Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Sponsored
When you started building a product, you didn’t dream of endless admin and organization tasks. You wanted to make something people would love. Meet Height [11], an AI project tool to manage the work for you, invisibly, automatically, and autonomously.
ControlVectors are a way to cheaply steer semantic generation output. They are cheaper to train than LoRA but can still be quite effective. This library makes it easier.
This is a Ray-based implementation of RLHF for Llama-style models, which includes Mistral. It contains a number of PPO stabilization tricks to improve performance.
The 3D Diffuser Actor combines diffusion policies and 3D scene representations to improve robot manipulation. This AI-driven approach allows robots to understand and interact with their environment more effectively.
AdamW is commonly used to decouple learning rate and weight decay. However, the common implementation in Pytorch doesn’t explicitly do this. This post discusses how to tune these parameters.
Ever wondered how OpenAI is able to innovate and build at such a fast pace? In this podcast, Logan Kilpatrick, OpenAI's head of developer relations, discusses OpenAI's unique culture, the importance of high agency and urgency, and the company's decision-making framework for launching products.
A new Temporal Dilated Video Transformer (TDViT) has been developed to improve the analysis of dense video tasks, such as frame-by-frame video object detection.
OpenLLMetry-JS is a set of extensions built on top of OpenTelemetry that gives you complete observability over your LLM application.
A list of full clusters that are available for rent by the hour.
Ensure your emails land in the inbox, not spam.
Sponsored
Learn how to successfully build generative AI pilot applications. Analysis and recommendations for product and IT leaders.
Anthropic is testing Prompt Shield, a technology designed to redirect U.S. users of its chatbot Claude seeking political and voting information to authoritative sources like TurboVote.
OpenAI has been reportedly developing two types of AI agent software for over a year. The first type can be used to automate complex tasks by taking over a customer's device. The second AI agent class handles web-based tasks and can gather public data. It is unclear when the company plans to release these agents.
Google released a new MoE model that matches Gemini 1.0 Ultra in performance but scales up to 1m tokens in context while using less compute due to its smaller size. It is natively multimodal.
It is often challenging to know which examples should be used when aligning language models using preference data. This work suggests a surprisingly robust baseline - choose the 1,000 longest examples.
Diffusion models can be repurposed for their broad ‘knowledge’ of the world as they get better at synthesizing images and videos. This paper found a phenomenal 0.02 bits per pixel compression. The key trick here was to measure perceptual similarity along the way and resend an original video frame as needed.
Researchers have created OpenMathInstruct-1, a new dataset for training open-source Large Language Models in math, matching the performance of closed-source models. This breakthrough, featuring 1.8 million problem-solution pairs, opens the door for more accessible and competitive math instruction AI tools.
Andrej Karpathy released a minimal, clean, and extensible implementation of the byte pair encoding used in language model tokenizers.
GPTScript is a new scripting language that automates interactions with OpenAI large language models. The project's ultimate goal is to create a fully natural language-based programming experience.
These models, which look similar to Llama 2, are trained on 3T tokens and excel at a number of tasks. The Qwen team has released chat versions and quants. Excitingly, the models seem to excel at reasoning, math, and code.
A list of 30 papers that relate to the newly released Sora video model.
Over 75% of public market trades (AUM of $1T+) are now driven by data-driven algorithms, a revolution initiated by hedge funds in the 90s. Today, that data-driven approach is rapidly infiltrating venture capital, with projections indicating that over 75% of VC deal reviews will involve AI and data analytics by 2025, transforming how investments are sourced, evaluated, and managed.
How we talk to each other, collaborate, and undertake creative tasks has meaningfully shifted since we introduced software. We’re starting to see the beginnings of another meaningful shift with AI. How substantial this shift will be is being underestimated. Startups born with AI integration in their products from day one will have a huge advantage over existing companies adding it on top of their existing products.
OpenAI has achieved an annual revenue run rate exceeding $2 billion propelled by the immense success of ChatGPT, making it one of the fastest-growing tech firms. With strong interest from enterprise clients looking to adopt Generative AI, OpenAI aims to more than double this figure in 2025.
OpenAI CEO Sam Altman is working to secure US government approval for his chip project as it risks raising national security and antitrust concerns.
Chat with RTX now supports multiple file formats and can import content directly from YouTube playlists for convenient content querying.
* Gain industry-specific GenAI & MLOps insights * Overcome upskilling challenges with tailored premium courses * Dive deep into the evolving AI career landscape
Sponsored
Feeling overwhelmed by the relentless pace of innovation in AI? Struggle no more! Join INTERVIEW KICKSTART’S ADVANCED GENERATIVE AI AND MLOPS MASTERCLASS [5] on FEB 18 AT 5 PM PST, led by Parivesh Priye (AI Scientist, FAANG), and unlock the door to cutting-edge skills and career clarity.
OpenAI released a new video generation model with stunning quality. It is currently available for red teamers to test.
Lambda Labs raised a $320M Series C to continue building a powerful GPU cloud.
The United States Patent and Trademark Office has issued guidance stating that only natural persons who significantly contribute to an invention's conception can be named as inventors. The guidance explicitly excludes AI systems from inventorship despite their role in the creative process.
Meta used large language models to write tests for its code base. It found substantial improvements in test coverage and overall code quality.
Meta’s VP & Chief AI Scientist Yann LeCun has previously said that we need more data-efficient self-supervised models for general intelligence. This work is one step towards that, with models trained on video to understand aspects of the world. The models are publicly available.
Researchers have developed a new video compression method using diffusion models that generate high-quality video frames while maintaining low data rates.
The Google team has released code to help with atmospheric modeling. The code is used as the core in DeepMind’s recent weather modeling tools.
This is a Python script for plotting the intermediate layer outputs of Mistral 7B. It produces a 512 x 256 image representing the output at every layer of the model.
Researchers have introduced CLIP-MUSED, a novel method for decoding visual information from the brain across different individuals.
Upstash aims to be a crucial part of AI applications. It offers serverless solutions for vector databases, caching, and rate limiting. Upstash is set to become an even more integral component of AI technology infrastructure with its new vector database.
SLEB is a new method to streamline large language models by pruning redundant transformer blocks.
Cohere’s non-profit research lab C4AI released a model called Aya, a new state-of-the-art open-source massively multilingual research LLM that covers 101 languages – including more than 50 previously underserved languages.
Sam Altman is the sole owner of OpenAI’s venture capital fund.
RoCo is a robust cache omission policy that enhances the efficiency of large language models in limited-resource environments.
Teach AI to do anything in the browser.
from actual humans which filters out tons of junk data. * ACCURATE AND UP TO DATE: The index is refreshed with tens of millions of pages daily [5]. * AFFORDABLE: free for up to 2,000 queries/mo [5]; paid plans start from $5/mo.
Sponsored
Looking for high quality, ethically-sourced data for your AI applications? The Brave Search API [5] is an affordable way for developers to access an independent index of the Web [5] with over 20 billion webpages.