Welcome back!

Today's issue of The Summary is dedicated to unpacking the main highlights from the Google I/O 2024 Keynote. From the unveiling of Project Astra to the video creation capabilities of Veo and the upgraded Gemini models, Google showcased significant advancements. Here's what we found..

Today’s Summary:

• 🚀 Google's Astra competes with ChatGPT-4o

• 🎥 Veo text-to-video AI

• 💡 Upgrades in Gemini models

• 🌍 Claude 3 launch in Europe

• 🧠 Ilya Sutskever, OpenAI Co-Founder, Resigns

• 🛠️ 3 new tools

TOP STORY

Google's Project Astra strikes back at ChatGPT-4o

The Summary: At Google I/O 2024, Google unveiled Project Astra, a real-time AI assistant that can see, hear, and understand the world around you. In the demo, Astra identified objects, answered questions about code, and even helped find missing items - all through natural conversation. Astra combines multimodal AI with optimizations to run with low latency on phones and smart glasses. Google positions Astra as a step toward the long-envisioned universal AI assistant to assist in daily life.

Key details:

• Astra uses the advanced multimodal Gemini Ultra, trained on text, image, audio, and video data

• Operates with low latency through sophisticated model and infrastructure optimizations

• Astra was demonstrated working in real-time on both a phone and prototype smart glasses

• The announcement came just one day after OpenAI’s reveal of its own multimodal demo for ChatGPT

Why it matters: Google is in a race to reassert its leadership in AI, after ChatGPT captured global attention. The development of multimodal AI that can fluidly perceive and interact with the real world represents a critical new frontier for AI companies. If the technology meets its potential, assistants like Astra could become true productivity multipliers and everyday aids.

VIDEO GENERATION

Google announces Veo, realistic text-to-video AI

The Summary: Google announced Veo, its latest generative AI model based on latent diffusion transformers, capable of producing high-quality 1080p videos up to a minute long from text prompts. Veo demonstrates an impressive understanding of cinematic terms and visual styles.

Veo aims to empower filmmakers, animators, and video creators by providing creative control. Google plans to invite select creators to test Veo and intends to integrate some features into YouTube Shorts.

Key details:

• Veo can generate coherent videos in various visual styles from text, image, or video prompts

• It comprehends cinematic concepts like "timelapse", "aerial shots", and "film noir"

• The generated videos maintain consistent characters, objects, and scenes across frames

• All videos will be watermarked to indicate they are AI-generated

Why it matters: Veo is a new development in Google's generative AI capabilities, offering realistic and coherent outputs. It positions itself as a strong competitor to OpenAI’s Sora. Responsible development, guided by ethical partnerships, will be essential as these models become more powerful.

GEMINI

Upgrades to Gemini AI Models, mega context window

The Summary: Google announced substantial upgrades to its Gemini AI models during the keynote. Gemini 1.5 Pro now supports a 2 million token context, enabling it to process about 3000 pages of text or 2 hours of video, the largest context window to date. Additionally, Google introduced a lightweight Gemini 1.5 Flash model optimized for speed.

Key details:

• Gemini 1.5 Pro boasts quality improvements in translation, coding, and reasoning

• The new Gemini 1.5 Flash model is optimized for efficiency and speed, with API cost lower than GPT 3.5-Turbo

• Both models are multimodal with a 1M token context (2M tokens for 1.5 Pro on a waitlist)

• Gemma2 27B, a new open-source model, performs nearly as well as Meta’s much larger Llama3-70B (MMLU benchmark: 75 vs 79.2).

Why it matters: The massive context window allows Gemini models to tackle large-scale data tasks previously beyond the reach of AI. These upgrades, along with the introduction of specialized models for speed and the open-source Gemma 2, present new opportunities for building efficient AI applications.

QUICK NEWS

Quick news

• Claude 3, one of the highest-performing and most capable AI models, is now available in Europe

• Breaking: OpenAI Co-Founder Ilya Sutskever departs; Jakub Pachocki named as the new Chief Scientist

TOOLS

🥇 New tools

• Fynk - AI contract management

• Stunning - The fastest way to build a website

• Wegic - AI web designer