AI Newsletter #002 (2023/05/29-2023/06/04)

News of the week

Asus to Sell Nvidia AI Servers You Can Install in Your Office
Asustek Computer Inc. is launching a service that utilizes generative artificial intelligence while maintaining control of data. To support this, they have partnered with Nvidia, whose stock has risen due to demand for their AI-training chips.

UAE’s Falcon 40B Dominates Leaderboard: Ranks 1 globally in latest Hugging Face independent verification of open-source AI models
Last week, the UAE’s Technology Innovation Institute (TII) launched Falcon 40B, the country’s first large-scale open-source AI model with 40 billion parameters. It quickly became the most powerful AI model on Hugging Face’s Open Large Language Model (LLM) Leaderboard.

Intel Discloses New Details On Meteor Lake VPU Block, Lays Out Vision For Client AI
Intel is using Computex to showcase their vision for the next generation of systems, which will include a Vision Processing Unit (VPU) derived from Movidius’ third-generation design. This VPU will be present in all Meteor Lake SKUs and will provide additional, energy-efficient performance for tasks such as dynamic noise suppression and background segmentation.

DeepMind AI Supercharges YouTube Shorts Exposure by Auto-Generating Descriptions for Millions of Videos
DeepMind and YouTube have developed a new AI model, Flamingo, to improve the searchability of YouTube Shorts videos. This technology has already been implemented on hundreds of thousands of videos, and YouTube plans to make it available for all Shorts videos, making them easier to find for viewers worldwide.

EU officials to meet OpenAI CEO again in June over AI laws
Just days after trading remarks over the implementation of AI laws in Euope, the EU’s industry commissioner, Thierry Breton, is set to meet OpenAI CEO Sam Altman in June to discuss AI regulations further, according to a Reuters report.

Products of the Week

Momentum Page
Page is a stupid-simple no-code landing page builder. With Page you can launch your product website in seconds, then distribute it with social media, create and feature your best content, engage your audience, grow and.. sell

Siit AI, powered by GPT-4, instantly answers all employees questions based on your internal knowledge bases, Notion and Confluence, directly via Slack and Teams. This is a revolution in your employee experience.

Quest AI
Build your ideas faster than ever with the power of AI/GPT. Got a wireframe (or a Figma design)? Generate working ReactJS. Clean and extendable code based on your design system. Export as JS/TS, React or NextJS. Transform your design-to-dev workflow!

CapeChat automatically encrypts your documents and redacts any sensitive data. It’s powered by the ChatGPT API, so you get the best language model while preserving your privacy.

Research of the Week

VOYAGER: An Open-Ended Embodied Agent with Large Language Models
The team from Nvidia developed VOYAGER, the first LLM-posered agent in Minecraft that continuously explores the world, acquires diverse skills, and makes novel discoveries without human intervention. It obtains 3.3X more unique items, travels 2.3X longer distances, and unlocks key tech tree milestones up to 15.3X faster than prior SOTA.

If you are interested in LLM application in gaming, you might be interested in this work:
Project page:

Let’s Verify Step by Step
To train more reliable models to perform complex multi-step reasoning, we can turn either to outcome supervision, which provides feedback for a final result, or process supervision, which provides feed back for each intermediate reasoning step.
Researchers at OpenAI have conducted a study comparing outcome supervision and process supervision for training large language models. The study found that process supervision significantly outperforms outcome supervision in training models to solve problems from the challenging MATH dataset. The researchers also showed that active learning significantly improves the efficacy of process supervision. They released the PRM800K dataset, which contains 800,000 step-level human feedback labels used to train their best reward model, to support related research.
If you are interested in training large language model to perform complex multi-step reasoning, this is a must read:

Chain-of-Thought Hub: A Continuous Effort to Measure Large Language Models’ Reasoning Performance
The authors obeserved that complex reasoning is likely to be a key differentiator between LLMs. This work proposes Chain-of-Thought Hub, an open-source evaluation suite on the multi-step reasoning capabilityies of LLMs. It compile a suite of callenging reasoning benchmarks to track the progress of LLMs.
The results shows that:
(1)Model scale clearly correlates with reasoning capabilities.
(2)As of May 2023, Claude-v1.3 and PaLM-2 are the only two models that are comparable with GPT-4, while open-sourced models still lag behind.
(3)LLaMA-65B performs closely to code-davinci-002, indicating that with successful further development such as reinforcement learning from human feedback (RLHF), it has great potential to be close to GPT-3.5-Turbo.
If you are interested in LLM performance evaluation, this research will provide you great information:

Want to receive TechNavi Newsletter in email?