TLDR AI 2023-09-15

Microsoft has open-sourced EvoDiff, a novel AI model that the company claims can generate high-fidelity, diverse proteins given a protein sequence.

Led by Lightspeed, the team wants to build real world scoring to help understand how useful LLMs are for enterprises.

MLPerf has released results from two benchmark suites: MLPerf Inference v3.1, which showcases record participation and performance gains, and MLPerf Storage v0.5, which assesses storage system performance for ML training workloads. The Inference benchmark suite introduced a large language model and updated recommender tests, reflecting emerging AI trends.

Alignment helps make language models more helpful and harmless. It can sometimes hurt performance, but in general is a positive. Alignment is expensive and requires substantial alignment data. However, if you allow the model to rewind during generation after evaluating its own output, it can improve alignment performance on the frozen model by up to 81%.

Mixture of experts (MoEs) is a cool way to increase the capacity of a model without increasing per token runtime. However, they're still tricky to run quickly and fine-tune. Researchers find that if you modify dense model parameters efficiently fine tuning to work with MoEs you can dramatically reduce tuning costs without hurting performance too much.

Researchers have developed a new model called HAMUR to improve how models make suggestions or "recommendations" across multiple subjects or domains, like music, books, or movies. Unlike older methods that mix up information between these domains, HAMUR uses a special technique to keep data separate and more flexible.

DeepFakeFace (DFF) is a collection of fake celebrity photos made with advanced technology designed to help us get better at telling real pictures from fake ones.

A comprehensive guide to building RAG-based LLM applications for production.

JPEG images are everywhere, but they don't work well with deep learning because you can't easily tweak them during training. This project reviews existing solutions and introduces a new method that fixes these problems, making JPEG images fully adjustable and compatible with deep learning systems.

Coqui has released weights for its xtts model, which can clone voice parameters and synthesize in different languages.

Researchers suggested a global AI ethics agreement a few years ago, emphasizing its feasibility despite philosophical differences. As the U.S.-China AI competition grows, challenges like misperceptions and language barriers persist in achieving global AI regulation.

Scientists have developed RETFound, an AI tool capable of diagnosing and predicting the risk of developing multiple health conditions, from ocular diseases to heart failure to Parkinson’s disease, from just retinal images.

Summit is an AI life coach that helps users achieve personal and professional goals. Summit helps to break down big goals into bite-sized actions, holds users accountable for achieving them, and supports users with personalized coaching along the way.

Build custom ChatGPT-like chatbots for your website instantly.