJOBS.COM (PRODUCT)

The aijobs.com platform is a simple marketplace specifically designed for matchmaking AI jobs with top AI talent.

ELON MUSK TO PAUSE X'S AI TRAINING ON SOME EU DATA

Irish Data Protection Commissioner Des Hogan has initiated legal action against an unnamed company, challenging its handling of EU users' personal data and potentially impacting its AI chatbot's data processing practices under GDPR scrutiny.

INTEL IS BRINGING GPUS TO CARS

Intel has announced the Arc A760A, a discrete GPU for cars aimed at enhancing in-vehicle entertainment with AI-powered features such as voice and gesture recognition.

UBER HIGHLIGHTS AUTONOMOUS VEHICLE EFFORTS NOW THAT TESLA'S IN ITS REARVIEW MIRROR

Uber's Q2 results emphasized its growing AV segment, highlighting a 6x rise in autonomous trips year-over-year and partnerships with AV leaders like Waymo and Alphabet. CEO Khosrowshahi expressed confidence in Uber's global AV acquisition strategy, distancing from a winner-take-all approach. The presentation also hinted at future AV developments, including a collaboration with BYD on autonomous-capable EVs.

MISTRAL: BUILD, TWEAK, REPEAT

La Plateforme introduces customizations for LLMs, including Mistral Large 2 and Codestral, offering developers the ability to fine-tune models with specific domain knowledge. An alpha release of 'Agents' provides layered, complex workflows integrated with Mistral Large 2 capabilities. Mistralai SDK has hit a stable 1.0 release for Python and Typescript, improving usability and consistency.

ZICO KOLTER JOINS OPENAI'S BOARD OF DIRECTORS

Zico Kolter, a Professor at Carnegie Mellon University and expert in AI safety and robustness, has joined OpenAI's Board of Directors and its Safety and Security Committee. His extensive research in AI safety, alignment, and model robustness will enhance OpenAI's efforts to ensure AI benefits humanity.

TREE ATTENTION

This is an extension of Ring Attention, which allows for extremely long context by scaling across multiple GPUs. The researchers derive an energy function to inform how models should be sharded.

REDUCING BIAS IN LARGE LANGUAGE MODELS

Bias-aware low-rank adaptation (BA-LoRA) is a new technique for fine-tuning LLMs that addresses bias propagation from pre-training data.

TEMPORAL EVENT FORECASTING WITH MULTIMODAL LLMS

Researchers explore the potential of images in improving temporal event forecasting using LLMs. Their proposed framework, MM-Forecast, identifies two crucial roles of images — highlighting and complementing textual data.

SQLITE-VEC

A vector database built on top of the powerful SQLite technology. It can handle millions of queries and has a nice API for vectors.

PUFFERLIB

PufferLib is a wrapper and accelerator for reinforcement learning libraries to unify the interface. It is extremely fast and contains a number of useful baselines.

LOW-RESOURCE LANGUAGE MODELS

Trans-tokenization is a cross-lingual strategy that enhances language models for low and mid-resource languages by leveraging high-resource language data.

DON'T PIVOT INTO AI RESEARCH

Scale has become the predominant driver of performance in the field of AI and machine learning. The substantial capital required has led to market consolidation, with only a few suppliers capable of employing productive ML researchers. This dynamic mirrors the historical consolidation seen in chip design, suggesting a future downturn in the prestige and salaries for ML roles as supply outstrips demand. Aspiring ML professionals should carefully assess their career motivations in light of these industry trends.

SONOVA'S AI HEARING AIDS OFFER CRYSTAL-CLEAR SPEECH IN NOISY PLACES

Sonova has introduced Phonak Audéo Sphere, a hearing aid featuring AI and dual-chip technology that promises a 53x boost in speech understanding in noisy environments. Developed over years, the platform combats the key challenge for hearing aid users — clarity in noise — using the DEEPSONIC chip with advanced DNN capabilities. With this leap in tech, Sonova aims to significantly improve the quality of life for the hearing impaired.

APPLE INTELLIGENCE MAY COME TO EU AFTER ALL

Apple's AI features in the upcoming macOS Sequoia will be available for Mac users in the EU, unlike on iOS and iPadOS 18, per the latest beta release notes. The EU exclusion, which is due to Digital Markets Act compliance issues, does not apply to Macs. Mac users will likely be able to access Apple Intelligence if they set their system to U.S. English.

WAYMO IS EXPANDING ITS ROBOTAXI SERVICE AREAS IN SAN FRANCISCO AND LOS ANGELES

Waymo is expanding its robotaxi service areas in San Francisco and Los Angeles, indicating confidence in its autonomous vehicle technology.

INTEL REPORTEDLY GAVE UP A CHANCE TO BUY A STAKE IN OPENAI IN 2017

Intel reportedly passed on acquiring a stake in OpenAI, which is now a significant player in AI, during 2017-2018 due to then-CEO Bob Swan's hesitation on AI market readiness.

HEY AI (PRODUCT LAUNCH)

Hey AI leverages AI-to-AI based engagements to disrupt the traditional dating app.

3️⃣ Save your virtual spot [6] at API Summit by Kong (FREE)

HUGGING FACE ACQUIRES XETHUB

Hugging Face serves and stores a lot of data, most of it in LFS. XetHub has written its own, powerful, alternative for scaling Git repositories.

HUMANE'S DAILY RETURNS ARE OUTPACING SALES

Humane's AI Pin faced a significant return rate post-launch with negative reviews, leading to approximately 7,000 of the sold units remaining with customers. Despite raising over $200 million from tech luminaries, Humane struggled with over $1 million in returns and internal leadership turnover. Amidst financial woes, there's speculation about a potential HP acquisition and new debt fundraising efforts.

GPT-4O SYSTEM CARD

Aligning a voice system is challenging. This post highlights the ongoing efforts to ensure the multimodal model is safe and helpful.

QWEN2-MATH

The Qwen team has released a 72B math model that outperforms all open and closed models on MATH. It also outperforms Llama-3.1-405B on a number of reasoning benchmarks. It is limited to English only, with multilingual models coming soon.

SEMANTIC SEGMENTATION IN PATHOLOGY

SAM2-PATH is a new framework designed to enhance semantic segmentation in digital pathology.

A MULTILINGUAL DATASET FOR SPOKEN LANGUAGE UNDERSTANDING

Speech-MASSIVE is a new multilingual Spoken Language Understanding (SLU) dataset. It offers a speech-based counterpart to the MASSIVE textual corpus.

PYTORCH FLEXATTENTION

PyTorch has released a new API that allows any type of attention variant to be coded and compiled to Triton. This allows for better portability, performance, and research velocity on types of attention.