TLDR AI 2024-02-20

Gemini for Google Workplace 💻, OpenAI $80B valuation 💰, enhanced speech translation 🗣️

Headlines & Launches

Gemini Business And Gemini Enterprise Plans For Google Workspace Are Coming (2 minute read)

Google is introducing "Gemini Business" and "Gemini Enterprise" plans for Workspace customers, offering access to its advanced AI model, Gemini 1.0 Ultra, with enterprise-grade data protection. It will not use employee conversations for model training.

OpenAI valued at $80 billion in deal with Thrive Capital (2 minute read)

OpenAI has reportedly reached a valuation of over $80 billion following a deal with VC firm Thrive Capital. This valuation marks a nearly 3x increase in less than 10 months, making OpenAI the 3rd highest-valued tech startup in the world. The deal involves a share sale under a "tender offer," allowing employees to cash out their shares in the company.

Magic raises $117m to continue code generation models (3 minute read)

Magic is building an AI software engineer and has raised more to help with this process.
Research & Innovation

Making CLIP’s Visually Understandable (16 minute read)

SpLiCE is a new method that makes CLIP's complex visual data easier to understand.

KV Quant - Scaling to 10M tokens (22 minute read)

Quantizing the KV cache, which is a technical detail of the Transformer architecture, allows it to use less memory at inference time. Quantization is the shrinking of the floating point precision with minimal loss in quality.

Enhanced Speech Translation (18 minute read)

ZeroSwot is a new method for Speech Translation (ST) that overcomes the lack of data and differences between speech and text. It uses unique techniques to train a speech encoder with only speech recognition data, allowing it to work with a multilingual translation model.
Engineering & Resources

3D Diffuser for Smarter Robots (GitHub Repo)

The 3D Diffuser Actor combines diffusion policies and 3D scene representations to improve robot manipulation. This AI-driven approach allows robots to understand and interact with their environment more effectively.

Library for ControlVectors (GitHub Repo)

ControlVectors are a way to cheaply steer semantic generation output. They are cheaper to train than LoRA but can still be quite effective. This library makes it easier.

OpenRLHF (GitHub Repo)

This is a Ray-based implementation of RLHF for Llama-style models, which includes Mistral. It contains a number of PPO stabilization tricks to improve performance.

How to jointly tune LR and WD in AdamW (12 minute read)

AdamW is commonly used to decouple learning rate and weight decay. However, the common implementation in Pytorch doesn’t explicitly do this. This post discusses how to tune these parameters.

Enhancing Dense Video Analysis Transformers (28 minute read)

A new Temporal Dilated Video Transformer (TDViT) has been developed to improve the analysis of dense video tasks, such as frame-by-frame video object detection.

Inside OpenAI (68 minute podcast)

Ever wondered how OpenAI is able to innovate and build at such a fast pace? In this podcast, Logan Kilpatrick, OpenAI's head of developer relations, discusses OpenAI's unique culture, the importance of high agency and urgency, and the company's decision-making framework for launching products.
Quick Links

Unspam AI (Product)

Ensure your emails land in the inbox, not spam.

List of GPU clusters for rent (2 minute read)

A list of full clusters that are available for rent by the hour.

OpenLLMetry (GitHub Repo)

OpenLLMetry-JS is a set of extensions built on top of OpenTelemetry that gives you complete observability over your LLM application.
The most important AI, ML, and data science news in a free daily email.
Join 500,000 readers for