Search the archive

EXPERIMENTING WITH LOCAL ALT TEXT GENERATION

Mozilla has trained a model that can run in the browser and automatically generate alt text for web images.

RANDOM COLOR ERASING BOOSTS MODEL ROBUSTNESS IN COMPUTER VISION

Researchers have developed a new learning strategy called Random Color Erasing to tackle color bias in computer vision. This method balances the importance of color and other features by selectively removing color information from training data, leading to models that perform better in complex scenarios like wide-area surveillance and industrial applications.

account

VIDEO GENERATION WITH MORA

Mora is an innovative video generation model that leverages multiple visual AI agents to challenge Sora, the leading model by OpenAI.

BOOSTING LLM'S ABILITY TO FOLLOW COMPLEX INSTRUCTIONS

Conifer improves LLMs' understanding of complex instructions by introducing a specialized dataset and a progressive learning approach.

FABRIC FEEDBACK TO IMPROVE IMAGE GENERATION QUALITY

FABRIC is an inference time, training-free method for improving image quality by using a scoring function or image examples to direct output quality.

KL IS ALL YOU NEED

KL divergence is a fast, cheap, and powerful way to measure a kind of distance between things. It is used copiously in traditional and modern AI. This post explores the powerful concept visually and mathematically.

BEYOND THE CLOUD TO EDGE AI

AI is moving to multimodal and the edge. How will that change our use of it and where it is effective?

EX-OPENAI BOARD MEMBER REVEALS WHAT LED TO SAM ALTMAN'S BRIEF OUSTING

Former OpenAI board member Helen Toner revealed new details about what led to CEO Sam Altman's ousting in November in a recent interview. The board had apparently learned about ChatGPT's release on Twitter. Toner says that Altman had lied to the board multiple times. Altman was reportedly withholding information and misrepresenting things that were happening in the company for years. His lies made it difficult for the board to make decisions, leading them to think that he wasn't the right person to lead the company to AGI.

LLM MERGING CHALLENGE

Sakana AI is sponsoring the LLM Merging challenge at NeurIPS this year.

TRIBECA TO SCREEN SORA-GENERATED FILMS

The 2024 Tribeca Festival will feature "Sora Shorts," showcasing five AI-generated short films created using OpenAI's text-to-video model Sora.

7 RETENTION STRATEGIES FOR AI COMPANIES

A guide for founders and product leaders that includes examples of how companies like Perplexity, Civit, Lapse, Omnivore, and others are leveraging them to boost retention.

Up your game with LMNT ↗️

PERPLEXITY PAGES

The Perplexity search engine has announced a new product for creating long-lasting research artifacts that are shareable and created with AI.

WONDER DYNAMICS ACQUIRED BY AUTODESK

The AI VFX company with Hollywood ties has been acquired by software maker Autodesk for an undisclosed amount.

ANTHROPIC'S AI NOW LETS YOU CREATE BOTS TO WORK FOR YOU

Anthropic is introducing a "tool use" feature for its Claude AI chatbot, enabling users to create personalized assistants that can interact with any external API. This feature can analyze data, provide product recommendations, track orders, offer technical support, and even process images for applications like interior design.

YUAN 2.0-M32: MIXTURE OF EXPERTS WITH ATTENTION ROUTER

Yuan 2.0-M32 is a Mixture of Experts model with 40B parameters, 3.7B active at any time. It performs close to Llama 3 70B despite using 1/19th the compute. It was trained on 2T tokens and seems surprisingly powerful for its size.

SCALING LAWS AND COMPUTE-OPTIMAL TRAINING BEYOND FIXED TRAINING DURATIONS

The original scaling laws papers used a Cosine learning rate schedule which doesn't achieve optimal loss if the Cosine period isn't aligned with the total training steps. This makes it challenging to train enough models to create meaningful scaling laws. This paper introduces the idea of a constant learning rate with a cool down as a way to reduce GPU costs for scaling law creation.

ULTRA-HD IMAGE DERAINING

Researchers have introduced a new dataset called 4K-Rain13k that features 13,000 pairs of 4K resolution images to tackle the challenge of deraining ultra-high-definition (UHD) images.

SDXL SCRIBBLE CONTROLNET

Scribble ControlNet allows you to draw arbitrary lines as conditioning information for image generation. It was trained on substantially more post training images than other ControlNets and boasts powerful performance.

VIDEO GENERATION WITH TRANSFORMERS

EasyAnimate is a method that adapts the DiT framework for advanced 3D video generation using transformers. This project incorporates a motion module block to capture temporal dynamics, ensuring smooth motion transitions and consistent frames.

BOOSTING LANGUAGE MODELS WITH SELF-EXPLORING OPTIMIZATION

Self-Exploring Language Models (SELM) is a method that enhances preference optimization in LLMs using online feedback.

TEXT2VIDEO TURBO

Consistency Distillation applied to video models dramatically reduces the number of steps needed to generate content.

AI WON'T EAT YOUR JOB, BUT IT WILL EAT YOUR SALARY

AI threatens not just the existence of jobs but the skill premium associated with tasks, potentially reducing the wages for skilled labor. AI could re-bundle job tasks and lower the barriers to task performance, leading to commoditization and decreased ability to charge a premium salary. As AI evolves, especially through AI agents, managerial advantages may also diminish, challenging the human-in-the-loop advantage and further eroding skill premiums.

PATRONUS AI RAISES $17M SERIES A

Patronus builds automated security for language models.

NEURALINK RIVAL SETS BRAIN-CHIP RECORD WITH 4,096 ELECTRODES ON HUMAN BRAIN

Precision Neuroscience claims to have set a new record for the number of neuron-tapping electrodes placed on a living human's brain at 4,096.

GOOGLE ADDS AI-POWERED FEATURES TO CHROMEBOOK

Google announced new AI features for Chromebook Plus devices, including a writing assistant, wallpaper creator, and easy access to the Gemini chatbot.

Read the blog post >>

MISTRAL'S NEW AI NON-PRODUCTION LICENSE

Mistral is working to find a balance between openness and business success. It has a new license aimed at striking that balance. It will continue releasing other projects under Apache 2.0 as well as the new MNPL license.

SONIC: A LOW-LATENCY VOICE MODEL FOR LIFELIKE SPEECH

A new model from the creators of SSMs, Mamba, and sub quadratic Transformer variants. Importantly, their new company Cartesia has trained a lightning-fast, realistic voice generation model. This points to their desire to move into the assistants' space.

VOX MEDIA AND THE ATLANTIC SIGN CONTENT DEALS WITH OPENAI

OpenAI has signed licensing deals with The Atlantic and Vox Media, allowing their content to train its AI models and be shared in ChatGPT with proper attribution.

MISTRAL'S CODE MODEL

Mistral released a 22B code model with 32k context. It is a powerful and fast model with broad performance across many programming languages. It has open weights and is available via Mistral's platform.

PRISM: SMALL MOLECULE FOUNDATION MODEL

PRISM is a foundation model trained on 1.2 billion small molecule mass spectra to enhance mass spectrometry analysis in drug discovery. It uses self-supervised learning to predict molecular properties from complex mixtures without prior annotations, significantly advancing the field of metabolomics. The model improves the identification and understanding of largely unknown natural molecules, accelerating the discovery of new therapeutics by leveraging massive, previously untapped spectral data.

DIFFUSION MODELS FOR INVERSE PROBLEMS

DMPlug is a novel plug-in method that uses pretrained diffusion models (DMs) to solve inverse problems (IPs). Unlike existing interleaving methods, DMPlug views the reverse diffusion process as a function, addressing both manifold feasibility and measurement feasibility effectively.

ENHANCED SUPER-RESOLUTION EFFICIENCY

PatchScaler is a diffusion-based method for single image super-resolution (SR) that significantly enhances inference efficiency.

3D UNDERSTANDING WITH ENHANCED SEGMENTATION

Reason3D is a novel multimodal large language model designed for comprehensive 3D environment understanding.

SCALE PRIVATE LEADERBOARD

Scale has made a private leaderboard for evaluating language models. The ordering isn't too surprising, but one interesting finding is the Llama 3 70B is extremely good at following instructions, often better than Claude Opus.

OPENAI SIGNS 100K PWC WORKERS TO CHATGPT'S ENTERPRISE TIER

OpenAI has signed PwC as its largest enterprise customer for ChatGPT, covering 100,000 users. PwC is the first partner to resell OpenAI's enterprise offerings.

in-person in SF. Get $200 of with TLDR24 discount code!

APPLE'S AI PLANS INVOLVES ‘BLACK BOX' FOR CLOUD DATA

Apple's upcoming AI features for iOS 18 aim to enhance privacy by processing data in a secure "black box" within its own hardware.

MIDJOURNEY'S HARDWARE TEAM

David Hold, founder of Midjourney, has assembled a new hardware team.

BOOK A FREE CONSULTATION SESSION TODAY. GET IN TOUCH NOW