Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
OpenAI formed a Safety and Security Committee after announcing the training of its new foundation model. This committee will be tasked with issuing recommendations to the board about actions to take as model capabilities continue to improve.
Jan Leike, a former OpenAI researcher who resigned over AI safety concerns, has joined Anthropic to lead a new "superalignment" team focusing on AI safety and security. Leike's team will address scalable oversight, weak-to-strong generalization, and automated alignment research.
Microsoft unveiled new features for Copilot at Build 2024, including Team Copilot for team collaboration, custom AI Agents to automate workflows, and Copilot Extensions and Connectors for easy customization. These enhancements aim to improve productivity and business process efficiency. The updates are currently in limited private preview, with broader availability expected later in 2024.
Scaling laws are a way to predict how well models will perform at certain sizes with a certain amount of data. They are expensive to obtain. This paper explores using gzip compression ratio as a strong signal to predict a data-dependent scaling law.
A few weeks ago, a new optimizer from Meta made the rounds as a promising alternative to Adam. This paper outlines the method in more detail, including the online updates portion of the method. In general, it seems like a promising result, especially when the total number of desired training steps is unknown at the beginning of training.
Researchers have improved transformer performance on arithmetic tasks by adding embeddings that encode each digit's position relative to the start of the number.
Sponsored
Implement AI to get product and customer sentiment insight based on everyday customer interactions.
without manual work. * _Ask Syncly AI_ to IDENTIFY WHAT DRIVES NEGATIVITY in customer experience. * With _Customer AI, _MONITOR HAPPY/UNHAPPY CUSTOMERS TO PREVENT CHURN.
One of the best depth models is now available as a pipeline in diffusers. This tutorial discusses how to use the model, what you can do with it, and a trick to make it work seamlessly with videos by conditioning on the first frame's latents.
Andrej Karpathy has released an update to LLM C, a single self contained GPT-2 implementation aimed at reproducing the 2019 suite of models. With this most recent update, the library is able to train the smallest of these models in just 90 minutes. It runs end to end and has minimal dependencies.
Content-Style Decoupled Makeup Transfer (CSD-MT) is a novel method for enhancing makeup transfer tasks without relying on real target images.
Vista is a driving world model designed to improve prediction accuracy and action controllability for autonomous driving.
Key members of OpenAI's safety team, including leaders of the superalignment team tasked with ensuring AI alignment, have departed partly due to lost faith in CEO Sam Altman. These departures follow internal conflicts and concerns about OpenAI's safety priorities and the company's direction under Altman's leadership. OpenAI is updating its offboarding documents after criticisms regarding potential equity cancellations for those who refuse to sign non-disparagement agreements.
This article discusses different approaches to AI regulation, contrasting model-level, use-level, and conduct-level frameworks. It argues that conduct-level regulation, which applies existing laws to new technology without excessive specificity, is preferable to use-level regulation, which could result in impractical restrictions and unnecessary complications for AI deployment. The recent EU AI Act's restrictions on AI's ability to infer emotions is cited as an example of the pitfalls of a use-level approach.
OpenAI reversed a policy that would have revoked vested equity from former employees who didn't sign a non-disparagement agreement.
An open-source large action model framework for developing AI web agents.
Opera browser has integrated Google's Gemini AI models into its Aria AI extension, providing more up-to-date information and conversational responses.
xAI has announced a $6 billion Series B funding round led by top investors to advance its AI systems, including the Grok-1 models with enhanced capabilities. The funding will fuel product go-to-market strategies, infrastructure, and R&D.
Apple's AI strategy is focused on practical AI features to improve its products, such as Project Greymatter, a set of AI tools that will be integrated into key Apple apps.
Anthropic researchers have unveiled a method to interpret the inner workings of its large language model, Claude Sonnet, by mapping out millions of features corresponding to a diverse array of concepts. This interpretability could lead to safer AI by allowing specific manipulations of these features to steer model behaviors. The study demonstrates a significant step in understanding and improving the safety mechanisms of AI language models.
Researchers have introduced the Hierarchical Semantic Decoding with Counting Assistance (HDC) framework to improve the Generalized Referring Expression Segmentation (GRES). Unlike previous methods, HDC transfers complementary modality information across granularities and aggregates semantic correspondences for better multi-level decoding.
Scribble2Scene is a new approach for semantic scene completion that reduces the need for extensive labeling.
A new Semantic and Spatial Adaptive (SSA) classifier has been introduced to tackle limitations in semantic segmentation. This innovative approach utilizes coarse masks to guide the adjustment of prototypes, enhancing fine-grained recognition and clarifying mask boundaries.
Sponsored
It's mind-boggling to think about just how many websites ask for your information (shopping, online subscriptions, even airport WiFi…). Surfshark's Alternative ID [10] keeps your email address and name private, minimizes spam emails, and keeps your information out of data breaches. Sign up today [10] for 80% off + 3 months free on the 12-month plan!
Building a RAG system requires you to know how well the retrieval portion is working. This toolkit gives a wide variety of powerful measures of performance for information retrieval.
RSBuilding introduces a new approach to interpreting buildings from remote sensing images, combining building extraction and change detection into one model.
This project introduces a new efficient large language and vision model called Meteor that uses multifaceted rationales to improve understanding and answering capabilities.
Bootloader Studio, a Vietnam-based affective AI and spatial computing startup pioneering the role of emotional intelligence in AI, has emerged from stealth with $5M in seed funding led by Antler Elevate. The company's first consumer product is slated to launch on Apple Vision Pro this year, laying the foundation for a new type of AI model: personalized, user-owned emotional models that allow users to dictate how they interact with businesses, doctors, friends, and more.
As the AI landscape rapidly evolves, AI startups must embrace a marketing-led approach to differentiate themselves and become category-defining leaders. By embracing a marketing-led approach, AI startups can shape the industry narrative, drive market adoption, and establish themselves as visionary leaders in their space.
While some experts believe AGI may be within reach, the consensus is uncertain due to the lack of a coherent theoretical model of general intelligence and an understanding of why or how LLMs function so effectively. The discussion underlines the vast unknowns surrounding AGI, drawing parallels to theology and the empirical approach of the Apollo Program, acknowledging both the potential benefits and the risks associated with AGI.
A helpful guide for creating neural circuit diagrams to explain pieces of AI models.
The AI wearable startup Humane, founded by ex-Apple employees, is seeking a sale of its business for $750 million to $1 billion.
InstructAvatar is a new model that creates emotionally expressive 2D avatars using text guidance.
Sponsored
Fortune 50 companies like AWS and Google Cloud and fast growing startups like Vercel and Secureframe work with TLDR to reach our audience of over 4 million tech-savvy decision makers with highly customized campaigns (like this one!).
Silicon Valley AI firms are fighting to partner with Character.ai, a fast-growing role-playing startup founded by AI pioneer Noam Shazeer. This comes as many large firms are pouring money into smaller companies.
Scarlett Johansson alleges that OpenAI created a voice for ChatGPT that mimics hers without consent, leading her to seek legal advice. OpenAI has since paused using the voice and is discussing the issue with her representatives. The situation highlights concerns over the use of celebrity likenesses in AI applications.
A new research paper details the mapping of AI model Claude 3 Sonnet's inner workings, revealing "features" activated by concepts like the Golden Gate Bridge. By adjusting these features' strengths, researchers can direct Claude's responses to incorporate specific elements, demonstrating a novel method of modifying large language models. The research aims to enhance AI safety by precisely adjusting model behaviors related to potential risks.
The Lyft team used online reinforcement learning, rewarded by driver future earnings, to match drivers with riders. They were able to dramatically improve in real time and made an estimated $30m/year extra for riders.
Evaluating language models is hard and details on how to do it are hard to find outside of the largest firms. This paper shows a reproducible and powerful set of evaluation criteria. It includes a valuable discussion of perplexity evaluation in the appendix.
Researchers introduce a new approach to customize diffusion models for generating identity-preserving images from user-provided references. Unlike traditional methods that require extensive domain-specific training, this technique uses classifier guidance to steer diffusion models without additional training.
Mistral released an official repository to fine-tune its models.
LoRA-Ensemble is a parameter-efficient deep ensemble method for self-attention networks. This technique, which extends Low-Rank Adaptation (LoRA) for implicit ensembling, allows for accurate and well-calibrated predictions without the high computational cost of traditional ensemble methods.
Modular norm is a new method for normalizing weight updates in neural networks that scales training efficiently across different network sizes.
MobileNet is a computer vision model with extreme performance and speed. It can run on edge devices. This blog post outlines the new model and some modern changes that went into its creation.
This project investigates whether language models use multi-dimensional features for computation, challenging the linear representation hypothesis.
GPT-4o, OpenAI's latest AI model, bridges real-time communication between humans and machines, extending capabilities beyond text to include vision and audio. The AI revolution introduces a new wave of human-to-AI and eventual AI-to-AI interactions, likely impacting the dynamics of our social behaviors and business models. As this technology progresses, its effect on human communication will unfold, potentially catalyzing the creation of innovative companies and software solutions.
Thanks to recent advances from Mozilla's Llamafile project, it is now possible to run inference for the flagship model from Mistral at 20 tokens per second on a commodity CPU.
Google's AI search feature is causing frustration by adding an uninvited three-second delay to search results, disrupting the user experience with unwanted information.
Large language models struggle with truly innovative brainstorming, often converging to consensus-based ideas rather than generating novel concepts.