Parse notes
Sakana.ai is building a world-class research lab in Tokyo to build smaller and more efficient foundation models. The company was founded by ex-Google researchers David Ha and Llion Jones, who are responsible for breakthroughs such as Transformers, World Models, and LoRA. It has raised a 30m seed round from Brandon Reeves at Lux Capital and Jon Chu at Khosla Ventures to pioneer this effort and put Tokyo on the AI map.
A strong 3B model that outperforms the 7B CodeLlama model and is small enough to run locally on MacBooks.
OpenAI is forming a new Collective Alignment team to integrate public input into its AI models.
This blog explores the current dominant paradigm for taking a pretrained language model and adding multimodal capabilities. Namely, aligning the embeddings between an encoder and the text model.
Hard problems have been thought to require hard data, but that data is hard to label and noisy. This work shows that fine-tuning models on related, but easy, data dramatically improves their generation capabilities on hard problems. Another data point to support fine-tuning as eliciting knowledge instead of adding knowledge.
Researchers have developed a novel approach called Mutual Distillation Learning For Person Re-identification (MDPR) that enhances person re-identification by combining two different techniques.
Sponsored
Microsoft is marketing that its Direct Lake architecture on Fabric is the ideal way to run direct queries on cloud storage in Power BI. Does the data support their claim? This webinar [12] will dive deep into AtScale’s TPC-DS benchmark report [12] that tests the scalability, cost effectiveness and manageability of Microsoft Fabric Architecture.
One of the most viral uses of generative image models is personal stylized image generation. Historically, this required training a LoRA or using DreamBooth. Now, with the use of ID embeddings, you can dramatically improve quality while reducing computational cost, with just a single image.
Following the recent trend of repurposing diffusion models for various other vision tasks, this work improves the stability of training for super resolution and upscaling.
FILCO is a new technique designed to improve AI systems, especially in areas like answering questions and checking facts, by giving them better context. It works by choosing the most helpful information and training models to filter out less relevant details.
Dataset to help align models to write correct Haiku’s.
The "Wait Calculation" concept advises delaying actions, like a space journey to Barnard's Star or starting certain projects, until technology sufficiently advances to save significant time and effort. However, this approach must be balanced against potential learning losses and the unpredictability of technological progress.
This project improves the Model Parameter Randomisation Test, a key tool in eXplainable AI (XAI), by introducing two new versions: Smooth MPRT and Efficient MPRT.
Sponsored
Join DeepBrain AI and Runway’s joint webinar to see the latest in generative AI video — bringing video to life with just text input.
Meta recently announced a suite of music and sound generation models that sound incredible. It has now released the models with inference code.
Tiny Narrations is a text-to-speech read version of the popular (and small) Tiny Stories dataset. It uses XTTS2 on the SF Compute H100 cluster.
An AI assistant that helps sales teams with personalized prospecting and efficient meeting preparation.