TLDR AI 2024-04-16

OpenAI Japan 🇯🇵, Adobe Generative Video 📹, Video Diffusion Models 📺

Headlines & Launches

Introducing OpenAI Japan (2 minute read)

OpenAI has opened an office in Japan. ChatGPT Enterprise is being adopted in Japan by businesses such as Daikin, Rakuten, and TOYOTA Connected. The Japanese government, focusing on AI policies for societal benefit, has collaborated to share ChatGPT best practices.

Adobe's Working On Generative Video (3 minute read)

Adobe is developing an AI model to generate video, which will be integrated into Premiere Pro later this year. It will feature object addition, removal, and generative extend capabilities. Adobe is collaborating with third-party vendors and addressing deepfake concerns through Content Credentials.

OpenAI and Meta Reportedly Preparing New AI Models Capable of Reasoning (3 minute read)

OpenAI and Meta are teasing the next iterations of their AI models, expected to feature enhanced reasoning and planning capabilities. Dubbed GPT-5 and Llama 3, the models aim to advance toward artificial general intelligence, with vague release timelines and application details. The tech community remains skeptical given the history of overhyped AI promises with limited substantive evidence.
Research & Innovation

Pile-T5 (12 minute read)

T5 is a workhorse of modern AI. Eleuther has retrained it with a modern tokenizer and for much longer. The result is a greatly improved base model for encoding tasks.

Deep dive: Video Diffusion Models (23 minute read)

This post explores how to train diffusion models to generate video, how to adapt image models, and even how to generate video from an image model without additional training.
Engineering & Resources

3D Understanding with Any2Point (GitHub Repo)

Any2Point is a new method that efficiently translates capabilities from vision, language, and audio models into the 3D domain without losing spatial geometries.

3D Human Motion Capture (3 minute read)

Addressing the challenges in monocular egocentric 3D human motion capture, especially under difficult lighting and fast movements, researchers introduce a novel method using a fisheye event camera.

Zero-shot Dialogue State Tracking in LLMs (GitHub Repo)

Researchers have developed a new approach called Mixture of Prefix Prompt Experts (MPPE) to enhance zero-shot dialogue state tracking, which transfers knowledge to unseen domains without needing new dataset annotations.

The Agony and Ecstasy of Reviewing Humane's Ai Pin (23 minute read)

Humane's Ai Pin, an AI-powered wearable, has many technical issues, from music streaming to AI assistant malfunctions. Despite its ambition to create an ambient computing experience, the first-generation device suffers from performance lags and a lack of essential features, though future software updates are promised. The Ai Pin is designed with attention to detail but struggles to replace traditional smartphones, positioning itself as a companion device for a more present and less screen-focused lifestyle.

You can't build a moat with AI (4 minute read)

AI differentiation is challenging, but the key lies not in AI models like LLMs, which are becoming commoditized, but in the unique data fed into these models. Effective data engineering is crucial as it directly impacts AI performance, with applications requiring integration of customer-specific data to provide accurate responses. Thus, creating a competitive edge in AI applications hinges on innovative data use rather than the AI technology itself.
Quick Links

Github2File (GitHub Repo)

A quick utility that dumps GitHub repositories into files that can be used with long context models that cannot browse the web.

Google launches Code Assist, its latest challenger to GitHub's Copilot (3 minute read)

Google's Gemini Code Assist, an AI code completion tool for enterprises, features a one million-token context window for more accurate code suggestions and supports on-premises codebases across multiple platforms.

TikTok may add AI avatars that can make ads (2 minute read)

Internal tests have shown these AI avatars are not yet as effective as human influencers in driving sales.
The most important AI, ML, and data science news in a free daily email.
Join 500,000 readers for