Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Time Series Diffusion Embedding (TSDE) is the first diffusion-based approach for time series representation learning. TSDE segments time series data and uses dual-orthogonal Transformer encoders with a crossover mechanism to generate informative embeddings.
Sponsored
Reach software developers, AI enthusiasts, executives, and other tech professionals reading TLDR every day. TLDR offers 10 interest-based newsletters to help you get in front of your target audience. Learn more [12] about running your first campaign with us.
The Liger Kernel, released surprisingly by LinkedIn, is an efficient set of kernels for language model training. It increases throughput by 20% and decreases memory usage by almost 60% for the popular Llama models. Importantly for practitioners, it only requires 3 lines of code change and integrates with many popular modeling frameworks.
pgvectorscale builds on pgvector with higher performance embedding search and cost-efficient storage for AI applications. It is around 28x faster than competitive and commonly used vector stores.
GenderCARE is a comprehensive framework aimed at assessing and mitigating gender biases. It introduces innovative criteria for evaluating gender bias, emphasizing inclusivity, diversity, and objectivity.
Traditional base language models are quite good at playing chess, averaging around 1800 elo. However, chat models tend to struggle with a precipitous drop in performance. This post explores how to tune chat models like GPT-4o to play chess through a combination of prompting and fine-tuning.
A vulnerability in Slack AI was discovered where attackers can exfiltrate data from private channels they aren't members of via indirect prompt injection. It allows attackers to use public channel messages to manipulate the LLM into divulging private information, such as API keys, in query responses. This issue, alongside a phishing attack vector, persists even after Slack AI's August 14th update incorporating channel and DM files, significantly expanding the risk surface area for such exploits.
Generalized SAM (GSAM) is a new method for fine-tuning the Segment Anything Model (SAM) with variable-size images more efficiently.
Google has released a new SigLIP model that uses a shape optimized vision transformer model architecture.
GaussianOcc is a method for fully self-supervised and efficient 3D occupancy estimation using surround views.
Meta has introduced Llama 3.1, an advanced open-source LLM that brings new capabilities like synthetic data generation and model distillation.
Sponsored
YC-backed Greptile's code review bot [4] summarizes changes in every pull/merge request, and leaves comments that highlight potential issues and unforeseen consequences.
Cursor raised a Series A to continue building its AI-powered coding IDE.
Perplexity AI, an AI-assisted search startup, plans to introduce ads on its search app in Q4 following increased usage and a recent $1 billion valuation.
Google's Gemini AI is a key feature in the Pixel 9 phones, offering users a range of AI-powered functionalities like task assistance, image editing, and screenshot management. However, its performance as a full-fledged assistant is mixed, with occasional glitches and some features from Google Assistant not fully integrated. Despite these issues, unique capabilities such as document summarization and creative photo "reimagining" tools are intriguing benefits for Pixel users.
AI21 has released new versions of its hybrid Transformer and State space model architecture.
This post is a long in-depth summary of what DeepMind is working on in its AGI safety and alignment research efforts.
The new biorecap R package helps researchers manage the overwhelming volume of bioRxiv preprints by using locally run large language models to retrieve and summarize recent publications.
Sponsored
Compliance is the next industry to be disrupted by AI. That's because Compliance work requires expertise, attention to detail, and the patience to handle repetitive tasks with accuracy — making AI a perfect fit. Trustero's patented AI for GRC [11] helps teams complete internal audits in minutes, answer security questionnaires in seconds, and identify compliance and security gaps as they arise. Want to try
Microsoft has open sourced model, code, and checkpoints for a high fidelity atmospheric prediction model.
Researchers have developed a new framework called NuSegDG to enhance the generalizability of nuclei segmentation across different medical images.
Pano2Room is a new method that reconstructs high-quality 3D indoor scenes from a single panoramic image, addressing challenges in single-view 3D scene synthesis.
Computer vision advancements are contributing to the fight against malaria through applications like VectorCam, which allows quick identification of mosquito species and data collection. The app, developed with support from the Gates Foundation, can distinguish malaria-transmitting species and help refine strategies for disease control. These innovations in mosquito surveillance are crucial for the strategic deployment of insecticides and other mitigation measures.
This post shares insights on designing and evaluating robust AI safety measures, drawing parallels between AI control and fields combating insider threats. It highlights the granular control developers can exert over AIs compared to employees, while also noting that unlike humans, AI duplicitousness can be consistently pervasive. Although inspired by various fields like physical security and safety engineering, AI control primarily differs due to its adversarial nature and lack of complex system interactions.
This repository provides a comprehensive overview of embodied learning, a promising approach in robotic manipulation that emphasizes physical interaction and perceptual feedback.
AMD is advancing the AI PC market with its Ryzen AI 300 chips, integrating NPUs to enhance AI-powered features like Microsoft's Recall.
Google is enhancing its AI-powered Gemini writing tools in Gmail, introducing a new "Polish" feature for refining drafts alongside existing options like "Formalize" and "Elaborate".
Ideogram has launched a new model which has superior text rendering and image generation capabilities.
Zed AI is a hosted service that integrates AI-powered coding into the Zed text editor, letting developers use LLMs while maintaining full control over their code. Zed has partnered with Anthropic to allow fast editing with Claude.
Nvidia's AI system ACE will power in-game character interactions in Mecha Break, a mech battle game launching in 2025. It uses GPT-4o and a mix of on-device and cloud-based AI for natural language processing and vocal generation. Initial demonstrations reveal lackluster NPC engagement, with concerns over response delays and uninspiring dialogue.
Amazing new model from Meta that does both next token prediction and diffusion on interleaved text and images. It matches benchmark performance on text and images with previous generation models like Dalle 2 and Llama 2.
It is an industry secret that pretraining models on code helps them generalize in other tasks that require reasoning. This paper from Cohere thoroughly studies that problem and confirms that code is broadly useful as a building block of reasoning.
This paper introduces a new optical flow estimation method that achieves high accuracy while significantly reducing computational costs.
Sponsored
Advancements in AI and robotics open a new horizon for the $81.8BN ice cream industry. Dice Cream [10] is building the world's first autonomous ice cream parlor [10] - serving 144 combinations of ice cream at nearly 1 scoop per second. If you act quickly, you can get an equity stake [10] in Dice Cream at the current valuation.
Repository of state-of-the-art methods for pruning language models that is well maintained with many supported algorithms.
Researchers have introduced a new approach to detecting visual misinformation that focuses on the original meta-context of images, a crucial aspect often overlooked in automated methods.
PA-LLaVA is a domain-specific language-vision assistant designed for pathology image analysis.
An in-depth read into the new Phi 3.5 models recently launched by Microsoft, including the MoE and vision model.
Andreessen Horowitz's latest consumer AI report ranks the top 100 generative AI apps, dividing them into the top 50 AI web products and the top 50 AI mobile apps based on consumer patterns of engagement. The report provides detailed insights into emerging categories, new industry challengers, and trends.
FinalSpark's "Neuroplatform" integrates human-brain organoids into computing, offering a rentable biocomputing platform aimed at reducing AI's energy usage. Challenges include standardizing production and extending organoid lifespan beyond 100 days. Alternatives like cellular computing and fungal networks are also explored for tasks beyond silicon-based systems' capabilities.
This detailed blog post presents seven fundamental rules of causal inference that encompass the relationship between causal mechanisms and observed correlations, explained via causal graphs, logic terms, and R code simulations.
This article discusses the predictability of LLMs and their limitations in creating humor, arguing that while LLMs excel in generating contextually appropriate text, their predictive nature makes them ill-suited for comedic writing, which relies on unexpectedness.
Global use of generative AI has nearly doubled in the past six months, with 75% of knowledge workers now utilizing it.
🕺 WHO'S ATTENDING: Industry experts from DataStax, Priceline, PayPal, NVIDIA, IBM, WorkSpan, and Unstructured
Sponsored
Enjoy great food, drinks, and networking while learning from the successful AI journeys of top companies. Hear about the tools used, lessons learned, and wins on their path to production.
OpenAI has launched fine-tuning for GPT-4o, allowing developers to customize the model for specific use cases with their own datasets. It is offering 1 million free training tokens per day through September 23.
OpenAI has entered a multi-year licensing agreement with Condé Nast, allowing it to display content from the publisher's brands, such as Vogue and The New Yorker, within its products, including ChatGPT and SearchGPT.
Meta FAIR researchers have introduced the Self-Taught Evaluator, a method to train evaluative LLMs without human annotations, potentially enhancing the efficiency and scalability of LLM assessment. Using the LLM-as-a-Judge concept, it iteratively generates and refines responses to create a training dataset, demonstrating improved performance on benchmarks like RewardBench. This technique could enable enterprises to leverage unlabeled data for LLM tuning while acknowledging the importance of a well-aligned seed model and the limitations of benchmarks.