Search the archive

SUPERHUMAN AI (PRODUCT)

AI-powered email for high-performing teams.

PERPLEXITY AI REPORTEDLY RAISING ADDITIONAL MONEY AT SIGNIFICANTLY HIGHER VALUATION CAP THAN $520M

Perplexity AI is in talks to raise new funds at a valuation exceeding its previous $520 million, drawing investors' interest due to its growth in the AI sector and its competition with major firms like OpenAI and Google.

LE CHAT

Mistral AI has launched 'le Chat Mistral,' a new multilingual conversational assistant that utilizes its Mistral models, with an enterprise version for businesses.

ANTHROPIC CLAUDE 3

Anthropic has trained three new models in the Claude 3 family, the strongest of which matches GPT4’s reported benchmark results. It is also a multimodal model and performs well on vision tasks. Importantly, Claude's coding ability is substantially improved with this release.

CHATGPT CAN READ ITS ANSWERS OUT LOUD

OpenAI introduced a Read Aloud feature for ChatGPT, allowing it to verbally respond to users in 37 languages with five voice options and enhancing its multimodal capabilities and usability for on-the-go interactions across web and mobile platforms.

SUPPORTING LONGER CONTEXT LENGTH IN LLMS

Researchers have developed a new technique called Resonance RoPE to help LLMs better understand and generate text in longer sequences than they were originally trained on. This method, which improves on the existing Rotary Position Embedding (RoPE) system, enhances model performance on long texts without extra computing effort.

UNDERSTANDING OBJECT RELATIONS IN IMAGES

The All-Seeing Project V2 introduces the ASMv2 model, which blends text generation, object localization, and understanding the connections between objects in images.

A NEW COMPLEX BENCHMARK FOR AI SYSTEMS

A new dataset called GPQA presents a real challenge with 448 tough multiple-choice questions in biology, physics, and chemistry. Even domain experts struggle, scoring around 65% accuracy, while non-experts only manage 34%. Advanced AI systems like GPT-4 achieve just 39% accuracy. This dataset is aimed at developing methods for supervising AI outputs in complex scientific questions.

production. Catch the latest episode on Spotify now

CLAUDE 3 OPUS FOR VIDEO SUMMARIZATION

Andrej Karpathy issued a long context challenge to make a blog post from a recent video of his. Claude 3 was able to perform this task with some data pre-processing help. The resulting blog post is high quality and interesting.

ENHANCED IMAGE RESTORATION

Dual-domain strip attention mechanism is a new method that significantly improves image restoration tasks.

OPEN SORA PLAN

This project aims to reproduce OpenAI’s text-to-video model Sora. It is currently building out the basic components and searching for interested open-source contributors to assist.

ENHANCING MODELS’ CONFIDENCE WITH SURE

SURE is a novel method that combines various techniques to improve the reliability of uncertainty predictions in deep neural networks, especially for image classification tasks.

AGGREGATOR’S AI RISK

The influence of the printing press on the dissemination of information and the formation of nation-states mirrors the impact of the Internet, particularly through Aggregators like Google and Meta. However, the emergence of generative AI challenges the Aggregator model by providing singular answers that reflect embedded worldviews, potentially undermining the universal appeal that underpins Aggregator economics and suggesting a shift towards personalized AI to maintain their dominance.

IS SYNTHETIC DATA THE KEY TO AGI?

The effectiveness of large language models is primarily influenced by the quality of their training data. Projections suggest that high-quality data will be scarce by 2027. Synthetic data generation emerges as a promising solution to this challenge, potentially reshaping internet business models and highlighting the importance of equitable data access and antitrust considerations.

for Mac

AI RESEARCH INTERNSHIP SEARCH AS A CS PHD STUDENT

This post outlines the process of getting an AI internship. It provides helpful preparation information for coding and research type questions.

PITCHLEAGUE (PRODUCT)

Your personal AI pitch deck coach.

ADOBE REVEALS A GENAI TOOL FOR MUSIC

Adobe's Project Music GenAI Control is a platform designed to generate and customize music from text descriptions or reference melodies.

👉 Start Now 👉

ELON MUSK SUES OPENAI

Elon Musk, co-founder of OpenAI, is suing the company claiming - among other things - that GPT-4 constitutes AGI and is no longer subject to the Microsoft agreement. He also argues for more transparency and openness.

OPENAI WANTS TO MAKE A WALKING, TALKING HUMANOID ROBOT SMARTER

Humanoid robots could be working alongside human manufacturing workers in only a few short years. Figure, a prominent company in the humanoid robot space, recently raised $675 million and struck a new agreement with OpenAI to develop next-generation AI models for its robots. The funding, which was reportedly over $150 million more than the amount Figure had initially sought, will help scale up the company's AI training, robotic manufacturing, and add on new engineers, speeding up the company's timeline for making its product available commercially. OpenAI's technology will help improve Figure's robots' semantic understanding, making them more useful in work scenarios.

HYPERATTENTION: USED IN GEMINI

It is widely believed, and unofficially confirmed, that Gemini’s amazing 1m+ token context window uses HyperAttention to achieve such success.

WHY DO LEARNING RATES TRANSFER?

A theoretical paper that attempts to explain the success of MuP hyperparameter transfer. Its authors find that the largest eigenvalue of the training loss Hessian is independent of the width and depth of the network.

GPT-4 ONLY SUCCEEDS ON 14% OF AGENT TASKS

The community is very excited about the potential of Agents to tackle a variety of digital workloads. However, even the best general-purpose models struggle to complete tasks where humans succeed 70%+ of the time. It is becoming clear that we may need specially trained models for these tasks.

FINETUNE GEMMA 2.4X FASTER WITH UNSOLTH

Unsloth is a project that aims to rewrite important kernels for training and inferencing language models. It has released an example with code for Google’s open-source Gemma model.

3D GEOMETRIC VISION MADE EASY

This project supports constructing 3D representations, in GLB form, from a loosely related collection of images, meaning you can take a few pictures of a location and reconstruct it for 3D uses.

ENHANCING LATENT SPACE SMOOTHNESS IN TEXT-TO-IMAGE DIFFUSION MODELS

A new approach named Smooth Diffusion addresses the challenge of latent space smoothness in text-to-image diffusion models. This method ensures a consistent and gradual change in images in response to minor adjustments in input.

SURFACE NORMAL ESTIMATION

DSNIE is a method that dramatically improves monocular surface normal estimation which can be used in a number of computer graphics applications.

VISUAL PLACE RECOGNITION

CricaVPR introduces a novel approach for improving visual place recognition by focusing on the relationships between multiple images, even under different conditions.

MAGICLAB'S HUMANOID CAN TOAST MARSHMALLOWS, FOLD CLOTHES, AND DANCE

MagicLab has unveiled MagicBot, a humanoid robot that can roast marshmallows and dance. The robot has exceptional balance and dexterity - it has the capability to execute somersaults. MagicBot has sub-millimeter positioning accuracy and a load capacity reaching several kilograms. Videos of the robot in action are available in the article.

EMISSARY SIMPLE BENCHMARKING INFRASTRUCTURE (PRODUCT)

Emissary is releasing a benchmarking platform that enables developers to evaluate prompts and manage model migration and regression on their own terms rapidly and reliably. Generate and expand test sets, define custom metrics, and collaboratively rate outputs manually teams through automated LLM-assisted eval or outsource rating to Emissary.

COHERE OPENS NEW YORK OFFICE

Cohere, a language modeling company known for its strong embeddings, has opened a New York City office to improve collaboration with senior leadership at large companies.

AMAZON TO SPEND $1 BILLION ON STARTUPS THAT COMBINES AI WITH ROBOTS

Amazon's $1 billion industrial innovation fund is set to increase investments in companies that integrate artificial intelligence and robotics, aiming to enhance efficiency across its logistics network.

FIGURE RAISES $675M FOR ROBOTICS

Figure, a robotics company, has raised $675M from several large investors, including OpenAI, to continue building humanoid robotics.

MODULAR MAX DEVELOPERS EDITION

Modular created a toolkit, MAX, to dramatically simplify and speed up AI deployments. It is now available.

GENIE: GENERATIVE INTERACTIVE ENVIRONMENTS

Google DeepMind's Genie is a foundation world model trained on Internet videos that can generate an endless variety of playable (action-controllable) worlds from synthetic images, photographs, and even sketches.

VIDEO SEGMENTATION WITH A UNIFIED ARCHITECTURE

UniVS is a unified architecture for video segmentation that tackles the challenges of various segmentation tasks with a novel prompt-based approach. By leveraging prompt features as queries and introducing a target-wise prompt cross-attention layer, UniVS eliminates the need for heuristic inter-frame matching, seamlessly adapting to different video segmentation scenarios.

IMAGE SYNTHESIS WITH COARSE-TO-FINE LATENT DIFFUSION

The Coarse-to-Fine Latent Diffusion (CFLD) method introduces a new approach to Pose-Guided Person Image Synthesis that overcomes the limitations of existing models by avoiding overfitting through a deep semantic understanding of images.

LLM QUANTIZATION: A COMPREHENSIVE STUDY

Post-training quantization can make large language models like OPT and LLaMA2 more memory- and compute-efficient.

power of LLMs without compromising privacy. Start free

3D MAPPING WITH LIGHTWEIGHT NEURAL NETWORKS

This project introduces a new approach to 3D mapping and localization that uses a lightweight neural network to process both point and line features, significantly enhancing pose accuracy with less memory and computational demand.

LEARNING TO GENERATE INSTRUCTION TUNING DATASETS

Bonito is an open-source model designed to convert unannotated text into tailored training datasets, enhancing the adaptability of large language models to specialized tasks without requiring pre-existing annotations.

EXPLORING VIDEO UNDERSTANDING WITH LLMA

This repository contains a collection of useful resources focused on video understanding using large language models.

MIST TEXT TO SPEECH

Rime has launched a new text-to-speech model with powerful conversational abilities. Unlike previous models, this one can include “ums” and realistic pauses.