TLDR AI 2023-09-25

Haystack (10,785 GitHub stars) offers everything you need to quickly design and build scalable, API-driven LLM backends.

Secoda, an AI-driven data search and cataloging platform startup, has secured $14 million in funding led by Craft Ventures. The Ontario-based company aims to make finding and understanding company data as seamless as a Google search. Secoda has seen a 250% YoY revenue growth, attracting notable customers like Remitly and Y Combinator.

Meta is reportedly preparing to launch "Gen AI Personas," a generative AI chatbot aimed at younger users with various personas designed to engage them.

Microsoft’s Bing is gaining support for OpenAI’s new DALLE-E 3 model, more personalized answers in search and chat, and tools that will watermark images as being AI-generated.

This paper introduces EmoPP, a model that incorporates emotional cues for more accurate prosodic phrasing predictions. The authors emphasize the significance of prosodic phrasing in achieving natural-sounding Text-to-Speech.

This paper proposes a model that can create super-realistic virtual humans in any pose you can think of. This is the first model of its kind trained only on fake data, but it can still make highly accurate human shapes that could be useful for science and video games.

Researchers have introduced TCOVIS, an online video instance segmentation technique that prioritizes temporal consistency.

Sidekick is a superhuman-style AI programming toolkit for the terminal and VSCode.

This repository introduces FGFusion, a new method that combines data from cameras and Lidars in a detailed way for self-driving cars. Instead of just using high-level details, it captures both big-picture and fine details to create a more accurate 3D view of the surroundings.

Retrieval is one way to improve generation quality. However, it is often slow to embed queries and search. This notebook shows some clever tricks to help improve retrieval performance.

Join the most in-depth (and most fun!) technical conference in AI.

This time, it’s all about solving technical challenges in taking LLMs from proof-of-concept to production apps. Featuring 40+ expert speakers hailing from industry heavyweights (NVIDIA, Google, etc.), as well as under-the-radar stealth startups.

Also: Musical interludes, guided meditation, and prompt injection games.

October 3rd | Online | 100% free

The newest version of Flash Attention is much faster while improving model context length and performance. It is now in the process of being natively supported in the HuggingFace ecosystem.

When language models are trained on “A is B”, they fail to automatically infer “B is A”.

Course on how to do low resource machine learning efficiently at the edge. Lectures and homework included.

Turn selfies into professional headshots. Create realistic AI photos of yourself that look like they were taken by a professional photographer.

DALL-E 3’s amazing ability to create art from simple text puts the future of prompt engineering in jeopardy.

Unify your data so that you can seamlessly search and discover across apps and files.